Community Finding with Applications on Phylogenetic Networks [Extended Abstract]

1
Abstract
With the advent of high-throughput sequencing methods, new ways of visualizing and analyzing increasingly amounts of data are
needed. Although some software already exist, they do not scale well or require advanced skills to be useful in phylogenetics.
The aim of this thesis was to implement three community finding algorithms – Louvain, Infomap and Layered Label Propagation
(LLP); to benchmark them using two synthetic networks – Girvan-Newman (GN) and Lancichinetti-Fortunato-Radicchi (LFR); to test
them in real networks, particularly, in one derived from a Staphylococcus aureus MLST dataset; to compare visualization frameworks
– Cytoscape.js and D3.js, and, finally, to make it all available online (mscthesis.herokuapp.com).
Louvain, Infomap and LLP were implemented in JavaScript. Unless otherwise stated, next conclusions are valid for GN and LFR.
In terms of speed, Louvain outperformed all others. Considering accuracy, in networks with well-defined communities, Louvain was
the most accurate. For higher mixing, LLP was the best. Contrarily to weakly mixed, it is advantageous to increase the resolution
parameter in highly mixed GN. In LFR, higher resolution decreases the accuracy of detection, independently of the mixing
parameter. The increase of the average node degree enhanced partitioning accuracy and suggested detection by chance was
minimized. It is computationally more intensive to generate GN with higher mixing or average degree, using the algorithm developed
in the thesis or the LFR implementation. In S. aureus network, Louvain was the fastest and the most accurate in detecting the
clusters of seven groups of strains directly evolved from the common ancestor.
Community Finding | Phylogenetic Networks | Data Visualization | Web Application
1. Introduction
According to The Economist, data has become the most
important resource in the world [1]. New ways to storage,
analyze and visualize increasing amounts of information are
becoming a pressing need.
Interactions among different elements of real systems can
be seen as networks [2]. From their topology or pattern of
connections, functional knowledge can be extracted.
Although empirical analysis can be useful in small networks,
for big data it is indispensable the use of algorithms to
uncover their properties. Consequently, mathematical models
called graphs are used to simplify their representation. Graph
theory provides all the tools available to analyze them. This
representation implies all the elements of the network and
their relations must be acquired. Generally, the closer to the
human population is the analysis, the more difficult it gets to
gather data. The two main reasons are the lack of non-invasive
acquisition methods or privacy issues [3]. In fact, in health,
legislation is becoming progressively stricter in respect to
patient’s data protection [4]. On one hand, it is fundamental
to assure ethical use of medical data, on the other, it may
constitute a hurdle for investigating the causes of many
pathologies. In the next lines, the discussion will focus in the
specific topic of infectious diseases.
As stated by the World Health Organization, Antimicrobial
resistance (AMR) is in the top ten of major threats demanding
attention in 2019 [5]. In the last years, AMR has been
potentiated by the use of drugs intended to treat diseases
caused by pathogenic organisms. Depending on the organism,
it is subdivided in antiviral (virus), antibiotic (bacteria),
antifungal (fungi) or antiparasitic (parasites) resistance.
Although resistance is a natural evolutionary process, in which
the most competent strains get selected over time, the
frequency of use, or even misuse, of antimicrobial drugs is
accelerating the growth of resistance rates for several
microbial species [6]. This is particularly relevant not only
among the human population, but also in farms where cattle
are being given preventive doses of antibiotics as growth
adjuvants. Moreover, the frequency people are travelling
around the world is facilitating the spread of resistant
organisms. Due to all of this and the high mortality of
Luís Rita1,3,4
Supervision Team: Alexandre Francisco1,2
, João Carriço3
, Vítor Borges4
(1) Instituto Superior Técnico, Universidade de Lisboa, Lisboa, Portugal
(2) INESC-ID, R. Alves Redol 9, 1000-029 Lisboa, Portugal
(3) Instituto de Microbiologia, Instituto de Medicina Molecular, Faculdade de Medicina, Universidade de Lisboa, Lisboa, Portugal
(4) Departamento de Doenças Infeciosas, Instituto Nacional de Saúde Dr. Ricardo Jorge, Lisboa, Portugal
Community Finding with Applications on Phylogenetic Networks

2
infectious diseases (Figure 1), it is fundamental to empower
central health authorities with scalable and easy-to-use tools
to allow their intervention in short time scales [7].
Figure 1 Infections are still one of the major causes of death (*).
Data collected by WHO between 2000-2015 from a population of
over 90 000 people, in WHO member states. Adapted from [8].
Next Generation Sequencing methods, developed in the
last years, have been providing us massive quantities of
microbiological data [9]. In fact, due to the high discriminative
power of whole genome sequencing methods, it has been
possible to infer more precise evolutionary relationships
among strains. Consequently, allelic profiles with higher
number of genes keep increasing and, with them, more
complete phylogenetic trees or networks depicting
evolutionary distances among strains can be assessed. All of
this has been useful in the multiple stages of disease
prevention: vaccine design, assess the pathogenicity of a
sequenced strain, detect outbreaks and in infection control.
As well as diagnosis, enhancing the detection of mixed
infections (caused by multiple genetically distinct organisms)
[10].
Generally, the goal of this thesis was to extract functional
knowledge from the topological structure of phylogenetic
networks. For instance, to be able to infer properties like
resistance to antimicrobials, based on the similarity and
characteristics of the other organisms in the network.
Contrarily to random networks, real ones present an
asymmetrical distribution of edges. From this asymmetry,
clusters of nodes can be identified. By definition, communities
are regions in the network with higher density of edges.
Nonetheless, this definition is not restrictive enough to
undoubtedly identify them in graphs. This raises the following
question: how can we detect communities if it is not clear
what we are looking for? To answer this question, community
finding theory has been developed in the past years. In spite
there has been a considerable advance in terms of accuracy
and speed of detection, some algorithms perform better than
others in different networks.
1.1. Objectives
The aim of the thesis was to implement three community
finding algorithms in JavaScript – Louvain, Infomap and
Layered Label Propagation (LLP); to benchmark them, in terms
of accuracy and speed, against two synthetic networks –
Girvan-Newman (GN) and Lancichinetti-Fortunato-Radicchi
(LFR) networks; to perform additional testing using an
Amazon, Zachary’s Karate Club and Staphylococcus aureus
Multilocus Sequence Typing (MLST) Single Locus Variant (SLV)
network; to compare different visualization frameworks in
terms of their tools and robustness to big data- D3.js (using
SVG and Canvas elements) and Cytoscape.js, and to create a
web application to allow network visualization (before and
after running community finding algorithms) and plot all data
obtained from the benchmark tests.
1.2. Structure Outline
In Methodology, the implemented community finding
algorithms, implemented benchmark tools, real test data and
data visualization libraries are introduced. In Results,
community finding algorithms are analyzed in terms of
accuracy and speed. The speed of generation of benchmark
networks is determined based on the different properties of
the networks. Different web visualization frameworks are
compared. And a S. aureus MLST SLV network is partitioned
using different algorithms and input parameters. In
Conclusion, a retrospective analysis on whether the goals
were attained or not is made, main difficulties, future work to
be performed in INSaFLU web application is detailed and a set
of prospective improvements to the work already done is
suggested.
2. Methodology
2.1. Community Finding
Louvain Algorithm
This algorithm, which was implemented in the thesis, is
divided in 2 phases: Modularity Optimization and Community
Aggregation [11]. Together, they are considered 1 pass. After
the first step is completed, the second follows. Both are
executed until there are no more changes in the network and
maximum modularity is achieved.
Modularity Optimization – the algorithm will randomly
order all nodes in the network such that, one by one, it will
remove and insert them in a different community 𝐶. This will
continue until no significant increase of modularity (input
parameter) is verified:
∆𝑀 = [
𝛴𝑖𝑛 + 𝑘𝑖,𝑖𝑛
2𝑚
− (
𝛴𝑡𝑜𝑡 + 𝑘𝑖
2𝑚
)
2
]
− [
𝛴𝑖𝑛
2𝑚
− (
𝛴𝑡𝑜𝑡
2𝑚
)
2
− (
𝑘𝑖
2𝑚
)
2
]
(1)
Let 𝛴𝑖𝑛 be the sum of the weights of the links inside 𝐶, 𝛴𝑡𝑜𝑡
the sum of the weights of all links to nodes in 𝐶, 𝑘𝑖 sum of the
weights of all links incident in node 𝑖, 𝑘𝑖,𝑖𝑛 the sum of the
weights of links connecting node 𝑖 and nodes in the
community 𝐶 and 𝑚 is the sum of the weights of all edges in
the graph.

3
Community Aggregation – After finishing the first step, all
nodes belonging to the same community are merged into a
single giant node. Links connecting giant nodes are the sum of
the ones previously connecting nodes from the same different
communities. This step also generates self-loops which are
the sum of all links inside a given community before being
collapsed into one node.
Infomap Algorithm
Similarly to Louvain, Infomap algorithm was also
implemented. It partitions the network using a different
quality function – minimum description length [12]:
𝐿(𝑀) = ( ∑ 𝑞 𝑚
𝑚∈𝑀
) 𝑙𝑜𝑔 ( ∑ 𝑞 𝑚
𝑚∈𝑀
)
− 2 ∑ 𝑞 𝑚 𝑙𝑜𝑔(𝑞 𝑚)
𝑚∈𝑀
− ∑ 𝑝 𝛼 𝑙𝑜𝑔(𝑝 𝛼)
𝛼∈𝑉
+ ∑ (𝑞 𝑚 + ∑ 𝑝 𝛼
𝛼∈𝑚
)
𝑚∈𝑀
𝑙𝑜𝑔 (𝑞 𝑚
+ ∑ 𝑝 𝛼
𝛼∈𝑚
)
(2)
Let 𝑞 𝑚 be the exit probability of a module 𝑚 and 𝑝 𝛼 the
relative weight 𝑤 𝛼 that is computed dividing the total weight
of the edges connected to 𝛼 by twice the total weight of all
links in the graph.
Layered Label Propagation Algorithm
LLP algorithm development was based in Label Propagation
(LP) [13]. The latter starts by assigning a different community
to each node in the network and then, iteratively, it modifies
the respective community based on the dominant one from
the nodes in the immediate neighborhood. In each iterate,
this step is executed for all nodes, in a previously randomized
order.
LLP not only takes into account the nodes in the
neighborhood, but also the ones in the remaining network
[14]. The iterative process is the same, the difference resides
in the value it optimizes. In this case, the assigned community
is the one that maximizes:
𝑘𝑖 − 𝛾(𝑣𝑖 − 𝑘𝑖) (3)
Being 𝑘𝑖 the number of nodes with label 𝜆𝑖 in the
neighborhood of a given node and 𝑣𝑖 the total number in the
whole graph labeled the same way.
This approach, which was also implemented, buffers the
number of nodes belonging to a certain community (in the
node’s neighborhood), by its usual presence across the whole
network.
Table 1 Time complexity of community finding algorithms.
Algorithm Time Complexity Reference
Louvain 𝑂(𝐿) [11]
Infomap 𝑂(𝑁𝑙𝑜𝑔𝑁) [12]
LP 𝑂(𝐿) [13, 15]
LLP 𝑂(𝐿) [14]
2.2. Benchmark
Different community finding algorithms partition the same
graph in different ways. Consequently, it is important to
understand which algorithms are the most indicated.
An accurate way to gain this insight is by partitioning
networks whose community structure is known. GN and LFR
benchmark networks are presented in the next sub-sections,
as well as NMI which estimates the similarity between the
original and the detected partitions.
Community finding algorithms were not only tested in
terms of accuracy, but also speed. All benchmark tests were
repeated 10 times. The mean values and corresponding
confidence intervals at 95% were included in Results.
Girvan-Newman Network
In GN network, 128 nodes are divided in 4 groups, each one
with exactly 32 nodes [16]. Then, a mixing parameter, chosen
by the user, defines the probability of nodes from different
communities being connected. This probability is given by:
𝜇 =
𝑘 𝑒𝑥𝑡
𝑘 𝑖𝑛𝑡 + 𝑘 𝑒𝑥𝑡
(4)
Each node is connected to 16 others (Figure 2). The
algorithm hereby implemented not only allows the mixing
parameter to vary, but also the average degree. This led to an
increased number of benchmark tests using these networks.
Figure 2 GN synthetic network. 𝑁 = 128, 𝜇 = 0.1 and 𝑘 = 16.
Represented using D3.js and SVG.
Lancichinetti-Fortunato-Radicchi Network
LFR benchmark networks (Figure 3) [17, 16] try to better
approximate the real ones. This means that not only they

4
consider a power-law distribution of community sizes with
coefficient 𝜁 (5), but also a power-law distribution of node
degrees with coefficient 𝛾 (6):
𝑝 𝑁 𝑐
~𝑁𝑐
−𝜁
(5)
𝑝 𝑘~𝑘−𝛾
(6)
Figure 3 LFR synthetic network. 𝑁 = 1000, 𝜇 = 0.1, 𝑘 𝑎𝑣𝑔 = 15,
𝑘 𝑚𝑎𝑥 = 50, 𝑐 𝑚𝑖𝑛 = 20 and 𝑐 𝑚𝑎𝑥 = 50. Represented using D3.js
and SVG.
Normalized Mutual Information
The clustering quality of the community finding algorithms
was tested using Normalized Mutual Information (NMI) [18].
An algorithm was developed, in JavaScript, in the thesis and
tested in several networks. It receives an input of 2 arrays,
with the same length, and returns the corresponding NMI:
𝑁𝑀𝐼(𝑌, 𝐶) =
2 × 𝐼(𝑌; 𝐶)
[𝐻(𝑌) + 𝐻(𝐶)]
(7)
𝐼(𝑌; 𝐶) = 𝐻(𝑌) − 𝐻(𝑌|𝐶) (8)
NMI is dependent on the mutual information 𝐼, the
conditional entropy 𝐻(𝑌|𝐶) and the entropy of the labeled
𝐻(𝑌) and clustered set 𝐻(𝐶).
2.3. Real Test Data
The higher the number of tests performed using the
implemented algorithms, the higher the certainty they are
working properly for the broadest number of networks. This
way, disconnected networks (sampled from an Amazon
network), small sized ones (Zachary’s Karate Club) and a
phylogenetic network (S. aureus MLST SLV) were additionally
considered. Nevertheless, no benchmark tests were
performed using those.
Amazon Network
This network was obtained after crawling Amazon website
[19]. Each node represents a given product available in the
store and a connected pair means they are frequently bought
together. The original network contains 334 863 nodes and
925 872 links. Only the first 5 000 links along the respective
nodes were used in the tests (Figure 4).
Figure 4 Amazon product co-purchasing network sampled from a 5
000 links graph. Represented using D3.js and SVG.
Zachary’s Karate Club Network
This is a social network representing 77 relations between 34
individuals from a university karate club [20].
Node 1 is the instructor and 34 is the president (Figure 5).
Dark and light blue nodes represent the split inside the club
that was created due to a conflict between them. Each link
connects 2 members when they used to meet regularly
outside the club, before splitting.
Figure 5 Zachary’s karate club network. Represented using D3.js
and SVG.
Staphylococcus aureus
To test the application of the community finding algorithms in
a phylogenetic network, an MLST database of allelic profiles
for S. aureus was used (Figure 6) [21]. The MLST profiles
contain information about seven loci (arcC, aroE, glpF, gmk,
pta, tpi and yqiL) and, at the time of download, it had 5199
MLST profiles (ST). A graph was created by linking each ST to
all of its SLVs and the largest component was used for the
benchmark of the community finding algorithms. The choice
of this test network was based on the fact it can emulate
similar patterns to the ones that are created using cgMLST or
wgMLST profiles, which have many more loci than MLST.

5
Figure 6 S. aureus MLST SLV network. Represented using D3.js and
Canvas.
2.4. Visualization Frameworks
In the visualization application, the user can opt by visualizing
every network using D3 or Cytoscape JavaScript libraries. In
the case of the first framework, there is the possibility of
visualizing data through a Canvas or an SVG element. All
networks were plotted using a force directed implementation.
This mode requires the developer to set features like the
gravity, repulsion or spring constant which will affect how the
network is displayed. Depending those parameters, node and
link overlapping should be reduced in order to get an
enhanced view of the graph (Figure 7).
Figure 7 Zachary's Karate Club network represented using D3.js
Canvas (top-left), D3.js SVG (top-right) and Cytoscape.js (bottom).
Regarding benchmark plotting, D3.js was used. Its
flexibility in terms of visual representation, performance and
the number of practical examples available on the web,
determined this choice.
3. Results
3.1. Web Application
In order to facilitate the analysis and visualization of the
results after the execution of the previous algorithms, a web
application was implemented (mscthesis.herokuapp.com).
Its backend was conceived using Node.js and it runs in a
Heroku server which is linked to a GitHub repository
(github.com/warcraft12321/Thesis) containing all the
implementations and documents related to the thesis. A
Digital Object Identifier (DOI) was attributed to this repository
using Zenodo (zenodo.org/badge/latestdoi/162063699). An
image of the app is available at Docker Hub
(cloud.docker.com/u/warcraft12321/repository/docker/warc
raft12321/thesis). The following actions are strictly executed
in the cloud:
• Generation of GN and LFR benchmark networks. The
latter are generated using a binary from a C++
implementation [22]. This was possible after
importing a package from NPM which simulates the
command line in the server;
• The real test networks had their data properly
processed in the cloud, so that it could be sent and
displayed in the frontend;
• Once the execution of Louvain, Infomap and LLP
algorithms is not performed in the user’s browser,
this may raise some privacy issues. Although, it was
not feasible to run Infomap in the user side and still
keep the web application functional in all devices;
• The communication between the app and NPM was
established in the server, so that it is possible to
import and plot the statistics from each package
uploaded to NPM in the application;
• Finally, all benchmark data was obtained after
running the algorithms in the cloud.
The frontend is both static and dynamic. The static
counterpart uses HTML, CSS and JavaScript. The dynamic
execution is managed by Node.js, which runs in the server-
side.
3.2. Community Finding Algorithms
In terms of accuracy, Louvain outperformed all others for
weakly mixed networks (𝜇 < 0.4). It is followed by Infomap,
LP and LLP (γ = 0.5) for 𝜇 < 0.3 or Infomap, LLP and LP (γ =
0.5) for 0.3 ≤ 𝜇 ≤ 0.4. By increasing 𝜇, LLP, which is able to
detect communities based on the panorama of the whole
network, consistently becomes the mosteffective for 𝜇 ≥ 0.5.
Comparing Louvain and Infomap when 𝜇 ≥ 0.5, their
performance is similar. Based on the CI 95%, it is not possible
to state one performs better than the other. Another point to
highlight is that Louvain and Infomap are very sensitive,
consequently, unreliable detecting communities in networks
with 𝜇 ≈ 0.5. LLP and LP do not present any specific value in
which community detection is steeply affected. These
conclusions are all valid to GN and LFR networks (Figure 8 and
Figure 9).

6
Figure 8 Clustering quality of each algorithm: Louvain, Infomap, LP
and LLP in terms of the mixing parameter in GN networks, with the
remaining properties as in Figure 2.
Figure 9 Clustering quality of each algorithm: Louvain, Infomap, LP
and LLP in terms of the mixing parameter in LFR networks, with the
remaining properties as in Figure 3.
Having considered only two gamma values in LLP, a more
precise analysis was needed to check whether there could be
others achieving better results.
In the case of GN network (Figure 10), it is possible to
detect communities with higher accuracy in networks less
mixed. The performance decreases whenever we increase 𝛾
for 𝜇 < 0.5. When 𝜇 ≥ 0.5, the NMI between the detected
partition and the one in the original network is higher as long
as we keep increasing 𝛾. Another point to highlight is the
possibility of detecting communities with higher precision in
networks with higher 𝜇 using appropriate 𝛾 than in others
with lower 𝜇 but with a non-optimal 𝛾. This is valid for
networks with 𝜇 > 0.5 (Figure 11).
Figure 10 GN network (Figure 2) after running LLP. When 𝛾 is closer
to 0, the coarse structure of the network with few, big and sparse
communities is highlighted. As γ grows, fine-grained structure is
unveiled. Communities become smaller and denser. Represented
using D3.js and SVG.
Figure 11 LLP algorithm accuracy in terms of the resolution
parameter. Analysis performed for mixing parameters 0 – 1 (blue-
brown) in GN network, with the remaining properties as in Figure 2.
In LFR networks, increasing gamma always leads to a
decrease in the NMI (Figure 12).
Figure 12 LLP algorithm accuracy in terms of the resolution
parameter. Analysis performed for mixing parameters 0 – 1 (blue-
brown) in LFR network, with the remaining properties as in Figure 3.
The influence of the average node degree was tested to
check whether it affects the capacity of each algorithm to
identify communities.
In the case of GN networks, for 𝜇 ≤ 0.5, an increased
average degree enhances the capacity of Louvain, Infomap, LP
and LLP to differentiate communities. For 𝜇 > 0.5, NMI
clustering quality is lower as higher is the average node
degree. Assuming it is not expected the algorithms to identify
the original communities for 𝜇 ≈ 1, this suggests that higher
average degrees tends to decrease detection by chance. This
observation can be applied to GN and LFR networks (Figure
13).

7
Figure 13 LLP algorithm accuracy in terms of the mixing parameter.
GN network with average node degrees: 15, 20 and 25. The
remaining properties are similar to the ones in Figure 2.
In terms of speed, Louvain algorithm performed
significantly better than any other. One important property
that contributes to its efficiency is that it only calculates
modularity variation at each iteration of Modularity
Optimization phase (1). In Infomap, minimum description
length is recalculated at the end of each iteration of the
optimization step (2). Due to this computationally intensive
step, Infomap presented the poorest performance in the
benchmark tests. Based on Table 1, it was expected the
difference of execution time between Infomap and all others
to be logarithmic. Due to inefficient steps in the
implementation, this was not achieved. Comparing LLP and LP,
the fastest was the one that needed to optimize the simplest
equation – LP (Figure 14). Although, time complexity appears
to be the same, as predicted in Table 1. Again, using the same
table, it was expected LLP and LP to execute about as fast as
Louvain. The small difference observed in Figure 14 might be
explained by some non-optimized step in the implementation
of LLP/LP.
Figure 14 Time that Louvain, Infomap, LLP and LP took to finalize
the analysis in terms of the size of the LFR networks. The remaining
properties are similar to the ones in Figure 3.
3.3. Benchmark Networks Algorithms
Previous GN and LFR networks were generated so the
previous tests could be performed. The algorithm to create
the former was implemented in the thesis. The time it takes
to generate such networks increases exponentially with the
average node degree and the mixing parameter (Figure 16).
Becoming unfeasible to execute it for average degrees higher
than 20, in an ordinary computer. Its performance was
compared to the Fortunato’s implementation which is highly
scalable and allows the generation of networks with much
higher number of links in shorter times (Figure 15). It was
verified an approximately linear increase in the time of
generation of networks with progressively higher mixing
parameters, contrarily to the one implemented in the thesis.
Just like before, generating networks with more edges,
requires additional computational power.
Figure 15 Time needed to generate GN networks (using LFR
implementation) in terms of their mixing parameter. Analysis
performed in networks with average node degrees: 15, 20 and 25.
The remaining properties are similar to the ones in Figure 2.
Figure 16 Time needed to generate GN networks (using thesis
implementation) in terms of their mixing parameter. Analysis
performed in networks with average node degrees: 15 and 20. The
remaining properties are similar to the ones in Figure 2.
Additional tests were performed in LFR networks. The
execution time was tested against the number of nodes and

8
the mixing parameter of the network. In both cases, it was
verified an approximately linear relation (Figure 17 and Figure
18). The number of nodes was notconsidered in GN networks,
once they have a fixed number of elements per community
and of communities. In spite only the mixing parameter and
the number of nodes were analyzed, the user can still choose
to adjust nodes maximum degree, the exponents of the nodes
degree distribution and communities size distribution, the
maximum/minimum number of nodes per community, the
number of overlapping nodes per community (covers were
not considered in the thesis) and the number of memberships
for the overlapping nodes [22].
Figure 17 Time needed to generate LFR networks in terms of their
number of nodes. Analysis performed in networks with average
node degrees: 15, 20 and 25. The remaining properties are similar
to the ones in Figure 3.
Figure 18 Time needed to generate LFR networks in terms of their
mixing parameter. Analysis performed in networks with average
node degrees: 15, 20 and 25. The remaining properties are similar
to the ones in Figure 3.
Every community finding, benchmark network generator,
accuracy measure and related algorithms were made
available in NPM (Table 2). This way, they can be used in any
user-specific application.
Table 2 Algorithms implemented in the thesis were made available
in NPM.
Algorithm Package Name URL
Louvain
louvain-
algorithm
npmjs.com/package/louvain-
algorithm
Infomap infomap npmjs.com/package/infomap
LLP
layered-label-
propagation
npmjs.com/package/layered-
label-propagation
GN Network
Generator
girvan-newman-
benchmark
npmjs.com/package/girvan-
newman-benchmark
NMI
normalized-
mutual-
information
npmjs.com/package/normalized-
mutual-information
Hamming
Distance
hamming-dist
npmjs.com/package/hamming-
dist
3.4. Visualization Frameworks
D3.js (SVG) allowed the user to visualize all networks used in
the web application, as well as to represent the benchmark
plots with the highest resolution possible (limited by density
of pixels in the user’s device). In terms of performance, it is
the second fastest.
D3.js (Canvas) is the fastest among the analyzed
frameworks. Nevertheless, when compared to the previous,
the nodes and links from the network lose resolution.
Cytoscape.js is advisable when network analysis will be
performed along with its representation. There is a
considerable number of algorithms, including community
finding ones, that can facilitate the study of the graph. The
lack of working examples and the slow processing of network
data made Cytoscape.js the less efficient. Representation of S.
aureus SLV network and others with similar or higher number
of nodes/edges overflows the web application.
3.5. Staphylococcus aureus
Louvain, Infomap and LLP algorithms were executed in the
biggest component of the SLV graph obtained from the MLST
dataset of S. aureus.
Louvain algorithm was executed using three modularity
variation thresholds: 0.1, 0.02 and 0.01. It is possible to verify
that high modularity partitions are rapidly achieved (2 passes
needed to obtain final result), as predicted in Introduction.
This characteristic makes modularity-based algorithms fast.
Using Infomap algorithm, several minimum description
length thresholds were considered: 0.1, 0.02 and 0.01. For all
values, every small dense group of nodes descending from the
same ancestor was separated in different communities. Thus,
it was not further considered.
LLP algorithm considered three values for the resolution
parameter: 0, 0.5 and 1. Similarly to Infomap, it partitioned
the network in several small and dense communities.

9
From the benchmark analyses of Louvain, Infomap and LLP
against GN network (Figure 8 and Figure 9), it was predictable
Louvain would perform better than the others, once it run in
a network with well-defined communities. After adjusting the
stopping parameter of Louvain, such that the communities
identified were the most pertinent – nodes from densely
connected regions descendent from the same ancestor
belonging to the same community, a final partition was
obtained (Figure 19). It is important to note that some clusters
analogous to the one highlighted with a square were
subdivided by the algorithm. One explanation is the presence
of high recombination rates between strains of different sub-
groups.
By uploading to PHYLOViZ Online the generated metadata
file with the nodes labeled accordingly to the inferred
communities, one may observe the original phylogenetic tree
with the same communities identified.
Figure 19 Partition obtained after running Louvain algorithm
(considering different thresholds for the modularity variation) on
Clonal Complex 0 (as identified in PHYLOViZ 2) of S. aureus MLST
SLV network. Represented using D3.js and Canvas.
4. Conclusion
Louvain, Infomap and LLP algorithms were implemented in
JavaScript. Unless otherwise stated, next conclusions are valid
for GN and LFR networks. In terms of speed, Louvain
outperformed all others. In terms of accuracy, in networks
with well-defined communities, Louvain was the most
accurate. For higher mixing, LLP was the best. Contrarily to
weakly mixed, it is advantageous to increase the resolution
parameter in highly mixed GN networks. In LFR networks,
higher resolution decreases the accuracy of detection,
independently of the mixing parameter. The increase of the
average node degree enhanced partitioning accuracy,
suggesting detection by chance was minimized. It is
computationally more intensive to generate GN networks
with higher mixing or average degree, using the algorithm
developed in the thesis or the LFR implementation. In the S.
aureus MLST SLV network, Louvain was the fastest and the
most accurate in detecting the clusters of 7 groups of strains
evolved from the same ancestor. The lower the modularity
variation threshold, the better the detection.
Along with the three community finding algorithms, GN
benchmark network generator, NMI and hamming distance
algorithms were made available in NPM. The web application
is hosted in the Heroku cloud platform. An image containing
all the required modules to run the app in a local machine can
be download from Docker Hub. All the implementations and
a complete roadmap of the thesis was made available in
GitHub. A Digital Object Identifier (DOI) was attributed to this
repository using Zenodo.
One of the main difficulties encountered was running
community finding algorithms in the user-side. A feature that
would rise less privacy concerns. It was necessary to run the
algorithms in the server, so that the application would not
crash. It was not possible to reach, in the Infomap
implementation, the computational efficiency predicted by its
time complexity. The online visualization of networks with
more than 10 000 nodes was unfeasible, even using the
framework that showed the best performance – D3.js (using
Canvas element).
During the next months, some of the tools developed in
this thesis will be implemented in INSaFLU. New visualization
frameworks for displaying phylogenetic trees, along with the
uploaded metadata, will enhance the traceability of the
evolutionary path among influenza strains. A geographical
map dynamically varying with time will allow the user to
precise the moment and location at which each sample was
obtained and enhance flu surveillance during the critical
season. An interactive histogram displaying each nucleotide
belonging to a consensus sequence obtained from a sample
of influenza virus with the SNPs highlighted, as well as with
their effects in the phenotype of the strain specified. New flu
surveillance dynamic reports will allow the visualization of
increased quantities of data in a minimalistic way.
As future improvements, the implemented algorithms can
be optimized in terms of memory and speed. Network and
chart visualization may run faster in a more efficient
implementation. The previous updates will make the web
application to load quicker and run smoother. To benchmark
community finding algorithms with a more representative set
of networks (spanning a wider range of topological

10
properties) would make possible to infer, with a higher
certainty, if any of the community finding algorithms
consistently performs better in phylogenetic networks. The
integration of community finding functionalities developed in
the thesis in a phylogenetic/surveillance-oriented analysis
web application like PHYLOViZ Online or INSaFLU would be
beneficial.
References
.
[1] "The world’s most valuable resource is no longer oil, but
data," The Economist, 6 May 2017. [Online]. Available:
https://www.economist.com/leaders/2017/05/06/the-
worlds-most-valuable-resource-is-no-longer-oil-but-
data. [Accessed 28 May 2019].
[2] A.-L. Barabási, "Network Science Book," [Online].
Available: http://networksciencebook.com. [Accessed
15 May 2019].
[3] H. G. S. Patil, A. N. Babu and P. S. Ramkumar, "Non-
invasive data acquisition and measurement in bio-
medical technology: An overview," in Maximizing
Healthcare Delivery and Management through
Technology Integration, IGI Global, 2016.
[4] "Health," EUROPEAN DATA PROTECTION SUPERVISOR,
[Online]. Available: https://edps.europa.eu/data-
protection/our-work/subjects/health_en. [Accessed 9
June 2019].
[5] "Ten threats to global health in 2019," World Health
Organization, [Online]. Available:
https://www.who.int/emergencies/ten-threats-to-
global-health-in-2019. [Accessed 4 June 2019].
[6] "Antimicrobial resistance," World Health Organization,
15 February 2018. [Online]. Available:
https://www.who.int/en/news-room/fact-
sheets/detail/antimicrobial-resistance. [Accessed 29
May 2019].
[7] Z. A. Memish, S. Venkatesh and A. M. Shibl, "Impact of
travel on international spread of antimicrobial
resistance," International Journal of Antimicrobial
Agents, vol. 21, no. 2, pp. 135-142, 2003.
[8] "Top 10 Leading Causes of Death Globally," [Online].
Available:
https://www.theatlas.com/charts/HkLaDreuW.
[Accessed 12 May 2019].
[9] B. Ribeiro-Gonçalves, A. P. Francisco, C. Vaz, M. Ramirez
and J. A. Carriço, "PHYLOViZ Online: web-based tool for
visualization, phylogenetic inference, analysis and
sharing of minimum spanning trees," Nucleic Acids
Research, vol. 44, no. 1, pp. 246-251, 2016.
[10] Y. Motro and J. Moran-Gilad, "Next-generation
sequencing applications in clinical bacteriology,"
Biomolecular Detection and Quantification, vol. 14, p.
1–6, 2017.
[11] V. D. Blondel, J.-L. Guillaume, R. Lambiotte and E.
Lefebvre, "Fast unfolding of communities in large
networks," J. Stat. Mech. (2008) P10008, p. 12, 2008.
[12] M. Rosvall, D. Axelsson and C. T. Bergstrom, "The map
equation," The European Physical Journal Special Topics,
vol. 178, no. 1, pp. 13-23, 2009.
[13] N. Raghavan, R. Albert and S. Kumara, "Near linear time
algorithm to detect community structures in large-scale
networks," Physical Review E 25th Anniversary
Milestones, vol. 76, no. 3, 2007.
[14] P. Boldi, M. Rosa, M. Santini and S. Vigna, "Layered label
propagation: a multiresolution coordinate-free ordering
for compressing social networks," in WWW '11
Proceedings of the 20th international conference on
World wide web, 2011.
[15] L. Šubelj, "Label propagation for clustering," in Advances
in Network Clustering and Blockmodeling, New York,
Wiley, 2018.
[16] M. Girvan and M. E. J. Newman, "Community structure
in social and biological networks," Proceedings of the
National Academy of Sciences of the United States of
America, vol. 99, no. 12, pp. 7821-7826, 2002.
[17] A. Lancichinetti, S. Fortunato and F. Radicchi,
"Benchmark graphs for testing community detection
algorithms," Physical review. E, Statistical, nonlinear,
and soft matter physics., vol. 78, no. 4, 2008.
[18] A. Lancichinetti, S. Fortunato and J. Kertesz, "Detecting
the overlapping and hierarchical community structure
of complex networks," New Journal of Physics, vol. 11,
2009.
[19] J. Yang and J. Leskovec, "Defining and Evaluating
Network Communities based on Ground-truth," in
Proceedings of 2012 IEEE International Conference on
Data Mining (ICDM), 2012.
[20] W. Zachary, "An Information Flow Model for Conflict and
Fission in Small Groups," Journal of anthropological
research, vol. 33, 1976.
[21] "Staphylococcus aureus MLST Databases," PubMLST, 5
June 2019. [Online]. Available:
https://pubmlst.org/saureus/. [Accessed 5 June 2019].
[22] A. Lancichinetti and S. Fortunato, "Benchmarks for
testing community detection algorithms on directed
and weighted graphs with overlapping communities,"
Physical review. E, Statistical, nonlinear, and soft matter
physics., vol. 80, 2009.

Community Finding with Applications on Phylogenetic Networks [Extended Abstract]

More Related Content

What's hot (20)

Similar to Community Finding with Applications on Phylogenetic Networks [Extended Abstract] (20)

More from Luís Rita (20)

Recently uploaded (20)

Community Finding with Applications on Phylogenetic Networks [Extended Abstract]