970572

research-article2020
EVB0010.1177/1176934320970572Evolutionary BioinformaticsYu et al

Hypergraph Clustering Based on Game-Theory for Evolutionary Bioinformatics

Volume 16: 1–8

Mining Microbial High-Order Interaction Module © The Author(s) 2020

Article reuse guidelines:
sagepub.com/journals-permissions
Limin Yu1,2,3, Xianjun Shen1,2,3 , Jincai Yang1,2,3, Kaiping Wei1,2,3, DOI: 10.1177/1176934320970572
https://doi.org/10.1177/1176934320970572

Duo Zhong1,2,3 and Ruilong Xiang1,2,3

1School of Computer, Central China Normal University, Wuhan, China. 2Hubei Provincial Key

Laboratory of Artificial Intelligence and Smart Learning, Central China Normal University,
Wuhan, Hubei, China. 3National Language Resources Monitoring and Research Center for
Network Media, Central China Normal University, Wuhan, Hubei, China.

ABSTRACT: Microbial community is ubiquitous in nature, which has a great impact on the living environment and human health. All these effects
of microbial communities on the environment and their hosts are often referred to as the functions of these communities, which depend largely
on the composition of the communities. The study of microbial higher-order module can help us understand the dynamic development and evo-
lution process of microbial community and explore community function. Considering that traditional clustering methods depend on the number
of clusters or the influence of data that does not belong to any cluster, this paper proposes a hypergraph clustering algorithm based on game
theory to mine the microbial high-order interaction module (HCGI), and the hypergraph clustering problem naturally turns into a clustering game
problem, the partition of network modules is transformed into finding the critical point of evolutionary stability strategy (ESS). The experimental
results show HCGI does not depend on the number of classes, and can get more conservative and better quality microbial clustering module,
which provides reference for researchers and saves time and cost. The source code of HCGI in this paper can be downloaded from https://
github.com/ylm0505/HCGI.

Keywords: Microbial higher-order module, game-theory, hypergraph clustering, evolutionary stability strategy

RECEIVED: June 6, 2020. ACCEPTED: October 12, 2020. Declaration of conflicting interests: The author(s) declared no potential
conflicts of interest with respect to the research, authorship, and/or publication of this
Type: Machine Learning Models for Multi-omics Data Integration – Original Research article.

Funding: The author(s) disclosed receipt of the following financial support for the CORRESPONDING AUTHOR: Xianjun Shen, Hubei Provincial Key Laboratory of Artificial
research, authorship, and/or publication of this article: This work was supported by the Intelligence and Smart Learning, Central China Normal University, Wuhan, Hubei 430000,
National Natural Science Foundation of China [61532008, 61872157, 6192008]; the China. Email: [email protected]
Self-determined Research Funds of CCNU from the Colleges’ Basic Research and
Operation of MOE [CCNU19QD003]; and the National Language Commission Key
Research Project [ZDI135-61].

Introduction infer the interaction between species by measuring the popula-

There are many kinds of microorganisms, which exist in human, tion dynamics of a complex community.11-14 Some studies
plant, soil, ocean and other environments. Microbial commu- mainly focus on the interaction between two species and the
nity can be defined as a collection of microorganisms that exist ability of two interaction models to predict network func-
at the same time and may interact with each other. They exist in tions.15,16 Although the interaction between the two is enough
a certain spatiotemporal habitat. The term microbial commu- to describe the important impact of ecology in some cases,
nity was first coined by Lederberg and McCray1 By definition, some studies have emphasized the need to merge higher-order
microbial community is ubiquitous in nature, which has a huge relationships, mainly because of their role in population
impact on the living environment and human health.2-4 All dynamics.5,17,18
these impacts of microbial community on the environment and Tsai et al19 developed a rule-based microbial network (RMN)
its host are usually called the functions of these communities, algorithm, which uses a triple relationship similar to the activation
which largely depend on the composition of the community, inhibitor target model to build a gene regulatory network. Based
that is to say, on the species and quantity of the existing species. on this principle, RMN algorithm was developed to describe the
In essence, the interaction between species involves multiple relationship between every three microorganisms as a cooperation
species, which often occurs in a high-order combination way, competition target. Guo and Boedicker20 found that the interac-
that is, the interaction between two species is regulated by one tion between species has an impact on the activity of the whole
or more other species,5 and the high-order interaction often community, and studied the interaction network of four microbial
dominates the function of microbial community, for example, communities through theoretical models. The results show that
higher-order functional interaction dominates the hydrolysis the interaction between multiple species determines the overall
rate of amylase in biological communities.6 metabolic rate of the microbial community.20 Bairey et al5 found
In order to quantify how the interaction between species that the high-order interaction has a lower bound to diversity,
affects the overall community function, two interactions while the pairwise interaction exerts an upper bound. When the
between species were systematically measured through experi- pairwise interaction is dominant and the number of species
ments, such as cross feeding between E. coli growth nutrient increases beyond the threshold, the feasibility decreases. On the
bacteria7,8 and naturally isolated bacteria.9,10 Other studies contrary, when the four-way interaction is dominant, the

Creative Commons Non Commercial CC BY-NC: This article is distributed under the terms of the Creative Commons Attribution-NonCommercial
4.0 License (https://creativecommons.org/licenses/by-nc/4.0/) which permits non-commercial use, reproduction and distribution of the work without
further permission provided the original work is attributed as specified on the SAGE and Open Access pages (https://us.sagepub.com/en-us/nam/open-access-at-sage).
2 Evolutionary Bioinformatics 

community becomes sensitive to species removal. In the case of assume that there are paired similarities between microorganisms,
mixing, when the pairwise interaction and the high-order interac- but many microorganisms in the real environment involve more
tion are dominant, the community becomes sensitive to species complex and multipath relationships. In such networks of high
removal. When combined, the intermediate species is the most order relationships, modeling pairs only results in the loss of infor-
feasible, and the removal and addition of species will destroy the mation in the original data. The graph model of pairwise relations
stability of the community. Therefore, the combination of high- does not consider higher order information, and cannot represent
order interaction and pairwise interaction determines the most the whole ecological network. Microorganisms diversity depends
stable range of diversity. These findings emphasize the importance heavily on the stabilization of higher-order interactions. It is
of high-order species interaction in determining the diversity of important to study the higher- order interaction of microorgan-
natural ecosystems. In order to study the contribution of pairwise isms to understand the diversity and stability of microbial commu-
interaction and higher-order interaction to community function, nity. Hypergraph clustering refers to the process of extracting the
Alicia Inspired by the complex genetic interaction, etc. Used the most consistent group from a group of objects by using high-order
ecological function landscape to separate and quantify the single, (rather than pairwise) similarity. Therefore, hypergraph clustering
paired and higher-order interactions in the microbial community is applied to study the high-order interaction network module of
function, screened the experimental results through the enzyme microorganisms.
dynamics theory, and found that the higher-order functional In this paper, first, the microbial abundance data were pro-
interaction dominated the hydrolysis rate of amylase in our com- cessed, from the perspective of logical relationship, the infor-
munity.6 That is to say, it is very important to build a meaningful mation entropy is used to determine the high-order interaction
microbial network to study the high-order interaction between relationship of microorganisms, and then the high-order rela-
microorganisms. tionship hypergraph of microorganisms is constructed. Then, a
From the perspective of modules, we can intuitively under- kind of microbial logical relationship is selected for further
stand the interaction information and the contribution of net- study, and game theory is introduced to transform hypergraph
work modules to the overall network, so as to study the possible clustering problem into clustering game problem. Finally, we
functions of modules formed by specific microorganisms in the use Baum-Eagon theorem to generate the clustering module
microbial community. Girvan and Newman21 proposed a method with iterative elimination strategy until the remaining objects
to detect the community structure, based on the centrality index cannot form a high-order relationship.
to find the community boundary, and applied it to study many
different social networks and biological networks, focusing on the Materials and Methods
least central edge and the most central edge between communi- The dataset
ties, not by adding the strongest edge to an initial empty vertex set
In this paper, the available microbial abundance data of healthy
to build the community, but by gradually removing edges from
humans based on 16s rRNA comes from the human microbi-
the original graph to build a community can help you understand
ome project (HMP), which is V13 high quality files processed
the composition of other complex datasets.
with the mothur software package. The data covered 18 of the
Copeland et al22 used the network analysis method to study
5 main body parts. Data source web site is http://hmpdacc.org/
the relationship between the abundance of each genus in the leaf
HMMCP/.
microbial community, determined two highly connected clusters,
and found functional groups. Aleksej considering the possibility
Hypergraph clustering based on game theory
of high-order interaction involving more than two species,
extended the concept of binary symbiosis to at most four species In the previous work, we used to find the maximum degree of
at the same time, and proposed a model for large microbial com- modularity to determine the number of clusters before the
munities. The experimental results showed that although there premise of clustering.24 In this paper, we studied the hyper-
was obvious resource competition at the level of the whole com- graph clustering problem from the perspective of game theory
munity, microbial communities still had interdependent meta- to analyze the high-order module analysis of microorganisms,
bolic groups, and they occur repeatedly in different habitats, which can automatically generate clusters without specifying
emphasizing that metabolic dependence is the main driving force the number of clusters.
of species symbiosis, and suggesting that the cooperative group is In the process of biological evolution, animal adaptability
a recurring module in the microbial community structure.23 That is formed in the interaction with their living environment. In
is to say, the study of higher-order microbial modules can help us the competition, animals finally choose the evolutionary sta-
to understand the dynamic development and evolution process of bility strategy (ESS), which is adopted by most members of
microbial communities, help us to understand the functional the population and will not be eroded by other strategies. In
composition of communities, and predict the potential function- our framework, the clustering problem is seen as a non-coop-
ality of community modules. erative game between multiple microorganisms, and an
The traditional clustering problem only studies the pairing rela- important concept in game theory is equilibrium, in which
tionship between two microorganisms. Most current algorithms the performance of each player in a population does not
Yu et al 3

exceed the population average return when Nash equilibrium 1

is reached. Bulo et al25 proved that the problem of finding π ( v1 ,…, vk ) =  k !
( )
 s {v1 ,…, vk } if {v1 ,…, vk } ∈ E ,
(4)
these equilibrium points (clusters) is equivalent to solving a  0 else ,
polynomial optimization problem with linear constraints, and 
1
used an algorithm based on Baum-Eagon inequality to solve where is a coefficient.
this problem. We apply it to the analysis of the microbial high k!
Given a population x ∈ ∆ paly clustering game, we can
order module in this paper.
( )
obtain the average population payoff u X k , which represents
the average similarity of the objects that make up the cluster. The
Game theory. A game can be represented as Γ = (P, S, π) .
P = {1,…, k } is the set of players in game, and S = {1,…, n } is
( )
average payoff u e i , X k −1 of i ∈V in population x represents
the average similarity of object i with respect to the cluster.
a set of pure strategies. p is a payoff function, Each player has The problem of extracting ESSs of our hypergraph cluster-
their own set of strategies and payoff functions. The evolution ing game can be transformed into the problem of finding strict
of populations occurs because we hypothesize that there is a local solutions of (5). Consider the following nonlinear optimi-
selection mechanism, similar to Darwin’s evolutionary process, zation problem:
that spreads the fittest survival strategies in a population to the
detriment of the weakest.
Given the set of all possible states of the population D:
maximize f (x ) = ∑s(e )∏x
e∈E i∈e
i x∈∆ (5)

  By the Karush-Kuhn-Tucker (KKT) conditions there exist

∆ = x ∈  n :

∑
i∈S
xi = 1 and xi ≥ 0 for all i ∈ S 

(1)
µi ≥ 0, i ∈ S , λ ∈  so that for all i ∈ S ,

xi represents the fraction of i-strategists in the population. ∇f ( x )i + µi − λ

(6)
If y ( ) ∈ ∆ determine which strategy represents the ith player 1
( )
i
= u e i , X k −1 + µi − λ = 0 and µi xi = 0,
will adopt to play the game Γ . The average payoff obtained by k
the agents can be defined as: ∇ is gradient operator, we can see for all i ∈ S ,
( ) ( )
u e , X k −1 ≤ u X k . In addition, it prove that any strict local
i

( )= ∑
k
(1) (k)
u y , …, y π ( s1 ,…, sk ) ∏ y s( j )
j
(2) maximizer x ∈ ∆ of (5) is a ESS of Γ .
The hypergraph clustering problem of extracting ESSs
( s1 ,…, sk )∈s k j =1
(evolutionary stability) can be transformed into finding a strict
This function insensitive to order of inputs. We suppose solution of (5). Baum and Eagon26 are introduced to find
that randomly pick people from x ∈ ∆ to play game Γ , the ESSs.21 For maximizing polynomial function in probability
k
population average payoff is calculated by u X , X^k is a ( ) domain, The Baum-Eagon inequality is an effective iterative
shortcut for a sequence (x,x,. . .,x), and the average payoff method.
that an i-strategist obtains in a population is calculated by
( )
Theorem 1 (Baum-Eagon). Defined p(x ) as a homogene-
u e i , X k −1 . x ∈ ∆ is in equilibrium when the distribution ous polynomial in the variable xi with non-negative coeffi-
of the strategy does not change any more, it is in an equilib- cients, let x ∈ ∆ . Define the mapping t = Φ ( x ) , we can obtain
rium state. ti as follow:
Indeed, x ∈ ∆ is a Nash equilibrium if
xi ∂ i P ( x )
( i
u e ,X k −1
) ≤ u ( X ) , for all i ∈ S.
k (3) ti =
∑
n
x j ∂ j P (x )
, i = 1,…, n (7)
j =1

Hypergraph clustering based on game theory. H = (V , E, s) is a Then P ((Φ )) > P (x ) unless Φ ( x ) = x , Φ is the growth trans-
hypergragh clustering problem, V is a finite set of vertices, formation of the polynomial P.
V = {1,…, n} is a collection of microbes in various parts of the Using the above method, we iteratively find a cluster and
body, a node corresponds to a microorganism, which is also a set remove it from the object set, then repeat the process on the
of objects to be clustered. E is a set of hyperedges, s : E →  is rest of the objects until the remaining nodes cannot form a
a weight function that assigns a real value to a hyperedge, and it hyperedge after eliminating the nodes in the cluster.
also is a similarity to the set of objects in E. And Γ = (P, S, π) is
the corresponding clustering game. The payoff function p intui- The data processing. The traditional microbial network construc-
tively does this by rewarding k players based on the similarity of tion is based on graph-theory. In this paper, eight logical rela-
the items they play. As a result, assuming that ( v1 ,…, vk ) ∈V k tionships (see Table 1) among microorganisms are calculated
is group of objects chosen by k players, the payoff function can based on the method of entropy. Bowers et al27 proposed a com-
be defined as: putational method based on genomic data to identify the detailed
4 Evolutionary Bioinformatics 

Table 1. Description of the logical relationship between microbes and the total number of microbes present in the human body.

Type Wayne figure Logic description Total

type1 C is present if and only if both A and B are present 655

type2 C is present if A is absent or B is absent 4

type3 C is present if A is present or B is present 1890

type4 C is present if A is absent and B is absent 1

type5 C is present if A is present (absent) and B is absent (present) 354

type6 C is present if A is absent (present) or B is present (absent) 93

type7 C is present if one of either A or B is present 1334

type8 C is present if both A and B are present(absent) 36

relationships between proteins. Based on information entropy, expressed microorganisms, that is, remove the rows that meet
Bowers et al calculates correlations between proteins and identi-
fies eight logical relationships, which reveals many previously ∑m (i , j ) < 4 , we obtained the final microbial abundance matrix.
j
unknown higher-order relationships. We applied it to construct Select a logical relationship. We select type1. Based on type1, we
microbial higher-order networks. calculate the uncertainty coefficient of U (C | A ) , U (C | B ) ,
Table 1 describes logical relationships among microorgan- U (C | f ( A , B )) by formula (8), which predict microorganism C
isms and the total number of each logical relationship in human through two other microorganism A and B, and f ( A , B ) is the
body based on the selected dataset. Venn diagram and related logical combination and represents whether there is a logical rela-
logic statements in the Table 1 illustrate eight different logical tionship between microorganism A and microorganism B.
relationships, which represent different co-occurrence relation- Moreover, it is also required that neither microorganism a nor
ships among three Microorganisms. In other words, these microorganism B can predict microorganism C independently.
describe the possible dependence of the existence of C on the
existence of A and B. It can be seen from Table 1 that the most
( )
The maximum fraction value of U C | f ( A , B ) is obtained, and
we select triplets that satisfy the uncertainty between two microor-
common logical relationships among microorganisms in ganisms described C is weak (U (C | A ) < 0.3,U (C | B ) < 0.3 )
human body are type1, type3 and type7. The discussion of logi- and the uncertainty of describing C based on logical correlation is
cal relations with less relations is of little significance. We can strong (U (C | f ( A , B )) > 0.5 ).
choose type1, type2 or type3, but the number of type 3 rela-
tionships is large, and the interaction between microorganisms
 H ( X ) + H ( Y ) − H ( X , Y )
is more complex, which is not easy to analyze, so type 1 is U (X | Y ) =   (8)
finally selected. H (X )
Figure 1 shows the calculation process of high-order logical
relationship of microorganisms. M ( p, s ) represents microbial H ( X ) and H ( Y ) in equation (8) denote the entropy of
abundance matrix. The element value m ( i , j ) of M ( p, s ) repre- individual distribution, H ( X , Y ) denotes the entropy of joint
sents the abundance data of microbial i and sample j . The value distributions. U is between 0 and 1, when U is 1, X is a
of data that satisfies m ( i , j ) > 0 in M ( p, s ) is set to 1, We can deterministic function of Y , and when U is 0, X is com-
get a matrix with values of 0 and 1, then remove the weakly pletely independent of Y .
Yu et al 5

Figure 1. The calculation process of high-order logical relationship of microorganisms. (A) Extract microbial abundance data from open source websites
converted into 0 to 1 microbial abundance matrix. (B) Calculate the number of 8 high order logical relationships in the presence of microorganisms. (C)
Select type1. (D) Calculate the uncertainty coefficient U ( X |Y ) as the initial hyperedge weight by entropy. (F) The final microbial hyperedge similarity
matrix Ws is calculated through intra-class scatter matrix.

The maximum fraction value of uncertainty coefficient Where e is the natural logarithm function, δ is a positive
obtained by the triplet is taken as the initial weight of the parameter, we can obtain the new hyperedge similarity matrix
hyperdge. based on the scatter matrix s( e ) .

Construction of weight matrix s(e). Weight each hyperedge Results and Analysis
with a positive value, we can defined it with s(e ) . In our previ- Evaluation index
ous work, based on the idea of hypergraph clustering and the
Joint entropy. The joint entropy ( JE) is a measurement uncer-
idea that the samples are as dense as possible, and the smaller
tainty evaluation method to measure the uncertainty related to
the dispersion degree within the samples, the better the classi-
a set of variables. Take JE as a standard of measuring the infor-
fication ability of the categories, we proposed that the method
mation contained in a cluster. Joint entropy is a measure of
of reconstruct the microbial hyperedge similarity based on the
uncertainty associated with a set of variables, which is used to
intra-class scatter matrix to the analyze high-order microbial
measure the amount of information contained in the same
network modules.20
cluster of microorganisms. The smaller the joint entropy, the
M={ m1 , m2 , m3 ,…, mn } denotes a collection of microorgan-
smaller the amount of expression information. X 1 , X 2 ,…, X n
isms, where mi is dimensional vector, N j is the number of
denotes microorganisms, p( X 1 , X 2 ,…, X n ) denotes is the
samples j and M j is a set of samples j . M can be consider
probability that these microbes are present in the sample. The
as a relationship matrix of microorganisms nodes and hyper-
definition of JE can be expressed as:
edges. The rows of the matrix represent the hyperedges, the
column s of the matrix represent the microbes, if the microbes JE = HX 1 , X 2 ,…, X n =
node j belongs to the hyperedge i , then the value of M (i , j )
is the initial weight of the hyperedge. µj is the mean vector of − ∑ p ( x , x ,…, x ) logp ( x , x ,…, x )
x1 , x2 ,…, xn
1 2 n 1 2 n
(11)

the hyperedge j , We define the intra-class scatter matrix as:

where HX 1 , X 2 ,…, X n is the amount of information that
I= ∑ (m − µ )(m − µ )
m∈M j
j j
T
(9) were transferred by the average value of random variables per
batch { X 1 = x1 , X 2 = x 2 ,…, X n = xn } .

1
µj =
Nj ∑ m.
m∈M j
Total correlation. Total correlation (TC)28 is derived from
mutual information generalization. Mutual information is an
information measure in information theory, which can be
Trace (﹒) is the average measure of characteristic variance. regarded as the information contained by one random variable
The smaller trace (﹒) is, the smaller the characteristic variance is about another random variable. TC is an effective method to
and the higher the tightness within the class is. So, We can measure the independence between a set of microbes and to
construct the weight of the hyperedge in the following way: evaluate the degree of microbial interdependence within the
same module. When the total correlation is close to 0, the
 eI 
( )
s e j = 1/ (trace 
 δ ) (10) microorganisms in the module are statistically independent.
The lower the total correlation, the lower the interdependence,
 
6 Evolutionary Bioinformatics 

Table 2. The clustering analysis and comparison of mid vagina based on type1.

Algorithm HCGI HCIS

Cluster JE TC JE TC

1 5.869237 1.726926 5.869237 1.726926

2 4.539703 4.502064 6.036468 6.45178

3 3.068427 1.553164 4.984123 8.21099

4 6.318696 5.685631 3.644253 2.288655

5 5.860217 4.757338 7.025616 12.27174

6 4.902228 0.78664 – –

7 3.556522 0.433895 – –

Sum 34.11503 19.44566 27.5597 30.95009

Table 3. The clustering analysis and comparison of intestines tract based on type1.

Algorithm HCGI HCIS

Cluster JE TC JE TC

1 6.219162 3.650304 4.793256 0.568318

2 5.492254 1.661277 6.219162 3.650304

3 5.114142 0.8542 6.439209 5.489505

4 8.542772 21.46201 – –

Sum 25.36833 27.62779 17.45163 9.708126

thus ensuring the quality of the high-order interaction in the Tables 2 and 3 show that cluster analysis and comparison of
module. mid vagina and intestines tract based on type 1. We can see that
the two methods end up with different clustering numbers on
TC ( X1 , X 2 ,…, X n ) the same body part. From Table 2, we can observe that the sum
n (12) of JE generated by HCGI algorithm is higher than that of
= ∑H ( X ) − H ( X , X ,…, X )
i 1 2 n HCIS, while the sum of TC is lower than that of HCIS, which
i =1 indicates that the results generated by HCGI clustering algo-
rithm express more total information, and the correlation
Comparison of algorithms between the generated clusters is weaker. But in Table 3, the
In this paper, the joint entropy, total correlation and modularity are correlation between the generated clusters by HCIS is weaker,
taken as the indexes to evaluate the clustering performance of the this is because the nodes removed by the two methods are not
algorithm, the hyperspectral clustering based on game theory and consistent, so it is not good to use the above evaluation indica-
intra-class scatter matrix (HCGI) and hypergraph clustering based tors to evaluate the clustering performance of the two methods.
on intra-class scatter matrix (HCIS).24 They were applied to clus- However, according to joint entropy and total correlation of
ter analysis and comparison of microorganisms in 18 parts of each cluster, we can judge the amount of expressed information
human body. Based on the clustering results, the JE and TC of each of the cluster and the connection strength between nodes.
cluster were calculated. we take the sum of JE of the subclusters and Next, through the case analysis of mid vagina, we can visualize
the sum of TC as an final index to evaluate the performance of an the clustering results of mid vagina, and specifically analyze the
algorithm. The sum of JE represents the total expression informa- advantages and disadvantages of two methods.
tion of the microbial cluster, and the sum of TC represents the total
dependence within the microbial community. Visual analysis
We selected the mid vagina where the clustering effect did Cytoscape29 has proven to be a high-level platform for visuali-
not change much after the addition of intra-class scatter matrix zation and analysis of biological networks. We used cytoscape
and intestines tract where HCIS did not work well in to com- to analyze the clustering results of HCGI on mid vagina and
pare and analyze in the previous work.24 find the differences between nodes within each cluster.
Yu et al 7

Figure 2. The cluster results of HCIS on mid vagina.

Figure 3. The cluster results of HCGI on mid vagina.

Figure 3 shows the clustering results of HCGI algorithm on can find that HCGI divides more clusters than HCIS algo-
mid vagina. The nodes in the box inside the black dotted box rithm. However, it can be observed that the cluster generated
belong to the same cluster. HCGI algorithm divides the nodes by HCGI algorithm is contained in the cluster generated by
on mid vagina into 7 clusters. The blue nodes that are not HCIS algorithm. The nodes eliminated by HCGI algorithm
framed by the black dotted line do not belong to any of the are also nodes connected sparsely to the network graph, and the
clusters. Figure 2 shows the clustering results of HCIS on mid completeness of the higher-order organization is still main-
vagina. HCIS divides the nodes on mid vagina into 5 clusters. tained in the process of iterative elimination. Based on the
In Figure 2, the blue square nodes do not belong to any of the framework of game theory, we can get better and more con-
clusters. Compare the clustering results of Figures 2 and 3, we servative microbial modules.
8 Evolutionary Bioinformatics 

Conclusion References
Studying microbial community interactions is challenging 1. Lederberg J, Mccray AT. Ome SweetOmics–a genealogical treasury of words.
Scientist. 2001;15:22-27.
because the microorganisms in a community often interact 2. Baker BJ, Jillian FB. Microbial communities in acid mine drainage. Fems Micro-
with each other in many ways, not just in pairs. The basic way biol Ecol. 2003;44:139-152.
3. Andersen R, Chapman SJ, Artz RRE. Microbial communities in natural and
to detect the biological significance of biological networks is disturbed peatlands: a review. Soil Biol Biochem. 2013;57:979-994.
based on the module perspective. 4. Fuhrman JA. Microbial community structure and its functional implications.
Nature. 2009;459:193-199.
On the basis of hyperedge similarity matrix reconstructed 5. Bairey E, Kelsic ED, Kishony R. High-order species interactions shape ecosys-
by intra-class scatter matrix in our previous work, this paper tem diversity. Nat Commun. 2016;7:12285.
6. Sanchez-Gorostiaga A, Bajić D, Osborne ML, Poyatos JF, Sanchez A. High-
introduces game theory and proposes hypergraph clustering
order interactions dominate the functional landscape of microbial consortia. Pre-
based on game theory (HCGI) algorithm for mining micro- print. Posted online May 29, 2018. bioRxiv 333534. doi:10.1101/333534
bial higher-order module research. Hypergraph clustering 7. Wintermute EH, Silver PA. Emergent cooperation in microbial metabolism.
Mol Syst Biol. 2010;6:407.
problem were converted to non-cooperative multi-player 8. Mee MT, Collins JJ, Church GM, Wang HH. Syntrophic exchange in synthetic
game clustering problems, and get the optimal clustering microbial communities. Proc Natl Acad Sci USA. 2014;111:E2149-E2156.
9. Freilich S, Zarecki R, Eilam O, et al. Competitive and cooperative metabolic
module by finding countermeasures balance. There is no interactions in bacterial communities. Nat Commun. 2011;2:1-7.
need to specify the cluster number in advance, and one mod- 10. Vetsigian K, Jajoo R, Kishony R. Structure and evolution of Streptomyces inter-
action networks in soil and in silico. PLoS Biol. 2011;9:e1001184.
ule is obtained after each iteration until the remaining object 11. Mounier J, Monnet C, Vallaeys T, et al. Microbial interactions within a cheese
set cannot build the higher-order relationship. The experi- microbial community. Appl Environ Microbiol. 2008;74:172-181.
12. Berry D, Widder S. Deciphering microbial interactions and detecting keystone
mental results show that the introduction of game theory
species with co-occurrence networks. Front Microbiol. 2014;5:219.
makes the clustering module more conservative and the 13. Fisher CK, Mehta P. Identifying keystone species in the human gut microbiome from
quality of the module better, which indicates that the higher- metagenomic timeseries using sparse linear regression. PloS One. 2014;9:e102451.
14. Needham DM, Fuhrman JA. Pronounced daily succession of phytoplankton,
order module of hypergraph clustering based on game theory archaea and bacteria following a spring bloom. Nat Microbiol. 2016;1:1-7.
is effective. 15. Faust K, Raes J. Microbial interactions: from networks to models. Nat Rev
Microbiol. 2012;10:538-550.
16. Stein RR, Bucci V, Toussaint NC, et al. Ecological modeling from time-series
Future Work inference: insight into dynamics and stability of intestinal microbiota. PLoS
Comput Biol. 2013;9:e1003388.
This paper is mainly based on information entropy to build non 17. Grilli J, Barabás G, Michalska-Smith MJ, Allesina S. Higher-order interactions
directional high-order microbial interaction relationship. The stabilize dynamics in competitive network models. Nature. 2017;548:210-213.
18. Goldford JE, Lu N, Bajić D, et al. Emergent simplicity in microbial community
interaction relationship in real biological network may be direc- assembly. Science. 2018;361:469-474.
tional relationship, such as metabolic relationship. Therefore, we 19. Tsai K-N, Lin S-H, Liu W-C, Wang D. Inferring microbial interaction network
from microbiome data using RMN algorithm. BMC Syst Biol. 2015;9:54.
can analyze and construct the possible high-order relationship 20. Guo X, Boedicker JQ. The contribution of High-Order metabolic interactions to
between directed microorganisms from the perspective of the global activity of a four-species microbial community. PLoS Comput Biol.
2016;12:e1005079.
metabolism.
21. Girvan M, Newman MEJ. Community structure in social and biological net-
With the development of time series data, many dynamic works. Proc Natl Acad Sci USA. 2002;99:7821-7826.
models have been used in microbial networks. Dynamic 22. Copeland JK, Yuan L, Layeghifard M, Wang PW, Guttman DS. Seasonal com-
munity succession of the phyllosphere microbiome. Mol Plant Microbe Interact.
models such as dynamic Bayesian models can be used to con- 2015;28:274-285.
struct high-order interaction networks among 23. Zelezniak A, Andrejev S, Ponomarova O, Mende DR, Bork P, Patil KR. Meta-
bolic dependencies drive species co-occurrence in diverse microbial communi-
microorganisms. ties. Proc Natl Acad Sci USA. 2015;112:6449-6454.
24. Yu L, Shen X, Jiang X, Yang J, Yang Y, Zhong D. Hypergraph clustering based
on intra-class scatter matrix for mining higher-order microbial module. In:
Author Contributions 2019 IEEE International Conference on Bioinformatics and Biomedicine (BIBM),
LY and XS designed and implemented the computing frame- San Diego, CA, 18-21 November, 2019.
25. Bulò SR, Pelillo M. A game-theoretic approach to hypergraph clustering. Adv
work. LY and XS analyzed the results and wrote the manu- Neural Inf Process Syst. 2009;35:1571-1579.
script. LY, XS, JY, KW, DZ and RX revised the manuscript. 26. Baum LE, Eagon JA. An inequality with applications to statistical estimation for
probabilistic functions of Markov processes and to a model for ecology. Bull Am
LY prepared the computational codes and carried out. All the Math Soc. 1967;73:360-363.
authors wrote, reviewed and approved the final manuscript. 27. Bowers PM, Cokus SJ, Eisenberg D, Yeates TO. Use of logic relationships to
decipher protein network organization. Science. 2004;306:2246-2249.
28. Watanabe S. Information theoretical analysis of multivariate correlation. IBM J
Res Dev. 1960;4:66-82.
ORCID iD 29. Otasek D, Morris JH, Bouças J, Pico AR, Demchak B. Cytoscape automation:
Xianjun Shen https://orcid.org/0000-0001-5714-8848 empowering workflow-based network analysis. Genome Biol. 2019;20:185.