João C.

Setubal
Waldeyr Mendes Silva (Eds.)

Advances
LNBI 12558

in Bioinformatics and
Computational Biology
13th Brazilian Symposium on Bioinformatics, BSB 2020
São Paulo, Brazil, November 23–27, 2020
Proceedings

123
Lecture Notes in Bioinformatics 12558

Subseries of Lecture Notes in Computer Science

Series Editors
Sorin Istrail
Brown University, Providence, RI, USA
Pavel Pevzner
University of California, San Diego, CA, USA
Michael Waterman
University of Southern California, Los Angeles, CA, USA

Editorial Board Members

Søren Brunak
Technical University of Denmark, Kongens Lyngby, Denmark
Mikhail S. Gelfand
IITP, Research and Training Center on Bioinformatics, Moscow, Russia
Thomas Lengauer
Max Planck Institute for Informatics, Saarbrücken, Germany
Satoru Miyano
University of Tokyo, Tokyo, Japan
Eugene Myers
Max Planck Institute of Molecular Cell Biology and Genetics, Dresden,
Germany
Marie-France Sagot
Université Lyon 1, Villeurbanne, France
David Sankoff
University of Ottawa, Ottawa, Canada
Ron Shamir
Tel Aviv University, Ramat Aviv, Tel Aviv, Israel
Terry Speed
Walter and Eliza Hall Institute of Medical Research, Melbourne, VIC, Australia
Martin Vingron
Max Planck Institute for Molecular Genetics, Berlin, Germany
W. Eric Wong
University of Texas at Dallas, Richardson, TX, USA
More information about this subseries at http://www.springer.com/series/5381
João C. Setubal Waldeyr Mendes Silva (Eds.)
•

Advances
in Bioinformatics and
Computational Biology
13th Brazilian Symposium on Bioinformatics, BSB 2020
São Paulo, Brazil, November 23–27, 2020
Proceedings

123
Editors
João C. Setubal Waldeyr Mendes Silva
University of São Paulo Instituto Federal de Goiás
São Paulo, Brazil Formosa, Goiás, Brazil

ISSN 0302-9743 ISSN 1611-3349 (electronic)

Lecture Notes in Bioinformatics
ISBN 978-3-030-65774-1 ISBN 978-3-030-65775-8 (eBook)
https://doi.org/10.1007/978-3-030-65775-8
LNCS Sublibrary: SL8 – Bioinformatics

© Springer Nature Switzerland AG 2020

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the
material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now
known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book are
believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors
give a warranty, expressed or implied, with respect to the material contained herein or for any errors or
omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
 Preface

The Brazilian Symposium on Bioinformatics (BSB) is an international conference that

covers all aspects of bioinformatics and computational biology. This volume contains
the accepted papers for BSB 2020, held virtually during November 23–27, 2020.
As in past years, the special interest group in Computational Biology (CEbiocomp)
of the Brazilian Computer Society (SBC) organized the event. A Program Committee
(PC) was in charge of reviewing submitted papers; this year, the PC had 46 members.
Each submission was reviewed by three PC members. There were two submission
tracks: full papers and short papers. One of us (Setubal) supervised the full paper track,
and the other (Mendes da Silva) supervised the short paper track. In the full paper track,
20 papers were accepted; 5 papers were accepted in the short paper track. All of them
are printed in this volume and were presented orally at the event. In addition to the
technical presentations, BSB 2020 featured the following invited speakers, with the
respective talk titles: David Roos (University of Pennsylvania, USA), “Cross-silo
integration and interrogation of complex biomedical datasets: from genomics to epi-
demiology and back again”; Bas E. Dutilh (Utrecht University, The Netherlands),
“Metagenomics: from illuminating “dark matter” to modeling the microbiome”; and
Jens Stoye (Bielefeld University, Germany), “Searching for Genomic Variants in
Multiple Genomes at Once”.
Based on PC reviews, we chose the authors of the paper by Gabriel Oliveira et al.,
“COVID-19 X-ray Image Diagnostic with Deep Neural Networks”, as the recipients
of the BSB 2020 Best Paper Award. Although not part of the selection criteria, we were
happy to see that it turned out that this paper addresses one of the many topics arising
from the current COVID-19 pandemic.
BSB 2020 was made possible by the dedication and work of many people and
organizations. We would like to express our thanks to all PC members, as well as to the
external ad-hoc reviewers. Their names are listed in the pages that follow. We are also
grateful to the local organizers and volunteers for their valuable help; to Raquel
Minardi for helping out with social media; to the sponsors for making the event
financially viable; to the developers of EasyChair (full papers) and JEMS (short
papers), which were the systems we used to handle submissions; and Springer for
agreeing to publish this volume and their staff for working with us on its production.
Finally, we would like to thank all authors for their time and effort in submitting their
work and the invited speakers for having accepted our invitation.

November 2020 João C. Setubal

Waldeyr Mendes Silva
 Organization

Conference Chair
Daniel Cardoso Moraes de Fluminense Federal University, Brazil
Oliveira

Program Committee Chairs

João Carlos Setubal University of São Paulo, Brazil
Waldeyr C. Mendes da Instituto Federal de Goiás, Brazil
Silva

Steering Committee
Daniel Cardoso Moraes de Fluminense Federal University, Brazil
Oliveira
João Carlos Setubal University of São Paulo, Brazil
Luis Antonio Kowada Fluminense Federal University, Brazil
Natália Florencio Martins Empresa Brasileira de Pesquisa Agropecuária, Brazil
Ronnie Alves Instituto Tecnológico Vale, Brazil
Sérgio Lifschitz Pontifícia Universidade Católica do Rio de Janeiro,
Brazil
Sérgio Vale Aguiar Campos Federal University of Minas Gerais, Brazil
Tainá Raiol Fundação Oswaldo Cruz, Brazil
Waldeyr Mendes Instituto Federal de Goiás, Brazil

Program Committee
Said Sadique Adi Federal University of Mato Grosso do Sul, Brazil
Nalvo Almeida Federal University of Mato Grosso do Sul, Brazil
Ronnie Alves ITV Sustainable Development, Brazil
Deyvid Amgarten University of São Paulo, Brazil
Marilia Braga Bielefeld University, Germany
Laurent Brehelin Laboratoire d’Informatique, de Robotique et de
Microélectronique de Montpellier, France
Marcelo Brigido University of Brasília, Brazil
Sérgio Campos Federal University of Minas Gerais, Brazil
Andre Carvalho University of São Paulo, Brazil
Ricardo Cerri Federal University of Sao Carlos, Brazil
Luís Cunha Fluminense Federal University, Brazil
Alberto Davila FIOCRUZ-RJ, Brazil
Ulisses Dias University of Campinas, Brazil
viii Organization

Zanoni Dias University of Campinas, Brazil

Luciano Antonio University of São Paulo, Brazil
Digiampietri
Bas E. Dutilh Utrecht University, The Netherlands
Andre Fujita University of São Paulo, Brazil
Ronaldo Fumio Hashimoto University of São Paulo, Brazil
Luis Kowada Fluminense Federal University, Brazil
Sergio Lifschitz PUC-RJ, Brazil
Felipe Prata Lima Federal Institute of Alagoas, Brazil
Fabrício Martins Lopes Federal University of Technology of Paraná, Brazil
Felipe A. Louza Federal University of Uberlândia, Brazil
Fábio Henrique Viduani Federal University of Mato Grosso do Sul, Brazil
Martinez
Natalia Martins Embrapa Genetic Research and Biotechnology, Brazil
João Meidanis University of Campinas, Brazil
Raquel Melo-Minardi Federal University of Minas Gerais, Brazil
Jhonatas Monteiro University of São Paulo, Brazil
Jefferson Morais Federal University of Pará, Brazil
Daniel de Oliveira Fluminense Federal University, Brazil
Sergio Pantano Institut Pasteur de Montevideo, Uruguay
Alexandre Rossi Paschoal Federal University of Technology of Paraná, Brazil
José Patané Butantan Institute, Brazil
Daniel Pinheiro University of São Paulo, Brazil
David Pires Butantan Institute, Brazil
Rommel Ramos Federal University of Pará, Brazil
Mariana Federal University of Rio Grande do Sul, Brazil
Recamonde-Mendoza
Marcel da Câmara Institut Curie, France
Ribeiro-Dantas
Fernando Luís Barroso Da University of São Paulo, Brazil
Silva
Thais Amaral e Sousa Instituto Federal de Goiás, Brazil
Marcilio de Souto University of Orleans, France
Kleber Padovani De Souza Federal University of Pará, Brazil
Sadegh Sulaimany University of Kurdistan, Iran
Guilherme Telles University of Campinas, Brazil
Alessandro Varani São Paulo State University, Brazil
Maria Emília Machado University of Brasília, Brazil
Telles Walter

Additional Reviewers

Eloi Araújo
Daniel Saad Nogueira Nunes
Lucas Oliveira
 Organization ix

Funding

Fundação de Amparo à Pesquisa do Estado do Rio de Janeiro (FAPERJ), Brazil

Sponsor

Sociedade Brasileira de Computação (SBC), Brazil

 Contents

A Classification of de Bruijn Graph Approaches for De Novo

Fragment Assembly. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Elvismary Molina de Armas, Maristela Holanda, Daniel de Oliveira,
Nalvo F. Almeida, and Sérgio Lifschitz

Redundancy Treatment of NGS Contigs in Microbial Genome Finishing

with Hashing-Based Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
Marcus Braga, Kenny Pinheiro, Fabrício Araújo, Fábio Miranda,
Artur Silva, and Rommel Ramos

Efficient Out-of-Core Contig Generation . . . . . . . . . . . . . . . . . . . . . . . . . . 25

Julio Omar Prieto Entenza, Edward Hermann Haeusler,
and Sérgio Lifschitz

In silico Pathogenomic Analysis of Corynebacterium Pseudotuberculosis

Biovar Ovis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
Iago Rodrigues Blanco, Carlos Leonardo Araújo,
and Adriana Carneiro Folador

Assessing the Sex-Related Genomic Composition Difference Using

a k-mer-Based Approach: A Case of Study in Arapaima gigas (Pirarucu) . . . 50
Renata Lilian Dantas Cavalcante, Jose Miguel Ortega,
Jorge Estefano Santana de Souza, and Tetsu Sakamoto

COVID-19 X-ray Image Diagnostic with Deep Neural Networks . . . . . . . . . 57

Gabriel Oliveira, Rafael Padilha, André Dorte, Luis Cereda,
Luiz Miyazaki, Maurício Lopes, and Zanoni Dias

Classification of Musculoskeletal Abnormalities with Convolutional

Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
Guilherme Tiaki Sassai Sato, Leodécio Braz da Silva Segundo,
and Zanoni Dias

Combining Mutation and Gene Network Data in a Machine Learning

Approach for False-Positive Cancer Driver Gene Discovery . . . . . . . . . . . . . 81
Jorge Francisco Cutigi, Renato Feijo Evangelista,
Rodrigo Henrique Ramos, Cynthia de Oliveira Lage Ferreira,
Adriane Feijo Evangelista, Andre C. P. L. F. de Carvalho,
and Adenilso Simao
xii Contents

Unraveling the Role of Nanobodies Tetrad on Their Folding and Stability

Assisted by Machine and Deep Learning Algorithms . . . . . . . . . . . . . . . . . . 93
Matheus Vitor Ferreira Ferraz, Wenny Camilla dos Santos Adan,
and Roberto Dias Lins

Experiencing DfAnalyzer for Runtime Analysis of Phylogenomic

Dataflows. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
Luiz Gustavo Dias, Marta Mattoso, Bruno Lopes,
and Daniel de Oliveira

Sorting by Reversals and Transpositions with Proportion Restriction . . . . . . . 117

Klairton Lima Brito, Alexsandro Oliveira Alexandrino,
Andre Rodrigues Oliveira, Ulisses Dias, and Zanoni Dias

Heuristics for Breakpoint Graph Decomposition with Applications

in Genome Rearrangement Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
Pedro Olímpio Pinheiro, Alexsandro Oliveira Alexandrino,
Andre Rodrigues Oliveira, Cid Carvalho de Souza, and Zanoni Dias

Center Genome with Respect to the Rank Distance . . . . . . . . . . . . . . . . . . . 141

Priscila Biller, João Paulo Pereira Zanetti, and João Meidanis

ImTeNet: Image-Text Classification Network for Abnormality Detection

and Automatic Reporting on Musculoskeletal Radiographs . . . . . . . . . . . . . . 150
Leodécio Braz, Vinicius Teixeira, Helio Pedrini, and Zanoni Dias

A Scientometric Overview of Bioinformatics Tools in the Pseudomonas

Putida Genome Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
Alex Batista Trentin, Beatriz Thomas Metzner, Glecio Oliveira Barros,
and Deborah Catharine de Assis Leite

Polysome-seq as a Measure of Translational Profile from Deoxyhypusine

Synthase Mutant in Saccharomyces cerevisiae . . . . . . . . . . . . . . . . . . . . . . 168
Fernanda Manaia Demarqui, Ana Carolina Silva Paiva,
Mariana Marchi Santoni, Tatiana Faria Watanabe,
Sandro Roberto Valentini, and Cleslei Fernando Zanelli

Anti-CD3 Stimulated T Cell Transcriptome Reveals Novel ncRNAs

and Correlates with a Suppressive Profile. . . . . . . . . . . . . . . . . . . . . . . . . . 180
Manuela M. do Almo, Isabel G. Sousa,
Waldeyr Mendes Cordeiro da Silva, Thomas Gatter, Peter F. Stadler,
Steve Hoffmann, Andrea Q. Maranhão, and Marcelo Brigido

A Simplified Complex Network-Based Approach to mRNA

and ncRNA Transcript Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
Murilo Montanini Breve and Fabrício Martins Lopes
 Contents xiii

A Systems Biology Driven Approach to Map the EP300 Interactors Using

Comprehensive Protein Interaction Network . . . . . . . . . . . . . . . . . . . . . . . . 204
Shivananda Kandagalla, Maria Grishina, Vladimir Potemkin,
Sharath Belenahalli Shekarappa, and Pavan Gollapalli

Analyzing Switch Regions of Human Rab10 by Molecular Dynamics

Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
Levy Bueno Alves, William O. Castillo-Ordoñez, and Silvana Giuliatti

Importance of Meta-analysis in Studies Involving Plant Responses

to Climate Change in Brazil . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
Janaina da Silva Fortirer, Adriana Grandis,
Camila de Toledo Castanho, and Marcos Silveira Buckeridge

A Brief History of Bioinformatics Told by Data Visualization . . . . . . . . . . . 235

Diego Mariano, Mívian Ferreira, Bruno L. Sousa, Lucianna H. Santos,
and Raquel C. de Melo-Minardi

Computational Simulations for Cyclizations Catalyzed by Plant

Monoterpene Synthases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
Waldeyr Mendes Cordeiro da Silva, Daniela P. de Andrade,
Jakob L. Andersen, Maria Emília M. T. Walter, Marcelo Brigido,
Peter F. Stadler, and Christoph Flamm

Oncogenic Signaling Pathways in Mucopolysaccharidoses . . . . . . . . . . . . . . 259

Gerda Cristal Villalba Silva, Luis Dias Ferreira Soares,
and Ursula Matte

Natural Products as Potential Inhibitors for SARS-CoV-2 Papain-Like

Protease: An in Silico Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
Jesus Alvarado-Huayhuaz, Fabian Jimenez, Gerson Cordova-Serrano,
Ihosvany Camps, and Wilmar Puma-Zamora

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271

 A Classification of de Bruijn Graph
Approaches for De Novo Fragment
Assembly

Elvismary Molina de Armas1 , Maristela Holanda2 , Daniel de Oliveira3 ,

Nalvo F. Almeida4 , and Sérgio Lifschitz1(B)
1
Depto. Informática, PUC-Rio, Rio de Janeiro, Brazil
{earmas,sergio}@inf.puc-rio.br
2
Depto. Ciência da Computação, UNB, Brası́lia, Brazil
[email protected]
3
Instituto de Computação, UFF, Niterói, Brazil
[email protected]
4
Faculdade de Computação, UFMS, Campo Grande, MS, Brazil
[email protected]

Abstract. Research in bioinformatics has changed rapidly since the

advent of next-generation sequencing (NGS). Despite the positive impact
on cost reduction, assembling the generated reads remains a challenge.
This paper presents in detail the main ideas related to de novo assembly,
the technologies involved, and theoretical concepts about the de Bruijn
graph structure. We also explain the existing approaches to minimize
the memory requirements for de Bruijn graph construction. Finally, we
propose a comparative view of several solutions, including the k -mers
codification and the data structures used to represent and persist them.

1 Introduction
The field of biological research has changed rapidly since the advent of mas-
sively parallel sequencing technologies, known as next-generation sequenc-
ing (NGS) [10,18]. Some commercial DNA sequencing platforms include the
Genome Sequencer from Roche 454 Life Sciences (www.my454.com), the Genome
Analyzer platform from Illumina (www.illumina.com), the SOLiD System
from Applied Biosystems (www.appliedbiosystems.com), the Pacific Bioscience
(PacBio) sequencers (https://www.pacb.com), and the Oxford MinION, which
uses Nano Pore Sequencing technology (https://nanoporetech.com).
These platforms’ vital characteristic is that they do not rely on Sanger chem-
istry [41] as first-generation machines did. With their arrival in the market in
2005 and the fast development since then, they have drastically lowered the cost
per sequenced nucleotide and increased throughput by orders of magnitude [36].
Their performance dramatically increased the numbers of generated reads (many
hundreds of thousands or even millions of reads) in a relatively short time [25]
with good genome coverage.
c Springer Nature Switzerland AG 2020
J. C. Setubal and W. M. Silva (Eds.): BSB 2020, LNBI 12558, pp. 1–12, 2020.
https://doi.org/10.1007/978-3-030-65775-8_1
2 E. M. de Armas et al.

NGS has brought a relevant impact in various areas such as genomics,

transcriptomics, metagenomics, proteogenomics, gene expression analysis, non-
coding RNA discovery, SNP detection, and protein binding sites identifica-
tion [18]. The genome assembly problem arises because it is impossible to
sequence a whole genome directly in one read using current sequencing tech-
nologies.
For assemblies with no reference genome, called de novo, assembling a large
genome (>100 Mbp) using short readings remains a challenge. Some successful
approaches based on the use of de Bruijn graph computational data structure
have been developed for de novo assembly. The construction and use of de Bruijn
graph demand a large amount of main memory and execution time because of
the large number of elements (nodes and edges) to process.
In the next sections, we will give more details about de novo assembly, the
technologies involved, and theoretical concepts about the de Bruijn graph struc-
ture. We also explain a categorization of the existing approaches that minimize
memory requirements for de Bruijn graph construction. Further, a comparative
view of several solutions shows the codification to represent the k -mers and the
data structures used to store them.

2 Genome Assembly
Genome assembly may be defined as the computational process of reconstructing
a whole-genome using numerous short sequences called reads up to the chromo-
somal level. An assembly is a hierarchical data structure that maps the sequence
data to a putative reconstruction of the target genome. However, the vast major-
ity of sequenced genomes are made available only in draft format, having as a
result only contigs, which are continuous stretches of DNA sequence, or scaffolds,
which are chains of contigs with additional information about their relative posi-
tions and orientations.
There are some challenges around the use of NGS data that brings diffi-
culties to obtain the assembly. DNA sequencing technologies share the funda-
mental limitation that read lengths are much shorter than even the smallest
genomes. The process of determining the complete DNA sequence of an organ-
ism’s genome overcomes this limitation by over-sampling the target genome with
short reads from random positions. Also, assembly software is challenged by
repeated sequences in the target genome. Genomics regions that share per-
fect repetitions may be indistinguishable, mainly if they are longer than the
reads. The repetition resolution is more difficult in front of sequencing errors.
Therefore it is necessary to further increase sequence accuracy by sequencing
individual genomes in a large number of times, increasing the sequenced reads
coverage.
In terms of computational complexity, the assembly may require High Per-
formance Computing (HPC) platforms for large genomes and the processing
of larger volumes of data. Algorithms developed for these HPC platforms are
typically complex and depend on pragmatic engineering and heuristics. Heuris-
tics help overcome complicated repetition patterns in real genomes, random and
 de Bruijn Graph Approaches for Fragment Assembly 3

systematic errors in factual data, and real computers’ physical limitations. Also,
the implementations and results are tied to a suitable parameter instantiation.
In case of de novo assembly, using a k -mer based algorithms (see Sect. 3), the
selection of k value is vital.

3 The de Bruijn Graph De Novo Assembly Approach

There are two main categories of NGS assemblers can: Overlap-Layout-

Consensus approach (OLC) [35], based on an overlap graph, and de Bruijn Graph
(DBG) approach, which relies on some form of k -mers as vertices of the graph.
The de Bruijn graph was defined outside the realm of DNA sequencing to repre-
sent strings from a finite alphabet. The nodes represent all possible fixed-length
strings. The edges represent suffix-to-prefix perfect overlaps [34]. In the context
of genome assembly, the de Bruijn graph could be defined as follows:

Definition 1. A read r with length m represents a genome substring, over the

alphabet Σ = {A, T, C, G}. Each character of the alphabet represents one of the
four nitrogenous bases present in DNA: adenine (A), guanine (G), cytosine (C)
and thymine (T ).

Definition 2. A k-mer is a substring over a read with specific k length.

The k -mer is a string whose length is k, 1 < k < m. k defines the minimum
length of a substring that two reads must share to define an overlap, linking two
reads in the graph traverse. Using a larger k value involves more accuracy to
discover repeated regions in the genome, but also increases the chances of loose
overlaps in reads, causing the loss of links in the graph. Consequently, it is not
easy to estimate the right k value for the best assembly. The total number of
k -mers present in one read is equal to m−k +1, while the total number of k -mers
present in n reads is (m − k + 1)n. The unique k -mers space for k value is 4k .

Definition 3. A de Bruijn graph Gk (V, E) represents overlaps between k-mers,

in which:

– The set of vertices is defined by V = S = {s1 , s2 , ..., sp }, where S is a set of

unique k-mers over a given set of reads.
– The set of edges is defined by E = {e1 , e2 , ..., eq }, where e = (si , sj ) if and
only if the k − 1-th suffix of si matches exactly the k − 1-th prefix of sj . si
and sj must be adjacent k-mers in at least one read.

The life cycle of DBG for genome assembly can be summarized in two steps.
First, construction involves the generation of all k -mers to generate a node per
distinct k -mer and an edge between two nodes if these k -mers have a k − 1
overlapped in at least one read. In the second step, the processing is carried on
by simplifying the graph and traverse it to generate contiguous genome regions
called contigs.
4 E. M. de Armas et al.

4 de Bruijn Graph Based Assemblers

There are some assemblers that use DBG approach, for example Velvet [48,49],
ALLPATHS [5], ABySS [45], SOAPdenovo [29], SPAdes [1] and Contrail
(https://sourceforge.net/projects/contrail-bio/). However, the de Bruijn app-
roach has as a drawback that de Bruijn graph may require an enormous amount
of memory (several gigabytes of RAM). Besides, the construction and analy-
sis of a de Bruijn graph are not easily parallelized [42]. As a result, de Bruijn
assemblers such as Velvet and ALLPATHS, which have been used successfully
on bacterial genomes, do not scale to larger genomes. For example, these pro-
grams would require several terabytes of RAM for human-sized genomes to store
their de Bruijn graphs [42] and memory requirements may be higher for more
complex genome organisms, as is the case of many plants.
This high memory consumption problem is expected to worsen in the future
because the NGS data generation rate has exceeded expectations based on
Moore’s law [23], meaning that the amount of raw data is expected to grow
much faster than the capacity of available memory [25].

4.1 Main Classification of Approaches

There are several techniques to reduce the memory footprint for the assembly
process. These approaches can be divided into two general groups. The first
group considers solutions that reduce the amount of data as a cost of not exact
representations of DBG. The goal is to reduce the data to process as much
as possible by removing or sampling until it is possible to execute the assembly
with available resources. However, it may affect the final quality of the assem-
bly. In that group, we also include those approaches that use probabilistic data
structures as Bloom Filters (BF) [2].
In that way, they can use a fixed amount of memory, independently on the
number of items to be processed. However, there is no accurate measure to guide
the reduction of the data. The memory consumption during the assembly process
is highly sensitive to data and the value of k. Thus the success of the reduction
is only validated when the assembly is achieved with the available memory. The
second group considers solutions that increase the memory resources for the
same amount of data for an exact representations of DBG, through some
techniques like partitioning or distributing, in main or external memory. We
discuss these approaches in detail in the following sections.

4.2 General Strategies to Reduce Memory Footprint for DBG

Construction
More specifically, alternating between the aforementioned groups we have some
techniques to reduce the memory requirements for the assembly process, which
can be examined through the following categories:
Pre-processing Techniques: First, there are pre-processing techniques such
as Diginorm in [46], Quake [24], ALLPATHS-LG error corrector [22], which try
 de Bruijn Graph Approaches for Fragment Assembly 5

to reduce the input size removing redundant information and errors before the
assembly process itself start.
Optimized Data Structures for Graph Representation: To minimize the
memory requirements during the k -mer unique identification process, indexes
for identifying duplicate k -mers may be used. Hash tables in memory have also
been successfully used for many assemblers, such as in [29,45,48], to identify
duplicate k -mers. However, for a large amount of NGS data, hash tables do not
work well because they may not fit in memory. Suffix-array is a data structure
also used to compute overlaps. The FM-index [44] has also been used to allow
the compressed representation of input reads and fast computation of overlaps
in string graph (equivalent to overlap graph), but it is not tested yet in the
construction of de Bruijn graph.
Succinct data structures have also been explored to represent de Bruijn graph
[3,4,11]. In [11] a succinct bitmap is also used to compress the representation
of de Bruijn graph, but overall its need for space will continue to increase as
the graph becomes “bigger”. Other approaches are based on the idea of sparse-
ness in genome assembly [47], where only a subset of k -mers present in the
dataset is stored. Bloom Filters (BF) have been arduously explored as solution
to deal with DBG computationally demanding for a not exact representations
of DBG [9,33,50]. They are used to store vertices (k -mers), while the edges are
implicitly deduced by querying the Bloom filter for the membership of all possi-
ble extensions of a k -mer. However, this approach does not correspond exactly to
the edges contained in the reads. Some works have been focused on mechanisms
that avoid false positives using Bloom Filters [9,40].
An extra-compacted de Bruijn Graph structure is introduced in [12]. It rep-
resents intermediate states of the DBG during its generation through a series of
iterations in which the number of k -mers to be processed is iteratively reduced.
The extra-compacted de Bruijn Graph nodes represents the unique dk -mers with
length equal or less than d, while edges corresponds to unique edges formed by
adjacent dk -mers, whose sharing k − 1-length overlap in at least one read.
Extending the Computational Resources: Some solutions proposed the use
of cloud-based resources to overcome the memory requirements limitations. In
[25] were designed a set of assembler experiments using the GAGE datasets
and a group of program assemblers in virtual machines in the Amazon AWS
environment. The financial analysis reveals that the cost for assembly increases
as the complex of genomes to be assembled, because such an operation requires
more expensive virtual machines, and the assembly may be executed for several
hours. Other solutions are based or combined with parallelization techniques as
BCALM2 [8], and the use of GPU and other memory systems like Gerbil [19]
and k -mer counter FPGA-based solution [32].
External Memory Approaches: Partition assembly algorithms were also pro-
posed for external processing. For example, the Minimum Substring Partitioning
(MSP) [28] technique allows us to split the input reads into subsequences and
distributes then into this disk partitions, then processing one disk partition at
6 E. M. de Armas et al.

a time. Moreover, the k -mers partitions can be processed in a distributed man-

ner, as well as the Contrail proposed to avoid memory bottleneck. Moreover,
BCALM1 [7] and BCALM2 [8] propose a construction of a compacted DBG
using partition and disk distribution approaches, such as DSK k -mer counter.
Among the techniques used for partitioning were found hash functions over the
k -mer minimum substring of p length and minimizers.
The construction of the graph embedded into a relational database man-
agement system (RDBMS) is another type of approach explored in [13–15,43].
Some indexes configuration based on B+-tree, hash over k -mer [15]) and k -mer
p-minimum substring [13] were tested for textitk-mer mapping process as part
of the DBG construction, in junction with an ad-hoc cost model to measure the
performance gained [14]. Although some optimizations are needed to improve
the execution time given by the index evaluation, the case study implemented
with PostgreSQL based on the Velvet assembly algorithm shows the feasibility of
using DBMS to manage I/O operations in the k -mer process mapping, allowing
incremental processing without reprocessing and recovery from failures [43].
The external memory approaches presented so far process from the beginning
the total number of k -mers, to following obtain the vertices of the graph (unique
k -mers) and corresponding edges. These imply carrying on the process with the
high level of redundancy present in those k-mers, which significantly impacts the
amount of memory needed, and the number of I/O operations.
An approach presented in [12] combines an external memory and optimized
data structures approach. Unlike the above techniques, it proposes the construc-
tion of an exact representation of de Bruijn Graph without the necessity of
process all k-mers. Through an iterative sequence of reductions, it is possible to
process the graph as much as possible in the main memory, and only when the
available main memory becomes insufficiently, will be using an external memory
solution. Then, large duplicate regions had already been identified, avoiding pro-
cessing a significant amount of duplicated k -mers in external memory, reducing
the number of I/O operations.

4.3 Specific Strategies to Reduce the Large Memory Consumption

There are a series of specific elements on which the studied solutions have focused
to reduce the large memory consumption in terms of implementation. As can be
seen, some of them correspond directly to one of the general approaches.
One of those elements is using data structures for fast lookup with lower
overhead to store the graph elements, especially for the set of unique k -mers that
correspond to the nodes. The most frequent data structure used is hash tables,
followed by search-trees data structure variations. Another essential element to
taking into account is the codification of k -mer for less memory per element.
We classified the approaches available in literature into two groups, those
based on a lossless compression, which is an exact representation of DBG, the
other based on lossy data compression. In the first group, we observed the clas-
sical 2-bits k -mer codification. For example, in DSK, Velvet, and ABYSS, they
remove parts of p1 and p2 length prefixes such as in KMC1, and some forms of
 de Bruijn Graph Approaches for Fragment Assembly 7

hash optimization as in Jellyfish with the use of a bijective function and Merac-
ulous, which use a lightweight hash (a combination of the hash family). In the
second group, we have seen best represented by Bloom Filters (BF) with false
positives (FP) or specialized probabilistic data structures based on BF.

Table 1. de Bruijn graph solutions comparison. Abbreviation used for the approach
classification (Cl.): KC (K-mer Counter), A (Assembler), S (Space efficient solution)
and E (external approach by disk distribution)

Approach Cl. Description of the approach Data structure to store k -mers

Jellyfish [31] KC Main memory. Hash table merging in Hash table using a quadratic
disk if not memory available reprobing function
BFCounter [33] KC BF in RAM. Two pass algorithm to BF. 4-bits per k -mer, Google
correct false positives sparsehash library
Meracolous [6] A Reliance on the linear U-U Novel lightweight hash
component of the graph
DSK [39] KC Disk distribution based on hashing Hash table
KMC1 [16] KC Disk distribution based on prefixes. Bins (disk files) according on
Sorting using the prefixes p1 and p2 of k -mers
least-significant-digit (LSD) radix
sort
Minia [9] S BF in RAM + set of FP on disk with BF for nodes. cPF structure for a set
a fixed amount of memory of critical FP
khmer [50] KC Count-Min Sketch in RAM. Not Count-Min Sketch to storing the
error correction is used frequency distributions of distinct
k -mers
Cascading BF [40] A BF + set of FP Cascading BFs to store FP k -mers
BCALM [7] E DSK + BCALM algorithm DBGFM to store no-branch path,
codified as FM index
KMC2 [17] KC Disk distribution based on k -mer Bins (disk files) according to the
signatures. Sorting using LSD radix related canonical minimizer. HT for
sort counting
KCMBT [30] KC Trie-based in-memory algorithm. Multiple Burst Trees (KCMBT)
Three phases algorithm
BCALM2 [8] E Based on BCALM + parallelization. Union-find data structure for
Three stages algorithm partitions with a minimal perfect
hash function based indexing
KMC3 [26] KC Disk distribution. Better balance of Bins to store k -mers
bin sizes and fast radix sort
HaVec [38] S BF + the quotient of the hash Vector to store the k -mers along
function division to verify FP with their neighbor information
Squeakr [37] KC Based upon a counting filter data CQF to store an approximation of a
structure CQFs multiset S of k -mers, maintaining a
false positive rate
FastEtch S Based upon Count-Min sketch Count-Min sketch stores an
algorithm [20, 21] probabilistic data structure approximated DBG with the subset
of nodes that are most likely to
contribute to the contig generation
step

Table 1 summarizes several approaches present in the literature, showing a

brief description of each one and classification based on the type of application: k -
mer counter, assembler, space-efficient solutions for DBG and disk distribution
as an external approach variant. It also presents the data structure used to
represent the DBG storing the k -mers with a specific codification.
8 E. M. de Armas et al.

Some variants for k -mer codification are used to reduce the amount of mem-
ory needed to represent and store each k -mer. DSK (Minia assembler) [39] and
ABYSS [45], for example, use a classical 2 bits representation for each nucleotide
base in k -mers. With the use of probabilistic data structures like Bloom Filters,
Minia [9], for example, allows approximated 1.44 log2 (16k/2.08) + 2.08 bits/k -
mer.
Other solutions use strategies based on hash tables, for example Jellyfish [31]
codifying a part of k -mer as the index of the table using a bijective hash function.
Meracolous [6] uses a recursive collision strategy with multiple hash functions
to avoid explicitly storing the k -mer themselves. khmer [50] uses a Count-Min
Sketch storing only counts, while k -mers must be retrieved from the original data
set, and HaVec [38] uses 5 bytes for each index in the hash table plus 2k bits for
k -mer. As another strategy, KMC1 [16], for example, stores k-mer without the
p1 and p2 prefixes.

4.4 k -mers Counters

The DBG construction implies a subroutine to identify distinct k -mers and get
their multiplicity. Identify distinct k -mers problem also has been touched by
counting k -mer tools [16,17,27,31,33,39]. Although k -mer counter tools aim
at generating histograms over k -mers distributions, their processes has some
similarities to the ones that get the vertices set of the DBG.
Identifying distinct k -mers have been approached by sorting [16,17], hash-
ing [27,31,39] or using Bloom Filters [33], combined sometimes with parallel
approaches to speed up the process [16,17]. Some of them [16,17,27,39] have
been focused on distribution the k -mers in disk partitions to counter them before,
loading in main memory each partition at time.
It is valid to note that k -mer counters have not the notion of vertices and
edges. Besides, to reduce the amount of data, they make some assumptions such
as do not count the k -mers with frequencies smaller or more significant than a
given value, which is not appropriate for the DBG construction.

5 Conclusions

Next-generation sequencing (NGS) data has significantly impacted several fields

of bioinformatics, greatly reducing costs. However, the genome assembly con-
tinues to be a challenge for genomic research since no technology is capable of
sequencing the whole-genome. Also, some aspects of NGS data makes the assem-
bly difficult, such as the error profiles for each NGS platform, the non-uniform
coverage of the target, hampering the resolution of genome repetitions.
The de Bruijn graph (DBG) approach for the de novo assembly is used
when there is no reference genome. The construction of DBG needs a high main
memory space and is responsible for the high computational cost. The most
critical parameter in the DBG is the k value, which impacts the assembly’s
accuracy, the number of vertices and edges, and the memory requirements.
 de Bruijn Graph Approaches for Fragment Assembly 9

Several techniques have been proposed to construct a DBG. This paper clas-
sifies those approaches into two main categories, one for an exact and complete
graph, another for non-exact DBG representation. Considering the solutions
based on probabilistic data structures used for an approximate DBG representa-
tion, the most explored data structure are the bloom filters. This data structure
allows vertices to be stored independently of their number, but at the cost of
false positives.
By revisiting the literature, we could find some solutions using external mem-
ory to construct a graph with an exact representation. They process all the k -
mers in external memory. A new algorithm building an accurate representation
without the necessity of process all k-mers have also been found.
This paper presented in detail the main theoretical and practical aspects
related to de novo assembly, particularly the de Bruijn graph structure. We
have enumerated the existing approaches to reduce the memory requirements
for DBG construction. Finally, we have proposed a comparative view of the
existing solutions in the literature.

Acknowledgments. The authors are partially supported by grants from FUNDECT,

FAPERJ, CNPq, CAPES and INCA, Brazilian Public Funding Agencies and Research
Institutes.

References
1. Bankevich, A., et al.: Spades: a new genome assembly algorithm and its appli-
cations to single-cell sequencing. J. Comput. Biol.: J. Comput. Mol. Cell Biol.
19(5), 455–477 (2012). https://doi.org/10.1089/cmb.2012.0021. https://pubmed.
ncbi.nlm.nih.gov/22506599
2. Bloom, B.H.: Space/time trade-offs in hash coding with allowable errors. Commun.
ACM 13(7), 422–426 (1970). https://doi.org/10.1145/362686.362692
3. Boucher, C., Bowe, A., Gagie, T., Puglisi, S., Sadakane, K.: Variable-order de
Bruijn graphs. In: Data Compression Conference Proceedings 2015 (2014). https://
doi.org/10.1109/DCC.2015.70
4. Bowe, A., Onodera, T., Sadakane, K., Shibuya, T.: Succinct de Bruijn graphs. In:
Raphael, B., Tang, J. (eds.) WABI 2012. LNCS, vol. 7534, pp. 225–235. Springer,
Heidelberg (2012). https://doi.org/10.1007/978-3-642-33122-0 18
5. Butler, J.: ALLPATHS: De novo assembly of whole-genome shotgun microreads.
Genome Res. 18(5), 810–820 (2008). https://doi.org/10.1101/gr.7337908
6. Chapman, J.A., Ho, I., Sunkara, S., Luo, S., Schroth, G.P., Rokhsar, D.S.: Mer-
aculous: de novo genome assembly with short paired-end reads. PLoS ONE 6(8),
e23501 (2011). https://doi.org/10.1371/journal.pone.0023501
7. Chikhi, R., Limasset, A., Jackman, S., Simpson, J.T., Medvedev, P.: On the rep-
resentation of de Bruijn graphs. In: Sharan, R. (ed.) RECOMB 2014. LNCS,
vol. 8394, pp. 35–55. Springer, Cham (2014). https://doi.org/10.1007/978-3-319-
05269-4 4
8. Chikhi, R., Limasset, A., Medvedev, P.: Compacting de Bruijn graphs from
sequencing data quickly and in low memory. Bioinformatics 32(12), i201 (2016).
https://doi.org/10.1093/bioinformatics/btw279
10 E. M. de Armas et al.

9. Chikhi, R., Rizk, G.: Space-efficient and exact de Bruijn graph representation based
on a Bloom filter. Algorithms Mol. Biol. 8(1), 22 (2013). https://doi.org/10.1186/
1748-7188-8-22
10. Claros, M.G., Bautista, R., Guerrero-Fernández, D., Benzerki, H., Seoane, P.,
Fernández-Pozo, N.: Why assembling plant genome sequences is so challenging.
Biology 1(2), 439 (2012). https://doi.org/10.3390/biology1020439
11. Conway, T.C., Bromage, A.J.: Succinct data structures for assembling
large genomes. Bioinformatics 27(4), 479–486 (2011). https://doi.org/10.1093/
bioinformatics/btq697
12. de Armas, E.M., Castro, L.C., Holanda, M., Lifschitz, S.: A new approach for de
Bruijn graph construction in de novo genome assembling. In: 2019 IEEE Interna-
tional Conference on Bioinformatics and Biomedicine, pp. 1842–1849 (2019)
13. de Armas, E.M., Ferreira, P.C.G., Haeusler, E.H., de Holanda, M.T., Lifschitz, S.:
K-mer mapping and RDBMS indexes. In: Kowada, L., de Oliveira, D. (eds.) BSB
2019. LNCS, vol. 11347, pp. 70–82. Springer, Cham (2020). https://doi.org/10.
1007/978-3-030-46417-2 7
14. de Armas, E.M., Haeusler, E.H., Lifschitz, S., de Holanda, M.T., da Silva, W.M.C.,
Ferreira, P.C.G.: K-mer Mapping and de Bruijn graphs: the case for velvet frag-
ment assembly. In: 2016 IEEE International Conference on Bioinformatics and
Biomedicine (BIBM), pp. 882–889 (2016). https://doi.org/10.1109/BIBM.2016.
7822642
15. de Armas, E.M., Silva, M.V.M., Lifschitz, S.: A study of index structures for K-
mer mapping. In: Proceedings Satellite Events of the 32nd Brazilian Symposium
on Databases. Databases Meet Bioinformatics Workshop, pp. 326–333 (2017)
16. Deorowicz, S., Debudaj-Grabysz, A., Grabowski, S.: Disk-based k-mer counting on
a PC. BMC Bioinform. 14(1) (2013). https://doi.org/10.1186/1471-2105-14-160
17. Deorowicz, S., Kokot, M., Grabowski, S., Debudaj-Grabysz, A.: KMC 2: fast and
resource-frugal k-mer counting. Bioinformatics 31(10), 1569 (2015). https://doi.
org/10.1093/bioinformatics/btv022
18. El-Metwally, S., Hamza, T., Zakaria, M., Helmy, M.: Next-generation sequence
assembly: four stages of data processing and computational challenges. PLoS Com-
put. Biol. 9(12), 1–19 (2013). https://doi.org/10.1371/journal.pcbi.1003345
19. Erbert, M., Rechner, S., Müller-Hannemann, M.: Gerbil: a fast and memory-
efficient k-mer counter with GPU-support. Algorithms Mol. Biol. 12(1), 9:1–9:12
(2017). https://doi.org/10.1186/s13015-017-0097-9
20. Ghosh, P., Kalyanaraman, A.: A fast sketch-based assembler for genomes. In: Pro-
ceedings of the 7th ACM International Conference on Bioinformatics, Computa-
tional Biology, and Health Informatics. BCB ’16, pp. 241–250. ACM, New York,
NY, USA (2016). https://doi.org/10.1145/2975167.2975192
21. Ghosh, P., Kalyanaraman, A.: FastEtch: a fast sketch-based assembler for genomes.
IEEE/ACM Trans. Comput. Biol. Bioinform. 16(4), 1091–1106 (2019). https://
doi.org/10.1109/TCBB.2017.2737999
22. Gnerre, S., et al.: High-quality draft assemblies of mammalian genomes from mas-
sively parallel sequence data. Proc. Natl. Acad. Sci. U.S.A. 108(4), 1513–1518
(2011). https://doi.org/10.1073/pnas.1017351108. 21187386[pmid]
23. Jackman, S.D., Birol, I.: Assembling genomes using short-read sequencing technol-
ogy. Genome Biol. 11(1), 202 (2010). https://doi.org/10.1186/gb-2010-11-1-202.
https://www.ncbi.nlm.nih.gov/pubmed/20128932, 20128932[pmid]
24. Kelley, D.R., Schatz, M.C., Salzberg, S.L.: Quake: quality-aware detection and
correction of sequencing errors. Genome Biol. 11(11), R116 (2010). https://doi.
org/10.1186/gb-2010-11-11-r116
 de Bruijn Graph Approaches for Fragment Assembly 11

25. Kleftogiannis, D., Kalnis, P., Bajic, V.B.: Comparing memory-efficient genome
assemblers on stand-alone and cloud infrastructures. PLoS ONE 8(9) (2013).
https://doi.org/10.1371/journal.pone.0075505
26. Kokot, M., Dlugosz, M., Deorowicz, S.: KMC 3: counting and manipulating k-
mer statistics. Bioinformatics 33(17), 2759–2761 (2017). https://doi.org/10.1093/
bioinformatics/btx304
27. Li, Y., Yan, X.: MSPKmerCounter: a fast and memory efficient approach for K-mer
Counting. arXiv e-prints (2015)
28. Li, Y., Kamousi, P., Han, F., Yang, S., Yan, X., Suri, S.: Memory efficient minimum
substring partitioning. PVLDB 6(3), 169–180 (2013). https://doi.org/10.14778/
2535569.2448951
29. Luo, R., et al.: SOAPdenovo2: an empirically improved memory-efficient short-
read de novo assembler. GigaScience 1(1), 1–6 (2012). https://doi.org/10.1186/
2047-217X-1-18
30. Mamun, A.A., Pal, S., Rajasekaran, S.: KCMBT: a k-mer counter based on mul-
tiple burst trees. Bioinformatics 32(18), 2783 (2016). https://doi.org/10.1093/
bioinformatics/btw345
31. Marcais, G., Kingsford, C.: A fast, lock-free approach for efficient parallel counting
of occurrences of k-mers. Bioinformatics 27(6), 764–770 (2011). https://doi.org/
10.1093/bioinformatics/btr011
32. McVicar, N., Lin, C., Hauck, S.: K-Mer counting using bloom filters with an
FPGA-attached HMC. In: 25th IEEE Annual International Symposium on Field-
Programmable Custom Computing Machines, FCCM 2017, Napa, CA, USA, 30
April–2 May 2017, pp. 203–210 (2017). https://doi.org/10.1109/FCCM.2017.23
33. Melsted, P., Pritchard, J.K.: Efficient counting of k-mers in DNA sequences using
a bloom filter. BMC Bioinform. 12(1), 333 (2011). https://doi.org/10.1186/1471-
2105-12-333
34. Miller, J.R., Koren, S., Sutton, G.: Assembly algorithms for next-generation
sequencing data. Genomics 95(6), 315–327 (2010). https://doi.org/10.1016/j.
ygeno.2010.03.001. 20211242[pmid]
35. Myers, E.W.: Toward simplifying and accurately formulating fragment assembly.
J. Comput. Biol. 2(2), 275–290 (1995). https://doi.org/10.1089/cmb.1995.2.275.
pMID: 7497129
36. Niedringhaus, T.P., Milanova, D., Kerby, M.B., Snyder, M.P., Barron, A.E.: Land-
scape of next-generation sequencing technologies. Anal. Chem. 83(12), 4327–4341
(2011). https://doi.org/10.1021/ac2010857
37. Pandey, P., Bender, M.A., Johnson, R., Patro, R.: deBGR: an efficient and near-
exact representation of the weighted de Bruijn graph. Bioinformatics 33(14), i133–
i141 (2017). https://doi.org/10.1093/bioinformatics/btx261
38. Rahman, M.M., Sharker, R., Biswas, S., Rahman, M.: HaVec: an efficient de Bruijn
graph construction algorithm for genome assembly. Int. J. Genom. 2017, 1–12
(2017). https://doi.org/10.1155/2017/6120980
39. Rizk, G., Lavenier, D., Chikhi, R.: DSK: k-mer counting with very low
memory usage. Bioinformatics 29(5), 652–653 (2013). https://doi.org/10.1093/
bioinformatics/btt020
40. Salikhov, K., Sacomoto, G., Kucherov, G.: Using cascading bloom filters to improve
the memory usage for de Brujin graphs. Algorithms Mol. Biol.: AMB 9, 2 (2014).
https://doi.org/10.1186/1748-7188-9-2
41. Sanger, F., Coulson, A., Barrell, B., Smith, A., Roe, B.: Cloning in single-stranded
bacteriophage as an aid to rapid DNA sequencing. J. Mol. Biol. 143(2), 161–178
(1980). https://doi.org/10.1016/0022-2836(80)90196-5
12 E. M. de Armas et al.

42. Schatz, M.C., Delcher, A.L., Salzberg, S.L.: Assembly of large genomes using
second-generation sequencing. Genome Res. 20(9), 1165–1173 (2010). https://doi.
org/10.1101/gr.101360.109
43. Silva, M.V.M., de Holanda, M.T., Haeusler, E.H., de Armas, E.M., Lifschitz, S.:
VelvetH-DB: Persistência de Dados no Processo de Montagem de Fragmentos
de Sequências Biológicas. In: Proceedings Satellite Events of the 32nd Brazilian
Symposium on Databases. Databases Meet Bioinformatics Workshop, pp. 334–341
(2017)
44. Simpson, J.T., Durbin, R.: Efficient construction of an assembly string graph using
the FM-index. Bioinformatics (Oxford, England) 26(12), i367–i373 (2010). https://
doi.org/10.1093/bioinformatics/btq217
45. Simpson, J.T., Wong, K., Jackman, S.D., Schein, J.E., Jones, S.J., Birol, I.: ABySS:
a parallel assembler for short read sequence data. Genome Res. 19(6), 1117–1123
(2009). https://doi.org/10.1101/gr.089532.108
46. Titus Brown, C., Howe, A., Zhang, Q., Pyrkosz, A.B., Brom, T.H.: A reference-
free algorithm for computational normalization of shotgun sequencing data. arXiv
e-prints arXiv:1203.4802 (2012)
47. Ye, C., Sam Ma, Z., Cannon, C., Pop, M., Yu, D.: Exploiting sparseness in de novo
genome assembly. BMC Bioinform. 13(Suppl. 6), S1 (2012). https://doi.org/10.
1186/1471-2105-13-S6-S1
48. Zerbino, D.: Velvet software. EMBL-EBI. https://www.ebi.ac.uk/zerbino/velvet/
(2016). Accessed 15 June 2019
49. Zerbino, D.R., Birney, E.: Velvet: algorithms for de novo short read assembly using
de Bruijn graphs. Genome Res. 18(5), 821–829 (2008). https://doi.org/10.1101/
gr.074492.107
50. Zhang, Q., Pell, J., Canino-Koning, R., Howe, A.C., Brown, C.T.: These are not the
K-mers you are looking for: efficient online K-mer Counting Using a Probabilistic
Data Structure. PLoS ONE 9(7), 1–13 (2014). https://doi.org/10.1371/journal.
pone.0101271
 Redundancy Treatment of NGS Contigs
in Microbial Genome Finishing
with Hashing-Based Approach

Marcus Braga1(B) , Kenny Pinheiro2 , Fabrı́cio Araújo2 ,

Fábio Miranda2 , Artur Silva2 , and Rommel Ramos2
1
Campus Paragominas, Universidade Federal Rural da Amazâonia,
Paragominas, Pará, Brazil
[email protected]
2
Instituto de Ciências Biológicas, Universidade Federal do Pará, Belém, Pará,
Brazil

Abstract. Repetitive DNA sequences longer than reads’ length produce

assembly gaps. In addition, repetition can cause complex and misassem-
bled rearrangements that creates branches in assembler graphs. Algo-
rithms must decide which way is the best. Incorrect decisions create
false associations, called chimeric contigs. Reads coming from different
copies of a repetitive region on genome may be wrongly assembled as
a unique contig, a repetitive contig. Furthermore, the growth of hybrid
assembling approaches using different sequencing platforms data, differ-
ent fragment sizes or even data from distinct assemblers are responsi-
ble for significantly increasing in the amount of generated contigs and
therefore subsequent redundancy on data. Thus, this work presents a
hybrid computational method to detect and eliminate redundant con-
tigs from microbial genome assemblies. It consists of two Hashing-Based
techniques: a Bloom Filter to detect duplicated contigs and a Locality-
Sensitive Hashing (LSH) to remove similar contigs. The redundancy
reduction facilitates downstream analysis and diminishes the required
time to finishing and curate genomic assemblies. The hybrid assembly of
GAGE-B dataset was performed with SPAdes (De Bruijn Graph) assem-
bler and Fermi (OLC) assembler. The proposed pipeline was applied to
the resulting contigs and the performance compared to other similar tools
such as HSBLASTN, Simplifier and CD-HIT. Results are presented.

Keywords: NGS contigs · Redundancy detection · Genome finishing ·

Bloom filter · LSH

1 Introduction
DNA sequencing is routinely used in various fields of biology. When Whole
Genome Sequencing is performed, the DNA is fragmented and the nucleotides are
sequenced. High-Throughput Sequencing Technology, also known as Next Gen-
eration Sequencing (NGS), allowed the parallelization of the sequencing process,
c Springer Nature Switzerland AG 2020
J. C. Setubal and W. M. Silva (Eds.): BSB 2020, LNBI 12558, pp. 13–24, 2020.
https://doi.org/10.1007/978-3-030-65775-8_2
14 M. Braga et al.

generating much more data than previous methods [1]. From the data generated
by NGS technologies, several new applications have emerged. Many of these
analyzes begin with the computational process of sequence assembly [2]. NGS
sequence assembly consists of grouping a set of sequences generated in sequenc-
ing, producing longer contiguous sequences, called contigs. These contigs are
joined together to form even larger known sequences, the scaffolds [3].
There are two general approaches to assembling NGS fragments: reference-
based and de novo approaches. In the first approach, a reference genome of a
related species is used as a guide to align the reads. De novo assembly is based
only on the overlapping reads to generate contigs [4]. These, in turn, may contain
gaps (regions not represented in the assembly).
New hybrid strategies have been developed to take advantage of each type
of assembly [5,6]. For example, hybrid strategies can combine reads and assem-
blies from different sequencing technologies and different assembly algorithms, or
use assemblies generated by different assemblers, combining the results (contigs
and/or scaffolds) produced by those tools to produce a new sequence [7].

1.1 DNA Repetitions and Contigs Redundancy

Repetitive sequences of DNA are present in all genomes. Repetitions have

always been technical challenges for sequence mapping and assembly tools. NGS
sequencing with short reads and high throughput made these challenges more
difficult. From a computational perspective, repetitions create ambiguities in
alignment and assembly which can lead to distortions and errors in the interpre-
tation of results [8].
Repetitions can include from only two copies to millions of copies, can range
in size from one to two bases (mono and dinucleotide) to millions of bases. The
best documented example of interspersed repetitions in the human genome is the
Alu repetitive element class, which covers approximately 11% of the genomes [9].
Repetitions that are sufficiently divergent do not present many computational
problems, therefore, for this work, a repetition is defined as a sequence of at least
100 bp in length that occurs two or more times in the genome and exhibits over
97% of identity for at least one other copy of itself. This definition excludes many
repetitive sequences, but includes those that present the main computational
challenges [10].
Repetitions may create gaps in de novo assembly. Besides creating gaps, rep-
etitions can erroneously collapse over one another and cause complex rearrange-
ments and assembly errors [11]. The degree of difficulty, in terms of correctness
and contiguity, that repetitions cause during genome assembly largely depends
on the read’s length.
Several new assemblers have emerged that address this problem, in particular
overlay-based and de Bruijn-based graphs. Both create graphs of different types
from the sequencing data. Repetitions cause ramifications in these graphs and
assemblers must guess which branch to follow. Incorrect decisions create false
associations, generating chimeric contigs and wrong copy numbers. If the assem-
 Redundancy Treatment of NGS Contigs with Hashing-Based Approach 15

bler is more conservative, it will break the assembly at these junction points,
leading to precise but fragmented assembly with very small contigs.
The central problem with the repetitions is that the assembler is unable
to distinguish from each other, which means that the regions flanking these
repetitions can be mistakenly assembled. It is common for an assembler to create
a chimera by joining two regions that are not close in the genome and, in the
end, the reads eventually align with the wrongly assembled genome [8].
There is a combination of strategies for solving repetitive DNA problems,
including the use of varying fragment libraries [12], post-processing software to
detect wrong assemblies [11], analysis of coverage statistics to detect and resolve
entanglements in DBGs.
NGS Technologies remain unable to generate one single sequence per chro-
mosome, instead, they produce a large and redundant set of reads, each read
being a fragment of the complete genome [13].
Assembly algorithms explore graphs through heuristics, selecting and travers-
ing paths and generating sequences as output. The set of contigs is hardly sat-
isfactory and usually needs to be postprocessed, most of the time to discard
very short contigs or contigs contained within other contigs. This procedure is
performed to reduce redundancy [14].
Specific errors inherent in sequencing platforms also affect the quality of the
generated assembly. In the case of the 454 and Ion Torrent platforms, for exam-
ple, errors in identifying homopolymer sequences may affect the construction of
contigs in the de Bruijn graph since k-mers derived from these regions may not
show agreement, resulting in a greater fragmentation of the assembly. Thus, it
is important to consider using different assembly tools and to give preference to
those that have greater tolerance to errors observed in the platform of interest.
The point is that when using a hybrid assembly approach, the problem of
redundant contigs persists and even increases. When assembling with different
assemblers using different methods, the contigs resulting from these assemblies
are merged, generating a much larger amount. These contigs, in general, repre-
sent different regions of the genome and lead to different gaps. The large number of
contigs generated by these hybrid assembly approaches require considerable com-
putational and human resources for analysis, especially for the identification of
assembly errors and the elimination of bases with higher probability of contigu-
ous edge error, which prevents the extent of overlap due to mismatch errors [15].

1.2 Computational Methods for Redundancy Detection

in Sequences

There are a large number of computational methods that can identify redundan-
cies in text sequences, such as biological sequences. String matching, or pattern
matching, or sequence matching, is a classic computational problem. Algorithms
of this nature are used to find matches between a standard input string and a
specified string. These methods can locate all occurrences of a character pattern
within another sequence [16].
16 M. Braga et al.

Several methods have been proposed to find similarities between sequences.

Some of these look for exact matches between sequences while others allow char-
acter insertions, deletions or substitutions trying to find the best possible align-
ment [17].
These algorithms can be divided into two main categories: exact string match-
ing and approximate string matching methods, which are a generalization of the
previous approach, which looks for similar or approximate patterns (sub-chains)
in a string [18].

Similarity Search. One form of searching for similarities is the Nearest Neigh-
bor Search and consists of locating data items whose distances to a query item
are the smallest in a large data set [19]. The search for similarities has become a
primary computational task in many areas, including pattern recognition, data
mining, biological databases, multimedia information retrieval, machine learning,
data compression, computer vision and statistical data analysis. The concept of
exact string matching rarely has meaning in these environments, while concepts
such as proximity, distance (similarity/dissimilarity) are much more beneficial
for this type of search [20].
Several methods have been developed to solve this kind of problem and many
efforts have been devoted to the approximate search. Hashing techniques have
been widely studied, especially Locality Sensitive Hashing, as a useful concept
for these categories of applications.
In general terms, hashing is an approach where it turns the data item into
a small numeric representation or, equivalently, a short code consisting of a
sequence of bits [21]. Hashing the nearest neighbor can be done in two ways:
by indexing data items through hash tables that store items with the same
code in the same hash bucket or by approximating the distance using what was
calculated with short codes [22].
The hashing approach to the approximate search aims to map the query
items to the destination items so that the approximate search by the nearest
neighbor can be performed efficiently and accurately using the destination items
and possibly a small subset of the raw query items. Target items are called hash
codes (also known as hash values) [23].

Locality Sensitive Hashing (LSH). The term Locality Sensitive Hashing was
introduced in 1988 [24] to designate a randomized structure able to efficiently
search for the nearest neighbor in large spaces. It is based on the definition of
LSH functions, a family of hash functions that maps similar input items with the
same hash code to a much higher probability than for different items. The first
specific LSH function, minHash, was invented by Broder [25] for the detection
and grouping of nearly duplicate web pages and is still one of the most studied
and applied LSH methods [20].
In conventional hashing, close items, which are similar, can be mapped/
scattered to different positions after hashing, but in LSH similar items remain
close even after hashing.
 Redundancy Treatment of NGS Contigs with Hashing-Based Approach 17

In LSH, we call candidate pairs to item pairs that have been mapped to
the same compartment. When the banding technique is applied, comparisons
are performed only on candidate pairs rather than on all pairs, as in a linear
search. If the goal is to find an exact match, techniques used to process data
such as MapReduce [26], Twitter Storm [27] can be used. These techniques are
based on parallelism, resulting in reduced time, however, these methods require
additional hardware. In most cases, only the most similar pairs are desired. The
level of similarity is defined by some threshold and the desired result is what is
known by searching for the nearest neighbor, and in these cases LSH model is
the best option. To apply LSH in different applications, it needs to be developed
according to the application domain [28].
Another important consideration is that false negatives and false positives
should be avoided as far as possible. It is said that there is a false negative when
the most similar pairs are not mapped to the same compartment. A false positive
happens when different pairs are mapped to the same compartment [29].

Bloom Filter. Bloom Filter is a probabilistic data structure that uses multiple
hash functions to store data in a large array of bits. It was introduced in 1970 by
Burton H. Bloom and is used in applications that perform membership queries
on a large set of elements [30].
A Bloom Filter (BF) is, therefore, a simple data structure that uses bit arrays
to represent a set and determine whether or not an element is present in it. False
positives are allowed, that is to say, with high probability, if an element is in the
set. On the other hand, false negatives are not possible, that is, it is possible
to know exactly if an element does not belong to the set. The achieved space
savings can often overcome this disadvantage of false positives if the probability
of error is controlled [31].
BF uses hash functions to map elements in an array, the filter. The member-
ship is tested by comparing the mapping results with the potential members of
the vector. An element is considered part of the set if and only if a hash function
maps that location to a key [32].

1.3 Contribution of This Work

Repetitions in the genome associated with small reads length are known to be
one of the most common reasons for fragmentation of a consensus sequence, as
reads that come from different copies of a repetitive region in the genome end
up not being properly identified and assembled in the same contig, known as
repetitive contigs. This problem is already addressed with a set of strategies
such as using miscellaneous fragment libraries, postprocessing software to detect
assembly errors, and coverage statistics analysis to detect and resolve entan-
glements in DBGs, however, none of these mechanisms completely solves the
problem. On the other hand, sequencing errors also affect the assembly, either
by generating further fragmentation of the assembly, or by resulting in abrupt
ends in graph pathways and ultimately contigs. As a result, hybrid assembly
18 M. Braga et al.

approaches are often used to try to achieve greater error tolerance. However, in
these approaches, a new problem arises, the production of an even larger amount
of redundant contigs.
In this paper, we present a computational method for detecting and elimi-
nating redundant contigs from microbial assemblies based on Bloom Filter and
LSH combination, allowing to minimize the computational effort in the genome
finishing step.

2 Biological Dataset and Assembly

The GAGE-B dataset [33] is used to evaluate large-scale genomic assembly algo-
rithms. GAGE-B data is originated from the genome sequencing of eight bacteria,
with size between 2.9 and 5.4 Mb and GC content between 33 and 69%. This data
is publicly available for Illumina sequencing. Some genomes were included which
HiSeq and MiSeq data were available also, resulting in 12 datasets (Table 1). All
GAGE-B datasets had underwent preprocessing steps such as adapter removal
and q10 quality trimming using the Ea-utils package [34].
The de novo assembly of the GAGE-B datasets were performed with SPAdes
[35] and Fermi [36] assemblers on an AMD Opteron (TM) Processor 6376 com-
puter with 64 CPUs, 1TB RAM, operating system CentOS release 5.10 Linux
OS version 2.6.18371.6.1.el5.

Table 1. Microbial genomes and sequenced reads from GAGE-B.

Species Genome size % GC NGS Read size Frag. size Coverage

(MB) content platform (pb) (pb)
A.hydrophila SSU 4.7 65 HiSeq 101 180 250x
B.cereus VD118 5.4 35 HiSeq 101 180 100–300
B.cereus ATCC 5.4 35 MiSeq 250 600 100x
10987
B.fragilis HMW 5.3 43 HiSeq 101 180 250x
615
M.abscessus 5.1 64 HiSeq 100 335 115x
6G-0125-R
M.abscessus 5.1 64 MiSeq 250 335 100x
6G-0125-R
R.sphaeroides 2.4.1 4.6 69 HiSeq 101 220 210x
R.sphaeroides 2.4.1 4.6 69 MiSeq 251 540 100x
S.aureus M0927 2.9 33 HiSeq 101 180 250x
V.cholerae 4.0 48 HiSeq 100 335 110x
CO1032(5)
V.cholerae 4.0 48 MiSeq 250 335 100x
CO1032(5)
X.axonopodis pv. 2.9 33 HiSeq 101 400 250x
Manihotis UA323
*Note: All datasets used paired-end reads from both ends of each fragment.
 Redundancy Treatment of NGS Contigs with Hashing-Based Approach 19

Prior to assembling with SPAdes, KmerGenie software [37] was used to esti-
mates the best k-mer length for each assembly. To predict the best K value, the
raw reads of each sample were used. First, we merged the forward and reverse
reads into a single file. Then the file containing the forward and reverse reads was
submitted to KmerGenie. For two samples, the assembly was performed with K
values arbitrated by SPAdes itself.
After the assemblies, the contigs generated by both assemblers were sub-
mitted to the QUAST software [38] to measure the assembly performance. In
addition to the number of contigs and the N50 value, the largest contig, com-
pleteness (genome fraction -%), misassemblies and number of genes (complete
and partial) metrics were also computed.

3 The Proposed Hybrid Model

In this application, a standard Bloom Filter was implemented on a bit array to

represent a set of fixed length contigs. For illustrative purposes, we assume that
the size of the BF is 20. All positions in the filter are started with the value
0. At first, the contigs are presented to the BF (insertion operation). For each
contig, BF uses hash functions (in this case 3) to determine positions along the
array. Each hash function generates hash codes, which will be the corresponding
positions in the filter. Since the functions are conventional, a uniform distribu-
tion of the positions generated along the structure occurs. Finally, each position
indicated by a hash code has its value changed from 0 to 1. The contig does not
have to be stored in the structure, only the indication of whether it is present
or not, through the value of bit 0 (missing) or 1 (present) in the set. This makes
BF fast and low on memory space usage.
To determine if a contig belongs to the set (test operation), the contig is
submitted to the same hash functions that generate the positions in the vector.
If the values in all generated positions are equal to 1, then that contig already
belongs to the set. In this case we have a duplication, i.e., an equal contig has
been detected and must be eliminated from the set. BFs are probabilistic struc-
tures, that is, if all positions correspond to 1, it can be said with high probability
the contig is part of the set. BFs are susceptible to false positives. Otherwise, if
the contig is submitted to BF and at least one of the generated hash codes maps
to a position with a 0 value, it can be stated with absolute certainty that the
contig does not belong to the set, since the BFs are not susceptible to false nega-
tives. These membership queries are possible because of uniform hash functions
properties.
The LSH approach used in this hybrid model to detect similar contigs is
also used to find similar items in documents and is divided into three steps:
Shingling, where contigs are divided into smaller k-size sequences (k-shingles or
k-grams); MinHashing, where smaller sequences are converted to small numeric
values (signatures) without losing their original similarity information; Locality-
Sensitive Hashing, where the similarity of signature (candidate) pairs is verified.
The proposed computational application, BFLSH, was developed in Java
and consists of two main modules. The first module implements a standard
20 M. Braga et al.

Bloom Filter with MD5, CRC32, and SHA-1 hash functions to do an exact string
match. Duplicate contigs are deleted and a new file is generated with the unique
sequences only. Then, an LSH approach performs approximate string match-
ing to identify similar contigs. The LSH function chosen was MinHash, which
behaves well with the Jaccard distance. The LSH method was implemented as
described in [39]. The contigs generated in the GAGE-B assemblies were sep-
arated into different files in FASTA format and submitted to the pipeline for
detection and removal of redundant copies.
To compare and measure the performance of the proposed model, three other
methods for reducing redundancy were implemented or used. The first, not yet
published, is based on HS-BLASTN [40], which uses the Burrows-Wheeler trans-
form as a basis for sequence alignment. The second was the Simplifier [13] and
finally the CD-HIT [41] was applied.
All programs were performed on an AMD Opteron (TM) Processor 6376
computer, 64 CPUs, 1TB RAM, CentOS release 5.10 operating system OS Linux
version 2.6.18371.6.1.el5. The similarity percentage used as a reference to con-
sider contigs as similar was 75%.

4 Results and Discussion

BFLSH was applied to the 12 GAGE-B datasets to reduce redundant contigs

and the results produced are shown in Table 2.

Table 2. BFLSH applied to the contigs of GAGE-B.

Species # Contigs # Contigs BFLSH # Contigs # Contigs BFLSH

SPAdes (%reduction) Fermi (%reduction)
A.hydrophila (HiSeq) 350 277 (20,85%) 1.600 371 (76,81%)
B.cereus (HiSeq) 4.248 3.484 (17,98%) 1.993 1.009 (49,37%)
B.cereus (MiSeq) 347 153 (55,90%) 386 129 (66,58%)
B.fragilis (HiSeq) 191 155 (18,84%) 1.028 538 (47,66%)
M.abscessus (HiSeq) 349 343 (1,71%) 1.204 246 (79,56%)
M.abscessus (MiSeq) 2.769 1.430 (48,35%) 11.172 10.897 (2,46%)
R.sphaeroides (HiSeq) 514 321 (37,54%) 4.482 1.863 (58,43%)
R.sphaeroides (MiSeq) 4.031 4.031 (0%) 1.572 727 (53,75%)
S.aureus (HiSeq) 702 655 (6,69%) 343 119 (65,30%)
V.cholerae (HiSeq) 247 130 (47,36%) 1.866 553 (70,36%)
V.cholerae (MiSeq) 4.731 2.584 (45,38%) 7.461 7.276 (2,47%)
X.axonopodis pv. Manihotis 305 212 (30,49%) 9.900 4.508 (54,46%)
(HiSeq)
* Bold and italic values represent for which assembly the method reduced the most contigs per
organism, proportionally.

In percentage terms, BFLSH achieved better results in eliminating redun-

dant contigs in SPAdes assemblies for the M.abscessus (MiSeq) and V.cholerae
(MiSeq) data. On the other hand, the method performed better on A.hydrophila
(HiSeq), B.cereus (HiSeq), B.cereus (MiSeq), B.fragilis (Hi-Seq), M.abscessus
 Redundancy Treatment of NGS Contigs with Hashing-Based Approach 21

(HiSeq), R.sphaeroides (HiSeq), R.sphaeroides (MiSeq), S.aureus (HiSeq),

V.cholerae (HiSeq) and X.axonopodis pv. Manihotis (HiSeq). The results of the
comparison between BFLSH and HSBLASTN, Simplifier and CD-HIT) in reduc-
ing SPAdes-generated contigs are shown in Table 3.

Table 3. Comparison between redundancy remove methods. SPAdes data.

Species Raw contigs BFLSH HS-BLASTN Simplifier CD-HIT

A.hydrophila (HiSeq) 350 277 299 346 275
B.cereus (HiSeq) 4.248 3.484 3.677 4.102 3.353
B.cereus (MiSeq) 347 153 215 321 120
B.fragilis (HiSeq) 191 155 178 190 156
M.abscessus (HiSeq) 349 343 344 349 346
M.abscessus (MiSeq) 2.769 1.430 1.632 1.783 780
R.sphaeroides (HiSeq) 514 321 366 490 303
R.sphaeroides (MiSeq) 4.031 4.031 3.130 2.310 1.088
S.aureus (HiSeq) 702 655 680 698 649
V.cholerae (HiSeq) 247 130 167 222 120
V.cholerae (MiSeq) 4.731 2.584 2.671 3.522 1.780
X.axonopodis pv. Manihotis 305 212 251 295 186
(HiSeq)

For the data generated by SPAdes, the most efficient method for decreasing
contiguity redundancy was CD-HIT, having obtained the largest reduction for 10
out of 12 datasets (Table 3, however BFLSH achieved the second lowest number
of contigs for nine organisms Table 3. For R.sphaeroides (MiSeq), BFLSH could
not detect any contiguous redundancy in SPAdes data, while the other meth-
ods found and considerably reduced the number of contiguous sequences. When
BFLSH was applied to the contigs generated with Fermi, there was a significant
elimination of redundancy (53.75%), compatible with the result obtained by the
other compared methods (HS-BLATN, Simplifier and CD-HIT).
It can be inferred that, for this organism and in the specific case of assembling
with SPAdes, the proposed Hashing-based technique failed to properly map the
actual similarities between the contigs into the Bloom Filter, that is, it was not
able to efficiently translate the proximity between the elements of the set of
contigs using the Jaccard similarity used in the LSH method.
Another hypothesis would be the usage of inappropriate parameters during
the application of the method for this isolated case. Bloom Filter’s parameter
variance, such as the number of hash functions, the size of the BF, or even the
amount of difference allowed to consider two items as similar can affect the engine
output. LSH methods may be more or less stringent depending on the parameter
used for similarity. In the case of this work, a value of 75% was used as default
for all experiments. One last possibility can still be raised: the difference in the
assembly approach used by SPAdes and Fermi. Since SPAdes is based on Bruijn
22 M. Braga et al.

Table 4. Comparison between redundancy removal methods. Fermi Data.

Species Raw contigs BFLSH HS-BLASTN Simplifier CD-HIT

A.hydrophila (HiSeq) 1.600 371 1.344 1.506 266
B.cereus (HiSeq) 1.993 1.009 1.298 1.689 777
B.cereus (MiSeq) 386 129 351 346 119
B.fragilis (HiSeq) 1.028 538 873 954 401
M.abscessus (HiSeq) 1.204 246 1.021 1.168 229
M.abscessus (MiSeq) 11.172 10.897 11.004 10.437 9.191
R.sphaeroides (HiSeq) 4.482 1.863 1.989 4.299 1.631
R.sphaeroides (MiSeq) 1.572 727 1.402 1.550 621
S.aureus (HiSeq) 343 119 290 253 99
V.cholerae (HiSeq) 1.866 553 1.514 1.788 425
V.cholerae (MiSeq) 7.461 7.276 7.384 7.014 6.291
X.axonopodis pv. 9.900 4.508 4.999 9.453 987
Manihotis (HiSeq)

graphs, its internal way of generating contigs uses kmers, while Fermi, being
from the OLC family of assemblers, does not. This factor may have in some way
affected the nature of the contigs generated by SPAdes for this organism, which,
combined with the other factors, may have led to unexpected yield for this case.
The results of the comparison between BFLSH and other methods for reduc-
ing Fermi-generated contigs are shown in Table 4.
In the comparison performed for the data generated by Fermi, the most
efficient method to reduce contigs redundancy was CD-HIT, having obtained
the largest reduction for the 12 datasets. However, BFLSH achieved the second
lowest number of contigs for twelve data sets in absolute terms.

5 Conclusion
A computational application has been developed capable of efficiently detecting
and eliminating redundant NGS contigs generated by de novo assemblers in
microbial genomes. The problem of redundant contigs generated in NGS data
assembly has been minimized with BFLSH.
For efficiency evaluation, the BFLSH was compared with three other distinct
methods and proved to be efficient. Bloom Filter and LSH techniques can effec-
tively be used to find similar items in biological sequences. The probabilistic
nature of hash functions makes the possibility of false negatives possible, but
these can be minimized with proper techniques.
One way to consider in the future is to adopt a parallelization approach,
which can significantly increase the time and memory efficiency, especially of
the LSH step.
 Redundancy Treatment of NGS Contigs with Hashing-Based Approach 23

Another alternative implementation for future thinking would be the use of

other LSH functions, such as Super-Bit, which uses the Cosine similarity, among
others [5].

References
1. Ambardar, S., Gupta, R., Trakroo, D., Lal, R., Vakhlu, J.: High throughput
sequencing: an overview of sequencing chemistry. Indian J. Microbiol. 56(4), 394–
404 (2016)
2. Nagarajan, N., Pop, M.: Sequence assembly demystified. Nat. Rev. Genet. 14(3),
157–167 (2013)
3. El-Metwally, S., Hamza, T., Zakaria, M., Helmy, M.: Next-generation sequence
assembly: four stages of data processing and computational challenges. PLoS Com-
put. Biol. 9(12), e1003345 (2013)
4. Martin, J.A., Wang, Z.: Next-generation transcriptome assembly. Nat. Rev. Genet.
12(10), 671–682 (2011)
5. Goswami, M., et al.: Distance sensitive bloom filters without false negatives. In:
Proceedings of the Twenty-Eighth Annual ACM-SIAM Symposium on Discrete
Algorithms, SODA ’17, USA, 2017, pp. 257–269. Society for Industrial and Applied
Mathematics (2017)
6. Tang, L., Li, M., Fang-Xiang, W., Pan, Y., Wang, J.: MAC: Merging assemblies by
using adjacency algebraic model and classification. Front. Genet. 10, 1396 (2020)
7. de Sousa Paz, H.E.: reSHAPE : montagem hibrida de genomas com foco em organ-
ismos bacterianos combinando ferramentas de novo. Dissertacao (2018)
8. Treangen, T.J., Salzberg, S.L.: Repetitive DNA and next-generation sequencing:
computational challenges and solutions. Nat. Rev. Genet. 13(1), 36–46 (2011)
9. Batzer, M.A., Deininger, P.L.: Alu repeats and human genomic diversity. Nat. Rev.
Genet. 3(5), 370–379 (2002)
10. Zavodna, M., Bagshaw, A., Brauning, R., Gemmell, N.J.: The accuracy, feasibility
and challenges of sequencing short tandem repeats using next-generation sequenc-
ing platforms. PLoS ONE 9(12), e113862 (2014)
11. Phillippy, A.M., Schatz, M.C., Pop, M.: Genome assembly forensics: finding the
elusive mis-assembly. Genome Biol. 9(3), R55 (2008)
12. Wetzel, J., Kingsford, C., Pop, M.: Assessing the benefits of using mate-pairs
to resolve repeats in de novo short-read prokaryotic assemblies. BMC Bioinform.
12(1), 95 (2011)
13. Bradnam, K.R., et al.: Assemblathon 2: evaluating de novo methods of genome
assembly in three vertebrate species. GigaScience 2(1), 2047–217X (2013)
14. Nagarajan, N., Pop, M.: Parametric complexity of sequence assembly: theory and
applications to next generation sequencing. J. Comput. Biol. 16(7), 897–908 (2009)
15. Ramos, R.T.J., Carneiro, A.R., Azevedo, V., Schneider, M.P., Barh, D., Silva, A.:
Simplifier: a web tool to eliminate redundant NGS contigs. Bioinformation 8(20),
996–999 (2012)
16. Galil, Z., Giancarlo, R.: Data structures and algorithms for approximate string
matching. J. Complex. 4(1), 33–72 (1988)
17. Pandiselvam, P., Marimuthu, T., Lawrance, R.: A comparative study on string
matching algorithm of biological sequences (2014)
18. Al-Khamaiseh, K., ALShagarin, S.: A survey of string matching algorithms. Int.
J. Eng. Res. Appl. 4, 144–156 (2014)
24 M. Braga et al.

19. Wang, J., Shen, H.T., Song, J., Ji, J.: Hashing for similarity search: a survey (2014)
20. Chauhan, S.S., Batra, S.: Finding similar items using lsh and bloom filter. In:
2014 IEEE International Conference on Advanced Communications, Control and
Computing Technologies, pp. 1662–1666 (2014)
21. Bender, M.A., et al.: Don’t thrash: how to cache your hash on flash. Proc. VLDB
Endow. 5, 1627–1637 (2012)
22. Slaney, M., Casey, M.: Locality-sensitive hashing for finding nearest neighbors [lec-
ture notes]. Signal Process. Mag. IEEE 25, 128–131 (2008)
23. Baluja, S., Covell, M.: Learning to hash: forgiving hash functions and applications.
Data Min. Knowl. Disc. 17(3), 402–430 (2008)
24. Indyk, P., Motwani, R.: Approximate nearest neighbors: towards removing the
curse of dimensionality. In: Conference Proceedings of the Annual ACM Sympo-
sium on Theory of Computing, pp. 604–613 (2000)
25. Broder, A.Z.: On the resemblance and containment of documents. In: Proceedings
of the Compression and Complexity of SEQUENCES 1997 (Cat. No.97TB100171),
pp. 21–29 (1997)
26. Jain, R., Rawat, M., Jain, S.: Data optimization techniques using bloom filter in
big data. Int. J. Comput. Appl. 142, 23–27 (2016)
27. Stephens, Z.D., et al.: Big data: astronomical or genomical? PLoS Biol. 13(7),
e1002195 (2015)
28. Andoni, A., Indyk, P.: Near-optimal hashing algorithms for approximate nearest
neighbor in high dimensions. Commun. ACM 51(1), 117–122 (2008)
29. Ding, K., Huo, C., Fan, B., Xiang, S., Pan, C.: In defense of locality-sensitive
hashing. IEEE Trans. Neural Netw. Learn. Syst. 29(1), 87–103 (2018)
30. Bloom, B.H.: Space/time trade-offs in hash coding with allowable errors. Commun.
ACM 13(7), 422–426 (1970)
31. Broder, A., Mitzenmacher, M.: Survey: Network applications of bloom filters: a
survey. Internet Math. 1, 11 (2003)
32. Naor, M., Yogev, E.: Tight bounds for sliding bloom filters. Algorithmica 73(4),
652–672 (2015)
33. Magoc, T., et al.: GAGE-B: an evaluation of genome assemblers for bacterial organ-
isms. Bioinformatics 29(14), 1718–1725 (2013)
34. Aronesty, E.: Comparison of sequencing utility programs. Open Bioinform. J. 7(1),
1–8 (2013)
35. Bankevich, A., et al.: SPAdes: a new genome assembly algorithm and its applica-
tions to single-cell sequencing. J. Comput. Biol. 19(5), 455–477 (2012)
36. Li, H.: Exploring single-sample SNP and INDEL calling with whole-genome de
novo assembly. Bioinformatics (Oxford, England) 28(14), 1838–1844 (2012)
37. Chikhi, R., Medvedev, P.: Informed and automated k-mer size selection for genome
assembly. Bioinformatics 30(1), 31–37 (2013)
38. Gurevich, A., Saveliev, V., Vyahhi, N., Tesler, G.: QUAST: quality assessment tool
for genome assemblies. Bioinformatics 29(8), 1072–1075 (2013)
39. Leskovec, J., Rajaraman, A., Ullman, J.D.: Mining of Massive Datasets. Cambridge
University Press, Cambridge (2014)
40. Chen, Y., Ye, W., Zhang, Y., Yuesheng, X.: High speed BLASTN: an accelerated
MegaBLAST search tool. Nucleic Acids Res. 43(16), 7762–7768 (2015)
41. Li, W., Godzik, A.: CD-HIT: a fast program for clustering and comparing large
sets of protein or nucleotide sequences. Bioinformatics 22(13), 1658–1659 (2006)
 Efficient Out-of-Core Contig Generation

Julio Omar Prieto Entenza , Edward Hermann Haeusler , and Sérgio

Lifschitz(B)

Departamento de Informática - (PUC-Rio), Rio de Janeiro, Brazil

[email protected]

Abstract. Genome sequencing involves splitting a genome into a set

reads that are assembled into contigs that are eventually ordered and
organized as scaffolds. There are many programs that consider the use
of the de Bruijn Graph (dBG) but they must deal with a high computa-
tional cost, mainly due to internal RAM consumption. We propose to use
an external memory approach to deal with the de Bruijn graph construc-
tion focusing on contig generation. Our proposed algorithms are based
on well-known I/O efficient methods that identify unitigs and remove
errors such as tips and bubbles. Our analytical evaluation shows that
it becomes feasible to generate de Bruijn graphs to obtain the needed
contigs, independently of the available memory.

1 Introduction
Genome sequencing is the process that determines the order of nucleotides within
a DNA molecule. Modern instruments splits a genome into a set of many short
sequences (reads) that are assembled into longer contiguous sequences, contigs,
followed by the process of correctly ordering contigs into scaffolds [18].
We may associate genome sequencing with the problem of finding a Hamil-
tonian Cycle through an Overlap Layout Consensus (OLC) assembly method.
Alternatively, it can be modeled as the problem of finding a Eulerian Cycle con-
sidering the de Bruijn Graph (dBG) based methods [14]. The latter can be seen
as a breakthrough for the research on genome assembly. This is due to the fact
that to find a Hamiltonian Cycle is an NP-complete problem [14].
When we handle actual dBGs, we must consider the existence of errors that
appear due to high-frequencies distortions on Next-generation Sequencing (NGS)
platforms. These errors induce the dBG size to be more prominent than the
overlap graph used in the OLC genome sequencing method using the same reads.
To remove the errors, we need to have some data structure representation of
the dBG. Current real-world datasets induce challenging problems as they have
already reached high volumes and will continue to grow as sequencing technolo-
gies improve [19]. As a consequence, the dBG increases the complexity leading
to tangles and topological structures that are difficult to resolve [16]. Also, the
graph has a high memory footprint for large organisms (e.g., sugarcane plants)
and it becomes worse due to the increase of the genome datasets. Therefore,
there are research works that focus on dealing with the ever-growing graph sizes
c Springer Nature Switzerland AG 2020
J. C. Setubal and W. M. Silva (Eds.): BSB 2020, LNBI 12558, pp. 25–37, 2020.
https://doi.org/10.1007/978-3-030-65775-8_3
26 J. O. P. Entenza et al.

for dBG-based genome assembly [3,15]. Any proposed solution must be aware
of the fact that the dBG is not entirely built before error pruning. After the
splitting of the genome in the first phase and, for a natural number k chosen
based on an empirical criterion, for each read of size m we will have m − k + 1
possible k-mers, which correspond to the nodes of the dBG. The number of k-
mers depends on the adjacency determined by the r ead itself. The splitting error
introduced in a given read may large increase the size of the subgraph it induces.
Roughly speaking, the dBG is a set of k-mers (subsequences of length k) [23]
linked to each other, according to information provided by their reads. Some
k-mers can come from different reads and the information about adjacency sup-
plied by any other read is processed only at dBG constructing phase. Thus, error
pruning should happen at this particular stage. The memory needed is so signif-
icant that we must use external memory to accomplish the dBG construction.
The different approaches that deal with the dBG size aim to design
lightweight data structures to reduce the memory requirements and to fit the
assembly graph into the main memory. Although it might be efficient, the amount
of memory increases according to the size of the dataset and the DNA of the
organism. While bacteria genomes currently take only a few gigabytes of RAM,
species like mammalian and plants require over tens to hundreds of gigabytes.
For instance, approximately 0.5 TB of memory is required to build a hash table
of 75-mers for the human genome [19].
We propose algorithms to simplify and remove errors in the de Bruijn graph
using external memory. As a result, it will be able to generate contigs using a
fixed amount of RAM, independently of the read dataset size. There are other
works addressing de Bruijn graph processing using external memory [8,11], but
they focus only on the constructing of large de Bruijn graphs efficiently with no
error prune considerations. To the best of our knowledge, this is the first proposal
of using an external memory approach focusing on the dBG simplification and
errors removal for contig generation during dBG construction.
We show an algorithm that provides out-of-core contraction of unambiguous
paths with an I/O cost of O(|E|/B), where E is the set of edges of the dBG and
B is the size of the partition loaded to the RAM each time. With the overhead
for creating the new partitions, the overall I/O complexity is O((sort(|E|) +
|E|/B) log P ath), where P ath is the length of the longest unambiguous path in
the dBG. For a machine with memory M , and a dBG satisfying |E| < M 2 /4B
sort(E) is performed with I/O complexity O(|E|/B) [9]. Summing up the I/O
complexity of this out-of-core contraction of paths is O((|E|/B) log P ath). The
out-of-core graph cleaning phase, by removing tips and bubbles, is performed with
a similar I/O complexity O((|E|/B +sort(|E|))logP ath). The creation of contigs
is performed by a full scan of the graph with I/O complexity O(|V |/B).

2 De Novo Assembly Using de Bruijn Graph

Given a de Bruijn Graph G = (V, E) for genomic sequence assembly each node
holds unique k-mers, and the edges reflect consecutive nodes if they overlap by
 Efficient Out-of-Core Contig Generation 27

k−1 characters. The assembly aims to construct a set of contigs from the dBG G.
Given a dBG as input, to generate contigs is equivalent to output all contiguous
sequences which represent unambiguous paths in the graph.
The use of the dBG to generate contigs consists of a pipeline: nodes enu-
meration, compaction, and graph cleaning. In the first step, a set of distinct
k -length substrings (k-mers) is extracted from the reads. Each k-mer becomes a
graph node. Next, all paths with all but the first vertex having in-degree 1 and
all but the last vertex having out-degree 1 (unitigs) are compacted into a single
vertex. Finally, the last step removes topological issues from the graph due to
sequencing errors and polymorphism [5].
The number of nodes in the graph can be huge. For instance, the size of
the genome of white spruce is 20 Gbp and generates 10.7 × 109 k-mers (with
k = 31) and needs 4.3 TB of memory [5]. Also, the whole genome assembly of
22 Gbp (bp - base pairs) loblolly pine generates 13 × 109 k-mers and requires
800GB of memory [5]. Theoretically speaking, a 1,000 Genomes dataset with 200
Terabytes of data can generate about 247 or nodes, 64–128 times larger than the
problem size of the top result in the Graph 500 list [15].
Next-generation sequencing platforms do not provide comprehensive read
data from the genome sequences. Hence, the produced data is distorted by high
frequencies of sequencing errors and genomic repeats [18]. Sequencing errors
compound this problem because each such error corrupts the correct genomic
sequence into up to k erroneous k-mers. These erroneous k-mers introduce new
vertices and edges to the graph, significantly expanding its size and creating
topological artifacts as tips and bubbles [23].
Different solutions have been proposed to address the memory issues in
genome assembly problem. One approach samples the entire k-mer set and per-
forms the assembly process over the selected k-mers [21]. Another approach
address to encode the dBG into efficient data structures such as light-weight
hash tables [3], succinct data structures [2] or bloom filters[6,17]. There are
research works based on distributed memory systems for processing power and
memory demanding resources [3,7,15].
Although their apparent differences, all of these approaches are based exclu-
sively on in-memory systems. Consequently, if the size of the graph exceeds the
amount of memory available, it will be necessary to increase the RAM. As in the
next future, the size of datasets will increase dramatically, and this situation will
stress the different systems [20]. There is a need for new approaches to process
all of this massive amount of information in a scalable way. We propose in this
work to use an external memory approach to process the dBG. To increase the
amount of RAM does not guarantee that the graph will always fit.

3 Overview of Our Proposed Approach

Our basic pipeline of de novo genome assembly could be divided into five basic
operations [23]: 1) dBG construction, which constructs a dBG from the DNA
reads; 2) Contraction of unambiguous paths, which merges unambiguous vertices
28 J. O. P. Entenza et al.

into unitigs; 3) Graph cleaning, which filters and removes errors such as tips and
bubbles; 4) Contigs creation, which create a first draft of the contigs and 5)
Scaffolds, which joins the previous contigs together to form longer contigs. In
this work, we face steps 2, 3, and 4 using an external memory approach. We
assume a de Bruijn graph exists, and it is persisted as an edge-list format.

Fig. 1. Graph Contraction. Dashed arcs represent the messages and the label
between brackets indicates if a vertex is a tail (t) or a head (h). a) A flipped coin
choose which node will be t or h. If a tail vertex has out-degree = 1 then it sends a
message to its neighbour. b) If a h vertex receives a message and its in-degree = 1 then
both vertices are merged. c) Shows the result after some repeated steps.

The Contraction of unambiguous paths simplifies the graph by merging

some nodes in the graph. Whenever a node with only one outgoing directed
edge points out to another node with only one incoming directed edge, these
two nodes are merged. These single nodes are called unitigs and are maximal if
they extend in either direction. Thus, the problem of compacting a de Bruijn
graph is to report the set of all maximal unitigs. Figure 1 shows how we simplify
the dBG into a compacted graph. All the remained nodes are maximal unitigs.
The algorithmic solution to this problem is straightforward in the in-memory
context. Let’s assume that each path representing a unitig is a linked list where
the head and the tail can be branching nodes (see Fig. 1a). Then, to obtain a
maximal unitig, we only need to visit each uc node and merge them with its
successor node into the new one.
Although the in-memory algorithm is straightforward, the use of this app-
roach is not efficient in the external memory because the number of I/O accesses
is linear concerning the number of nodes. Finding all the maximal unitigs is
analogous to apply the well-known graph edge contraction technique in external
memory on all the unambiguous paths. Given a graph G, the contraction of an
edge (u, v) is the replacement of u and v with a single vertex such that edges
incident to the new vertex are the edges other than (u, v) that were incident with
u or v [22]. As in our case, this algorithm can only apply on unambiguous paths.
Therefore, we divided our proposal into two steps: 1) select branching nodes ua
 Efficient Out-of-Core Contig Generation 29

as head/tail of each unitig paths; and 2) apply the graph contraction technique
over these paths until all nodes are maximal unitigs.
In Graph cleaning we aim to remove short dead-end divergences, called tips,
from the main path. One strategy consists of testing each branching node for all
possible path extensions up to a specified minimum length. If the length of the
path is less than a certain threshold (set by the user) then the nodes belonging
to this path are removed from the graph [7,23].
The tips removing process is analogous to traversal the paths from a branch-
ing node, ua , to a dead-end node ue . The graph does not fit into RAM, even
after the unitig process. We need to traverse the dBG in an I/O efficient way
to find and remove all tips. Our algorithm is based on an external memory list
ranging from the ue to ua nodes. However, we have to make two significant mod-
ifications: (i) the ranking is represented by each edge/node’s coverage to decide
which path will be removed, and (ii) as two of more dead-ends could reach the
same ua node, we need to keep in RAM a data structure to make the traversal
backward. This way, we eliminate the selected path from the branching node.
Bubbles are paths that diverge from a node then converge into another.
The process of fixing bubbles begins by detecting the divergence points in the
graph [23]. For each point, all paths from it are detected by tracing the graph
forward until a convergence point is reached. Some assemblers restricts the size
of the bubble to n nodes where k ≤ n ≤ 2k [7], others use a modified version of
Dijkstra’s algorithm [23]. To simulate the different external memory approaches,
we need to identify all branching nodes ua . We execute an I/O-efficient breath-
first search (BFS) from ua until we find a visited node. It means that there is
a bubble at some point in the search (vb ). Then, we select the branch that will
be kept and start another BFS in a backward direction (the start node is vb ).
Finally, we remove the other paths until we find back ua . After the execution
of steps 2 and 3, the Contigs creation step involves the output of all the contigs
represented by nodes.
Processing Out-of-Core Graphs. Many graph engines implement a vertex-
centric or “think like a vertex” (TLAV) programming interface. This paradigm
iteratively executes a user-defined program over vertices of a graph. The vertex
program is designed from the vertex’s perspective, receiving as input the data
from adjacent vertices and incident edges. Execution halts after a specified num-
ber of iterations, called supersteps, are completed. It is important to note that
each vertex knows the global supersteps. There is no other knowledge about the
overall graph structure but its immediate neighbors and the messages that are
exchanged along their incident edges [13].
The computation graph engines proceed in supersteps. For every superstep,
it loads one or more partitions p based on available RAM. Then, it processes
the vertices and edges that belong to p and saves the partitions back to disk.
A different subset of partitions is then loaded and processed until all of them
have been treated for the given superstep. The process is then repeated for the
next superstep until there are no remaining vertices to visit (see Algorithm 1
30 J. O. P. Entenza et al.

Algorithm 1. Out-of-core graph processing taken from GraphChi [9]

Input:
G = (V, E): a general graph
UpdateFunction: a user-defined program to apply over the vertices
1: procedure OutOfCore(G, U pdateF unction)
2: for each superstep ∈ Algorithm do  superstep is a global variable
3: for each p ∈ Partition do
4: p.load()  Load partition to RAM
5: parallel for each v ∈ p.vertices() do  Apply UpdateFunction
6: UpdateFunction(G,v)
7: p.save()  Save partition to disk

from GraphChi [9]). If the machine has sufficient RAM to store the graph and
metadata, all partitions can be kept in RAM, and there is no disk access [9].
Because all the operations related to partitions and parallel vertices processing
are fixed, from now we will only highlight the UpdateFunction and the number
of supersteps. For simplicity, UpdateFunction(G,u) means that we apply the
function over the vertex u in the graph G.

4 Contig Generation

The graph contraction algorithm (Algorithm 2) is based on I/O-efficient list rank-

ing algorithm based on graph contraction [4] but using a TLAV [12]. The output
is the distance of each node from the beginning of the list. In this case, it cor-
responds to k-mer concatenation. Thus, in our output, the beginning node will
have the unitig concatenation. Initially, we assign the k-mer as the rank of each
vertex. We then continue recursively: first, in one superstep, we find a maximal
independent set among all vertices that belong to a unitig (line 11). For each
of these, we flip a fair coin and vertices that flipped “tails” pick a neighbor
that flipped “heads” (if any) to contract with [1]. Later the vertices identified
as “tails” with precisely one outgoing edge send the required information to
its neighbors (lines 12 and 13). In the next superstep, the vertices marked as
head, with only one in-going edge, are merged with tail vertices, and the edge
information is updated (lines 17 to 22). The function PreprocessNewGraph cre-
ates new partitions from the removed and added vertices and edges (line 9).
This step implies merging nodes, and remove duplicate edges and update the
graph partitioning. Next, we recursively continue coarsening the graph until all
unambiguous nodes are collapsed.
I/O Analysis. Lines 11–22 can be done with a full scan over all the graph
partitions. We load a partition into RAM, we update the vertices and edge
values, and then write them to the disk. Then, we load the next partition and its
vertices and edges according to the saved partition. So, the I/O cost is O(|E|/B).
On the other hand, P reprocessN ewGraph (line 9) creates new partitions
from superstep i − 1 to be processed in the superstep i. To create the new
 Efficient Out-of-Core Contig Generation 31

Algorithm 2. Graph Contraction

Input: G = (V, E): a contracted dBG
Output: G : a graph with all no ambiguous paths contracted.
1: procedure ContractGraph(G)
2: superstep ← 0 and merge ← true
3: while merge = true do
4: merge ← f alse
5: M ergeN odes(G, u)  External Memory Context
6: superstep ← superstep + 1
7: if superstep is odd then
8: G ← P reprocessN ewGraph(G)
9: procedure MergeNodes(u)
10: f lag ← T ∨ H in randomly way
11: if superstep is even then
12: if f lag = true ∧ d+ (u) = 1 then
13: send({f lag, seq, neighbors, id}, j)  message to outgoing edge
14: else
15: if d− (u) = 1 then  merge nodes from incoming neighbor
16: m ← receive({f lag, seq, i})
17: if m.f lag = f lag then
18: add edge(u, m.neighbors)
19: seq ← glue(seq, m.seq)
20: delete(m.id)
21: merge ← true

partitions, first, it is necessary to divide the nodes by their ID and later sort
all the edges based on their destination vertex ID [9]. Thus, the I/O cost of the
process for the created graph is O(sort(|E|)).
Finally, at each superstep, Algorithm 2 contracts a constant fraction of the
vertices per iteration by graph contraction [22]. It expected O(log P ath) iter-
ations with high probability, where P ath is the longest unambiguous path
in the graph [1]. Hence, the overall expected I/O cost for simplifying the
graph is O((sort(|E|) + |E|/B) log P ath). If we assume |E| < M 2 /(4B) then
sort(|E|) = O(|E|/B). This condition can be satisfied with a typical value of M ,
say 8 GB, B in the order of kilobytes and a graph size smaller than a petabyte
[10]. On these conditions the I/O cost is O((|E|/B) log P ath) (Fig. 2).
To remove tips, we design a straightforward procedure in few supersteps
(Algorithm 3). At this point, all nodes are contracted. Thus, all terminal nodes
are potential tips, and they may be removed. In the first superstep, the vertices
having an in-degree or out-degree of zero and sequence’s length less than 2k are
identified as potential tips (line 10), and a message is sent to their neighbors
(line 11). In the next superstep (lines 14–17), the vertices that received the
messages, search for the maximal multiplicity among all neighbors and remove
those potential tips with multiplicity less than the maximal value. Removing
the tips generates new linear paths in the graph that will be contracted. Note
32 J. O. P. Entenza et al.

Fig. 2. Removing tips. Dotted lines show the sent messages. a) F and I are
marked as potential tips. Later, they send a message to their neighbors. b) C and D
receive the messages and check their multiplicity to eliminate the real tips. So, H and
I are removed. c) The graph is compressed to obtain the final result.
that once the initial set of tips are removed, it could produce other tips. Most
assemblers execute the removal tip algorithm a fixed number of times.

Algorithm 3. Tips removal

Input: G = (V, E): a dBG

Output: G : another graph with tips removed.
1: procedure Tips(G)
2: superstep ← 0
3: while tips = true do
4: tips ← f alse
5: RemoveT ips(G, u)
6: superstep ← superstep + 1
7: ContractGraph(G))
8: procedure RemoveTips(u)  External Memory Context
9: if superstep = 0 then
10: if u is ue ∧ |seq| ≤ 2k then  Identify all potential tips
11: send(id)  Send a message to its neighbor
12: else
13: maximal ← max(∀u.neighbors.multip)  Get max multiplicity
14: for all m ∈ receive(id) do  Identify the real tips and remove them
15: if (u, m.id).mult < maximal then
16: delete node(m.id)
17: tips ← true
 Efficient Out-of-Core Contig Generation 33

I/O Analysis. The function RemoveT ips needs only two supersteps to remove
any tip given a graph: one to identify all tips and other to removes all of them
(lines 8–17). This can be carried out with at most two full scans over all the
graph partitions. Thus, the I/O cost is O(|E|/B). Although RemoveT ips only
executes two supersteps, the graph contraction dominates the I/O cost (line
7). The I/O cost is O((|E|/B + sort(|E|)) log P ath), where Path represents the
longest path created after the tips are removed.
The primary approach to identify and remove bubbles is based on BFS
(breadth-first search) algorithm. As bubbles consist of paths with very different
multiplicities, those paths with low multiplicity are deleted and use the path
with the highest multiplicity to represent the bubble. In our algorithm, each
vertex manages its history, which makes it easy to control the different paths
and to pick up the right one.
The proposed algorithm has two stages: forward and backward. In the for-
ward stage (Algorithm 4) identifies all paths that form a bubble and select one of
them. On the other hand, the backward stage, (Algorithm 5), removes the redun-
dant paths and compacts the graph. Due to space limitations, we will illustrate
and explain both algorithms by examples. See Fig. 3 for the forward stage and
Fig. 4 for the backward stage.

Fig. 3. Forward Bubble Detection. The figures only show the id par in the sent
messages. a) A is a potential bubble beginning, so it sends messages to the outgoing
neighbors. b) B and E keep it and forward new messages updating the vertices IDs. c)
When H, in-degree ≥ 2, receives a message it selects that E belongs to a bubble. Later,
it marks the node for the backward stage.

I/O Analysis. In the forward stage, SelectP ath (line 4) iterates through the
bubbles in a constant number of times depend on a limit value. Also, only a
small number of messages are present in the graph, each one originating from
any ambiguous vertex whose out-degree is greater than 2. Moreover, each mes-
sage will be passed along an edge exactly once, as notifications are only sent
34 J. O. P. Entenza et al.

Algorithm 4. Forward stage bubble detection

Input:
G = (V, E): a compressed dBG and limit: max. length of the bubble path

Output: G : a graph with all bubbles selected.
1: procedure FindBuble(G)
2: superstep ← 0
3: while superstep ≤ limit do
4: SelectP ath(G, u))
5: superstep ← superstep + 1
6: procedure SelectPath(G,u)  External Memory Context
7: if superstep = 0 ∧ d+ (vertex) ≥ 2 then  Identify possible bubble start
8: outgoing.add(id, id, seq)
9: else
10: outgoing ← {} and incomming ← receive(id1, id2, seq)
11: for each m ∈ incomming do
12: if m.id ∈/ history then
13: m.seq ← glue(seq, m.seq) and m.id2 ← id
14: history.add(m) and outgoing.add(m)
15: else
16: for each h ∈ history do  All possible bubble ends
17: if h.id1 = m.id1 then
18: apply heuristic(m, h)
19: if m is bubble then
20: mark m.id2 and history.add(m)
21: else
22: mark h.id2
23: send(outgoing)  send message to all out-edges

along outgoing edges. This means that only in-edges are read, and the out-
edges are written. As this algorithm implies an external BFS traversal, the I/O
cost is O(BP ath(|V | + |E|)/B) ≈ O(BP ath|E|/B) where BP ath is the longest
length among all bubbles and |E| = O(|V |) because a dBG is a sparse graph.
Then the total I/O cost is O(limit ∗ BP ath ∗ |E|/B) = O(BP ath ∗ |E|/B).
On the other hand, the backward stage uses another BFS but in the oppo-
site direction to the forward phase, so the I/O cost is the same. Addition-
ally, it iterates through the set of vertices (lines 6–8) and executes a contrac-
tion on the resulting graph. Therefore, the I/O cost of the backward stage is
O((|E|/B + sort(|E|)) log BP ath).
At this point, the graph should be formed by contracted vertices. As each
vertex represents a contig, we can output them using a full scan over the graph.
Thus, the I/O complexity is O(|V |/B), where |V | is the number of contigs.
 Efficient Out-of-Core Contig Generation 35

Fig. 4. Backward Bubble Elimination. a) H is a bubble ending, so it sends messages

to the incoming neighbors communicating that E is marked. b) When E receives the
message, it is marked and does not send more messages because it detects that the
next node is the bubble. B does not send any message because its history does not
have the node E. c) The graph is simplified after all bubbles are removed.

Algorithm 5. Backward step bubble detection

Input:
G = (V, E):compressed dBG and limit: max. length of the bubble path

Output: G : a graph with all bubbles nodes marked.
1: procedure BackTrackPath(G)
2: superstep ← 0
3: while superstep ≤ limit do
4: RemoveP ath(G, u)
5: superstep ← iter + 1
6: for each u ∈ V do  External Memory Context
7: if u.mark = true then
8: delete(v)
9: CompressGraph(G))
10: procedure RemovePath(G,u)  External Memory Context
11: if superstep = 0 ∧ d− (u) ≥ 2 then  Identify all possible bubble ends
12: for each m ∈ history do
13: if m is marked then
14: outgoing.add(m.id, m.id2)
15: else
16: outgoing ← {}; incomming ← receive(id1, id2)
17: for each m ∈ incomming do  Find the closest common ancestor
18: if m.id2 = id then
19: dst ← history.pop(m.id)
20: if dst = id then
21: outgoing.add(m) and mark = true
22: send(outgoing)
36 J. O. P. Entenza et al.

5 Conclusions
In this paper, we have proposed out-of-core algorithms for dealing with contig
generation, one of the most critical steps of fragment assembly methods based
on de Bruijn graphs. Besides presenting these algorithms, we have made an I/O
analytical evaluation that shows that it becomes feasible to generate de Bruijn
graphs to obtain the needed contigs, independently of the available memory.
The I/O cost studies show that graph simplification is one of the most expen-
sive steps. Actually, we could expect it because this phase involves a more sig-
nificant number of vertices and edges. To deal with that, one could choose to do
the assembly without simplifying the graph. In this condition, a tip is a branch
with low coverage and not just a vertex. Hence, it will be necessary to apply a
list ranking from the dead-end to branching nodes.
Among other issues, we may cite that it is hard to estimate the number of
executions related to each phase. Primarily it depends on the number of nodes
in the graph, which itself depends on the properties of the read’s datasets. As
these values vary from one dataset and sequencing technology to another, the
assembly algorithms execute each step, a fixed and empirical number of times. As
future work, we may cite the evaluation of other graph simplification approaches
targeting erroneous and non-recognizable structures, such as X-cuts.

Acknowledgments. This work is partially supported by grants from CNPq, CAPES

and FAPERJ, Brazilian Public Funding Agencies.

References
1. Anderson, R.J., Miller, G.L.: A simple randomized parallel algorithm for list-
ranking. Inf. Process. Lett. 33(5), 269–273 (1990)
2. Bowe, A., Onodera, T., Sadakane, K., Shibuya, T.: Succinct de Bruijn Graphs. In:
Raphael, B., Tang, J. (eds.) WABI 2012. LNCS, vol. 7534, pp. 225–235. Springer,
Heidelberg (2012). https://doi.org/10.1007/978-3-642-33122-0 18
3. Chapman, J.A., et al.: Meraculous: de novo genome assembly with short paired-end
reads. PLoS One 6(8), e23501 (2011)
4. Chiang, Y.J., et al.: External-memory graph algorithms. Procs. ACM/SIAM Symp.
Discr. Algorithm. (SODA) 95, 139–149 (1995)
5. Chikhi, R., Limasset, A., Medvedev, P.: Compacting de Bruijn graphs from
sequencing data quickly and in low memory. Bioinformatics 32(12), i201–i208
(2016)
6. Chikhi, R., Rizk, G.: Space-efficient and exact de Bruijn graph representation based
on a bloom filter. Algorithm. Mol. Biol. 8(1), 22 (2013)
7. Jackman, S.D., et al.: Abyss 2.0: resource-efficient assembly of large genomes using
a bloom filter. Genome Res. 27(5), 768–777 (2017)
8. Kundeti, V.K., et al.: Efficient parallel and out of core algorithms for constructing
large bi-directed de Bruijn graphs. BMC Bioinf. 11(1), 560 (2010)
9. Kyrola, A., Blelloch, G., Guestrin, C.: Graphchi: large-scale graph computation on
just a PC. In: USENIX Symposium on Operating Systems Design and Implemen-
tation (OSDI), pp. 31–46 (2012)
 Efficient Out-of-Core Contig Generation 37

10. Kyrola, A., Shun, J., Blelloch, G.: Beyond synchronous: new techniques for
external-memory graph connectivity and minimum spanning forest. In: Gudmunds-
son, J., Katajainen, J. (eds.) Experimental Algorithms — SEA 2014. Lecture Notes
in Computer Science, vol. 8504, pp. 123–137. Springer, Cham (2014). https://doi.
org/10.1007/978-3-319-07959-2 11
11. Li, Y., Kamousi, P., Han, F., Yang, S., Yan, X., Suri, S.: Memory efficient minimum
substring partitioning. Proc. VLDB Endow. 6(3), 169–180 (2013)
12. Malewicz, G., et al.: Pregel: a system for large-scale graph processing. In: Process
ACM SIGMOD Intl. Conf. on Manage. Data, pp. 135–146 (2010)
13. McCune, R.R., Weninger, T., Madey, G.: Thinking like a vertex: a survey of vertex-
centric frameworks for large-scale distributed graph processing. ACM Comput.
Surv. (CSUR) 48(2), 25:1–25:39 (2015)
14. Medvedev, P., Georgiou, K., Myers, G., Brudno, M.: Computability of models for
sequence assembly. In: Giancarlo, R., Hannenhalli, S. (eds.) Algorithms in Bioinfor-
matics — WABI 2007. Lecture Notes in Computer Science, pp. 289–301. Springer,
Cham (2007). https://doi.org/10.1007/978-3-540-74126-8 27
15. Meng, J., Seo, S., Balaji, P., Wei, Y., Wang, B., Feng, S.: Swap-assembler 2: opti-
mization of de novo genome assembler at extreme scale. In: Proceedings of the
45th ICPP, pp. 195–204. IEEE (2016)
16. Miller, J.R., Koren, S., Sutton, G.: Assembly algorithms for next-generation
sequencing data. Genomics 95(6), 315–327 (2010)
17. Salikhov, K., Sacomoto, G., Kucherov, G.: Using cascading bloom filters to improve
the memory usage for de Brujin graphs. Algorithm. Mol. Biol. 9(1), 2 (2014)
18. Simpson, J.T., Pop, M.: The theory and practice of genome sequence assembly.
Ann. Rev. Genomics Hum. Genet. 16, 153–172 (2015)
19. Sohn, J., Nam, J.W.: The present and future of de novo whole-genome assembly.
Briefings Bioinf. 19(1), 23–40 (2016)
20. Stephens, Z.D., et al.: Big data: astronomical or genomical? PLoS Bio. 13(7),
e1002195 (2015)
21. Ye, C., Ma, Z.S., Cannon, C.H., Pop, M., Douglas, W.Y.: Exploiting sparseness in
de novo genome assembly. BMC (BioMed Central) Bioinf. 13, S1 (2012)
22. Zeh, N.: I/o-efficient graph algorithms. In: EEF Summer School on Massive Data
Sets (2002)
23. Zerbino, D.R., Birney, E.: Velvet: algorithms for de novo short read assembly using
de Bruijn graphs. Gen. Res. 821–829 (2008)
 In silico Pathogenomic Analysis
of Corynebacterium Pseudotuberculosis
Biovar Ovis

Iago Rodrigues Blanco , Carlos Leonardo Araújo ,

and Adriana Carneiro Folador(B)

Laboratory of Genomics and Bioinformatics, Center of Genomics and Systems Biology,

Institute of Biological Sciences, Federal University of Pará, Belém, Pará, Brazil
[email protected]

Abstract. Corynebacterium pseudotuberculosis is a pathogenic bacterium that

may transmit caseous lymphadenitis, veterinary infection that severely attacks
animals such as goats and sheep. It is known that the toxin Phosholipase D is
the major virulence factor associated with this disease. However, genomic com-
putational studies can reveal further information concerning pathogenicity mech-
anisms of bacteria. Through sequence analysis tools, it is possible to assess the
genomic bases of these mechanisms and to analyze similarities among the dif-
ferent strains of this species. Nitrate reductase-negative bacteria are classified in
the biovar ovis, able to transmit the infection. Thus, we developed an in silico
comparative pathogenomic analysis with genomes of 33 strains of C. pseudotu-
berculosis biovar ovis strains, which cause caseous lymphadenitis. Looking for the
identification of pathogenicity-related genes, virulence factors and composition
of pathogenicity islands, it was possible to computationally predict pathogenic-
ity potentials of target proteins and their respective biological processes during
infection, besides identification of prophage genome elements and prediction of
protein protein interactions.

Keywords: Pathogenomics · Corynebacterium pseudotuberculosis ·

Pathogenicity islands · Virulence factors

1 Introduction

1.1 Corynebacterium Pseudotuberculosis

The bacterial genus Corynebacterium is a member of the CMNR group (Corynebac-

terium, Mycobacterium, Nocardia and Rhodococcus), composed of microorganisms that
share structural and genetic characteristics, such as cell wall organization and G + C con-
tent varying from 47% to 74% [1]. In this genus, species can be classified as pathogenic
and non-pathogenic. Within the first group, some species stand out, such as C. pseudo-
tuberculosis, C. diphtheriae and C. ulcerans, responsible for the transmission of caseous

© Springer Nature Switzerland AG 2020

J. C. Setubal and W. M. Silva (Eds.): BSB 2020, LNBI 12558, pp. 38–49, 2020.
https://doi.org/10.1007/978-3-030-65775-8_4
 In silico Pathogenomic Analysis 39

lymphadenitis (CLA), ulcerative lymphangitis and diphtheria. These bacteria are capa-
ble of producing virulence factors (VF) and pathogenicity-related proteins, responsible
for causing diseases [2].
C. pseudotuberculosis can be divided into two biovars: equi and ovis, causers of seri-
ous infections. This discrimination depends on the pathogen’s ability to reduce nitrate:
positive nitrate reductase bacteria are members of biovar equi, responsible for causing
ulcerative lymphangitis in buffaloes and horses, and negative nitrate reductase bacteria
are classified in biovar ovis, etiological causers of CLA, forming wounds and abscesses
in internal organs of small ruminants, as goats and sheep [3–5].

1.2 Comparative Pathogenomics

After the advent of next generation sequencing (NGS) technologies, there has been an
exponential growth in studies related to the investigation of new organisms and their
biological processes. Consequently, the time and cost applied to the techniques were
reduced and a revolution was made in genomics [6]. The assessment of bacterial genomes
through worldwide available databases allows the investigation of genetic factors through
in silico researches – those that focus on computational predictions. Thus, it is possible
to reveal pathogenicity mechanisms and other biological processes underlying infection
by studying the genome of the pathogen [7, 8].
Several biological processes relate with pathogenicity in Corynebacterium, as adher-
ence, iron uptake, secretion of toxins, invasion and colonization in the host [9]. It is pos-
sible to predict genes related with these functions by obtaining amino acid sequences
and aligning them in online databases. Pathogenomic analyses seek to identify impor-
tant genetic determinants of pathogenicity, as VF and their roles in pathogenesis through
analysis of the pathogen’s genetic repertoire [10]. Comparative genomic analyses are
increasingly fast and promising, highlighting its recent advances.
The analysis of determinants of pathogenicity in bacteria is useful to determine which
genes are responsible for pathogenic functions and how they work. We can observe the
conservation of these genes among different strains of a same species, or even among
different species of a same genus [11]. Comparative pathogenomics can point out impor-
tant genetic factors among determined genomes. The identification of VF is so important
once significant antimicrobial resistance has been observed in pathogenic bacteria, what
is associated to bacterial adaptation factors and its ability to modify biological mecha-
nisms adapting to the host’s defense system [12]. Therefore, the identification of VF can
reveal notable information about this pathogen [13, 14].

1.3 Determinants of Pathogenicity

Genes acquired through horizontal transfer are frequently found in the genomes of
pathogenic bacteria, which may present different genetic characteristics from the rest
of the genome, such as low G + C content and pathogenic function of the genic reper-
toire [15]. These genes are frequently clustered in genomic regions called “pathogenic
islands” (PAI), which may suggest characteristics related to the pathogen’s lifestyle
and its infection mechanisms. The composition of PAI can reveal new genes related to
pathogenicity and protein-protein interactions among closely related proteins [9].
40 I. R. Blanco et al.

Through application of in silico approaches, it is possible to predict the composition

of these regions, as well as features of proteins constituting them. It is also feasible to
predict pathogenic potential of molecules, by means of computational calculation based
on amino acid sequences, evincing their roles in pathogenicity.
Other VF in C. pseudotuberculosis include toxic lipids associated to the cell wall,
which can mediate bacterial resistance to phagocyte attacks, besides neuroaminidases
and endonucleases, which play pivotal role in pathogenic processes [16]. Also, the inser-
tion of prophage DNA in bacterial genomes is sometimes observed in pathogenic bac-
teria, directly related to phage encoded virulence genes, what can be identified through
in silico tools.

2 Methods
2.1 Pan-Genomic Analysis of C. Pseudotuberculosis Biovar Ovis
Complete genomes of 33 strains of biovar ovis were downloaded and standardized for
nucleotides, amino acids and function files (.nuc,.pep and.function). The strains were:
1002, 1002B, 12C, 226, 267, 29156, 3995, 4202, C231, E55, E56, FRC41, I19, MEX1,
MEX25, MEX29, MIC6, N1, P54B96, PA01, PA02, PA04, PA05, PA06, PA07, PA08,
PAT10, PO22241, AN902, T1, VD57 and ft_2193-67. The software PGAP [11] gener-
ated orthologous gene clusters among strains. The Gene Family method was used with
minimum score of blastall 40, 80% identity, 90% coverage and e-value of 0.00001.

2.2 Prediction of Virulence Factors in Corynebacterium

The Virulence Factor Database (VFDB) provides an online platform to identify VF and
their conservation in various bacterial genomes. In order to identify VF, their classes
and sequences, we have used the VFAnalyzer tool, which resorts genomic data from the
NCBI database [17]. The genomes analyzed were compared to the reference genome of
the non-pathogenic strain C. glutamicum ATCC 13032.

2.3 Composition of Pathogenicity Islands (PAI)

GIPSy software was used to predict the existence and composition of PAI. It aligns with
complementary databases, such as VFDB and NCBI. The genome of C. pseudotuber-
culosis 12C was considered as query, using as reference (non-pathogenic) the genome
of C. glutamicum ATCC 13032. GIPSy provides composition, position of genes, G + C
content and G + C deviation of PAI [18].

2.4 Synteny in C. Pseudotuberculosis Genomes

The online tool SimpleSynteny [19] was used to compare genomes of C. pseudotuber-
culosis 12C, FRC41 and 1002, generating images of genomic loci. The tool requires
FASTA sequences of the genomes and target genes, showing their conservation among
genomes through BLAST, and their positions. We performed gapped-type BLAST with
e-value threshold of 0.001 and minimum query cutoff of 50%.
 In silico Pathogenomic Analysis 41

2.5 In Silico Prediction of Pathogenicity Potentials

We applied computational prediction to the obtained genes in order to predict whether

they have pathogenic potential or not. This analysis was performed through the MP3
tool [20], able to predict the pathogenic ability of genomes. MP3 uses the highly reliable
methods Support Vector Machine (SVM) and Hidden Markov Model (HMM).
Through each mentioned method, a score is calculated for proteins, classifying them
as pathogenic (P) or non-pathogenic (NP). To have reliable prediction, we considered
as reference for our analysis the proteins predicted as pathogenic by both methods.

2.6 Protein-Protein Interactions

VF sequences were submitted to the STRING database [21], which predicts direct and
indirect interactions among proteins through computational approach, generating a reli-
ability score. We analyzed those proteins with a predicted score above 0.7 and maximum
of 10 interactions. The networks were generated based on the conserved gene neighbor-
hood, co-ocurrence of genes and protein homology. C. pseudotuberculosis strain E19
was used as reference for the prediction by protein homology.

2.7 Identification of Prophages

The presence of phage DNA in the genome of C. pseudotuberculosis 12C was per-
formed with the web tool PHAST (Phage Search Tool), which contains an extensive
set of genes identified in prophages. The input genome must be inserted as a fasta file,
generating homologous annotated hits through BLAST with different databases and a
graphic visualization of the prophage genetic content [22].

3 Results and Discussion

3.1 Identification of Adherence Factors

The pathogenomic analysis of the genus revealed 4 major classes of VF, comprising 12
types of them. The complete analysis showed 36 genes associated with VF in Corynebac-
terium. Regarding their conservation, we noticed different results for the different strains
analyzed.
The first class of VF groups the adherence proteins, which are divided into Collagen-
Binding Proteins; SpaA-type pili; SpaD-type pili; SpaH-type pili and Surface-anchored
pilus proteins. For our considered dataset, however, collagen-binding proteins were not
conserved in any of the strains [23].
Pili are filamentary cellular structures that provide adhesive and invasive functions to
the pathogen after infection [24]. In C. diphtheriae, pili are formed by the sortase machin-
ery, being also present in other pathogenic genera such as Streptococcus, Actinomyces
and Clostridium. The three Sortase-mediated Pilus Assembly structures SpaA, SpaD
and SpaH have similar composition, and are essential for the attachment to host’s tissues
42 I. R. Blanco et al.

after infection. SpaA pilus is responsible for the adherence of bacteria to human pha-
ryngeal epithelial cells, while SpaB and SpaC are responsible for binding to pharyngeal
cells [25].
Of the surface-anchored pilus proteins SapA, SapD and SapE, initially characterized
as VF in C. jeikeium K411 genome [26], only SapA was conserved in the genomes of
biovar ovis and C. ulcerans FRC11. Some adherence-related VF in Corynebacterium
are present only in C. diphtheriae, as spaA, spaB, spaE, spaF, spaG, spaH, srtD, srtE,
sapD and sapE. It is suggestible that some of these adherence genes may have been
horizontally transferred to C. pseudotuberculosis [15]. Genes srtA, spaD, srtB, srtC,
sapA are conserved in most of strains, while spaC is shared among nearly half of them
(Table 1).

Table 1. Conservation of VF in genomes of C. pseudotuberculosis biovar ovis

VF Class VF Gene Cp 12C Cp FRC 41 Cp 1002 N. of

(CP011474) (NC_014329) (CP001809) strains
Adherence SpaA-type pili spaC Cp12C_2000; cpfrc_01902; Cp1002_1899; 16
Cp12C_2001; cpfrc_01903; Cp1002_1900; 13
Cp12C_2002 cpfrc_01901 Cp1002_1901 12
srtA Cp12C_2004 cpfrc_01905 Cp1002_1901 33
SpaD-type pili spaD Cp12C_1973 cpfrc_01874 Cp1002_1872 33
srtB Cp12C_1974 cpfrc_01875 Cp1002_1874 31
srtC Cp12C_1971 cpfrc_01873 Cp1002_1870 33
Surface-anchored sapA Cp12C_1968 cpfrc_01870 Cp1002_1867 30
pilus proteins
Iron ABC transporter fagA Cp12C_0036 cpfrc_00032 Cp1002_0030 30
uptake fagB Cp12C_0035 cpfrc_00031 Cp1002_0029 31
fagC Cp12C_0034 cpfrc_00030 Cp1002_0028 33
fagD Cp12C_0037 cpfrc_00033 Cp1002_0031 33
ABC-type heme hmuT Cp12C_0482 cpfrc_00455 Cp1002_0451 33
transporter hmuU Cp12C_0483 cpfrc_00456 Cp1002_0452 33
hmuV Cp12C_0484 cpfrc_00457 Cp1002_0453 31
Iron uptake and ciuA Cp12C_1041 cpfrc_00987 Cp1002_0981 33
siderophore ciuB Cp12C_1042 cpfrc_00988 Cp1002_0982 33
biosynthesis
ciuC Cp12C_1043 cpfrc_00989 Cp1002_0983 33
ciuD Cp12C_1044 cpfrc_00990 Cp1002_0984 30
ciuE Cp12C_1045 cpfrc_00991 Cp1002_0985 33
Regulation Diphtheria toxin dtxR Cp12C_1290 cpfrc_01219 Cp1002_1213 33
repressor
Toxin Phospholipase D pld Cp12C_0033 cpfrc_00029 Cp1002_0027 33

3.2 Identification of Iron Uptake Factors

The second class of VF groups the iron uptake proteins, which can be divided into ABC
transporters; ABC-type heme transporters; iron uptake and siderophore biosynthesis
 In silico Pathogenomic Analysis 43

proteins and siderophore-dependent iron uptake systems. ATP-binding cassette (ABC)

transport systems are known to be related to iron and manganese uptake in Corynebac-
terium [27], while ABC hemin transport proteins are known to be related to regulation
of VF in Yersinia pestis, Escherichia coli, Vibrio colerae and Shigella dysenteriae [28].
An operon associated with virulence in this genus is fagABCD, present in PAI in
C. pseudotuberculosis together with PLD, determinant factor for CLA [9]. All of these
orthologs were conserved in the prediction for these strains. In addition, the ABC-
type heme transporters, which encode ABC heme proteins in Corynebacterium are also
exhibited as VF in our analysis as the genes hmuT, hmuU and hmuV.
Genes related to iron uptake are often found in pathogenic bacteria, once they require
this element for living. Iron can be located in human cells and binds to heme and heme
proteins. After infection, pathogenic bacteria acquire host iron sources, especially heme,
heme proteins, transferrin and lactoferrin [29].
The iron uptake gene cluster ciuABCDE is related to ABC-type iron transport,
biosynthesis and siderophore production. It is reported that this gene cluster exhibits
iron-dependent manner regulation [30]. In our analysis, the entire cluster was found to
be present in all analyzed strains, indicating essential function of these genes on the
biological mechanisms of bacteria. Further, an additional group of genes in C. diphthe-
riae also relates to siderophore-dependent iron uptake: the irp6 operon, which acts as
promoter and may be repressed by DtxR metal ions, the diphtheria toxin repressor [31].

3.3 Identification of Regulation Factors

For Corynebacterium, the diphtheria toxin repressor DtxR stands out as a major regula-
tion factor. It is a metal ion-activated molecule that acts as a transcriptional regulator and
is conserved in C. efficiens, C. glutamicum, C. jeikeium, C. diphtheriae and all of our
analyzed strains. It is responsible for regulation of virulence genes in various pathogenic
bacteria [32]. It is suggested that dtxR genes in distinct species of Corynebacterium have
been acquired through horizontal gene transfer and having different functional roles in
distinct species, an indicative of selection for differing functions [33].

3.4 Identification of Toxin Factors

Only one toxin factor was identified in C. pseudotuberculosis genomes: PLD, the major
VF in this species, conserved in all biovar ovis strains. It is the main responsible for
CLA and is placed in PAI. It causes host macrophage cell death and may be regulated
by several environmental factors [34]. Further, C. ulcerans and C. diphtheriae have a
second VF in their genomes, DT, responsible for the different pathogeny diphtheria.

3.5 Prediction of Pathogenicity Islands

We noted pathogenicity-related genes in these genomes. Thus, we also identified the
complete genomic content of PAI. We were able to identify 11 PAI, with genes annotated
as VF, iron uptake proteins, toxins, etc. Some genes are still classified as hypothetical
proteins, emphasizing the need for further studies on the pathogen’s genome. In addition,
the conservation of major VF in three genomes is showed in Fig. 1.
44 I. R. Blanco et al.

Fig. 1. Synteny analysis of main C. pseudotuberculosis VF in the genomes of strains 12C, FRC41
and 1002. Strains are compared regarding to conservation and position of important VF clusters
in their genomes.

In PAI 1, some genes stand out, as those inserted in the fagABCD operon, followed
by the toxin pld gene. In PAI 3, some genes show relation with manganese uptake
and ABC transport, as mentioned. In PAI 4, many genes present association with iron
ABC transport, as the ciuABCDE cluster, experimentally proven to have virulent func-
tion [9]. In PAI 5, a genomic region is conserved, highlighting the virulent function
of the PTS sugar transporter pfoS, responsible for regulation of toxins in Clostridium
perfringens. Aditionally, the proline iminopeptidase protein was found, essential for vir-
ulence of Xanthomonas campestris and conserved in C. pseudotuberculosis [9]. In PAI
7, an adhesin-related protein is found, which may relate to the role of adherence in the
pathogenic process, once it is essential for the invasion of the bacteria in host cells.
Further, in PAI 8 we noted two adherence-related molecules: the collagen-binding
surface protein and the sortase srtA1. In PAI9, the iron dicitrate transport proteins phuC
and fecD are placed, as well as the sortase srtA2, conserving virulent function [23].

3.6 Prediction of Pathogenicity Potentials

The sequences inserted in PAI were submitted to the MP3 tool to predict their proba-
ble pathogenic roles. Pathogenicity potentials for all coding sequences of the complete
genome C. pseudotuberculosis 12C, composed by 2,220 coding sequences, were calcu-
lated. 683 were classified as pathogenic by both prediction methods (HMM and SVM),
adding an average value of 30.3% of proteins predicted as pathogenic in silico.
The number of pathogenic proteins in PAI, however, was different. The composition
of these loci is formed by low G + C content sequences, what has implied in a higher
pathogenic content prediction for most of these islands. We found a mean value of
48.6% of pathogenic proteins, 18.3% more than the complete genome. The pathogenic
content varied significantly, showing standard deviation of 16.3% among PAI regarding
to pathogenic protein content (Table 2).

3.7 Identification of Prophages

In order to complement the analysis, we looked for prophage sequences in the genome
of the pathogen. Three incomplete prophage regions have been identified. One of
 In silico Pathogenomic Analysis 45

Table 2. Pathogenic composition of PAI in C. pseudotuberculosis 12C genome

Genomic N. of genes Pathogenic genes G + C deviation

region
Complete 2,220 683 (30.7%) 10%
genome
PAI 1 13 7 (53.8%) 23%
PAI 2 12 4 (33.3%) 33%
PAI 3 21 9 (42.8%) 9%
PAI 4 14 11 (78.5%) 35%
PAI 5 15 7 (46.6%) 23%
PAI 6 7 3 (42.8%) 14%
PAI 7 7 2 (28.5%) 14%
PAI 8 17 7 (41.1%) 35%
PAI 9 18 13 (72.2%) 5%
PAI 10 6 2 (33.3%) 16%
PAI 11 8 5 (62.5%) 12%

the prophage regions was found to be common in Micobacterya, Burkholderia and

Synechococcus. It is composed of six coding sequences, with annotations similar to
enzymatic, structural and ABC transport functions.
Another prophage region, composed by 10 coding sequences shows homology to
two ribonucleotide reductase proteins, a 50S ribosomal protein and three hypothetical
proteins. The third prophage region shows homology to a plasmid partitioning protein,
an inner membrane translocase and some bacillus-related hypothetical proteins.
Especially, one of these proteins showed homology to phage tail genes, which are
reported to have multiple roles, and work as adhesion molecules for bacterial attachment
in the host tissue. These tail molecules interact with the cell surface and cell wall of the
host cell and are subjected to strong adaptative selection pressure [35].

3.8 Prediction of Protein-Protein Interactions

As observed, most of VF exhibit some relation regarding to biological mechanisms. We

tried to obtain a visual representation of the interaction among the main pathogenicity
determinants in C. pseudotuberculosis. Thus, we generated a PPI network for main
VF, including proteins of the ciuABCDE, hmuTUV and fagABCD clusters. Further, a
putative ABC transporter “AIG10663.1” is placed among them.
Despite being a computational prediction, we note through this method
experimentally-proved interactions through homology-based biochemical pathways
database data, conservation of structural protein domains and existing reports in the
literature [9, 36–39], what may be confirmed through experimental support (Fig. 2).
46 I. R. Blanco et al.

Fig. 2. PPI network of major VF in C. pseudotuberculosis

An experimental report had suggested the interaction between ciuABCDE and

fagABCD cluster genes for iron uptake [40]. Our study is in concordance with this find-
ing, suggesting an additional role of hmuTUV in the regulation of this mechanism and
pathogenicity. The identification of associations among pathogenic proteins is directly
associated to their role. Our results showed a high-conserved pathogenic network, indi-
cating they may work together in the pathogenicity mechanism, what might be supported
by complementary molecules [3].

4 Conclusion

This study proposed the investigation of pathogenic determinants of C. pseudotubercu-

losis biovar ovis to elucidate mechanisms of VF, genetic elements, PAI and orthology
among the strains. Thus, we have made use of in silico analyses to reveal these data,.
Comparative pathogenomics revealed four main classes of Corynebacteria VF,
which sum up to 36 genes. However, only 20 of these genes were observed in the
analyzed strains. These VF are essential for invasion, propagation and survival of the
pathogen after infection in the host cell. Not all of the Corynebacterium VF are present
in the analyzed strains, indicating different acquisition of them among strains.
Additionally, we showed the synteny of these VF, genomic loci and the presence
and composition of PAI in the reference genome of C. pseudotuberculosis 12C. The
majority of VF described are found in these regions and organized in gene clusters,
however in different genomic positions among distinct strains. Proteins composing PAI
were analyzed in silico for potentials of pathogenicity, finding a 18.6% higher potential
in these low G + C content regions in comparison to the complete genome.
 In silico Pathogenomic Analysis 47

Once the presence of mobile genetic elements is often found to be related with
pathogenicity in bacterial genome, we looked for presence of phage DNA in the C.
pseudotuberculosis genome, identifying three prophage regions, highlighting one phage
tail gene, which is known to be related with pathogen attachment to host cell after
infection. Finally, the main VF identified in this research were analyzed looking for PPI
networks, exhibiting high-confidence associations among ciuABCDE, hmuTUV and
clusters, suggesting combined roles in pathogenic function.

References
1. Dorella, F., Pacheco, L., Oliveira, S., et al.: Corynebacterium pseudotuberculosis: microbiol-
ogy, biochemical properties, pathogenesis and molecular studies of virulence. Vet. Res. 37(2),
201–218 (2016)
2. Araújo, C., Alves, J., Lima, A., et al.: The Genus Corynebacterium in the Genomic Era. Basic
Biology and Applications of Actinobacteria, Shymaa Enany, IntechOpen (2018)
3. Araújo, C., Blanco, I., Souza, L., et al.: In silico functional prediction of hypothetical proteins
from the core genome of Corynebacterium pseudotuberculosis. PeerJ 8, e9643 (2020)
4. Biberstein, E., Knight, H., Jang, S.: Two biotypes of Corynebacterium pseudotuberculosis.
The Vet. Rec. 89, 691–692 (1971)
5. Williamson, L.: Caseous lymphadenitis in small ruminants. The Vet. Clin. North Am. Food
Anim. Pract. 17(2), 359–371 (2001)
6. Van Dijk, E., Jaszczyszyn, Y., Naquin, D., et al.: The third revolution in sequencing technology.
Trends Genet. 34(9), 666–681 (2018)
7. Guedes, M., Souza, B., Sousa, T., et al.: Infecção por Corynebacterium pseudotuberculosis
em equinos: aspectos microbiológicos, clínicos e preventivos. Pesquisa Veterinária Brasileira
35(8), 701–708 (2015)
8. Weerasekera, D.: Characterization of virulence factors of Corynebacterium diphtheriae and
Corynebacterium ulcerans. thesis (2019)
9. Ruiz, J., D’afonseca, V., Silva, A., et al.: Evidence for reductive genome evolution and lateral
acquisition of virulence functions in two Corynebacterium pseudotuberculosis strains. PLoS
ONE 6(4), e18551 (2011)
10. Pallen, M., Wren, B.: Bacterial pathogenomics. Nature 449(7164), 835–842 (2007)
11. Zhao, Y., Wu, J., Yang, J., et al.: PGAP: Pan-genomes analysis pipeline. Bioinformatics 28(3),
416–418 (2012)
12. Quiroz-Castañeda, R.: Pathogenomics and molecular advances in pathogen identification. In:
Farm Animals Diseases, Recent Omic Trends and New Strategies of Treatment. IntechOpen
(2018)
13. Cassiday, P., Pawloski, L., Tiwari, T., et al.: Analysis of toxigenic Corynebacterium ulcerans
strains revealing potential for false-negative real-time PCR results. J. Clin. Microbiol. 46(1),
331–333 (2007)
14. Lo, B.: Diptheria. Medscape. https://emedicine.medscape.com/article/782051-print.
Accessed on 24 08 2020
15. Guimarães, L., Soares, S., Trost, E., et al.: Genome informatics and vaccine targets in
Corynebacterium urealyticum using two whole genomes, comparative genomics, and reverse
vaccinology. BMC Genom. 16, 5 (2015)
16. Collin, M., Fischetti, V.: A novel secreted endoglycosidase from Enterococcus faecalis with
activity on human immunoglobulin G and ribonuclease B. J. Biol. Chem. 279(21), 22558–
22570 (2004)
48 I. R. Blanco et al.

17. Liu, B., Zheng, D., Jin, Q., et al.: VFDB 2019: a comparative pathogenomic platform with
an interactive web interface. Nucleic Acids Res. 47(1), 687–692 (2019)
18. Soares, S., Geyik, H., Ramos, R., et al.: GIPSy: genomic island prediction software. J.
Biotechnol. 232, 2–11 (2016)
19. Veltri, D., Wight, M., Crouch, J.: SimpleSynteny: a web-based tool for visualization of
microsynteny across multiple species. Nucleic Acids Res. 44, 41–45 (2016)
20. Gupta, A., Kapil, R., Dhakan, D.B., et al.: MP3: a software tool for the prediction of pathogenic
proteins in genomic and metagenomic data. PLoS ONE 9, 4 (2014)
21. Szklarczyk, D., Gable, A., Lyon, D., et al.: STRING v11: protein–protein association net-
works with increased coverage, supporting functional discovery in genome-wide experimental
datasets. Nucleic Acids Res. 47(D1), 607–613 (2019)
22. Zhou, Y., Liang, Y., Lynch, K., et al.: PHAST: a fast phage search tool. Nucleic Acids Res.
39, 347–352 (2011)
23. Ton-That, H., Marraffini, L., Schneewind, O.: Sortases and pilin elements involved in pilus
assembly of Corynebacterium diphtheriae. Mol. Microbiol. 53, 251–261 (2004)
24. Ton-That, H., Schneewind, O.: Assembly of pili on the surface of Corynebacterium
diphtheriae. Mol. Microbiol. 50(4), 1429–1438 (2003)
25. Mandlik, A., Swierczynski, A., Das, A., et al.: Corynebacterium diphtheriae employs specific
minor pilins to target human pharyngeal epithelial cells. Mol. Microbiol. 64, 111–124 (2007)
26. Hansmeier, N., Chao, T., Daschkey, S., et al.: A comprehensive proteome map of the lipid-
requiring nosocomial pathogen Corynebacterium jeikeium K411. Proteomics 7(7), 1076–
1096 (2007)
27. Tauch, A., Kaiser, O., Hain, T., et al.: Complete genome sequence and analysis of the multire-
sistant nosocomial pathogen Corynebacterium jeikeium K411, a lipid-requiring bacterium of
the human skin flora. J. Bacteriol. 187(13), 4671–4682 (2005)
28. Schmitt, M., Drazek, E.: Construction and consequences of directed mutations affecting the
hemin receptor in pathogenic Corynebacterium species. J. Bacteriol. 183, 1476–1481 (2001)
29. Stojiljkovic, I., Perkins-Balding, D.: Processing of heme and heme-containing proteins by
bacteria. DNA Cell Biol. 21(4), 281–295 (2002)
30. Kunkle, C., Schmitt, M.: Analysis of a DtxR-regulated iron transport and siderophore
biosynthesis gene cluster in Corynebacterium diphtheriae. J. Bacteriol. 187(2), 422–433
(2005)
31. Qian, Y., Lee, J., Holmes, R.: Identification of a DtxR-regulated operon that is essential for
siderophore-dependent iron uptake in Corynebacterium diphtheriae. J. Bacteriol. 184(17),
4846–4856 (2002)
32. D’Aquino, J., Tetenbaum-Novatt, J., White, A., et al.: Mechanism of metal ion activation of
the diphtheria toxin repressor DtxR. Proc. Nat. Acad. Sci. U.S. Am. 102(51), 18408–18413
(2005)
33. Oram, D., Avdalovic, A., Holmes, R.: Analysis of genes that encode DtxR-like transcriptional
regulators in pathogenic and saprophytic corynebacterial species. Infect. Immun. 72(4), 1885–
1895 (2004)
34. McKean, S., Davies, J., Moore, R.: Expression of phospholipase D, the major virulence factor
of Corynebacterium pseudotuberculosis, is regulated by multiple environmental factors and
plays a role in macrophage death. Microbiology 153, 2203–2211 (2007)
35. Brüssow, H.: Impact of phages on evolution of bacterial pathogenicity. In: Bacterial
Pathogenomics, ASM Press (2007)
36. Guerrero, J., de Oca Jiménez, R., Dibarrat, J., et al.: Isolation and molecular characterization
of Corynebacterium pseudotuberculosis from sheep and goats in Mexico. Microbial Pathog.
117, 304–309 (2018)
 In silico Pathogenomic Analysis 49

37. de Sá, M.A., Gouveia, G., Krewer, C.: Distribution of PLD and FagA, B, C and D
genes in Corynebacterium pseudotuberculosis isolates from sheep and goats with caseous
lymphadenitis. Genet. Mol. Biol. 36(2), 265–268 (2013)
38. Galvão, C., Fragoso, S., de Oliveira, C., et al.: Identification of new Corynebacterium pseu-
dotuberculosis antigens by immunoscreening of gene expression library. BMC Microbiol.
17(1), 202 (2017)
39. Silva, W., Folador, E., Soares, S., et al.: Label-free quantitative proteomics of Corynebac-
terium pseudotuberculosis isolates reveals differences between Biovars ovis and equi strains.
BMC Genom. 18(1), 451 (2017)
40. Raynal, J., Bastos, B., Vilas-Boas, P., et al.: Identification of membrane-associated proteins
with pathogenic potential expressed by Corynebacterium pseudotuberculosis grown in animal
serum. BMC Res. Notes 1, 11 (2018)
 Assessing the Sex-Related Genomic
Composition Difference Using
a k-mer-Based Approach: A Case
of Study in Arapaima gigas (Pirarucu)

Renata Lilian Dantas Cavalcante1(B) , Jose Miguel Ortega2 ,

Jorge Estefano Santana de Souza1 , and Tetsu Sakamoto1
1
Bioinformatics Multidisciplinary Environment – BioME, IMD, Federal University
of Rio Grande do Norte – UFRN, Natal, Brazil
[email protected], [email protected]
2
Lab Biodados, Department of Biochemistry and Immunology, Federal University of
Minas Gerais – UFMG, Belo Horizonte, Brazil

Abstract. Arapaima gigas is the largest freshwater bony fish in the

world, in which adults could weigh 200 kg and measure 3 m in length.
Due to its large size and its low-fat meat, Arapaima gigas has quickly
become a species of special interest in fish-farming. One challenge faced
during their production is the lack of an efficient sexing methodology,
since their sexual maturation occurs late (around the third to the fifth
year) and the genetic mechanisms linked to their sex determination sys-
tem are not known yet. For a more sustainable management, it is of
paramount importance to seek an effective and non-invasive method to
differentiate sexually juvenile individuals of Arapaima gigas. For this, the
establishment of genetic markers associated with sexual differentiation
would be an advantageous tool. In this study, we proposed a k-mer based
approach to identify genome features with sex-determining properties.
For this purpose, we used genomic data from six adult representatives of
Arapaima gigas, three males and three females, and counted the k-mers
comprising them. As result, we found k-mers from repetitive regions with
high difference and disproportion in the count among individuals of the
opposite sex. These differences in the k-mer-based genomic composition
could indicate the existence of genetic factors involved in the sexing of
individuals in Arapaima gigas.

Keywords: Osteoglossiformes · Sexual differentiation · Amazon fish ·

Repetitive sequence · Molecular marker · Fish-farming

1 Introduction
Arapaima gigas, commonly known as “Pirarucu” or “Paiche”, is the largest bony
freshwater fish in the world. It belongs to the bonytongues (Order Osteoglos-
siformes) Arapaimidae family, and has a natural habitat in the Amazon Basin
c Springer Nature Switzerland AG 2020
J. C. Setubal and W. M. Silva (Eds.): BSB 2020, LNBI 12558, pp. 50–56, 2020.
https://doi.org/10.1007/978-3-030-65775-8_5
 Assessing the Sex-Related Genomic Composition Difference 51

[1]. Adult specimens may weigh around 200 kg and measure about 3 m [2,3].
Due to its large size, its flesh containing low-fat and low fishbone, along with
its physiological characteristic of emerging to the surface at intervals of 15 min
to assimilate oxygen, Arapaima gigas became a vulnerable species to overfish-
ing in the Amazon region [4] leading to greater surveillance of the marketing of
Pirarucu by the Brazilian government in the early 2000s [5,6].
Some studies show that the use of Pirarucu in intensive fish farming is facil-
itated, in part, by the physiological characteristics of the animal that guarantee
the rusticity of the species [7]. For example, the obligate air-breathing causes this
species to tolerate environments with low concentrations of dissolved oxygen in
the water [8]. In addition to the facility for captive management, attributes such
as the low content of fat, combined with the rapid growth of the species, which
weighs an average of 10 kg in its first year of life, add value for an intensification
in the commercial exploitation of the Arapaima gigas [5,9,10].
One of the problems related to its fishing exploitation and fish-farming is
that the genetic mechanisms linked to sex-differentiation in Arapaima gigas [11]
is not know yet. Since its sexual maturation occurs late, around the third to the
fifth year of life, and sexual dimorphism is not a strong feature of the species
[10], its sexing is yet performed using laborious procedures (e.g. laparoscopy and
transrectal ultrasound). In recent decades, the creation of Arapaima gigas in
captivity has been increasingly stimulated, either to develop research to better
know the particularities of the species or to exploit its economic potential [12,13].
For more sustainable management, both for captive breeding and for the study
of the species, it is of paramount importance to seek an effective and non-invasive
method to differentiate sexually juvenile individuals.
For this, the establishment of a molecular genetic marker related to sexual
differentiation would be an advantageous tool. Previous analysis of the Arapaima
gigas genome found no genes associated with the identification of the sex deter-
mination system of these individuals [14,15]. And chromosomal characterization
studies could not distinguishes cytologically a sex chromosome in Arapaima gigas
[16,17]. In this study, we proposed to asses the genomic composition of Arapaima
gigas using a k-mer-based approach to identify regions in excess or missing in
one of the sexes.

2 Materials and Methods

2.1 Sequencing and Data Processing
In this study, we used genomic data from six adult representatives of Ara-
paima gigas, three males (M15, M20 and M25) and three females (F15, F20
and F25). Four samples (M15, M20, F15 and F20) were collected from Biopro-
ject PRJEB22808 available in National Center for Biotechnology Information
(NCBI) database [14]. And the other samples (M25 and F25) were acquired
from another Bioproject nominated as PRJNA540910, also available in NCBI
database [15].The quality of the reads was verified with the help of FastQC (v.
0.11.4) [18], and low-quality reads were trimmed with the help of the Sickle
52 R. L. D. Cavalcante et al.

paired-end (v. 1.33) [19]. Data sets were processed in the sagarana HPC cluster,
CEPAD-ICB-UFMG.

2.2 k-mer Analysis

This analysis was performed in four steps: (1) k-mer count, (2) k-mer count
normalization, (3) k-mer filtering for repetitive regions and (4) k-mer count
comparison. The k-mer count was performed with the help of the tool K-mer
Counter (KMC, v.3.1.1) [20], a free software written in C++, whose premise
is to count k-mers (sequences of consecutive k nucleotide) in a given genome
sequencing file. The trimmed fastq files of the 6 representatives Arapaima gigas
were submitted as input data to KMC algorithm using the parameter -k, on the
k-mer size, as 23.
After KMC step, in order to normalize the data Quantile Normalization
(QN), a global arrangement method, which consists of a non-parametric methods
that makes two or more distributions identical on statistical properties [21], was
performed using an in-house Perl script (v.5.16.3). Lastly, to compare the average
of the normalized counts between male and female samples, we performed a T-
Test in R to identify which k-mers presented small p-value (p ≤ 0.05). For further
procedures, we used a in-house Perl script to select k-mers which were extracted
from a repetitive region with a repeat unit up to 8 nucleotides. K-mer with the
same repeat unit were merged and had their normalized count summed up.

3 Results and Discussion

After trimming low-quality sequences of genome sequencing data of 6 samples
of Arapaima gigas, we had per each file 9.1 million, 39 million, 38.1 million, 38.5
million, 14 million, 38 million reads remained for F15, F20, F25, M15, M20 and
M25, respectively. The KMC result of k-mer count showed a number of total
k-mers counted of 6 billion, 26 billion, 32 billion, 25 billion, 9 billion and 32
billion k-mers for F15, F20, F25, M15, M20 and M25 samples, respectively (See
Table 1).

Table 1. Stats of average of reads and total number of k-mer in six samples of Arapaima
gigas.

Samples Reads average Total no. of k-mers

F15 9.170.700 6.036.151.500
F20 39.041.211 26.057.501.877
F25 38.168.487 32.523.525.409
M15 38.590.767 25.269.745.617
M20 14.141.274 9.380.791.863
M25 38.030.714 32.336.285.757
 Assessing the Sex-Related Genomic Composition Difference 53

Because of the difference on the number of reads among samples, we used

the Quantile Normalization method (QN) to make them comparable. For further
analysis, we considered only k-mers comprised of repeating units of size 1 to 8 bp.
The average comparison of normalized k-mers count between individuals of the
opposite sex and the same sex showed some repeats which were more abundant
in one of the sex (Table 2).

Table 2. Repeat units with significant difference (p-value < 0.05) on the average of
the normalized k-mer count between female and male samples of Arapaima gigas.

Repeat unit F15 F20 F25 M15 M20 M25 p-value

AAAAC 10626.84 9914.83 10002.67 7261.5 7390.34 8410.68 0.007
AACAGCTG 162 139.16 161.17 84.85 108 110.67 0.009
AGAGCGG 71.49 78.33 66.51 87.51 92.01 96.99 0.011
AAGGC 31540.49 28319.83 17831.17 7457.32 9015.83 58.5 0.019
ACGTC 3440.67 4072.83 3104.49 1584.17 2348.67 2398.16 0.021
ACCACCAG 50.16 48.66 40.16 28.34 32.34 34.5 0.023
AAGC 11656.51 9888.33 9237.18 6352.5 6537.66 5921.67 0.025
ACACATCC 0 0 0 0.85 0.83 1.34 0.026
AAGGCC 2552.17 2221.5 1767.66 638.18 975.49 1480.83 0.026
ACCAGT 241.17 295.5 249.84 306.66 360.32 354.17 0.030
AAAAG 5775.17 5536.67 4985.32 3803.83 3697.84 4719.84 0.031
AGGG 11899.16 12294.5 10126.67 8220 8186.33 9562.84 0.031
ACATATC 88.34 75 86.83 71.17 54 53.67 0.033
AAACATT 95.32 100.32 99.18 113.68 110 104.34 0.035
AACC 10210.34 9663.99 10725.17 6843.5 6845.99 8788.84 0.035
AAAGTCAC 32.67 28.5 36.83 24.33 21.5 21.83 0.039
ACTGGC 0 0 81.33 135.01 166 96.99 0.039
ACCCAGGT 32.67 30.33 36 26.17 29 24.5 0.040
AACAACCC 78.99 72 61.5 90.68 92.01 95.67 0.040
AAAGTAAT 18.68 21.32 9.85 0 0 0 0.040
AACAC 12706 12649.66 11450.5 7673.34 9504 10389.66 0.042
AAGACATT 23.83 23 24.67 26.17 29 26.5 0.042
AATGATG 61 60 60.66 48.66 54 54.34 0.043
Repeat units considered as promissing candidates for sex differentiation are highlighted.

Arapaima gigas plays an important role in the economy of the north region
of Brazil [12,13]. Understand the mechanisms of sexual determination in fish are
essential for a sustainable management of ichthyofauna, either for commercial
or conservation purposes [22]. But, elucidating these mechanisms in fish is chal-
lenging, especially in Arapaima gigas case, because this species do not have a
typical sex chromosome in their genome [16,17] and the difference of the genome
sequences between samples of opposite sex seems to be minimal. Both Arapaima
54 R. L. D. Cavalcante et al.

gigas genome assemblies [14,15] did not find significant differences between the
genomic content of male and female samples.
In this context, we explored other genomic features in Arapaima gigas to find
some clues about the genetic factors involved in the sex determining system. For
this, we analysed the genomic composition of Pirarucu using k-mer based app-
roach. In this study, we have noticed the existence of k-mers from repeat regions
over or underrepresented in one of the sexes, indicating potential differences in
the genetic composition between males and females of Arapaima gigas.
The difference is not so expressive, which corroborates the reports that esti-
mate 0.01% [14] to 0.1% [15] of the genome of this species as linked to the
sexual determination. The sequences reported in this work are all part of repet-
itive sequences. Despite of their low complexity, repetitive regions have been
reported to have important role on sex determination [23]. In medaka, which
has the XY system, there is a large stretch of repetitive regions on the Y-specific
regions [24]. The chromosome Y of Pacific salmon also bears a specific repetitive
regions (OtY1) that is used as genetic marker to differentiate sex [25].
In this context, the repetitive sequences found in this study could be a com-
ponent that could be used to determine sex in individuals of Arapaima gigas. We
recognize, however, the necessity to perform analyses with a greater number of
samples to obtain a better statistical support for our results, as well as to sug-
gest bench trials for the validation of the in-silico analyses. Despite of that, the
k-mer-based method applied on this work has demonstrated to be an interesting
strategy to help us discover the sex-determination system in Pirarucu specimens
and can be extended to other species.

4 Conclusions

This short paper reports a few repetitive genome sequences that can be dif-
ferentiated in quantity in male and female of Arapaima gigas. Indicating that
k-mer-based methods is an interesting approach to assist us in unraveling the
sex-determination system in Arapaima gigas.

Acknowledgments. This work was supported by CAPES (Brazilian Federal Agency

for the Support and Evaluation of Graduate Education). To Bioinformatics Multi-
disciplinary Environment - BioME and CEPAD-ICB-UFMG for the computational
structure for the accomplishment of this work. To Dr. Manfred Schartl from University
of Würzburg - Germany that met our request to publish the reads of the Arapaima
gigas genome from Bioproject PRJNA540910. And to Dr. Sidney Emanuel Batista dos
Santos from Federal University of Pará - UFPA for contributing in the discussions of
this project.

References
1. Reis, R.E., Kullander, S.O., Ferraris, C.J.: Check List of the Freshwater Fishes of
South and Central America. Edipucrs, Porto Alegre (2003)
 Assessing the Sex-Related Genomic Composition Difference 55

2. Bezerra, R.F., Soares, M.C.F., Maciel Carvalho, E.V.M., Coelho, L.C.B.B.:

Pirarucu, Arapaima Gigas, the Amazonian Giant Fish is Briefly Reviewed. Fish,
Fishing and Fisheries Series, p. 41. Nova Science, New York (2013)
3. Nelson, J.S., Grande, T.C., Wilson, M.V.H.: Fishes of the World. John Wiley &
Sons, Hoboken (2016)
4. Goulding, M., Smith, N.J.H., Mahar, D.J.: Floods of Fortune: Ecology and Econ-
omy Along the Amazon. Columbia University Press, New York (1996)
5. Bayley, P.B., Petrere, M., Jr.: Amazon Fisheries: Assessment Methods, Current
Status and Management Options. Canadian Special Publication of Fisheries and
Aquatic Sciences/Publication Speciale Canadienne des Sciences Halieutiques et
Aquatiques (1989)
6. Wilson, D.E., Burnie, D.: Animal: The Definitive Visual Guide to the World’s
Wildlife. Dorling Kindersley, London (2001)
7. de Oliveira, E.G., et al.: Effects of stocking density on the performance of juvenile
pirarucu (arapaima gigas) in cages. Aquaculture 370, 96–101 (2012)
8. Bard, J., Imbiriba, E.P.: Piscicultura do pirarucu, arapaima gigas. Embrapa
Amazônia Oriental-Circular Técnica (INFOTECA-E) (1986)
9. de M. Carvalho, L.O.D., do Nascimento, C.N.B.: Engorda de pirarucus (arapaima
gigas) em associação com búfalos e suı́nos. Embrapa Amazônia Oriental-Circular
Técnica (INFOTECA-E) (1992)
10. Almeida, I.G., Ianella, P., Faria, M.T., Paiva, S.R., Caetano, A.R.: Bulked seg-
regant analysis of the pirarucu (arapaima gigas) genome for identification of
sex-specific molecular markers. Embrapa Amazônia Oriental-Artigo em periódico
indexado (ALICE) (2013)
11. Hrbek, T., Farias, I.P.: The complete mitochondrial genome of the pirarucu (ara-
paima gigas, arapaimidae, osteoglossiformes). Genet. Mol. Biol. 31(1), 293–302
(2008)
12. Fontenele, O.: Contribuicao para o conhecimento da biologia do pirarucu arapaima
gigas (cuvier), em cativeiro (actinopterygii, osteoglossidae) [Ceara, Brasil] (1955)
13. Imbiriba, E.P.: Potencial de criação de pirarucu, arapaima gigas, em cativeiro. Acta
Amazonica 31(2), 299 (2001)
14. Vialle, R.A., et al.: Whole genome sequencing of the pirarucu (arapaima gigas)
supports independent emergence of major teleost clades. Genome Biol. Evol. 10(9),
2366–2379 (2018)
15. Kang, D., et al.: The genome of the arapaima (arapaima gigas) provides insights
into gigantism, fast growth and chromosomal sex determination system. Sci. Rep.
9(1), 5293 (2019)
16. Marques, D.K., Venere, P.C., Galetti Jr., P.M.: Chromosomal characterization
of the bonytongue arapaima gigas (osteoglossiformes: Arapaimidae). Neotropical
Ichthyol. 4(2), 215–218 (2006)
17. de Oliveira, E.A., et al.: Cytogenetics, genomics and biodiversity of the South
American and African arapaimidae fish family (teleostei, osteoglossiformes). PLoS
ONE 14(3), e0214225 (2019)
18. Wingett, S.W., Andrews, S.: Fastq screen: a tool for multi-genome mapping and
quality control. F1000Research, p. 7 (2018)
19. Joshi, N.A., Fass, J.N.: Sickle: a sliding-window, adaptive, quality-based trimming
tool for fastq files (version 1.33) [software] (2011)
20. Kokot, M., Dlugosz, M., Deorowicz, S.: KMC 3: counting and manipulating k-mer
statistics. Bioinformatics 33(17), 2759–2761 (2017)
21. Hicks, S.C., Irizarry, R.A.: Quantro: a data-driven approach to guide the choice of
an appropriate normalization method. Genome Biol. 16(1), 117 (2015)
56 R. L. D. Cavalcante et al.

22. Martı́nez, P., Viñas, A.M., Sánchez, L., Dı́az, N., Ribas, L., Piferrer, F.: Genetic
architecture of sex determination in fish: applications to sex ratio control in aqua-
culture. Front. Genet. 5, 340 (2014)
23. Ezaz, T., Deakin, J.E.: Repetitive sequence and sex chromosome evolution in ver-
tebrates. Adv. Evol. Biol. 2014, 9 (2014)
24. Kondo, M., et al.: Genomic organization of the sex-determining and adjacent
regions of the sex chromosomes of medaka. Genome Res. 16(7), 815–826 (2006)
25. Devlin, R.H., Biagi, C.A., Smailus, D.E.: Genetic mapping of Y-chromosomal DNA
markers in pacific salmon. Genetica 111(1–3), 43–58 (2001)
 COVID-19 X-ray Image Diagnostic
with Deep Neural Networks

Gabriel Oliveira, Rafael Padilha(B) , André Dorte, Luis Cereda, Luiz Miyazaki,
Maurı́cio Lopes, and Zanoni Dias

Institute of Computing, University of Campinas, Campinas, SP, Brazil

[email protected]

Abstract. The COVID-19 pandemic impacted all spheres of our society.

The outbreak increased the pressure on public health systems, urging the
scientific community to develop and evaluate methods to reliably diag-
nose patients. Driven by their effectiveness in medical imaging analysis,
deep neural networks have been seen as a possible alternative to automat-
ically diagnose COVID-19 patients from chest X-rays. Despite promising
initial results, most analyses so far have been performed in small and
under-represented datasets. Considering this, in this work, we evaluate
state-of-the-art convolutional neural network architectures proposed in
recent years by the deep learning field on images from COVIDx [24], a
dataset consisting of 13, 975 chest X-ray from COVID-19, pneumonia,
and healthy patients. In our experiments, we investigate the effect of
data pre-processing steps and class unbalancing for this task. Our best
model, an ensemble of several networks, achieved an accuracy above 93%
in the testing set, showing promising results in a challenging dataset.

Keywords: COVID-19 diagnostic · Chest X-ray image analysis · Deep

learning

1 Introduction
With its impact on society, public health, and economy, the outbreak of COVID-
19 (SARS-CoV-2) has shaped how the year 2020 will be remembered. The scien-
tific community has devoted its efforts to trace the origin of the disease, study-
ing its effects on the human body, as well as evaluating treatment methods
[7,10,28]. Amidst the crisis, society directed its resources onto identifying how
artificial intelligence could aid the fight against the pandemic [13], such as pre-
dicting mortality and growth rates [23,25], analyzing the virus genome [18], and
discovering possible drugs [2].
As the increase in COVID-19 cases overwhelms healthcare systems world-
wide, finding accurate and efficient diagnosis methods is critical to prevent fur-
ther disease spread and treat affected patients. In this sense, encouraged by
recent successes of machine learning applied to medical imaging analysis [14]—
e.g., skin lesion classification [12], brain tumor segmentation [27], cardiac image
c Springer Nature Switzerland AG 2020
J. C. Setubal and W. M. Silva (Eds.): BSB 2020, LNBI 12558, pp. 57–68, 2020.
https://doi.org/10.1007/978-3-030-65775-8_6
58 G. Oliveira et al.

analysis [3]—, recent works assessed its performance in COVID-19 diagnostics.

They employed convolutional neural networks (CNN) to classify chest X-ray
images or CT scans as to belonging to COVID-19 or healthy patients. CNNs are
a type of neural network specialized in processing images, which focus on learn-
ing, directly from the input, the most discriminative features for the target task.
They achieved promising results in pneumonia and lung nodules classification
[17,20].
Training a CNN from scratch is challenging, as the optimization process
requires a considerable amount of labeled data, which are not always available in
medical problems. In this sense, most approaches rely on Transfer Learning [16]—
i.e., leveraging from the knowledge learned in a different task to better generalize
in the target problem—to fine-tune the network for COVID-19 diagnosis.
Narin et al. [15] evaluated three CNN architectures, pre-trained for object
recognition and fine-tuned to a dataset of 100 chest X-ray images (half of them
from COVID-19 patients). Apostolopoulos and Bessiana [1] expanded on the pre-
vious evaluation, including other networks and employing a dataset of almost
1500 images—224 from patients diagnosed with COVID-19, 700 with pneumo-
nia, and 500 healthy ones. Even though both works achieved good results, the
datasets used are very limited in size and patient representativity.
To address such issues, Wang et al. [24] created the COVIDx dataset, contain-
ing 14, 198 chest X-rays of 14, 002 patients—from which 573 were images from
394 COVID-19 patients. Additionally, they proposed COVID-Net, a lightweight
CNN architecture, that outperformed traditional networks trained in the same
task, such as VGG-19 and ResNet-50. Finally, authors audited the decisions
made by their network to validate its reliability.
In this work, we perform an extensive evaluation of CNN architectures pub-
lished in recent years by the deep learning research community. We use the
dataset collected by Wang et al. [24], comparing not only the performance of
each network but also their number of parameters, which directly relates to their
efficiency. Similar to previous works [1,15], we assess the effectiveness of Transfer
Learning, adapting models pre-trained in object recognition to the COVID-19
diagnosis problem. In our analysis, we consider two scenarios: one with three pos-
sible diagnostics—COVID-19, pneumonia, or healthy—and a binary scenario in
which we are interested in predicting whether the patient has COVID-19 or not.
We contrast these experiments with traditional machine learning classifiers, such
as Support Vector Machines (SVM) and Random Forest.
The remaining of the text is organized as follows. In Sect. 2, we describe
COVIDx dataset, presenting each class and example images. We present details
of our methodology in Sect. 3, describing the overall pipeline of our evaluation
and the data preparation used. In Sect. 4, we present the experimental evaluation
of different classification methods in this problem. Finally, in Sect. 5, we present
our final thoughts and draw possible research lines for future work.
 COVID-19 X-ray Image Diagnostic with Deep Neural Networks 59

2 Dataset
The dataset was obtained from COVIDx [24], which aggregates images from
several different datasets. Available data varies in resolution from 400 × 500 to
3520 × 4280 pixels and are grouped in three different classes: COVID-19, Pneu-
monia, and Normal (cases of no disease). Table 1 shows the patient and chest
radiography images distribution in the training and test sets. In the Pneumonia
and COVID-19 classes, there are different X-rays of the same patient. Figure 1
illustrates one example of each class.

Table 1. Distribution of patients and chest radiography images, considering Normal,

Pneumonia, and COVID-19 diagnostics for the training and test sets.

Patients Images
Train Test Train Test
Normal 7,966 100 7,966 100
Pneumonia 5,444 98 5,459 100
COVID-19 320 74 473 100
Total 13,730 272 13,898 300

Fig. 1. Examples of X-rays from the COVIDx dataset [24]. (a) Normal, (b) Pneumonia
and (c) COVID-19.

Some X-ray images in the dataset present artifacts and noise patterns (Fig. 2),
such as medical devices connected to the patient, contour and volume of the
breasts, or differences in size and shape of the lungs and rib cage due to the
patient being an adult or a child. Even though these patterns might be common
to the task, they do not directly relate to the diagnostic outcome and, hence,
with enough data a model should learn how to ignore them.
60 G. Oliveira et al.

Fig. 2. Examples of images with different patterns, such as medical devices connected
to the patient, contour and volume of the breasts and noise.

3 Methodology
Ultimately, our goal is to accurately classify if a chest X-ray image belongs
to a patient with COVID-19 or not. In our evaluation, we follow the pipeline
depicted in Fig. 3, in which an input image is preprocessed and then analyzed
by a classification model.

Fig. 3. Overview of our pipeline. An input image is preprocessed and then classified
considering two scenarios. In the Multi-class Classification, the model assesses if the
X-ray image belongs to a COVID-19, a Pneumonia, or a Normal patient. On the Binary
Classification, the model decides between a COVID-19 or a non-COVID outcome.

We consider two scenarios concerning the output of our models. In the first
one, we approach the problem as a multi-class classification with three possi-
ble outcomes—i.e., COVID-19, Pneumonia, or Normal. The rationale is that by
 COVID-19 X-ray Image Diagnostic with Deep Neural Networks 61

explicitly differentiating non-COVID cases, the model might better capture sub-
tle differences in each diagnostic. Whereas, in the second scenario, we consider a
binary classification between COVID-19 and non-COVID images, in other words,
Normal and Pneumonia are grouped into a single class.
In the next subsections, we detail how images are preprocessed before being
classified, as well as present the classification models used in our evaluation.

3.1 Data Preparation

Each model evaluated in our analysis expects an input image with particular
dimensions and a range of values. In this regard, an input image must be stan-
dardized before training and classification.
During training, we resize each image to the network expected input size
and normalize its pixel values accordingly. We also apply data augmentation
techniques to increase our training set. Considering we have a small dataset
that might not realistically represent the variety of images encountered in real
scenarios, as seen in Fig. 2, data augmentation is essential to artificially add
variability in training and improve model generalization.
We employed the following augmentation techniques: random rotation in
the range [−5◦ , 5◦ ], zoom of at most 10% of the image’s dimension, vertical
and horizontal shifts up to 10% of the image’s height and width, respectively.
Considering the set of all possible transformations, each image can generate up to
10,000 slightly altered versions, virtually increasing our training set. We present
in Fig. 4 examples of these transformations.
At testing time, before being fed to the classification model, we resize and
normalize each input image, without applying any augmentation technique.

Fig. 4. Data augmentation strategies applied during training. (a) Original image, (b)
rotation, (c) zoom, (d) vertical and (e) horizontal shifts.

3.2 Classification Model

In recent years, the research community has proposed several CNN architec-
tures, optimization methods, and regularization techniques. Besides focusing on
improving generalization and efficacy in particular tasks, some works also aimed
62 G. Oliveira et al.

at increasing the efficiency of such networks, allowing them to train faster and
run on low-powered devices.
In this work, we evaluate several architectures proposed in the past five years.
As training them from scratch would require large datasets, we use pre-trained
CNNs optimized on ImageNet [6], an object classification dataset with 14 million
images. The rationale behind this is that deep networks tend to learn similar
concepts in their initial and intermediate layers common to most visual tasks
[26]—from object recognition to medical imaging analysis.
With the knowledge previously obtained, we adapt the networks for the
COVID-19 diagnosis task. We remove their last fully-connected layer, respon-
sible for classifying in one of the ImageNet classes, and exchange it for a new
fully-connected layer with three or two output neurons—for the Multi-class or
Binary classification scenarios, respectively—activated by softmax operation.
As each architecture imposes different constraints on the learning process,
the final characteristics learned by each of them capture distinct and often com-
plementary aspects of the training data. Because of this, we investigate if an
ensemble of networks improves the performance over individual models.
We explore three fusion approaches. In the first one, we average the answers
of all CNNs, aiming for the mean consensus between them. Secondly, instead of
merely averaging their responses, we train a meta-classifier on top of the concate-
nated answers. This meta-model learns subtle relative patterns between fused
classifiers. Finally, we take a step back and extract features from the penultimate
layer of each network, optimizing a meta-classifier on their concatenation. This
meta-model aggregates the knowledge from intermediate features and learns how
to jointly classify them.
In comparison, we also train other machine learning classifiers as baselines for
this task. We evaluate SVM, Random Forest, XGBoost [4] and Logistic Regres-
sion. In addition to the pre-processing described in Subsect. 3.1, we serialize each
image before feeding them to each classifier.

4 Experimental Evaluation
In this section, we present the results of the different CNN architectures and
baseline classifiers evaluated in the COVIDx dataset.
To train the predictive models, we organized the original training set into
train and validation splits. Initially, we randomly sampled 473 X-ray images of
each class (the size of the smaller class), divided into 383 for training and 90 for
validation. This was done to mitigate the class unbalance of the dataset, making
a balanced sub-sample that allows us to train each architecture efficiently. Using
this reduced balanced set, we evaluated all models in terms of accuracy and
number of parameters for the Multi-class classification scenario.
Considering the top three models, we evaluated their performance apply-
ing data augmentation and using the whole unbalanced dataset. To do so, we
employed a stratified division of the original training set, using 80% for training
and 20% for validation. In this setup, the model has more images to learn from
 COVID-19 X-ray Image Diagnostic with Deep Neural Networks 63

and generalize, even though the training step requires more time. Each model
was evaluated in the Multi-class and Binary classification scenarios, as well as
their ensemble, combined under different strategies.

4.1 Model Evaluation

As baselines, we trained SVMs with linear, polynomial, and RBF kernels, Ran-
dom Forest, Logistic Regression, and XGBoost classifiers. The choice of hyperpa-
rameters for each technique was obtained through a grid search, using a 5-fold
cross-validation with the balanced training set. Before training, we serialized
each input image, transforming them into one-dimensional vectors.
For the evaluation of the CNN architectures, we applied Transfer Learn-
ing. For all architectures, we fixed the weights pre-trained on ImageNet dataset
and updated only the newly-added classification layer specific to our task. We
optimized the networks with cross-entropy loss, using an SGD optimizer with a
learning rate of 0.001, momentum of 0.9, and weight decay with a rate of 0.001.
Additionally, we employed early stopping, interrupting the training when the
validation loss stopped increasing for several epochs. We present in Table 2 the
results obtained by the baselines and the CNN architectures in the validation
set of the balanced sub-sample of the dataset.

Table 2. Accuracy in the balanced validation set and number of parameters for each
evaluated models.

Model Accuracy in validation (%) Number of parameters

ResNet50 [8] 87.41 25,636,712
EfficientNetB7 [22] 84.81 66,658,687
MobileNetV2 [9] 81.85 3,538,984
DenseNet121 [11] 80.74 8,062,504
MobileNet [9] 80.37 3,538,984
Random forest 79.25 –
XGBoost [4] 78.52 –
SVM-RBF 78.15 –
Xception [5] 77.78 22,910,480
SVM-Poly 75.93 –
InceptionV3 [21] 74.07 23,851,784
NASNetLarge [29] 72.22 88,949,818
ResNet50V2 [8] 70.37 25,613,800
Logistic regression 68.89 –
SVM-Linear 68.15 –

The best baseline methods were Random Forest, XGBoost, and SVM with
RBF kernel, with accuracies ranging from 79.25% to 78.15%. Even though their
64 G. Oliveira et al.

performances were superior to a few architectures, they are not suitable for
dealing with the spatial information of images in the same way as CNNs. By
reshaping the image as a vector, the models lose most of the spatial structure
between neighboring pixels, which often hinders performance. On the other hand,
logistic regression and linear SVM suffer from trying to correctly fit a linear
decision plane in the high dimensional space produced by the flattened image.
CNNs that were outperformed by the most baselines have an increased num-
ber of parameters, which indicates that network size does not directly relate to
the accuracy in the task. This is probably due to the lack of data required to
carefully train a network with such complexity for this problem. The exception
was EfficientNetB7, which accounts in second both in the number of parameters
and accuracy. Despite its size, it outperforms in the ImageNet recognition task
several architectures [22]—both with lower and higher number of parameters.
This highlights its capability of learning discriminative and rich features, which
is essential when adapting it to a task with limited data.
Our experiments show that ResNet50 architecture achieved the highest accu-
racy in the validation set, with an intermediate number of parameters among
the networks evaluated. We select ResNet50, EfficientNetB7 and MobileNetV2
as the base models to perform additional explorations.

4.2 Multi-class and Binary Classification of the Unbalanced Dataset

Once selected, we evaluated the top three models in the complete unbalanced
dataset. To do so, we split it into 80% for training and 20% for validation in a
stratified division. Due to the class unbalancing, the model would naturally give
greater importance to the over-represented classes, i.e., Normal and Pneumonia.
To overcome this issue, we assigned weights to each category according to their
sizes. The weight of a particular class is equal to the number of images of the
largest class divided by the number of samples in it.
Besides that, we also employed the data augmentation techniques from Sub-
sect. 3.1 to increase the diversity in our training. Considering the increased
amount of data, we unfroze the initial and intermediate layers of each model,
allowing the optimization process to freely update their weights. Each CNN
was trained with the same optimizer and hyperparameters in the Multi-class
and Binary classification scenarios. We report the balanced accuracy for each
method in the top part of Table 3.
The CNNs obtained high accuracy, with ResNet50 and MobileNetV2 outper-
forming EfficientNetB7 in both classification scenarios. Similar to the previous
experiment, all three CNNs exceeded the baseline Random Forest, which was
considerably affected by the class unbalancing, especially in the Binary setup.

4.3 Ensemble of CNNs

In addition to evaluating each separate model, we investigated their comple-

mentarity and explored how to combine their knowledge in this task. In this
 COVID-19 X-ray Image Diagnostic with Deep Neural Networks 65

analysis, we fused the answers—i.e., the output classification probabilities—and

the features from the penultimate layer of each CNN.
Besides combining the answers through average operation, we also trained
meta-classifiers on top of the concatenation of probabilities or deep features.
We evaluated SVMs (with linear, polynomial, and RBF kernels), Random For-
est, and shallow Multi-Layer Perceptrons (MLP) as meta-models. The choice of
hyperparameters—such as the number of hidden layers and units of the MLP,
or trees in the Random Forest—was done based on the balanced accuracy of the
validation set. We present the results in the Table 3.

Table 3. Balanced accuracy in the stratified validation and test sets of Multi-class
and Binary classification for individual CNNs and ensembles. We highlight the best
method and ensemble strategies in each scenario and evaluation set.

Method Multi-class Binary

Accuracy in Accuracy in Accuracy in Accuracy in
validation (%) test (%) validation (%) test (%)
Classifiers
ResNet50 96.81 90.66 96.13 91.99
EfficientNetB7 83.54 79.33 89.21 81.75
MobileNetV2 90.01 82.66 97.87 90.25
Random forest 65.30 62.33 50.55 55.00
Ensemble – Probabilities
Average 97.22 89.66 98.66 93.25
SVM-Linear 98.37 90.00 98.87 89.75
SVM-Poly 98.70 90.00 99.81 87.25
SVM-RBF 98.47 90.33 98.81 89.50
MLP 99.88 92.00 99.36 93.50
Random forest 98.68 90.33 99.41 84.75
Ensemble – Deep features
SVM-Linear 97.67 89.66 99.43 89.75
SVM-Poly 98.26 90.00 99.41 87.75
SVM-RBF 98.34 90.00 99.38 89.50
MLP 98.01 83.33 99.43 86.00
Random forest 95.35 88.00 96.82 84.74

Most ensemble strategies outperformed individual CNNs in the validation set,

highlighting the importance of combining the characteristics learned by different
models. In the Multi-class scenario, we improved from 96.81% accuracy with
ResNet50 to 99.88% with an MLP over the probabilities. Whereas in the Binary
classification, a similar gain was obtained, increasing from an accuracy of 97.87%
with MobileNetV2 to 99.81% with a polynomial SVM over the probabilities.
Our experiments show that fusing probabilities or deep features achieved
similar performance, with slight differences depending on the scenario or meta-
classifier. Despite that, working with probabilities proved to be more efficient and
more robust to overfitting due to the reduced input dimensionality compared to
the concatenated deep features.
66 G. Oliveira et al.

4.4 Test Set Evaluation

We evaluated all methods in the test set and present the results in Table 3. The
overall performance decreased considerably when comparing validation and test
sets for most methods. This was probably due to the severe class unbalance,
whose impact could not be entirely prevented by data augmentation and the use
of class weights during training.
Considering the individual models, ResNet50 obtained the highest accuracy
in both classification scenarios. Whereas, among the ensembles, the MLP opti-
mized over the probabilities outperformed other configurations. We also compare
their confusion matrix on Fig. 5. In all cases, the methods are slightly biased
towards classifying COVID-19 samples as Normal or Pneumonia, presenting the
majority of the false predictions in the bottom row of each matrix. Nonetheless,
the ensemble approach is still able to correctly classify 89% of COVID X-rays.

Multi-class Classification. Binary Classification.

Fig. 5. Confusion matrix for ResNet50 and the best ensemble strategy in both classi-
fication scenarios for the test set.

5 Conclusions and Future Work

With the outbreak of COVID-19, healthcare systems worldwide are looking for
more accurate and efficient diagnostic methods for the treatment of patients.
With the recent advancement of machine learning techniques applied to medical
imaging analysis, the automatic classification of X-ray images has become an
essential tool for aiding medical diagnoses and a possible approach to diagnosing
COVID-19 patients.
In this work, we evaluated several CNN architectures published in the deep
learning literature to classify chest X-ray images of COVID-19. The experiments
showed that transfer learning techniques can achieve good results by leveraging
the generalization of the initial layers in a different domain. To aid model gen-
eralization, we employed data augmentation strategies to artificially increase
available data. Besides that, we evaluated ensemble techniques applied to the
best obtained models, improving the results even further. We achieved an accu-
racy of 92.00% in the test set considering the classification between COVID-19,
 COVID-19 X-ray Image Diagnostic with Deep Neural Networks 67

Pneumonia, and Normal classes, while obtaining a 93.50% accuracy when dis-
tinguishing between COVID-19 and non-COVID.
As future work, additional preprocessing steps can be investigated, such as
segmenting the lung parts of the image, applying filters to reduce noise and
highlight particular artifacts, as well as using images with higher resolution.
In the post-processing step, explainability techniques, such as Grad-CAM [19],
could be used to further audit the network decisions and interpret them. Finally,
being a recent problem, datasets for this task are still limited in size and present
a considerable class unbalancing. Because of this, a continuous data collection is
essential for future research in this problem.

References
1. Apostolopoulos, I.D., Mpesiana, T.A.: Covid-19: automatic detection from X-ray
images utilizing transfer learning with convolutional neural networks. Phys. Eng.
Sci. Med. 43(2), 635–640 (2020). https://doi.org/10.1007/s13246-020-00865-4
2. Beck, B.R., Shin, B., Choi, Y., Park, S., Kang, K.: Predicting commercially avail-
able antiviral drugs that may act on the novel coronavirus (SARS-CoV-2) through
a drug-target interaction deep learning model. Comput. Struct. Biotechnol. J. 18,
784–790 (2020)
3. Bizopoulos, P., Koutsouris, D.: Deep learning in cardiology. IEEE Rev. Biomed.
Eng. 12, 168–193 (2018)
4. Chen, T., Guestrin, C.: XGBoost: a scalable tree boosting system. In: ACM Inter-
national Conference on Knowledge Discovery and Data Mining (ACM KDD), pp.
785–794 (2016)
5. Chollet, F.: Xception: deep learning with depthwise separable convolutions. In:
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1251–
1258 (2017)
6. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: a large-
scale hierarchical image database. In: IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), pp. 248–255 (2009)
7. Gao, Q., Bao, L., Mao, H., Wang, L., Xu, K., Yang, M., Li, Y., Zhu, L., Wang, N.,
Lv, Z., et al.: Development of an inactivated vaccine candidate for SARS-CoV-2.
Science 369(6499), 77–81 (2020)
8. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition.
In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp.
770–778 (2016)
9. Howard, A.G., et al.: MobileNets: efficient convolutional neural networks for mobile
vision applications. arXiv:1704.04861 (2017)
10. Huang, C., Wang, Y., Li, X., Ren, L., Zhao, J., Hu, Y., Zhang, L., Fan, G., Xu, J.,
Gu, X., et al.: Clinical features of patients infected with 2019 novel coronavirus in
wuhan, china. The Lancet 395(10223), 497–506 (2020)
11. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected
convolutional networks. In: IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), pp. 4700–4708 (2017)
12. Kawahara, J., Hamarneh, G.: Multi-resolution-tract CNN with hybrid pretrained
and skin-lesion trained layers. In: International Workshop on Machine Learning in
Medical Imaging, pp. 164–171 (2016)
68 G. Oliveira et al.

13. Lalmuanawma, S., Hussain, J., Chhakchhuak, L.: Applications of machine learning
and artificial intelligence for Covid-19 (SARS-CoV-2) pandemic: A review. Chaos,
Solitons & Fractals 139, 110059 (2020)
14. Litjens, G., et al.: A survey on deep learning in medical image analysis. Med. Image
Anal. 42, 60–88 (2017)
15. Narin, A., Kaya, C., Pamuk, Z.: Automatic detection of coronavirus dis-
ease (COVID-19) using X-ray images and deep convolutional neural networks.
arXiv:2003.10849 (2020)
16. Raghu, M., Zhang, C., Kleinberg, J., Bengio, S.: Transfusion: understanding trans-
fer learning for medical imaging. In: Advances in Neural Information Processing
Systems (NIPS), pp. 3347–3357 (2019)
17. Rajpurkar, P., et al.: CheXNet: radiologist-level pneumonia detection on chest X-
rays with deep learning. arXiv:1711.05225 (2017)
18. Randhawa, G.S., Soltysiak, M.P., El Roz, H., de Souza, C.P., Hill, K.A., Kari, L.:
Machine learning using intrinsic genomic signatures for rapid classification of novel
pathogens: Covid-19 case study. PLoS ONE 15(4), e0232391 (2020)
19. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Grad-
cam: Visual explanations from deep networks via gradient-based localization. In:
IEEE International Conference on Computer Vision (ICCV), pp. 618–626 (2017)
20. Shen, W., Zhou, M., Yang, F., Yang, C., Tian, J.: Multi-scale convolutional neural
networks for lung nodule classification. In: International Conference on Information
Processing in Medical Imaging (IPMI), pp. 588–599 (2015)
21. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the incep-
tion architecture for computer vision. In: IEEE Conference on Computer Vision
and Pattern Recognition (CVPR), pp. 2818–2826 (2016)
22. Tan, M., Le, Q.: EfficientNet: rethinking model scaling for convolutional neural
networks. In: IEEE International Conference on Machine Learning (ICML), pp.
6105–6114 (2019)
23. Tuli, S., Tuli, S., Tuli, R., Gill, S.S.: Predicting the growth and trend of COVID-
19 pandemic using machine learning and cloud computing. Internet of Things 11,
100222 (2020)
24. Wang, L., Wong, A.: COVID-Net: a tailored deep convolutional neural network
design for detection of COVID-19 cases from chest x-ray images. arXiv:2003.09871
(2020)
25. Yan, L., et al.: An interpretable mortality prediction model for COVID-19 patients.
Nat. Mach. Intell. 2, 1–6 (2020)
26. Yosinski, J., Clune, J., Bengio, Y., Lipson, H.: How transferable are features in deep
neural networks? In: Advances in Neural Information Processing Systems (NIPS),
pp. 3320–3328 (2014)
27. Zhao, J., Zhang, M., Zhou, Z., Chu, J., Cao, F.: Automatic detection and classifi-
cation of leukocytes using convolutional neural networks. Med. Biol. Eng. Comput.
55(8), 1287–1301 (2016). https://doi.org/10.1007/s11517-016-1590-x
28. Zhou, P., Yang, X.L., Wang, X.G., Hu, B., Zhang, L., Zhang, W., Si, H.R., Zhu,
Y., Li, B., Huang, C.L., et al.: A pneumonia outbreak associated with a new
coronavirus of probable bat origin. Nature 579(7798), 270–273 (2020)
29. Zoph, B., Vasudevan, V., Shlens, J., Le, Q.V.: Learning transferable architectures
for scalable image recognition. In: IEEE Conference on Computer Vision and Pat-
tern Recognition (CVPR), pp. 8697–8710 (2018)
 Classification of Musculoskeletal
Abnormalities with Convolutional Neural
Networks

Guilherme Tiaki Sassai Sato(B) , Leodécio Braz da Silva Segundo ,

and Zanoni Dias

Institute of Computing, University of Campinas, Campinas, Brazil

[email protected], [email protected]

Abstract. Computer-aided diagnosis has the potential to alleviate the

burden on medical doctors and decrease misdiagnosis, but building a suc-
cessful method for automatic classification is challenging due to insuffi-
cient labeled data. In this work, we investigate the usage of convolutional
neural networks to diagnose musculoskeletal abnormalities using radio-
graphs (X-rays) of the upper limb and measure the impact of several
techniques in our model. We achieved the best results by utilizing an
ensemble model that employs a support vector machine to combine dif-
ferent models, resulting in an overall AUC ROC of 0.8791 and Kappa of
0.6724 when evaluated using an independent test set.

Keywords: Deep learning · Musculoskeletal abnormalities · X-ray

1 Introduction
Musculoskeletal conditions are extensively present in the population, affecting
over 1.3 billion people worldwide [9]. These conditions often cause long-term
pain, directly and indirectly reducing the quality of life of those suffering from
it and their household [33,34]. In this setting, medical imaging as X-rays plays
an essential role as one of the main tools for abnormality detection.
Insufficient medical staff, along with the complexity of diagnosis, creates a
system prone to errors. False-negative diagnosis leads to untreated injuries and
symptoms such as chronic pain and further complications in the long term, and
false-positives diagnosis leads to unnecessary treatment. Computer-Aided Diag-
nosis (CAD) systems are used to counteract these problems, improving diagnos-
tic accuracy, assist decision-making, and reducing radiologists’ workload.
Computer-aided diagnosis has been a topic of research since the 1960s and
has significantly evolved due to advances in medicine itself and in computer sci-
ence. From a task standpoint, CAD has found application in a wide variety of
medical disorders. A few of the innumerous works include breast cancer [8,31],
lung cancer [19], and Alzheimer’s [6]. X-ray classification has been particularly
prevalent for the chest area [2,32]. Under the same scope of this work (muscu-
loskeletal abnormalities in the upper limb), 70 works were submitted under a
c Springer Nature Switzerland AG 2020
J. C. Setubal and W. M. Silva (Eds.): BSB 2020, LNBI 12558, pp. 69–80, 2020.
https://doi.org/10.1007/978-3-030-65775-8_7
70 G. T. S. Sato et al.

competition1 that took place using the same dataset as this work. The reported
achieved Cohen’s Kappa range from 0.518 to 0.843. However, no further infor-
mation, such as methodology, is generally available for these works.
Methodology-wise, some of the most successful classifiers include k-nearest
neighbors (KNN) [14,20], support vector machines (SVM) [7], random forests
[1,15], and neural networks [16,30]. In this work, we will evaluate the use of a
neural network classifier due to its promising performance, producing state-of-
the-art results in many other applications [4,26].
A wide range of techniques, such as deep learning, image processing, and com-
puter vision, are applied to interpret radiographic images automatically. How-
ever, deep learning models’ success is highly dependent on the amount of data
available, creating a particular challenge for medical images due to privacy con-
cerns and time-consuming labeling requiring experts. In this work, we present a
method to classify normal and abnormal X-rays from the upper limb by applying
convolutional neural networks and several machine learning techniques aiming
to improve the classification and offset the lack of data. The rest of this paper
is organized as follows. Section 2 describes the settings of this work and all the
experiments executed to reach a final classifier, and Sect. 3 evaluates this clas-
sifier using an independent test set and discusses the results obtained, as well
as alternative scenarios. Section 4 concludes the paper, including an overview of
the results obtained and future work.

2 Methodology
This section goes over the dataset used in this work and details the experi-
ments executed. Over the experiments, we explore different scenarios by apply-
ing machine learning and deep learning techniques and measure the impact of
the proposed changes when comparing to previously tested scenarios, in a path
to maximize the classifier robustness.

2.1 Dataset
In this work, we used the MURA: Large Dataset for Abnormality Detection
in Musculoskeletal Radiographs [23] dataset, which contains 40,005 radiographic
images labeled by radiologists. The MURA dataset is divided into 14,656 studies
from the body upper extremities – shoulder, humerus, elbow, forearm, wrist,
hand, and finger. Each study is labeled as either normal or abnormal.
The available data is divided into validation and training sets. For testing,
a third set was created to be used in a competition and, therefore, not publicly
available. To offset this fact and provide a realist measurement of our model’s
performance in a real-world scenario, we split the training data to create a test
set. Our goal was to create a test set the same size as the validation set. As the
provided validation set has 8.6% of the size of the training set, the new test set
was created by moving, for each body part, 8.6% of the studies from the training
to the test set. The number of items in each class is described in Table 1.
1
https://stanfordmlgroup.github.io/competitions/mura/.
 Classification of Musculoskeletal Abnormalities 71

Table 1. Distribution of the number of studies (images) contained in each class for
the three sets.

Train Validation Test

Normal Abnormal Normal Abnormal Normal Abnormal
Shoulder 1242 (3838) 1331 (3833) 99 (285) 95 (278) 122 (373) 126 (335)
Humerus 293 (608) 247 (549) 68 (148) 67 (140) 28 (65) 24 (50)
Elbow 997 (2677) 601 (1841) 92 (235) 66 (230) 97 (248) 59 (165)
Forearm 543 (1069) 257 (595) 69 (150) 64 (151) 47 (95) 30 (66)
Wrist 1993 (5237) 1218 (3670) 140 (364) 97 (295) 201 (528) 108 (317)
Hand 1365 (3702) 475 (1354) 101 (271) 66 (189) 132 (357) 46 (130)
Finger 1161 (2834) 600 (1805) 92 (214) 83 (247) 119 (304) 55 (163)

2.2 Experiments

The models used for classification were developed in Python, mainly using the
PyTorch framework [21]. Every sample from the dataset has its target class
defined among 14 classes (7 body parts, normal or abnormal). Before inputting
the images to the model, we normalized each image’s pixels values to the mean
and standard deviation of the ImageNet dataset [5] and resized to 224 × 224
pixels.
We trained each model over 40 epochs, with a batch size of 25. The samples
are initially shuffled and reshuffled before each epoch. Upon each epoch, the
network performance is measured with the scikit-learn library package [22]. To
measure the binary classification performance, the 14 class output is condensed
into two by ignoring the body part information. To determine the output of a
study consisting of several images, the probability distribution output for each
image in the study are averaged.

Experiment I: Fit, Pad, or Stretch? Our goal is to use pre-trained networks

as a baseline. However, pre-trained networks expect a square input, and our
images have a variable aspect ratio, so before testing an assortment of networks,
we need to decide how to transform our images: fit, pad, or stretch.
All three transformations come with pros and cons. “Fit” and “Pad” preserve
the image aspect ratio, while “Stretch” distorts it. On the other hand, “Fit” loses
some amount of information, which might be considerable if the image is very tall
or long, “Pad” fills part of the input with irrelevant information, while “Stretch”
includes the entire image. Figure 1 illustrates the three transformations.
To decide on the best method, we trained the ResNet-18 network on the three
options and compared the performance. Table 2 shows the results obtained. All
of them were pretty similar, but “Stretch” achieved the best Kappa. Further-
more, “Stretch” was the fastest operation on our tests. Therefore, we chose the
“Stretch” operation to be used in the subsequent experiments.
72 G. T. S. Sato et al.

(a) Original image [23] (b) Fit (c) Pad (d) Stretch

Fig. 1. Three transformations to change a rectangular image into square. Fit does not
deform but loses information. Pad also does not deform but adds irrelevant information.
Stretch includes all pixels but deforms the image.

Table 2. Performance on the validation data for each of the transformations.

Accuracy Balanced accuracy AUC ROC Kappa

Fit 0.8232 0.8140 0.8720 0.6373
Pad 0.8165 0.8054 0.8723 0.6222
Stretch 0.8274 0.8175 0.8617 0.6453

Experiment II: Pre-trained Networks. The use of pre-trained models pro-

vides us with a consolidated and validated architecture. These models have
demonstrated good results in many similar tasks [10]. In addition, it reduces
the training time required compared to the training necessary to achieve similar
results without pre-training.
Pre-trained models may, however, have a few caveats. The models were
trained using the ImageNet dataset, containing colored images that do not resem-
ble medical images. Therefore, the layer structure of the networks may be sub-
optimal for this task. Moreover, the ImageNet images contain three channels.
Consequently, our inputs must be reshaped to conform to this restriction [35].
We tested several networks among the best-performing ones, including
DenseNet [12], EfficientNet [17,29], Inception-v3 [28], Inception-v4 [3,27],
Inception-ResNet-v2 [3,27], ResNet [11], and VGG [25]. Table 3 presents the
results obtained on each model. All the models performed very similarly, with a
Kappa coefficient between 0.6450 and 0.6671.

Experiment III: Data Augmentation. To increase the model capacity to

generalize, we tested two data augmentation techniques: cropping and horizontal
flip. To verify the effect of these augmentations, we trained the DenseNet-161
network once using only the horizontal flip augmentation, doubling the number of
samples (Augmented Dataset A); and once using the original, five random crops
and its flips, resulting in a 12-fold increase in the size (Augmented Dataset B),
similar to the augmentation proposed by Krizhevsky et al. [13]. Table 4 compares
the results of the two experiments with the original dataset.
 Classification of Musculoskeletal Abnormalities 73

Table 3. Performance of each network on the validation data.

Accuracy Balanced accuracy AUC ROC Kappa

DenseNet-121 0.8365 0.8279 0.8874 0.6649
DenseNet-161 0.8374 0.8293 0.8892 0.6671
EfficientNet-B7 0.8274 0.8182 0.8873 0.6458
Inception-v3 0.8290 0.8197 0.8769 0.6491
Inception-v4 0.8315 0.8226 0.8783 0.6546
Inception-ResNet-v2 0.8324 0.8229 0.8864 0.6559
ResNet-18 0.8274 0.8175 0.8617 0.6453
ResNet-152 0.8299 0.8220 0.8780 0.6519
VGG-16 0.8357 0.8254 0.8877 0.6621
VGG-19 0.8282 0.8155 0.8840 0.6450

Table 4. Comparison of performance on the validation data using DenseNet-161. Aug-

mented Dataset A is the dataset using only horizontal flip, and Augmented Dataset B
is the dataset using both horizontal flip and crop.

Accuracy Balanced accuracy AUC ROC Kappa

Original dataset 0.8374 0.8293 0.8892 0.6671
Augmented dataset A 0.8465 0.8370 0.8929 0.6848
Augmented dataset B 0.8432 0.8355 0.8922 0.6792

From these results, we could conclude that, despite having more data, Aug-
mented Dataset B performs worse than Augmented Dataset A, but still better
than the original. Moreover, since training time is approximately linear in the
number of samples, using Augmented Dataset A results in longer training time.
Therefore, we trained other networks using Augmented Dataset B (Table 5).
These networks were chosen based on the results of Experiment II while avoid-
ing more than one network from the same “family.”

Experiment IV: Ensemble Model. An ensemble model combines multiple

models to make a final prediction, similar to consulting multiple opinions. This
will supposedly improve the model as if a single model performs poorly for a
sample, its prediction may be overridden by other models, improving the model
stability.
We used the five models from Experiment III and combined its predictions
using a set of techniques. First, we made a final prediction based on consensus,
i.e., the mode of the five predictions, ignoring the probabilistic distribution.
Next, we combined the results using a weighted average, using one weight for
each of the five models. Then we tested two techniques using neural networks: a
sparsely connected and a fully connected layer to make a final prediction. The
architecture of these layers is shown in Fig. 2. While these last two techniques
74 G. T. S. Sato et al.

Table 5. Performance of networks on the validation data for the networks trained
using a dataset augmented using only horizontal flip (Augmented dataset A).

Accuracy Balanced accuracy AUC ROC Kappa

DenseNet-161 0.8465 0.8370 0.8929 0.6848
EfficientNet-B7 0.8399 0.8314 0.8913 0.6719
Inception-ResNet-v2 0.8457 0.8369 0.8892 0.6836
ResNet-152 0.8349 0.8243 0.8931 0.6602
VGG-16 0.8374 0.8302 0.8967 0.6676

are also ultimately a weighted average, each element of the output array for
each model is now independently weighted. Also, while the mere five weights in
the simple weighted were obtained using an exhaustive search, the weights in
these two experiments were obtained using gradient descent. Finally, we used
a support vector machine (SVM) with a radial basis function (RBF) kernel to
ensemble the models. The SVM is implemented by scikit-learn [22].

Fig. 2. Schematic representation of the two neural network architectures used in this
experiment. The number of classes shown here is 4 for simplification, while the real
number of classes is 14. The number of models n is 5 in the experiment.

Figure 3 illustrates the final methodology used. The results obtained are pre-
sented in Table 6, comparing it with the average results obtained by the models
individually. All ensemble models were able to perform better than the average
performance of the single models. The single models present a hard to avoid
overfitting problem, leading to nearly 100% accuracy in the training set for all
the models, therefore making it impossible to find the best way to combine these
predictions using the training set since practically any combination will lead to
nearly 100% accuracy. We must then tune the parameters in the ensemble layer
using the validation set, which should inevitably lead to a higher drop in perfor-
mance when transitioning to the test set.
 Classification of Musculoskeletal Abnormalities 75

Table 6. Comparison of performance on the validation data. The Consensus model

does not output a probability distribution, only the final prediction, penalizing its AUC
ROC score.

Accuracy Balanced accuracy AUC ROC Kappa

Single model (average) 0.8409 0.8319 0.8926 0.6736
Consensus 0.8449 0.8349 0.8753 0.6811
Weighted average 0.8641 0.8572 0.9086 0.7222
Sparsely connected 0.8590 0.8504 0.9039 0.7110
Fully connected 0.8540 0.8467 0.9102 0.7015
SVM (RBF) 0.8699 0.8638 0.9208 0.7345

Fig. 3. Overview of the final classification methodology adopted. The X-ray [23] is
squared, processed by five neural networks, and the five independent predictions are
combined to reach a final prediction using a SVM.

3 Results and Discussion

The Gradient-weighted Class Activation Mapping (Grad-CAM) [18,24] is a

method to visualize areas of the input image that are important to reaching
the outputted prediction. Figure 4 shows the heatmaps generated by Grad-CAM
for each individual model and the ensemble model, demonstrating the improve-
ment caused by ensembling the models. In this example VGG-16, EfficientNet-B7
and Inception-ResNet-v2 predict “abnormal humerus” while DenseNet-161 and
ResNet-152 predict “normal humerus.” The Ensemble SVM model prediction is,
correctly, “abnormal humerus.” The incorrect models focus on a different part
of the image but are overridden by other models.
Since the Grad-CAM uses the information from the feature layer to build
the visualization and our ensemble models combine the individual models at the
output layer level, we have multiple feature layers, which would preclude the
use of Grad-CAM. To work around this, we built the ensemble visualization by
76 G. T. S. Sato et al.

(a) (b) (c) (d) (e) (f) (g)

Original VGG-16 EfficientNet-Inception- DenseNet- ResNet- Ensemble
[23] B7 ResNet-v2 161 152

Fig. 4. Grad-CAM heatmaps for each individual model and the ensemble model for an
abnormal humerus sample. Models VGG-16, EfficientNet-B7 and Inception-ResNet-v2
predict correctly, while DenseNet-161 and ResNet-152 predict incorrectly. The Ensem-
ble SVM model takes all models into consideration and outputs a correct final predic-
tion.

averaging the Grad-CAM outputted matrix (used to create the heatmap) for
the five models. Hence, the heatmap generated for the ensemble model is an
approximation and not specific to any of the models.
Figure 5 summarizes the Kappa coefficients for all the experiments. Experi-
ment II had an average Kappa of 0.6542; the SVM (RBF) model from Experi-
ment IV improved it to 0.7345, which is our best result.

0.75

0.70

0.65

0.60
Inc Effi seN 161

VG t-v2

IV Wei V C et-1 6
I P it
D I S ad
II Dens Net- h
1
II ientN t-161
Inc II I cept t-B7
n-R ption v3
II esNe v4

n-R tNet 61
Re sNet 2
sN -18

-A II GG 2
III De VG -16
III III- B D seN -19

III esN -B7

Sp ght ons 52

lly nn ge
IV sely Av sus

SV nne d
(R d
)
BF
I esN -1
Effi eN 12

Re t-v

5
II ense tretc

IV Co ecte
M cte
IF

Fu Co era
ep cien et-1
II et-1
-
-

G
G

ed en
ep nce ion

en et-

e
e
e

III -B
n

R
In
c

II

-B
tio
tio
II

ar
-B B
-
III

IV
II

Fig. 5. Summary of the Kappa coefficient for the experiments. The maximum value
was 0.7345.

To evaluate our model in a scenario closer to the real world, we will now
use our test set, defined in Subsect. 2.1, which has no overlap with the other
two sets. Table 7 shows the performance metrics achieved using the Ensemble
 Classification of Musculoskeletal Abnormalities 77

SVM model, which was the best performing model on the validation data. The
model’s performance expectedly decreased due to overfitting.
Our model performed worst on hands, with a Kappa of 0.4717, and best
on elbows, with a Kappa of 0.7921. The overall Kappa was 0.6724. In contrast,
human radiologists performed worst on fingers and best on wrists [23]. Our model
was able to outperform two of the three radiologists evaluated on the elbow
classification task, and all of them in the finger classification, but falls behind
on other body parts. The MURA competition has ended, so we cannot evaluate
our model on the official test set. As a consequence, the human radiologists
were evaluated on the official test set, and our model, using our test set, so the
performance comparison might not be completely accurate.

Table 7. Performance of the model Ensemble SVM on the test data, broken down by
inputted body part.

Accuracy Balanced accuracy AUC ROC Kappa

Shoulder 0.8266 0.8270 0.8778 0.6535
Humerus 0.8077 0.7976 0.8274 0.6061
Elbow 0.9038 0.8895 0.9023 0.7921
Forearm 0.7922 0.7755 0.8142 0.5578
Wrist 0.8835 0.8633 0.9142 0.7393
Hand 0.8090 0.7225 0.7759 0.4717
Finger 0.8563 0.8558 0.8901 0.6817
Overall 0.8484 0.8319 0.8791 0.6724

The performance difference between the test and validation sets is likely
explained due to the fitting of ensemble parameters using the validation set,
as described in Experiment IV. We can verify this by running the test set on
the Ensemble Consensus model, which does not have any extra parameters.
Table 8 shows the results of this test. These results are similar to the results
obtained in the validation set (by some metrics, even better), and despite being
the worst performer on Experiment IV, it performed better than the Ensemble
SVM model, the best ensemble model. Therefore, we could assume that if we
were able to train the single models in such a way to reduce overfitting, we
could train the ensemble parameters using the train set and reduce the gap
between validation and test results. Alternatively, splitting the data into four

Table 8. Performance of the model Ensemble Consensus on the test data. The model
has similar performance on the validation set.

Accuracy Balanced accuracy AUC ROC Kappa

0.8593 0.8339 0.8652 0.6899
78 G. T. S. Sato et al.

sets instead of three to fit the ensemble parameters using a separated set could
also be beneficial, but would further reduce an already limited dataset.

4 Conclusion

Our work proposed to explore this machine learning classification problem using
convolutional neural networks and related methods. The best setting found was
to stretch the input to a square, apply horizontal flip data augmentation, and
ensemble a variety of architectures using a support vector machine, reaching an
AUC ROC of 0.8791 and Kappa of 0.6724.
Although the overall result was lower than that obtained by human radiolo-
gists, we were still able to achieve promising results in some scenarios. However,
a transition to the clinical setting is challenging. Besides accuracy improvement
under a controlled scenario, the algorithm would need, for example, to han-
dle unexpected inputs, provide translation and rotation invariance, and include
explainability to mitigate automation bias.
The greatest hindrance in this work was the models’ overfitting, to which
further experimentation with other methods is needed to adequately address it,
such as early stopping and regularization. The development of larger datasets
would also provide a big leap on this matter.

References
1. Alickovic, E., Subasi, A.: Medical decision support system for diagnosis of heart
Arrhythmia using DWT and random forests classifier. J. Med. Syst. 40(4), 1–12
(2016). https://doi.org/10.1007/s10916-016-0467-8
2. Arzhaeva, Y., et al.: Development of automated diagnostic tools for pneumoconiosis
detection from chest X-ray radiographs (2019)
3. Cadene, R.: Pretrained models for Pytorch (2017). https://github.com/Cadene/
pretrained-models.pytorch
4. Chiu, C.C., et al.: State-of-the-art speech recognition with sequence-to-sequence
models. In: IEEE International Conference on Acoustics, Speech and Signal Pro-
cessing (ICASSP), pp. 4774–4778. IEEE (2018). https://doi.org/10.1109/ICASSP.
2018.8462105
5. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: a large-scale
hierarchical image database. In: IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), pp. 248–255. IEEE (2009). https://doi.org/10.1109/CVPR.
2009.5206848
6. Ding, Y., et al.: A deep learning model to predict a diagnosis of Alzheimer disease
by using 18F-FDG PET of the brain. Radiology 290(2), 456–464 (2019). https://
doi.org/10.1148/radiol.2018180958
7. Dolatabadi, A.D., Khadem, S.E.Z., Asl, B.M.: Automated diagnosis of coronary
artery disease (CAD) patients using optimized SVM. Comput. Methods Programs
Biomed. 138, 117–126 (2017). https://doi.org/10.1016/j.cmpb.2016.10.011
 Classification of Musculoskeletal Abnormalities 79

8. Garud, H., et al.: High-magnification multi-views based classification of breast

fine needle aspiration cytology cell samples using fusion of decisions from deep
convolutional networks. In: IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), pp. 76–81 (2017). https://doi.org/10.1109/CVPRW.2017.
115
9. Global Burden of Disease Collaborative Network: Global burden of disease study
2017 (GBD 2017) results (2018)
10. Habibzadeh, M., Jannesari, M., Rezaei, Z., Baharvand, H., Totonchi, M.: Auto-
matic white blood cell classification using pre-trained deep learning models: ResNet
and Inception. In: Proceedings of the 10th International Conference on Machine
Vision (ICMV), vol. 10696, p. 1069612. International Society for Optics and Pho-
tonics (2017). https://doi.org/10.1117/12.2311282
11. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition.
In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp.
770–778 (2016). https://doi.org/10.1109/CVPR.2016.90
12. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected
convolutional networks. In: IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), pp. 4700–4708 (2017). https://doi.org/10.1109/CVPR.2017.
243
13. Krizhevsky, A., Sutskever, I., Hinton, G.E.: ImageNet classification with deep con-
volutional neural networks. In: Advances in Neural Information Processing Sys-
tems, pp. 1097–1105 (2012)
14. Li, Q., Li, W., Zhang, J., Xu, Z.: An improved k-nearest neighbour method to
diagnose breast cancer. Analyst 143(12), 2807–2811 (2018). https://doi.org/10.
1039/C8AN00189H
15. Liu, X., et al.: Automated layer segmentation of retinal optical coherence tomog-
raphy images using a deep feature enhanced structured random forests classi-
fier. IEEE J. Biomed. Health Inform. 23(4), 1404–1416 (2018). https://doi.org/
10.1109/JBHI.2018.2856276
16. Ma, J., Wu, F., Zhu, J., Xu, D., Kong, D.: A pre-trained convolutional neural net-
work based method for thyroid nodule diagnosis. Ultrasonics 73, 221–230 (2017).
https://doi.org/10.1016/j.ultras.2016.09.011
17. Melas-Kyriazi, L.: EfficientNet PyTorch (2019). https://github.com/lukemelas/
EfficientNet-PyTorch
18. Nakashima, K.: Grad-CAM with PyTorch (2017). https://github.com/kazuto1011/
grad-cam-pytorch
19. Nishio, M., et al.: Computer-aided diagnosis of lung nodule classification between
benign nodule, primary lung cancer, and metastatic lung cancer at different image
size using deep convolutional neural network with transfer learning. PLoS ONE
13(7), e0200721 (2018). https://doi.org/10.1371/journal.pone.0200721
20. Panwar, M., Acharyya, A., Shafik, R.A., Biswas, D.: K-nearest neighbor based
methodology for accurate diagnosis of diabetes mellitus. In: Sixth International
Symposium on Embedded Computing and System Design (ISED), pp. 132–136.
IEEE (2016). https://doi.org/10.1109/ISED.2016.7977069
21. Paszke, A., et al.: PyTorch: an imperative style, high-performance deep learning
library. In: Advances in Neural Information Processing Systems, pp. 8026–8037
(2019)
22. Pedregosa, F., et al.: Scikit-learn: machine learning in Python. J. Mach. Learn.
Res. 12, 2825–2830 (2011)
23. Rajpurkar, P., et al.: MURA: large dataset for abnormality detection in muscu-
loskeletal radiographs. arXiv:1712.06957 (2017)
80 G. T. S. Sato et al.

24. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Grad-
CAM: visual explanations from deep networks via gradient-based localization. In:
IEEE International Conference on Computer Vision (ICCV), pp. 618–626 (2017).
https://doi.org/10.1109/ICCV.2017.74
25. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale
image recognition. arXiv:1409.1556 (2014)
26. Sun, Y., Liang, D., Wang, X., Tang, X.: DeepID3: face recognition with very deep
neural networks. arXiv:1502.00873 (2015)
27. Szegedy, C., Ioffe, S., Vanhoucke, V., Alemi, A.A.: Inception-v4, Inception-ResNet
and the impact of residual connections on learning. In: Proceedings of the 31st
AAAI Conference on Artificial Intelligence, pp. 4278–4284 (2017)
28. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the Incep-
tion architecture for computer vision. In: IEEE Conference on Computer Vision
and Pattern Recognition (CVPR), pp. 2818–2826 (2016). https://doi.org/10.1109/
CVPR.2016.308
29. Tan, M., Le, Q.V.: EfficientNet: rethinking model scaling for convolutional neural
networks. arXiv:1905.11946 (2019)
30. Tschandl, P., et al.: Expert-level diagnosis of nonpigmented skin cancer by com-
bined convolutional neural networks. JAMA Dermatol. 155(1), 58–65 (2019).
https://doi.org/10.1001/jamadermatol.2018.4378
31. Wang, H., Zheng, B., Yoon, S.W., Ko, H.S.: A support vector machine-based
ensemble algorithm for breast cancer diagnosis. Eur. J. Oper. Res. 267(2), 687–699
(2018). https://doi.org/10.1016/j.ejor.2017.12.001
32. Wang, X., Peng, Y., Lu, L., Lu, Z., Bagheri, M., Summers, R.M.: ChestX-ray8:
hospital-scale chest X-ray database and benchmarks on weakly-supervised classifi-
cation and localization of common thorax diseases. In: IEEE Conference on Com-
puter Vision and Pattern Recognition (CVPR), pp. 2097–2106 (2017). https://doi.
org/10.1109/CVPR.2017.369
33. Woolf, A.D., Pfleger, B.: Burden of major musculoskeletal conditions. Bull. World
Health Organ. 81, 646–656 (2003)
34. World Health Organisation: Musculoskeletal conditions fact sheet (2019). https://
www.who.int/news-room/fact-sheets/detail/musculoskeletal-conditions
35. Xie, Y., Richmond, D.: Pre-training on grayscale ImageNet improves medical image
classification. In: Leal-Taixé, L., Roth, S. (eds.) ECCV 2018. LNCS, vol. 11134, pp.
476–484. Springer, Cham (2019). https://doi.org/10.1007/978-3-030-11024-6 37
 Combining Mutation and Gene Network
Data in a Machine Learning Approach
for False-Positive Cancer Driver Gene
Discovery

Jorge Francisco Cutigi1,2(B) , Renato Feijo Evangelista2 ,

Rodrigo Henrique Ramos1,2 , Cynthia de Oliveira Lage Ferreira2 ,
Adriane Feijo Evangelista3 , Andre C. P. L. F. de Carvalho2 ,
and Adenilso Simao2
1
Federal Institute of Sao Paulo, Sao Carlos, SP, Brazil
[email protected]
2
University of Sao Paulo, Sao Carlos, SP, Brazil
3
Barretos Cancer Hospital, Barretos, SP, Brazil

Abstract. An increasing interest in Cancer Genomics research emerged

from the advent and widespread use of next-generation sequencing tech-
nologies, which have generated a large amount of digital biological data.
However, not all of this information in fact contributes to cancer studies.
For instance, false-positive-driver genes may contain characteristics of
cancer genes but are not actually relevant to the cancer initiation and
progression. Including this type of genes in cancer studies may lead to
identifying unrealistic trends in the data and mislead biomedical deci-
sions. Therefore, proper screening to detect this specific type of gene
among genes considered drivers is of utmost importance. This work is
focused on the development of models dedicated to this task. Support
Vector Machine (SVM) and Random Forest (RF) machine learning algo-
rithms were selected to induce predictive models to classify supposedly
driver genes as real drivers or false-positive drivers based on both muta-
tion data and gene network interactions. The results confirmed that the
combination of the two sources of information improves the performance
of the models. Moreover, SVM and RF models achieved a classification
accuracy of 85.0% and 82.4% over labeled data, respectively. Finally, a
literature-based analysis was performed over the classification of a new
set of genes to further validate the concept.

Keywords: Cancer bioinformatics · Driver genes · False-positive

driver · Complex networks · Machine learning

c Springer Nature Switzerland AG 2020

J. C. Setubal and W. M. Silva (Eds.): BSB 2020, LNBI 12558, pp. 81–92, 2020.
https://doi.org/10.1007/978-3-030-65775-8_8
82 J. F. Cutigi et al.

1 Introduction
Cancer is one of the main cause of death globally, being responsible for around
9.6 million deaths in 2018, according to the World Health Organization1 . It
is considered a complex disease and is caused by the accumulation of genetic
alterations in the human body cells, which are called genetic mutations.
The investigation of mutations is crucial for the understanding of cancer initia-
tion and progression. A single cell undergoes a vast number of mutations; nonethe-
less, not all of them lead to cancer. In this context, mutations can be classified in
two types: Passenger mutations, which comprehend the majority of mutations in
cancer cells and are not significant for the cancer progression, i.e., do not confer
a selective advantage to cells; and Driver mutations, which are a small group of
mutation significant for cancer, i.e., they give cancer cells a growth advantage.
Distinguishing between driver and passenger mutations is a long-standing
investigation line in Cancer Genomics. Many computational methods have been
developed on this topic [5,9], which are based on various types of data that are
currently available. Among them, mutation data analysis has taken a prominent
position after the advent of next-sequencing generation technologies (NGS) and
thanks to projects such as the TCGA (The Cancer Genome Atlas) [29] that
makes large collections of mutation data available. Gene interaction informa-
tion is also often explored and has an important role in many computational
methods [22]. This type of data provides essential information about complex
gene interactions among genes and their related proteins. Such interactions are
represented by complex networks, in which genes are nodes and edges connect
genes that are physically interacting or functionally related [16].
There are also computational methods that benefit from the simultaneous anal-
ysis of mutation and gene network data. HotNet [28], HotNet2 [17], Hierarchi-
cal HotNet [25], MUFFINN [3], nCOP [13], NetSig [12], and GeNWeMME [6] are
methods that employ network algorithms and mutation data analysis to find sig-
nificantly related mutations and to identify driver genes. Recently, machine learn-
ing algorithms, such as DriverML [11], LOTUS [4], and MoProEmbeddings [10]
have taken advantage of the massive volume of digital biological data to induce
predictive models able to suggest driver genes and find novel biological insights.
Although these methods have been extensively used for driver gene identifi-
cation, they can misclassify some genes as drivers, thus being necessary expert
curation to filter their findings [1]. It occurs because some genes present charac-
teristics of drivers, but are not actually involved in cancer initiation and progres-
sion. These genes are referred to as false-positive drivers and may even mislead
a biomedical decision or compromise the performance of models that consider
them as variables. The avoiding of the misclassification of false-positive-drivers
as drivers is still an ongoing challenge. Thus, the development of dedicated tools
for further screening these models’ findings so that possible misclassified genes
can be detected is required. In this work, the terms false-positive drivers and
false-drivers are used as synonymous.
1
www.who.int/news-room/fact-sheets/detail/cancer.
Machine Learning Approach for False-Positive Cancer Driver Gene Discovery 83

In this context, this work presents a machine learning-based approach to

induce predictive models able to classify driver gene candidates as real drivers
or false-drivers. Random Forest (RF) and Support Vector Machine (SVM) were
selected as the supervised learning methods. The proposed approach extracts
useful information from both mutation data and gene network interactions as
features for the models. The training was performed over a data set composed
of 876 labeled genes, created from the combination of somatic mutation data
of 33 types of cancer and centrality measures of a union of four gene inter-
actions networks. The models were compared, considering a set of evaluation
metrics. Experimental results show that the combination of mutation data and
gene interaction data can improve the predictive models’ prediction potential.
An automated literature-based validation, taken accounted for the number of
citations of classified genes, evidenced the potential of models suggesting false-
drivers genes.
This paper is organized as follows: Sect. 2 describes the machine learning-
based approach employed in this work, from data collection to the induction
of predictive models. Next, Sect. 3 presents a thorough evaluation of the mod-
els using classification metrics and automated literature-based analysis. Finally,
Sect. 4 summarizes the research findings and concludes with final considerations.

2 Method
This section describes the steps in the development of this research. Figure 1
shows a summary of the approach established for the research. In Step 1, cancer
mutation data, gene interaction networks, and gene labels are selected from
reliable and widely used sources. In Step 2, data is preprocessed, and features
are extracted. Finally, in Step 3, a hyper-parameters tuning is performed so that
optimized models can be induced and evaluated through stratified k-fold cross-
validation. Further assessment of the models’ applicability is performed over new
genes reported as drivers by other sources.

Fig. 1. An overview of the approach.

84 J. F. Cutigi et al.

2.1 Data Collection and Preprocessing

Mutation Data. Data of 33 types of cancer were selected based on a TCGA
Pan-Cancer study [1] and downloaded from the cBioPortal2 [2] using a web
API3 . The collection contains mutation data of two types: (1) Single Nucleotide
Variants (SNVs); and (2) Insertions and Deletions (InDels). The mutation data
for each type of cancer is structured in a tab-delimited text file referred as MAF
(Mutation Annotation Format)4 , with rows listing different somatic mutations
and columns containing more than one hundred entries of related information,
such as gene name (Hugo Symbol), type of mutation (Variant Classification),
and sample/patient id (Tumor Sample Barcode).
Each MAF file was subjected to a preprocessing routine. This process is
crucial in cancer analysis, given that the data contain information that should
be suppressed (e.g., a specific type of mutations) for exome analysis. More-
over, hypermutated samples should be removed since they are usually consid-
ered outliers and can bias the analyses. The preprocessing was performed in two
steps. First, a filter was applied and only nine specific somatic variants were
kept in MAF file, namely: Frame Shift Del, Frame Shift Ins, In Frame Del,
In Frame Ins, Missense Mutation, Nonsense Mutation, Nonstop Mutation,
Splice Site and Translation Start Site. These variants were selected
because they are non-silent mutations and from coding regions, i.e., they are
likely to be mutations that lead to a functional impact. Later, hypermutated sam-
ples were removed from the MAF file. Among the possible methods to identify
hypermutated samples, the one proposed by Tamborero et al. [27] was applied.
According to their criterion, a sample is hypermutated when it contains more
than (Q3 + 4.5 × IQR) somatic mutations, where Q3 is the third quartile, and
IQR is the interquartile range of the distribution of mutations across all data
samples. Finally, the 33 preprocessed MAFs were merged into a single MAF file.
Table 1 compares the composition of the mutation data with and without the
preprocessing.

Table 1. Mutation data before and after preprocessing routine.

Non-preprocessed mutation data Preprocessed mutation data

Patients 10429 9741
Genes 20072 19184
Mutations 2192073 1228126

Gene Interaction Network Data. The following four gene-networks that

use protein-protein (PPIs) as the main source of interactions were selected to
extract node measures: (1) ReactomeFI (Reactome Functional Interactions) -
2
https://www.cbioportal.org/datasets.
3
https://www.cbioportal.org/api/swagger-ui.html.
4
https://docs.gdc.cancer.gov/Data/File Formats/MAF Format/.
Machine Learning Approach for False-Positive Cancer Driver Gene Discovery 85

2019 version; (2) Binary interaction from HINT (High-quality INTeractomes) -

March 2020 version; (3) HPRD (Human Protein Reference Database) - Release 9;
and (4) HuRI (Human Reference Interactome) - 2020 release. ReactomeFI is an
extensive gene network built from curated pathways from many sources, whose
data are obtained mainly from reliable PPI and known pathways [14]. HINT is
a curated compilation of high-quality PPI, filtered systematically and manually
for the removal of low-quality/erroneous interactions [7]. HPRD is a classical and
curated human protein interactions, built from PPI, post-translational modifi-
cations, enzyme-substrate relationships, and disease associations [15]. HuRI is a
protein interaction network built from pairwise combinations of human protein-
coding genes involved in binary interactions [19].
The selected gene networks were treated as undirected and unweighted
networks. A union operation was applied to these networks, resulting in a
single network U GN . Such operation considers the interaction and nodes
of all networks, i.e., U GN = Reactome ∪ HIN T ∪ HP RD ∪ HuRI. For
example, if N1 is a network with node set {gi , gj , gk } and interaction set
{(gi , gj ), (gj , gk )}, and N2 is a network with node set {gi , gj , gk , gl } and inter-
action set {(gi , gj ), (gi , gk ), (gk , gl )}, then the union N1 ∪ N2 has node set
{gi , gj , gk , gl } and edge set {(gi , gj ), (gj , gk ), (gk , gl )}. Finally, after the union pro-
cess, a single network was constructed to be used in this research. The resulting
network U GN has 18959 genes and 372583 interactions.

Combined Data. In order to train supervised machine learning algorithms,

the data set needed to be properly structured. The samples in the unlabeled
data set are the genes, while the features are the measures extracted from the
mutation and gene network data, as summarized below:

– ten features were extracted for each gene from the MAF file to create a muta-
tion data set DSM U T . One feature is related to the gene’s coverage, i.e., the
number of patients in which the gene is mutated. Nine features were extracted
for each somatic variant, representing each specific somatic variant’s number
of mutations. Thus, the data set DSM U T is composed of 19184 samples and
ten features.
– ten features were extracted for each gene (node) in the network U GN to create
a data set DSGN . The features are the following centrality measures: degree,
betweenness, closeness, eigenvector, coreness, clustering coefficient, average
of neighbors’ degree, leverage, information, and bridging. Such measures con-
sider distinct aspects of the network structure and topology to characterize
the importance of a node, thus highlighting its central role according to each
of measures [21]. The data set DSGN is composed of 18959 samples and ten
features.

A combined data set was obtained by merging DSM U T and DSGN . Some of
the genes were not contained in both data sets, therefore only their intersection
was taken. The merging leads to the data set DSCOM B , composed of 16281
samples and 20 features.
86 J. F. Cutigi et al.

Gene Labels. A total of 250 genes listed as possible-false-positive drivers5 by

the Network of Cancer Genes (NCG) [24] were used as a reference to label genes
in the dataset as false-drivers (F D) for the induction of predictive models by
the supervised machine learning algorithms. For the drivers class D, two sets
DN CG and DCGC of known driver genes were extracted from the NCG6 and
Cancer Gene Census (CGC)7 [26], respectively. The DN CG set contains 711
known cancer drivers, while DCGC 723. A union of both sets was performed,
resulting in a known driver gene set D = DN CG ∪ DCGC of 729 genes. However,
a total of 49 genes in this set were also present in F D, therefore, they were
removed from D. The remaining 680 genes were then used as a reference to label
the drivers.
Finally, the genes in the data set DSCOM B contained in the F D and D lists
were extracted and properly labeled. Considering only the labeled samples, the
resulting data set DSCOM BL is composed of 876 samples, 20 numeric features,
and one class label (647 drivers and 229 false-drivers). The unlabeled data set
DSCOM BU is composed of 15405 samples. A z-score standardization was applied
to all features DSCOM BL , and the class imbalance was also addressed prior to
the training process, as described in Sect. 2.2.

2.2 Machine Learning Training Process

Supervised machine learning algorithms were trained with the data set
DSCOM BL to induce predictive models that classify genes as drivers or false-
drivers. Scikit-learn [23], a Python module for machine learning, was used in all
processes described in this section.

Predictive Models. Two machine learning algorithms were selected to induce

the predictive models: Support Vector Machine (SVM) and Random Forest (RF).
The models were induced using a stratified 5-fold cross-validation scheme with
under-sampling applied to every training portion of folds to avoid over-fitting
and address class-imbalance. The under-sampling was performed by randomly
removing samples from the majority class, i.e., the driver class. It is important to
note that both the under-sampling and the folds split procedures were repeated,
taking different random states on every new training process.

Hyper-parameter Selection. Different hyper-parameters sets were assessed

by training multiple models in a grid-search process, always using 5-fold cross-
validation with under-sampling and repeating 10 times in order to account for
the possible influence of randomness. SVM hyper-parameters C and gamma were
both varied from 1 to 10, in addition to the auto and scale standard options.
The linear, sigmoid and rbf kernels were considered. RF was tested with a
5
Version 6.0 – http://ncg.kcl.ac.uk/false positives.php.
6
Version 6.0 – http://ncg.kcl.ac.uk/download.php.
7
Version 91 – https://cancer.sanger.ac.uk/census.
Machine Learning Approach for False-Positive Cancer Driver Gene Discovery 87

total number of estimators in the range from 20 to 500. The maximum depth
was tested with limits from 3 to 20 layers, including the no restriction scenario.
Both gini and entropy criteria were evaluated. The possibilities for maximum
numbers of features considered to perform a split varied according to the fol-
lowing functions: None, auto, sqrt, log2. The mean prediction accuracy was
used as a metric for the comparison and was calculated over all folds and repeti-
tion for each model. The best model for each algorithm was selected for further
evaluation.

Evaluation and Validation Criteria. Accuracy, Precision, Recall, and F1-

score were also quantified to assess the selected models. These metrics are impor-
tant because they can help to identify possible systematic trends on the miss-
classifications. The receiver operating characteristic (ROC) curves were gen-
erated for the models trained, considering both mutation and gene network
features. They were then compared to the ROC curves obtained from models
trained using just a single source of features (i.e., either mutation or gene network
data). The models with a single source of features were also trained using the
same methodology and hyper-parameters selected through new grid-searches.
The areas under the ROC curves (AUC) were also calculated for comparison.
Further evaluation was conducted over the models’ classifications of unla-
beled genes that have been recently reported as drivers [20]. The findings were
compared to the literature through a systematic citation frequency analysis using
a literature-mined database of cancer genes [18]. Such analysis was designed to
bring additional evidence that the induced models are, in fact, capable of iden-
tifying potential false-positive-driver genes, thus validating the concept.

3 Results

3.1 Model Evaluation

The optimal hyper-parameters obtained from the grid-searches performed for the
three labeled data sets, with different features, are provided in Table 2. Because
some of the optimal parameters obtained for the SVM were in the limits of the
tested ranges, these were then expanded, but no significant improvement in per-
formance was observed. The data sets are referred to as follow. DSCOM BL :
labeled data set with features from mutation data and gene network data;
DSM U TL : labeled data set with features only from mutation data; and DSGNL :
labeled data set with features only from gene network data.
The models induced using these hyper-parameters were further investigated
using the metrics described in Sect. 2.2. Once again, the mean of each metric was
calculated over 30 repetitions of the whole training process. Table 3 shows the
calculated averages and standard deviations for each selected evaluation metric.
Both models trained with combined features presented satisfactory results,
without significant differences in performance based on the selected metrics.
Moreover, the construction of models using combined mutation and network
88 J. F. Cutigi et al.

Table 2. Optimal hyper-parameters for data sets containing different features.

SVM RF
DSCOM BL DSM U TL DSGNL DSCOM BL DSM U TL DSGNL
Kernel linear linear rbf Estimators 300 100 150
C 1 1 1 max. depth 10 12 5
gamma 1 1 10 max. features None auto auto
Criterion gini gini gini

features outperforms the other models induced with a single source of features.
This trend is also observed when analyzing ROC curves, presented in Fig. 2.
Additionally, DeLong test [8] was performed over the 30 repetitions to compare
ROC curves of DSCOM BL and DSM U TL . A p-value < 0.05 was observed in 24
and 27 tests for SVM and RF, respectively.

Table 3. Comparison between models induced by different features: DSCOM BL ,

DSM U TL , and DSGNL .

SVM RF
DSCOM BL DSM U TL DSGNL DSCOM BL DSM U TL DSGNL
Accuracy 0.850 ± 0.007 0.828 ± 0.006 0.704 ± 0.011 0.824 ± 0.007 0.785 ± 0.007 0.659 ± 0.009
Precision 0.916 ± 0.005 0.898 ± 0.004 0.784 ± 0.006 0.931 ± 0.004 0.923 ± 0.006 0.853 ± 0.009
Recall 0.877 ± 0.007 0.865 ± 0.009 0.827 ± 0.014 0.823 ± 0.009 0.774 ± 0.009 0.651 ± 0.018
F1 0.896 ± 0.005 0.881 ± 0.004 0.805 ± 0.008 0.874 ± 0.005 0.842 ± 0.006 0.738 ± 0.010

3.2 Concept Application and Validation

The results have shown that combining features extracted from mutation and
gene network data improves the predictive potential of the models induced by
machine learning algorithms. However, these results were obtained using the
labels defined based on the NCG and the CGC and, therefore, there could still
be open questions regarding the application of the models on the prediction for
currently unlabeled genes. Thus, the models were applied to a new set of driver
genes recently reported by the IntOGen8 (Integrative Onco Genomics) [20].
A subset of 131 driver genes from IntOGen that was also contained in the
unlabeled portion of DSCOM B consisted of a new data set for the validation.
The previously trained SVM and RF models were applied to classify this new
set of genes. Data were processed following the same steps applied to create the
set DSCOM BL . The SVM model detected 12 possible false-drivers (F D) genes
and 119 drivers (D), while the RF identified 15 F D and 116 D.
The consistency of these findings was further investigated based on current
literature reports using CarcerMine9 , a literature-mined database of drivers [18].
8
www.intogen.org/.
9
bionlp.bcgsc.ca/cancermine/ – query performed in October, 2020.
Machine Learning Approach for False-Positive Cancer Driver Gene Discovery 89

Support Vector Machine (SVM) Random Forest (RF)

Fig. 2. ROC curve comparison

CancerMine extracts literature evidence of cancer genes, classifying them as

drivers, oncogenes, and tumor suppressors genes. Among the 131 genes, 78 genes
(59.5%) have been reported as cancer genes in at least one research paper.
Figure 3 shows the frequency of citations found in CancerMine for genes
classified as D or F D by both models. For each model, the figure presents the
frequencies taking a threshold of at least one, three, and ten citations as drivers,
oncogenes, and tumor suppressors genes in research papers. Genes classified as D
were reported as cancer genes more often than the ones detected by the models
as F D. At higher thresholds, there are no longer citations of F D genes, while the
D genes are still fairly cited. This difference between the frequency in citations
for genes classified as D and F D is literature-based evidence that the models,
in fact, identified genes that are likely to be false-positive-drivers. Finally, when
comparing the two models’ results, the SVM findings were in higher agreement
with the literature.

Fig. 3. Frequency of citations of the classified genes and cancer genes.

90 J. F. Cutigi et al.

4 Discussion and Conclusion

This work proposed a machine learning approach to detect potential false-drivers
cancer genes. The natural application of this discovery tool is to avoid the mis-
classification of false-positive-drivers as drivers and possibly eliminate unnec-
essary further analysis. Detecting false-drivers is also crucial to prevent their
inclusion in data analyses or on the development of models, which could lead to
the identification of unrealistic patterns. For this task, ten measures from muta-
tion data and ten from gene interactions were extracted, and Support Vector
Machines and Random Forest models were induced using the combined source of
features. Data were properly preprocessed, and stratified k-fold cross-validation
was applied to the models’ training. Hyper-parameters-optimization was also
conducted. In general, both models achieved satisfactory classification perfor-
mance, benefited by the combination of mutation and gene interaction features.
Furthermore, the discovery over a new set of genes recently reported as drivers
were in agreement with the literature, based on a citation frequency analysis.
However, it is important to note that this concept has been implemented con-
sidering the currently available data, which is scarce and still under continuous
investigation. Therefore, it is expected that the proposed tool can be eventually
revisited and improved as new information becomes available.

Supplementary Information. Source codes, scripts of experiments and the

complete list of libraries and versions used in this work are available on the
following link: https://github.com/jcutigi/FalseDriverDiscovery BSB2020.

References
1. Bailey, M.H., et al.: Comprehensive characterization of cancer driver genes and
mutations. Cell 173(2), 371–385.e18 (2018). https://doi.org/10.1016/j.cell.2018.
02.060
2. Cerami, E., et al.: The cBio cancer genomics portal: an open platform for explor-
ing multidimensional cancer genomics data. Cancer Discov. 2(5), 401–404 (2012).
https://doi.org/10.1158/2159-8290.CD-12-0095
3. Cho, A., Shim, J.E., Kim, E., Supek, F., Lehner, B., Lee, I.: MUFFINN: cancer
gene discovery via network analysis of somatic mutation data. Genome Biol. 17(1),
129 (2016). https://doi.org/10.1186/s13059-016-0989-x
4. Collier, O., Stoven, V., Vert, J.P.: LOTUS: a single- and multitask machine learning
algorithm for the prediction of cancer driver genes. PLoS Comput. Biol. 15(9), 1–27
(2019). https://doi.org/10.1371/journal.pcbi.1007381
5. Cutigi, J.F., Evangelista, A.F., Simao, A.: Approaches for the identification
of driver mutations in cancer: a tutorial from a computational perspective.
J. Bioinform. Comput. Biol. 18(03), 2050016 (2020). https://doi.org/10.1142/
S021972002050016X. pMID: 32698724
6. Cutigi, J.F., Evangelista, A.F., Simao, A.: GeNWeMME: a network-based com-
putational method for prioritizing groups of significant related genes in cancer.
In: Kowada, L., de Oliveira, D. (eds.) BSB 2019. LNCS, vol. 11347, pp. 29–40.
Springer, Cham (2020). https://doi.org/10.1007/978-3-030-46417-2 3
Machine Learning Approach for False-Positive Cancer Driver Gene Discovery 91

7. Das, J., Yu, H.: HINT: high-quality protein interactomes and their applications in
understanding human disease. BMC Syst. Biol. 6, 92 (2012). https://doi.org/10.
1186/1752-0509-6-92
8. DeLong, E.R., DeLong, D.M., Clarke-Pearson, D.L.: Comparing the areas under
two or more correlated receiver operating characteristic curves: a nonparametric
approach. Biometrics 44, 837–845 (1988)
9. Dimitrakopoulos, C.M., Beerenwinkel, N.: Computational approaches for the iden-
tification of cancer genes and pathways. Wiley Interdisc. Rev.: Syst. Biol. Med.
9(1), e1364 (2017). https://doi.org/10.1002/wsbm.1364
10. Gumpinger, A.C., Lage, K., Horn, H., Borgwardt, K.: Prediction of cancer
driver genes through network-based moment propagation of mutation scores.
Bioinformatics 36(Supplement 1), i508–i515 (2020). https://doi.org/10.1093/
bioinformatics/btaa452
11. Han, Y., et al.: DriverML: a machine learning algorithm for identifying driver genes
in cancer sequencing studies. Nucleic Acids Res. 47(8), e45–e45 (2019)
12. Horn, H., et al.: NetSig: network-based discovery from cancer genomes. Nat. Meth-
ods 15, 61–66 (2018). https://doi.org/10.1038/nmeth.4514
13. Hristov, B.H., Singh, M.: Network-based coverage of mutational profiles reveals
cancer genes. Cell Syst. 5(3), 221–229 (2017)
14. Jassal, B., et al.: The reactome pathway knowledgebase. Nucleic Acids Res. 48(D1),
D498–D503 (2020)
15. Keshava Prasad, T.S., et al.: Human protein reference database-2009 update.
Nucleic Acids Res. 37(Database issue), D767–D772 (2009). https://doi.org/10.
1093/nar/gkn892
16. Kim, Y., Cho, D., Przytycka, T.M.: Understanding genotype-phenotype effects
in cancer via network approaches. PLoS Comput. Biol. 12(3), e1004747 (2016).
https://doi.org/10.1371/journal.pcbi.1004747
17. Leiserson, M.D.M., et al.: Pan-cancer network analysis identifies combinations of
rare somatic mutations across pathways and protein complexes. Nat. Genet. 47(2),
106–114 (2015). https://doi.org/10.1038/ng.3168
18. Lever, J., Zhao, E.Y., Grewal, J., Jones, M.R., Jones, S.J.: CancerMine: a
literature-mined resource for drivers, oncogenes and tumor suppressors in cancer.
Nat. Methods 16(6), 505–507 (2019)
19. Luck, K., et al.: A reference map of the human binary protein interactome. Nature
580, 1–7 (2020)
20. Martı́nez-Jiménez, F., et al.: A compendium of mutational cancer driver genes.
Nat. Rev. Cancer 20, 1–18 (2020)
21. Oldham, S., Fulcher, B., Parkes, L., Arnatkeviciute, A., Suo, C., Fornito, A.:
Consistency and differences between centrality measures across distinct classes of
networks. PLoS ONE 14(7), 1–23 (2019). https://doi.org/10.1371/journal.pone.
0220061
22. Ozturk, K., Dow, M., Carlin, D.E., Bejar, R., Carter, H.: The emerging potential
for network analysis to inform precision cancer medicine. J. Mol. Biol. 430(18),
2875–2899 (2018)
23. Pedregosa, F., et al.: Scikit-learn: machine learning in Python. J. Mach. Learn.
Res. 12, 2825–2830 (2011)
24. Repana, D., et al.: The Network of Cancer Genes (NCG): a comprehensive cat-
alogue of known and candidate cancer genes from cancer sequencing screens.
Genome Biol. 20(1), 1 (2019). https://doi.org/10.1186/s13059-018-1612-0
92 J. F. Cutigi et al.

25. Reyna, M.A., Leiserson, M.D.M., Raphael, B.J.: Hierarchical HotNet: identify-
ing hierarchies of altered subnetworks. Bioinformatics 34(17), i972–i980 (2018).
https://doi.org/10.1093/bioinformatics/bty613
26. Sondka, Z., Bamford, S., Cole, C.G., Ward, S.A., Dunham, I., Forbes, S.A.: The
COSMIC cancer gene census: describing genetic dysfunction across all human can-
cers. Nat. Rev. Cancer 18(11), 696–705 (2018)
27. Tamborero, D., et al.: Comprehensive identification of mutational cancer driver
genes across 12 tumor types. Sci. Rep. 3, 2650 (2013)
28. Vandin, F., Upfal, E., Raphael, B.J.: Algorithms for detecting significantly mutated
pathways in cancer. J. Comput. Biol. 18(3), 507–522 (2011). https://doi.org/10.
1089/cmb.2010.0265. pMID: 21385051
29. Weinstein, J.N., et al.: The cancer genome atlas pan-cancer analysis project. Nat.
Genet. 45(10), 1113 (2013)
 Unraveling the Role of Nanobodies Tetrad
on Their Folding and Stability Assisted
by Machine and Deep Learning Algorithms

Matheus Vitor Ferreira Ferraz1,2 , Wenny Camilla dos Santos Adan2 ,

and Roberto Dias Lins1,2(B)
1 Department of Fundamental Chemistry, Federal University of Pernambuco,
Recife, PE 50670-560, Brazil
2 Department of Virology, Aggeu Magalhães Institute, Oswaldo Cruz Foundation,
Recife, PE 50670-420, Brazil
[email protected]

Abstract. Nanobodies (Nbs) achieve high solubility and stability due to four con-
served residues referred to as the Nb tetrad. While several studies have highlighted
the importance of the Nbs tetrad to their stability, a detailed molecular picture of
their role has not been provided. In this work, we have used the Rosetta package to
engineer synthetic Nbs lacking the Nb tetrad and used the Rosetta Energy Func-
tion to assess the structural features of the native and designed Nbs concerning
the presence of the Nb tetrad. To develop a classification model, we have bench-
marked three different machine learning (ML) and deep learning (DL) algorithms
and concluded that more complex models led to better binary classification for
our dataset. Our results show that these two classes of Nbs differ significantly in
features related to solvation energy and native-like structural properties. Notably,
the loss of stability due to the tetrad’s absence is chiefly driven by the entropic
contribution.

Keywords: Camelid antibodies · Rosetta Energy Function · Machine learning

1 Introduction
Ever since their discovery, single-domain binding fragment of heavy-chain camelid anti-
bodies [1], referred to as nanobodies (Nbs), have gained considerable attention in transla-
tional research as therapeutic and diagnostic tools against human diseases and pathogens
[2]. Along with its small size (15 kDa) and favorable physical-chemical properties (e.g.,
thermal and environmental stabilities), Nbs display binding affinities equivalent to con-
ventional antibodies (cAbs) [1, 3]. Moreover, its heterologous expression in bacteria
allows overcoming cAbs production pitfalls, such as high production cost and need of
animal facility [4, 5]. Hence, Nbs are considered as a promising tool against numerous
diseases. A variety of Nbs is currently being investigated under pre-clinical and clinical
stages against a wide range of viral infections [6, 7].

© Springer Nature Switzerland AG 2020

J. C. Setubal and W. M. Silva (Eds.): BSB 2020, LNBI 12558, pp. 93–104, 2020.
https://doi.org/10.1007/978-3-030-65775-8_9
94 M. V. F. Ferraz et al.

The general structural topology of the Nbs is depicted in Fig. 1. It is characterized by

a core structure composed of a pair of β-sheets, built from 4 and 5 antiparallel β-strands
linked by loops and a disulfide bridge. In contrast to cAbs, which contains six variable
loops, Nbs display three highly variable loops H1, H2, and H3. These loops correspond
to the Complementary Determining Region (CDR), which is responsible for antigenic
binding and recognition, hence providing the target specificity of the Nbs. The overall
structure of the Nbs is maintained by four conserved portions, termed as the framework.
A significant difference regarding Nbs and cAbs arises from the lack of the variable light
chain, and as a consequence, the light-heavy domains interface. To compensate for this
loss, four highly conserved residues referred to as the Nb Tetrad are found to replace
the nonpolar side chains with polar ones [8, 9]. The Nb tetrad comprises the residues
Y/P37, E44, R/C45 and G47. Presumably, these substitutions increase hydrophilicity
and solubility of the Nbs, being crucial for their stability [10].

Fig. 1. Cartoon representation of the overall topology of an Nb (PDB ID: 3DWT) [11]. The Nb
domain consists of 9 β-sheets linked by loop regions, 3 of these constitute the CDR region and
are colored in green, blue, and red. The framework region separated by the hypervariable loops
are colored in silver. The Nb tetrad residues are highlighted in yellow. (Color figure online)

These four residues’ presence is a hallmark characteristic of Nbs as it has been shown
by several sequence alignments studies [12, 13]. The high conservation of these residues
indicates an evolutionary-driven constraint, and it highlights their pivotal role in Nbs
structure. To ascertain that changes in the Nb tetrad would negatively impact the Nb
folding, we have previously designed a Nb by altering the tetrad residues. The obtained
chimera presented low expression yields and the absence of a well-defined globular
three-dimensional structure due to aggregation (unpublished data). On the contrary,
attempts to “camelize” human/murine Abs by grafting the Nb tetrad to the Ab heavy
chain’s corresponding position has resulted in structural deformations of the framework
β-sheet, leading to scarce stability and aggregation [14]. Although it has been described
a phage-display library derived from llamas that has produced a set of stable and soluble
Nbs devoid of the Nb tetrad [15], these Nbs are unusual, and their stability should be
explained in the light of an alternative mechanism.
 Unraveling the Role of Nanobodies Tetrad 95

Given the Nb’s tetrad importance in maintaining its folded structure and stability,
these residues can be considered key to engineer novel Nbs. It has been shown that molec-
ular dynamics simulations are not sufficient to capture the lower stability of aggregating
Nbs, and it does not elucidate the structural and thermodynamic features role of the Nb
tetrad to the structures of the Nbs [16]. In this study, we seek to identify the impact
of the Nb tetrad from a molecular perspective. To gain insight into the thermodynamic
contributions to the folded Nbs, we have used the Rosetta Energy Function compo-
nents combined with machine learning to identify whether there are differences in the
structural pattern of natural Nbs and the corresponding Nbs without the presence of
the tetrad sidechains, by replacing them for methyl groups (Alanine). We benchmarked
two machine learning (ML) models (Support Vector Machine [17] and Random Forest
[18]) and one deep learning (DP) model (Artificial neural network by Multilayer Per-
ceptron [19]) to evaluate the performance of these algorithms in effectively capture the
differences among the classes from the multivariate nature of the data.

2 Computational Details
2.1 Dataset Preparation
A total of 30 non-redundant X-ray derived Nb structures, with resolution lower than 3 Å,
were retrieved from the Protein Data Bank (PDB). To alleviate bad atomic contacts, the
nearest local minimum in the energy function was achieved by geometry-minimizing
their initial coordinates using the Rosetta package v. 3.10 [20] and the linear-Broyden-
Fletcher-Goldfarb-Shanno minimization flavor conditioned to the Armijo-Goldstein
rule. The minimization protocol was carried out in a stepwise fashion, where the
sidechain angles were initially geometry-minimized, followed by full rotamer pack-
ing and minimization of the orientation of sidechain, backbone, and rigid body. To
enhance sampling, χ1 and χ2 rotamers angles were used for all residues that pass an
extra-chi cutoff of 1. Hydrogen placement was optimized during the protocol. For each
of the minimized structures, the Nb tetrad residues were identified and replaced by ala-
nine using the RosettaScripts in the four positions, and the obtained structures were
geometry-minimized accordingly to the previously described protocol. Thus, the final
dataset consisted of 60 instances.
To evaluate folding propensity, the Nb structures were scored using the all-atom
Rosetta Energy Function 2015 (REF2015) [21] to calculate the energy of all atomic
interactions within the proteins. The REF2015 possesses 20 terms and these were used
as the features. The terms can be found in the GitHub (https://github.com/mvfferraz/
NanobodiesTetrad), and a detailed description of each term can be found in reference
[21]. The score function is a model parametrized to approximate the energy for a given
protein conformation. Thus, it consists of a weighted sum of energy terms expressed as
mathematical functions based on fundamental physical theories, statistical-mechanical
models, and protein structures observations. The Rosetta package is a state-of-art prime-
tool to the modeling and design of proteins, and its empirical energy function successfully
allows for a valid assessment of the relative thermodynamic stability of folded proteins.
The weights for each energy term were kept as default. The parsed command lines, PDB
codes, and dataset are available in the GitHub.
96 M. V. F. Ferraz et al.

2.2 Classification Methods

All the algorithms were written using Python v. 3, and the Scikit Learn Library [22] was
employed in conjunction with the Pandas [23] and Numpy [24] packages. In addition,
Tensorflow [25] and Keras [26] were used for the NN algorithm. The dataset was split
as training (70%) and test (30%) sets. The data features vector was standardized using
preprocessing tools to normally distribute the data by scaling the data to a zero mean
and unit variance. Details of the code can be found in the GitHub.

Linear Discriminant Analysis (LDA). To identify whether the folding propensity of

the Nbs containing the Nb tetrad and those that do not, are linearly separable, a one-
component LDA was carried out [27]. LDA projects the input data to a linear subspace
constituted of directions to maximize the separation between the classes and minimize
the separation among a class. Bayes’ statistics are applied to fit conditional class densities
for each sample of the data. To select the significant variables, the ensemble learning
method of extremely randomized trees (Extra Tree) classifier [28] was used. The number
of estimators was kept as 100. The number of features to consider when searching for
the best split was assigned as 2, and the quality of a split was measured using the entropy
criterion. The LDA was solved using eigenvalue decomposition and was performed by
fitting the data and then transforming it without additional parameters. The weights of
the LD were used to detect which features are responsible for separating the classes
explicitly. To compare if two means were statically different, two-tailed paired t-test
was used (GraphPad Prism 8 [29]). Differences were considered statistically significant
for a p-value such that p < 0.05, at the 95% confidence level.

Support Vector Machine (SVM). SVM consists of a non-probabilistic binary linear

classifier, and wherein classification is performed by the construction of a set of hyper-
planes in a high-dimensional space. SVM seeks to find a line of separation between the
hyperplanes from each class. This line is optimally drawn for maximizing the distance
between the closest points regarding each class. C-Support Vector Classification (SVC)
was used with a linear Kernel with C = 1 hyperparameter, identified with a grid-search
over pre-defined values for C (0.001, 0.01, 0.1, 1, 10, 100) and different types of Kernel
(Linear and Radial basis function). The linear Kernel, K, is defined as a function of the
vectors in the input space, x and y, as K(x, y) = xT y, for x, y ∈ Rd .

Random Forest (RF). RF is a meta estimator that builds a number of decision trees on
bootstrapped training samples and uses averaging from random samples for each split in
a tree. All parameters were implemented as the default, save by the criterion to measure
the split’s quality, set as entropy envisioning information gain.

Neural Network (NN). TensorFlow library was used in conjunction with the Keras
high-level application programming interface. The classification was performed using
the Multi-layer Perceptron (MLP) Classifier with 100 hidden layers. An MLP is a feed-
forward artificial NN class, which learns a function f (·) : Rm → Rn by training on a
dataset with m input dimensions and n output dimensions, and it contains hidden layers
in between the input and output layer. Each hidden layer contains a weight propagated
for each posterior layer as a weighted linear summation and followed by a non-linear
 Unraveling the Role of Nanobodies Tetrad 97

activation function g(·) : R → R. The weight optimization was conducted by stochastic

gradient descent, and the step-size for updating the weights was defined as 0.01. Max-
imum iterations number was set as 500, or until it reaches convergence by considering
the default tolerance.

Diagnostic Performance Evaluation. Four performance measures were assessed. The

accuracy of the models was computed using the 10-fold cross-validation. To verify the
model’s performance, the confusion matrix, and the Receiver Operating Characteris-
tic (ROC) curve were evaluated along with the models’ learning curve. For a binary
classification task, precision, recall, and f1-score are defined according to the assigned
classification (true positive (tp), true negative (tn), false positive (fp) and false negative
(fn)) as described by Eqs. 1–3. For a detailed description of each metric, see [30].

tp
Precision = (1)
tp + fp
tp
Recall = (2)
tp + fn
2 tp
f1 = = (3)
recall−1 + precision −1
tp + 2 (fp + fn)
1

3 Results and Discussion

3.1 Features Selection

The Rosetta energy terms are convenient mathematical approximations to the physics
that governs protein structure and stability. The Rosetta Energy Function (REF) ranks
the relative fitness of several amino acid sequences for a given protein structure, and it
is capable of predicting the threshold for protein stability by discriminating native-like
from non-native structures in a decoy [31]. The functional form relies upon pairwise
decomposability of energy terms. The decomposition limits the number of energetic
contributions to 1/2N (N − 1), where N is the atom’s number in the system.
When using Rosetta energy function to calculate the score of a protein, i.e., the relative
energy for a given conformation reasoned by specific parameters of the Hamiltonian, it
yields a total of 20 energetic terms [21]. A feature selection was performed to reduce the
effects of noise or irrelevant variables to construct the models. A feature was considered
relevant and non-redundant if it presented a feature importance score higher than 0.05
(Fig. 2). From the obtained split, a total of 7 features were filtered:

• fa_dun: the probability of a given rotamer is a native-like state based on Dunbrack’s

statistics for a given φ and ψ angles;
• hbond_sc: energy for the sidechain-sidechain hydrogen bond;
• lk_ball_wtd: asymmetric solvation energy;
98 M. V. F. Ferraz et al.

Fig. 2. Feature selection based on REF terms using Extra trees classifier. Feature importance
greater than 0.05 was regarded as a relevant feature. The red dashed line represents the threshold
for a given feature to be filtered. (Color figure online)

• fa_intra_sol_xover4: Lazaridis-Karplus solvation energy for intra-residue interac-

tions;
• fa_sol: Lazaridis-Karplus solvation energy model based on Gaussian exclusion;
• ref : An approximation to the relative energies of the unfolded-state ensembles;
• fa_elec: Coulombic electrostatic potential.

The short descriptions of the terms were retrieved from [21]. As can be seen, almost
half of the selected terms are related to the system’s solvation properties. Since the
replacement of hydrophilic residues for alanine increases the hydrophobic content of
the Nbs lacking Nb tetrad, these structural differences have potentially been captured
by the REF. These observations corroborate the well-described importance of the Nbs
tetrad for solubility.

3.2 Linear Separability of the Data

LDA was used to evaluate whether the filtered features’ combination can discriminate
natural Nbs from Nbs lacking Nb tetrad. Since we have two classes, the LDA was
performed in a one-dimensional fashion. LDA is a supervised dimensionality reductor
that identifies the attributes that mostly account for the classes’ variance. From fitting
a Gaussian density to each class, a single LD was able to separate the class linearly.
Figure 3A shows the one-dimensional separability for the classes. In general, natural
Nbs lead to a negative value for the LD, and the contrary is observed for the Nbs that
lack the Nb tetrad.
To investigate the features that account for the most separation between the classes,
the LD loadings were assessed. The loadings indicate the contribution of each feature in
predicting class assignment and are shown in Fig. 3B. A higher weight (relatively to their
modulus) are fa_dun, lk_ball_wtd, and ref . These results highlight the importance of the
Nbs tetrad to solubility and stability of Nbs. The lk_ball_wtd consists of the orientation-
dependent solvation of polar atoms when assuming the ideal water geometry. As already
stated, the REF was able to capture the solvation contribution to the Nb tetrad presence.
 Unraveling the Role of Nanobodies Tetrad 99

Fig. 3. Discriminant features assessed by LDA. (A) One dimensional LDA. The bars in red are
the LD values for the Nbs lacking the Nb tetrad whereas the bars in blue consist of the natural
Nbs; (B) Loading of each feature used to calculate de LDs; (C–E) ref energy, lk_ball_wdt, fa_dun
energy terms, respectively, for each class. The p-value < 0.05 indicates significant differences in
the means. (Color figure online)

Moreover, the other two features are related to native-like conformations properties.
Thus, these results show that the Nb tetrad potentially impacts Nbs’ backbone φ and
ψ angles distribution as it is found in the Dunbrack’s library of rotamers. Given the
importance of the torsion angles for protein folding, a putative explanation for the Nbs
tetrad’s role in maintaining the structure arises from a geometrical issue. It must be noted
that regarding the geometric features, this effect is unlikely to be an artifact from the
modeling, since the replacement by alanine residues are not expected to cause significant
structural changes, due to the small size of its sidechain, it can be positioned and matched
for in any part of the protein (except when replacing tightly buried glycine residues).
To ensure the average of these features were statistically different between the classes,
a two-tailed t-test was used. Figure 3(C–E) shows the distribution of the data for each
class, along with averages and standard deviations. All three features presented a p-value
< 0.05, indicating that these averages are statistically different.

3.3 Classification Algorithms

Given that the selected features can discriminate between the natural Nbs and Nbs that
lack the Nb tetrad, classification algorithms were employed to compare the different
performance in capturing the classes’ structural differences. In this benchmark, machine
100 M. V. F. Ferraz et al.

learning (SVM and RF) and deep learning (ANN-MLP) were assessed regarding their
binary classification performance. SVM is an instance-based learning model, and RF is
an ensemble method. MLP is a class of NN, and here it has been employed more than
three hidden-layers, and therefore, consists of a DL approach.
All the models have been prepared with the same data and training set. All the
20 features were taken into account to carry out the classifications, since using the
selected features from extra trees classifier resulted in poor performances (Data not
shown for conciseness). Since it is a small dataset, it is prone to suffer from overfitting
the data (high variance). Thus, we performed several performance evaluations. Initially,
the models were compared regarding their threshold metrics. Threshold metrics are
useful for diagnosing classification prediction errors. Initially, the scores (Fig. 4A),
which are directly associated with a combination of the precision and the recall values,
were calculated using two approaches: 1) Evaluation was performed considering the
initial training/test set; 2) A 10-fold cross-validation was employed. In the latter flavor
of evaluating the estimator performance, the training set is split into k sets, and the metrics
are calculated in a loop for the different generated sets. The performance is then measured
by the average of each k-fold cross-validation. The SVM model presented a remarkable
performance in properly assigning the classes, with an accuracy of 0.94 for the initial test
set, and an accuracy average of 0.80 when considering ten different subsets. Followed
by SVM, MLP also presented good metrics, even though with a slightly lower value.
From the three models, the one with the poorest metrics was the RF algorithm. The two
formers are more complex and robust models, so that the classification task is likely not
trivial, in such a way, a simpler algorithm will not capture the main differences between
the classes. The algorithms were compared using the confusion matrix (Fig. 4B–D).
The diagonal elements of the matrixes consist of the number of true label classification,
whereas, off-diagonal elements represent the mislabeled classifications. The SVM and
MLP algorithms outperformed the RF model. The performance metrics are summarized
in Table 1 and demonstrate the SVM and MLP algorithms’ efficacy for our dataset.
To identify how much the models can benefit from adding more data, learning curves
were plotted. Two learning curves were constructed: 1) Train learning curve: calculated
based on the training set and diagnosis how well the model is learning, and 2) Validation
learning curve: calculated based on a hold-out validation set and diagnosis how well the
model is generalizing. Figure 5D–F shows the learning curve for the models. For SVM,
the training curve modestly decreases as more samples are added, and the learning curve

Table 1. Threshold performance metrics for each binary classification model

Model Nb Precision Recall F1-score Instances Accuracy

SVM Natural 0.90 1.00 0.95 9 0.94
 Tetrad 1.00 0.89 0.94 9
RF Natural 0.83 0.56 0.67 9 0.72
 Tetrad 0.67 0.89 0.76 9
MLP Natural 0.82 1.00 0.90 9 0.89
 Tetrad 1.00 0.78 0.88 9
 Unraveling the Role of Nanobodies Tetrad 101

Fig. 4. Threshold metrics for the models’ performance (A) Validation scores for the SMV, RF and
MLP models considering the accuracy for the train/test split and for the 10-fold cross-validation;
(B–D) Confusion matrix for SVM, RF, and MLP

increases until reaching a plateau at a score of nearly 0.80. As can be seen, the model
fits the data well, but its generalization has a slightly lower value for the score. Thus, the
SVM model might be slightly overfitted. However, its learning capability is progressively
increased as more samples are added, indicating that the set number is small. For MLP,
a similar trend is observed. However, for the same number of samples, SVM acquires
a higher score for the learning curve, suggesting a better model’s performance. These
results indicate that one source of difficulty for classifying using this dataset resides in
the small number of samples.
Furthermore, it shows that the algorithm’s training and learning process is not
straightforward, given that MLP presents a higher score for the training, proposing that
the more complex fitting to the data is required. The RF model did not reflect sensitivity
to increasing the number of samples, and a decrease in the learning curve is observed.
Thus, the RF model does not benefit from increasing the dataset, and its overfitting can-
not be attributed solely to the small size of the dataset, but rather to the simplicity of the
algorithm over a complex classificatory task.
These information show that SVM and MLP have the potential to classify between
the classes. Such a model is of fundamental relevance for a myriad of protein design
algorithms that rely on Monte Carlo sampling. Since a large number of decoys are usually
generated, identifying the Nbs that possess native-like characteristics is of enormous
advantage to time and resources saving for experimental characterization. From our
benchmarking, the RF model is not a proper model to learn from the data. Besides SVM
102 M. V. F. Ferraz et al.

Fig. 5. Assessment of the models’ performance through their characteristic curves. (A–C) ROC
curves for SVM, RF, and MLP models, respectively; (D–F) Learning curves and the validation
score as function of the number of training examples for SVM, RF, and MLP, respectively

having a slight advantage over MLP, the latter is a promising alternative since its training
curve perfectly fits the training data, and its increasing learning curve is a promising
indicator of its potential. The SVM presented a satisfactory performance, and from
searching for different parameters combination, a considerable gain in the predictivity
capacity might be observed. The ML and DL algorithms’ performance confirms that
there are traits that allow for the discrimination of Nbs containing the tetrad or not.
Our results show that abolishing the tetrad associates with loss of folding stability in
agreement with literature data. It is captured by the ref term, which in turn is shown to
have significant contributions from the solvation energy and torsional dihedral motion
terms. Therefore, the loss of Nbs stability due to eliminating the tetrad is a mostly
entropic-driven phenomenon.

4 Conclusions
We have compared the structural features, calculated by the REF’s energy term, of
natural Nbs containing the Nbs tetrad and a synthetic set of Nbs lacking the tetrad. Data
mining analyses revealed that the two classes of nanobodies differ mainly by folding and
solvation features, corroborating with previous studies suggesting the tetrad’s importance
for stability and solubility. This work’s findings expand the knowledge on the impact
of the Nbs tetrad from a molecular-level perspective by highlighting the importance of
entropic contributions to their stability.

Acknowledgements. This work has been funded by FACEPE, CAPES, CNPq, and FIOCRUZ.
We acknowledge the LNCC for the availability of resources and support.
 Unraveling the Role of Nanobodies Tetrad 103

References
1. Muyldermans, S.: Nanobodies: natural single-domain antibodies. Annu. Rev. Biochem. 82,
775–797 (2013)
2. Mir, M.A., Mehraj, U., Sheikh, B.A., Hamdani, S.S.: Nanobodies: The “Magic Bullets” in
therapeutics, drug delivery and diagnostics. Hum. Antib. 28, 29–51 (2020)
3. Vincke, C., Muyldermans, S.: Introduction to heavy chain antibodies and derived Nanobodies.
Methods Mol. Biol. 911, 15–26 (2012)
4. Morrison, C.: Nanobody approval gives domain antibodies a boost. Nat. Rev. Drug. Discov.
18, 485–487 (2019)
5. Jovčevska, I., Muyldermans, S.: The Therapeutic potential of Nanobodies. BioDrugs 34(1),
11–26 (2019). https://doi.org/10.1007/s40259-019-00392-z
6. Beghein, E., Gettemans, J.: Nanobody technology: A versatile toolkit for microscopic imag-
ing, Protein–Protein interaction analysis, and protein function exploration. Front. Immunol.
8, 771 (2017)
7. Konwarh, R.: Nanobodies: Prospects of expanding the Gamut of neutralizing antibodies
against the novel coronavirus, SARS-CoV-2. Front. Immunol. 11, 1531 (2020)
8. Revets, H., De Baetselier, P., Muyldermans, S.: Nanobodies as novel agents for cancer therapy.
Expert. Opin. Biol. Ther. 5, 111–124 (2005)
9. Muyldermans, S.: Single domain camel antibodies: Current status. J. Biotechnol. 74, 277–302
(2001)
10. Barthelemy, P.A., et al.: Comprehensive analysis of the factors contributing to the stability
and solubility of autonomous human VH domains. J. Biol. Chem. 283, 3639–3654 (2008)
11. Vincke, C., Loris, R., Saerens, D., Martinez-Rodriguez, S., Muyldermans, S., Conrath, K.:
General strategy to humanize a camelid single-domain antibody and identification of a
universal humanized nanobody scaffold. J. Biol. Chem. 284, 3273–3284 (2009)
12. Mitchell, L.S., Colwell, L.J.: Comparative analysis of nanobody sequence and structure data.
Proteins 86, 697–706 (2018)
13. Kunz, P., et al.: Exploiting sequence and stability information for directing nanobody stability
engineering. Biochim. Biophys. Acta Gen. Subj. 1861, 2196–2205 (2017)
14. Rouet, R., Dudgeon, K., Christie, M., Langley, D., Christ, D.: Fully human VH single domains
that rival the stability and cleft recognition of camelid antibodies. J. Biol. Chem. 290, 11905–
11917 (2015)
15. Tanha, J., Dubuc, G., Hirama, T., Narang, S.A., MacKenzie, C.R.: Selection by phage dis-
play of llama conventional V(H) fragments with heavy chain antibody V(H)H properties. J.
Immunol. Methods 263, 97–109 (2002)
16. Soler, M.A., de Marco, A., Fortuna, S.: Molecular dynamics simulations and docking enable
to explore the biophysical factors controlling the yields of engineered nanobodies. Sci. Rep.
6, 34869 (2016)
17. Hearst, M.A., Dumais, S.T., Osuna, E., Platt, J., Scholkopf, B.: Support vector machines.
IEEE Intell. Syst. Appl. 13, 18–28 (1998)
18. Breiman, L.: Random forests. Mach. Learn. 45, 5–32 (2001)
19. Pal, S.K., Mitra, S.: Multilayer perceptron, fuzzy sets, classifiaction. IEEE. Trans. Newural.
Netw. 3(5), 683–697 (1992)
20. Leaver-Fay, A., et al.: ROSETTA3: An object-oriented software suite for the simulation and
design of macromolecules. Methods Enzymol. 487, 545–574 (2011)
21. Alford, R.F., et al.: The Rosetta all-atom energy function for macromolecular modeling and
design. J. Chem. Theory Comput. 13, 3031–3048 (2017)
22. Pedregosa, F., et al.: Scikit-learn: Machine learning in Python. J. Mach. Learn. Res. 12,
2825–2830 (2011)
104 M. V. F. Ferraz et al.

23. McKinney, W.: Data structures for statistical computing in python. In: Proceedings of the 9th
Python in Science Conference, pp. 56–61. Austin (2010)
24. Harris, C.R., et al.: Array programming with NumPy. Nature 585, 357–362 (2020)
25. Abadi, M., et al.: Tensorflow: Large-scale machine learning on heterogeneous distributed
systems. arXiv preprint arXiv:1603.04467 (2016)
26. Gulli, A., Pal, S.: Deep learning with Keras. Packt Publishing Ltd, Birmingham (2017)
27. Fisher, R.A.: The use of multiple measurements in taxonomic problems. Annals Eugen. 7,
179–188 (1936)
28. Geurts, P., Ernst, D., Wehenkel, L.: Extremely randomized trees. Mach. Learn. 63, 3–42
(2006)
29. Prism, G.: Graphpad software. San Diego, CA, USA (1994)
30. Powers, D.M.: Evaluation: From precision, recall and F-measure to ROC, informedness,
markedness and correlation. J. Mach. Learn. Technol. 2, 37–63 (2011)
31. Cunha, K.C., Rusu, V.H., Viana, I.F., Marques, E.T., Dhalia, R., Lins, R.D.: Assessing protein
conformational sampling and structural stability via de novo design and molecular dynamics
simulations. Biopolymers 103, 351–361 (2015)
 Experiencing DfAnalyzer for Runtime
Analysis of Phylogenomic Dataflows

Luiz Gustavo Dias1(B) , Marta Mattoso2 , Bruno Lopes1 ,

and Daniel de Oliveira1
1
Fluminense Federal University, IC/UFF, Niterói, Brazil
[email protected], {bruno,danielcmo}@ic.uff.br
2
Federal University of Rio de Janeiro, COPPE/UFRJ, Rio de Janeiro, Brazil
[email protected]

Abstract. Phylogenomic experiments provide the basis for evolutionary

biology inferences. They are data- and CPU-intensive by nature and aim
at producing phylogenomic trees based on an input dataset of protein
sequences of genomes. These experiments can be modeled as scientific
workflows. Although workflows can be efficiently managed by Workflow
Management Systems (WfMS), they are not often used by bioinformati-
cians, which traditionally use scripts to implement their workflows. How-
ever, collecting provenance from scripts is a challenging task. In this
paper, we specialize the DfAnalyzer tool for the phylogenomics domain.
DfAnalyzer enables capturing, monitoring, debugging, and analysing
dataflows while being generated by the script. Additionally, it can be
invoked from scripts, in the same way bioinformaticians already import
libraries in their code. The proposed approach captures strategic domain
data, registering provenance and telemetry (performance) data to enable
queries at runtime. Another advantage of specializing DfAnalyzer in the
context of Phylogenomic experiments is the capability of capturing data
from experiments that execute either locally or in HPC environments. We
evaluated the proposed specialization of DfAnalyzer using the SciPhy-
lomics workflow and the proposed approach showed relevant telemetry
scenarios and rich data analyses.

Keywords: Scientific workflow · Provenance · Dataflow analysis

1 Introduction
Over the past years, several categories of experiments in the bioinformatics
domain became more and more dependent on complex computational simula-
tions. One example are the Phylogenomic analyses that provide the basis for
evolutionary biology inferences, and they have been fostered by several tech-
nologies (e.g., DNA sequencing methods [10] and novel mathematical and com-
putational algorithms). This leads to a high submission rate of protein sequences
This work was partially supported by CNPq, CAPES, and FAPERJ.
c Springer Nature Switzerland AG 2020
J. C. Setubal and W. M. Silva (Eds.): BSB 2020, LNBI 12558, pp. 105–116, 2020.
https://doi.org/10.1007/978-3-030-65775-8_10
106 L. G. Dias et al.

to databases such as UniProt, which now contain millions of sequences [18] that
can be used in such analyses. Due to the need of processing such volume of data,
the execution of this type of experiment became data- and CPU-intensive, thus
requiring High-Performance Computing (HPC) environments to process data
and analyze results in a timely manner.
A phylogenomic analysis experiment may be modeled as a workflow [4]. A
workflow is an abstraction that allows for the user (e.g., a bioinformatician)
to compose a series of activities connected by data dependencies, thus creat-
ing a dataflow. These activities are typically legacy programs (e.g., MAFFT,
BLAST, etc). Bioinformaticians often fall back on Workflow Management Sys-
tems (WfMSs), such as Galaxy [2], Pegasus [5] and SciCumulus [12], to model
and manage the execution of these workflows, but they have to rewrite the work-
flow into their languages and restrictions to their execution environments. Thus,
despite collecting provenance data [6] (the derivation history of a data product,
starting from its original sources - e.g., a dataset containing many DNA and
RNA sequences), which is a key issue to analyze and reproduce results (and
allied with domain-specific data and telemetry data, it provides an important
framework for data analytics), WfMSs are often not used by bioinformaticians.
This way, several bioinformaticians prefer to implement their experiments
using scripts (e.g., Shell, Perl, Python) [9]. More recently, bioinformaticians
also started to explore efficient Data-Intensive Scalable Computing (DISC) sys-
tems and migrate their data- and CPU-intensive experiments to frameworks
like Apache Spark (https://spark.apache.org/), e.g., SparkBWA [1] and ADAM
(https://adam.readthedocs.io/en/latest/).
Collecting provenance data (and domain-specific data) from scripts and DISC
frameworks is challenging. There are several alternatives to WfMSs focused on
capturing and analyzing provenance from scripts and DISC frameworks that can
be applied in phylogenomic analyses [7,15]. However, they also present some lim-
itations. The first one is that some approaches require specific languages (e.g.,
noWorkflow [15] works only with Python) and specific versions of the framework
(e.g., SAMbA-RaP [7] requires a specific version of Apache Spark). Flexibility to
define the level of granularity is also an issue. In general, automatic provenance
data capturing generates fine-grained provenance, which commonly overwhelms
bioinformaticians with a large volume of undesired data to analyze (e.g., access
to files and databases). On the other hand, automatic capturing coarse-grained
provenance may not provide enough data for analysis (even existing WfMS pro-
vide non-flexible level of granularity). In addition, capturing domain-specific
and telemetry data is also an issue in these approaches. We consider that If this
integrated database (provenance, domain-specific data and telemetry data) is
available, bioinformaticians can focus on analyzing just relevant data, reproduce
the results and also observe a specific pattern to infer that something is not
going well in the script at runtime, deciding to stop it or change parameters.
To address these issues, DfAnalyzer [17] was recently proposed to provide
an agnostic way for scientists to define the granularity of the provenance data
capture. The DfAnalyzer provenance database can be queried at runtime (i.e.,
 Experiencing DfAnalyzer for Runtime Analysis of Phylogenomic Dataflows 107

during the execution of the experiment), using a W3C PROV compliant data
model. This portable and easy-to-query provenance database is also a step towards
the reproducibility of bioinformatics experiments, regardless how they are imple-
mented, since DfAnalyzer can be coupled to scripts, Spark applications and exist-
ing WfMSs. In this paper, we specialize DfAnalyzer to the context of phylogenomic
experiments and present the benefits for bioinformatics data analyses and debug-
ging. Thus, the main contribution of this paper is an extension of DfAnalyzer to
collect telemetry (performance) data and its customization for the bioinformatics
domain.
The remainder of the paper is organized as follows. Section 2 provides back-
ground concepts, describing the DfAnalyzer tool and discusses related work.
Section 3 introduces the extensions and customizations in DfAnalyzer tool and
presents the evaluations in a case study with a phylogenomic experiment. Finally,
Sect. 4 concludes the paper and points future directions.

2 Background
2.1 DfAnalyzer: Runtime Dataflow Analysis of Scientific
Applications Using Provenance
DfAnalyzer [17] is a W3C provenance compliant system that captures and stores
provenance during the execution of experiments, regardless of how they are
implemented. DfAnalyzer is based on a formal dataflow representation to regis-
ter the flow of datasets and data elements. It allows for analyzing and debugging
dataflows at runtime. One important characteristic of DfAnalyzer is that it cap-
tures only relevant data (as defined by the user), thus avoiding overloading users
with a large volume of low level data. DfAnalyzer captures “traditional” prove-
nance data (e.g., data derivation path) but also domain-specific data through
raw data extraction, e.g., a DNA sequence or the e-value. These characteristics
are essential since experiments may generate massive datasets, while only a small
sub-set of provenance and domain data is relevant for analysis [14]. The original
architecture of DfAnalyzer has five components: (i) Provenance Data Extrac-
tor (PDE); (ii) Raw Data Extractor (RDE); (iii) Dataflow Viewer (DfViewer);
(iv) Query Interface (QI); and (v) Provenance Database. The first two compo-
nents are invoked by plugging calls on the script, while the other three have
independent interfaces for the user to submit data analyses at runtime.
After deploying the DfAnalyzer, users are required to identify relevant data
in their own script, model these data and instrument the code (add DfAna-
lyzer calls) in order that DfAnalyzer automatically captures data and populates
the database. It is worth noticing that the database tables are automatically
created based on the instrumentation performed in the code. This data identi-
fication is based on the following: Dataflow: a tag to identify the dataflow that
is being captured; Transformations: The data transformations that are part of
the dataflow; Tasks: The smaller unit of processing, a transformation may be
executed by several tasks; Datasets: Group of data elements consumed by tasks
and transformations. A transformation consumes an output produced by other
108 L. G. Dias et al.

transformation. Data elements: The attributes that compose datasets. They rep-
resent either domain-specific data or input parameters of the experiment. Such
information is inserted at the beginning of the script or program. Before insert-
ing the tags on the script, it is needed to map these concepts to the script’s
dataflow. Identifiying the dataflow on the script is essential to represent the
data transformations, dependencies, and data elements that need to be stored
in the provenance database.
Listing 1 shows an example of code instrumentation in DfAnalyzer. To instru-
ment the code, we have a 3-step process: (i) import DfAnalyzer packages in the
script; (ii) define the prospective provenance (i.e., the definitions of the dataflow
and transformations—lines 2 to 17); and (iii) define the retrospective provenance
(i.e., activities and data elements to capture—lines 20 to 27). When the scripts
start running each call sends to DfAnalyzer the prospective and retrospective
data to be stored in the database. Figure 1 presents a fragment of the prove-
nance database schema. Each dataset identified in the script has an associated
table in the database. It is worth noticing that this instrumentation is per-
formed only once and the script may be executed several times after that. There
are datasets, data elements, and data transformations that are typically used in
phylogenomic workflows. To avoid this repetitive step for bioinformaticians and
allow for a consistent data representation, this work provides specialized services
for DfAnalyzer users in this domain.

Fig. 1. A fragment of the provenance database schema

 Experiencing DfAnalyzer for Runtime Analysis of Phylogenomic Dataflows 109

2.2 Related Work

Hondo et al. [8] use provenance to the analysis of the transcriptome of the fun-
gus Schizosaccharomyces pombe in four different conditions. Like in DfAnalyzer,
they also adopt the PROV-DM model [11], but use different noSQL database
systems to represent provenance data. Despite the flexibility of PROV-DM, the
data granularity level chosen in [8] is coarse-grain, just capturing the transfor-
mation name, the program used to execute it, and the transformation execution
time. Afgan et al. [2] focus on providing support in RNA sequencing experi-
ments in Galaxy WfMS for non-specialists. They encapsulate the complexity of
the environment configuration and data analysis. Although the tool supports the
design and execution of workflows, provenance capture has a fixed granularity
level and domain-specific data is not captured.
Carvalho et al. [3] propose an approach that converts code from interactive
notebooks into workflows to capture provenance and identify the dataflow. The
approach identifies transformations and tasks automatically, e.g., functions are
transformations. Although this reduces the instrumentation effort, it is depen-
dent on the organization of the code, i.e., if the programmer does not use func-
tions, the identification of the transformations may be compromised. In addi-
tion, it does not capture domain-specific data neither supports parallel scripts
(which is very common in the bioinformatics domain). Pimentel et al. [15] pro-
pose the noWorkflow tool to automatically collect provenance from the execution
of Python scripts. noWorkflow is easy to deploy, but it is specific for Python
scripts and does not support parallel executions. Pina et al. [16] also special-
ized DfAnalyzer, but they have focused on fine-tuning parameters in scripts of
110 L. G. Dias et al.

Convolutional Neural Networks. It does not help capturing provenance in the

genomics domain neither telemetry data.

3 Evaluating DfAnalyzer for Phylogenomics Experiments

This section presents the extensions implemented in DfAnalyzer to be special-
ized for the phylogenomics domain and the evaluation with the SciPhylomics
workflow [13].

3.1 Specializing DfAnalyzer

DfAnalyzer already provides ways to capture provenance and domain-specific
data, but based on previous experiments, performance data are also very impor-
tant for data analytics. Bioinformaticians often have to analyze the domain-
specific data together with the performance data to evaluate the trade-off
between producing results with high quality and the time needed to produce
such results.
Therefore, in this subsection we explain the specialization of DfAnalyzer to
capture performance data. We have added a new component to the DfAnalyzer
architecture named Telemetry Data Extractor. This component is build on top
of the psutil library (version 5.7.2). Psutil is cross-platform and leads to retrieve
performance data of running processes. It also leads to capture system resources
usage (e.g., CPU, memory, etc.). Before the execution of each task or trans-
formation, the TDE component is invoked to monitor the resource usage. The
monitoring process ends when the task or transformation finishes execution. All
performance data are stored in a specific table that is associated to the Task
table presented in Fig. 1.

3.2 SciPhylomics Workflow

SciPhylomics [13] is a phylogenomic analysis workflow that aims to construct
phylogenetic trees comparing hundreds of different genomes. SciPhylomics is
composed by nine transformations. The first four transformations are associated
to phylogenetic analysis (or gene phylogeny): (i) multiple sequence alignment
(MSA), (ii) MSA conversion, (iii) search for the best evolutionary model, and
(iv) construction of phylogenetic trees. After the execution of the gene phylogeny
activities, the data quality analysis activity is executed. This data quality anal-
ysis allows for filtering results that do not meet a quality criteria. The last four
activities represent the phylogenomic analysis (or genome phylogeny): (vi) con-
catenation of MSA to obtain a “superalignment”, (vii) election of an evolutionary
model for the “phylogenomic tree” construction based on the “superalignment”,
(viii) construction of phylogenomic trees and (ix) the phylogenomic tree election.
SciPhylomics is data- and CPU-intensive, so it requires HPC environments
to produce results in a feasible time. In this evaluation, we implemented SciPhy-
lomics in a Python script that is build on top of Parallel Python library. Parallel
 Experiencing DfAnalyzer for Runtime Analysis of Phylogenomic Dataflows 111

Python is a python module which provides mechanisms for parallel execution of

python scripts on multiple processors/cores and clusters.
As aforementioned, the definition of the dataflow is essential for identifying
capturing relevant provenance, domain-specific and telemetry data. This way,
in this evaluation we considered a fragment of Sciphylomics composed of five
transformations associated to gene phylogeny, as presented in Fig. 2: (i) Sequence
Cleaning; (ii) Sequence Alignment; (iii) Sequence Conversion (iv) Model Gener-
ation; (v) Tree Generation. It is worth noticing that the transformations of the
dataflow were defined specifically for the context of this paper, being possible
to split these transformations into smaller ones (or merging them into a bigger
one).
After defining the transformations in the script, we need to list the programs
that execute them. In this fragment of SciPhylomics, each transformation may be
implemented by several programs (i.e., a variability). The gray transformations
presented in Fig. 2 are the variant transformations, e.g., data transformation (2)
can be executed by the applications MAFFT, ClustalW, and Muscle. Similarly,
in transformation (5) the programs RAxML and MrBayes can be used.
This characteristic is interesting for the analysis phase since we can com-
pare the results and performance of different programs that implement the same
transformation. After identifying SciPhylomics structure and the variabilities, we
instrumented the SciPhylomics python script according to Listing 1 to specify
the data transformations, tasks, data dependencies, datasets, and data elements.

Input Phylogenomic
1 2 3 4 5
Sequence Trees

Mafft
RaXML
Model
RemovePipe ClustalW ReadSeq
Generator
MrBayes
Muscle

Fig. 2. A fragment of the SciPhylomics dataflow

3.3 Setup of the Experiment

The fragment of SciPhylomics presented in Fig. 2 was executed varying the pro-
grams in transformations 2 and 5. We defined two different executions of SciPhy-
lomics (called Experiments A and B) depending on the chosen tree generation
program. In addition, for each experiment, we can vary the chosen MSA pro-
gram. Thus, six variants were performed with a dataset composed of 98 different
multifasta files (each variant was executed 10 times).
The machine configuration used to execute SciPhylomics was an AMD
FX(tm)-8150 8-Core Processor, 32 GB RAM, and 2 TB hard disk. The overview
of the parameter values used in the executions is in Table 1. The parameters
values were defined following an specialist. In special, Ngen is related to the
112 L. G. Dias et al.

number of cycles that the Markov Chain Monte Carlo (MCMC) algorithm is
executed, the main algorithm goal is to make small random changes in some
parameters to accept or reject according to the probability. Also, Nchains is
the parameter related to the number of different parallel MCMCMC chains
within a single execution run, printfreq is related to the frequency that the
information is shown on the screen, nruns how many independent analyses
are started simultaneously. Concerning the MSA programs, it was not nec-
essary to set parameters and the programs were executed in default mode.
More information are find on Muscle (www.ebi.ac.uk/Tools/msa/muscle/),
Mafft (mafft.cbrc.jp/alignment/software/), Clustalw (www.genome.jp/tools-
bin/clustalw), RAxML (http://cme.h-its.org/exelixis/web/software/raxml/)
and MrBayes (http://mrbayes.sourceforge.net/commref mb3.2.pdf) documenta-
tion.

Table 1. Parameters of the variants of SciPhylomics fragment

Experiment A Experiment B
Wf. step Program Parameter Wf. step Program Parameter
Clean RemovePipe Total Clean RemovePipe Total
Alignment Mafft Default Alignment Mafft Default
ClustalW Default ClustalW Default
Muscle Default Muscle Default
Converter ReadSeq Default Converter ReadSeq Default
Evolutive Model generator Default Evolutive Model generator Default
Model model
generator generator
Tree RaXML Default Tree MrBayes ngen 100000
generator generator
nchains 4
printfreq 1000
burnin 0
nruns 2
rates mrbayes 4

3.4 Data Analysis in Practice

In this Section, we aim at presenting the advantages of analyzing prove-
nance from phylogenomic analyses in practice using a series of analytical
queries. The experimental process was executed in 75.08h. For the analysis
presented in this section, we used six DfAnalyzer tables: dataflow (contains
information about the dataflow), data set (contains details of the datasets),
data transformation (contains details of the transformations), task (contains
details of the tasks), dataflow execution (contains statistics of the execution of
the dataflow), data transformation execution (contains statistics of the execu-
tion of each transformation). For more details about the DfAnalyzer database
schema please refer to [17].
The first analysis is related the evolutionary model generated by the fourth
transformation, implemented by ModelGenerator program. Let us assume that
 Experiencing DfAnalyzer for Runtime Analysis of Phylogenomic Dataflows 113

the user needs to analyze what are the models chosen for the input data. If the
bioinformatician executed SciPhylomics without DfAnalyzer, one has to open
each file to check this information, which is time-consuming, tedious, and error-
prone. By using DfAnalyzer, one has only to submit the query shown on List-
ing 2. In this query, the user wants to discover the number of times a specific
evolutionary model was used, but just when the length of the input sequence is
larger than 20.

The result of the query presented in Listing 2 is shown in Fig. 3. One can
state that WAG and RtREV models are the most common ones. WAG presented
higher likelihoods than any of the other models. This type of query can be
adapted to other attributes of the database, such as quality of the generated
tree, e-value, etc. Another performed analysis was related to the execution time
of each transformation and resource usage. Capturing this type of data can
impact the performance of the experiment since it is usually captured in a short
time interval and may introduce overhead. Thus, in this analysis we defined an
interval for capturing performance data (30 s window). The box-plots of the
execution time behaviour for the six variations of SciPhylomics are presented in
Fig. 4.

Fig. 3. Number of occurrences of each evolutionary model.

As presented in Fig. 4, there is a non-negligible variation in execution time

according to the tree generator program used in the experiments. The experi-
ments executed with RaXML (Fig. 4b) finished faster than those performed with
MrBayes (Fig. 4a). The execution time difference can be explained based on the
technique used for each program. MrBayes performs a Bayesian inference and
114 L. G. Dias et al.

Fig. 4. Execution time of SciPhylomics varying the MSA programs (ClustalW, Mafft
and Muscle) and tree generator programs (MrBayes and RaXML).

model choice across a wide range of phylogenetic and evolutionary models and is
more costly than RAxML. The programs execute different operations and gen-
erate different datasets that are composed of different attributes as well. While
RaXML does not execute the operation of “search for evolutionary model” (since
the ModelGenerator program is executed), MrBayes execute this operation, and
in this experiment variability the operation “search for evolutionary model” is
executed twice. These characteristics explain the difference in time execution in
both cases, and the resource usage as well, as shown on Fig. 5.

Fig. 5. SciPhylomics resource usage: memory and disk

 Experiencing DfAnalyzer for Runtime Analysis of Phylogenomic Dataflows 115

4 Conclusion
This paper presents an approach to capture provenance data from phylogenomic
dataflows without the need to rewrite the workflow to a specific programing lan-
guage or Workflow Management System engine. We specialized the novel DfAn-
alyzer tool in the context of Phylogenomic experiments so that in an existing
workflow it is possible to achieve a flexible granularity-level data capture and cre-
ate an integrated database composed of domain-specific, provenance and teleme-
try data. DfAnalyzer is prospective provenance based. By modelling prospective
provenance data, retrospective provenance is automatically captured while the
workflow executes, and the data created by its transformations are stored in a
relational database for further querying. Differently from other approaches, the
provenance is captured with flexible granularity, and the bioinformaticians can
specify what is important for their analysis and reduce the experiment cost in
different spheres.
Another advantage of applying DfAnalyzer in the context of Phylogenomic
experiments is the capability of capturing data from experiments that exe-
cute either locally or HPC environments, due to the fact that DfAnalyzer is
asynchronous and request-based, it can execute in different environments. This
asynchronous characteristic contributes that the instrumentation does not cause
delays in the workflow execution. In addition, we extended DfAnalyzer to cap-
ture telemetry (performance data). This way, users are allowed to perform anal-
yses based on both the provenance data and performance metrics. We evaluated
the proposed specialization of DfAnalyzer using the previously defined SciPhy-
lomics workflow and the proposed approach showed relevant telemetry scenarios
and rich data analyses. In future work, we intend to evaluate reproducibility in
experiments based on the analysis of the provenance database.

References
1. Abuı́n, J.M., Pichel, J.C., Pena, T.F., Amigo, J.: SparkBWA: speeding up the
alignment of high-throughput DNA sequencing data. PLoS ONE 11(5), e0155461
(2016)
2. Afgan, E., et al.: The galaxy platform for accessible, reproducible and collaborative
biomedical analyses: 2016 update. Nucleic Acids Res. 44(W1), W3–W10 (2016)
3. Carvalho, L.A.M.C., Wang, R., Gil, Y., Garijo, D.: NIW: converting notebooks into
workflows to capture dataflow and provenance. In: Tiddi, I., Rizzo, G., Corcho, Ó.
(eds.) Proceedings of Workshops and Tutorials of the 9th International Conference
on Knowledge Capture (K-CAP 2017), Austin, Texas, USA, 4 December 2017.
CEUR Workshop Proceedings, vol. 2065, pp. 12–16. CEUR-WS.org (2017)
4. de Oliveira, D.C.M., Liu, J., Pacitti, E.: Data-intensive workflow management: for
clouds and data-intensive and scalable computing environments (2019)
5. Deelman, E., et al.: Pegasus, a workflow management system for science automa-
tion. FGCS 46, 17–35 (2015)
6. Freire, J., Koop, D., Santos, E., Silva, C.T.: Provenance for computational tasks:
a survey. CS&E 10(3), 11–21 (2008)
116 L. G. Dias et al.

7. Guedes, T., et al.: Capturing and analyzing provenance from spark-based scientific
workflows with SAMbA-RaP. Future Gener. Comput. Syst. 112, 658–669 (2020)
8. Hondo, F., et al.: Data provenance management for bioinformatics workflows using
NoSQL database systems in a cloud computing environment. In: 2017 IEEE Inter-
national Conference on Bioinformatics and Biomedicine (BIBM), pp. 1929–1934.
IEEE (2017)
9. Marozzo, F., Talia, D., Trunfio, P.: Scalable script-based data analysis workflows
on clouds. In: WORKS, pp. 124–133 (2013)
10. Masulli, F.: Comput. Methods Programs Biomed. 91(2), 182 (2008)
11. Moreau, L., et al.: PROV-DM: the PROV data model. W3C Recommendation 30,
1–38 (2013)
12. Oliveira, D., Ocaña, K.A.C.S., Baião, F.A., Mattoso, M.: A provenance-based adap-
tive scheduling heuristic for parallel scientific workflows in clouds. JGC 10(3),
521–552 (2012)
13. de Oliveira, D., et al.: Performance evaluation of parallel strategies in public clouds:
a study with phylogenomic workflows. Future Gener. Comput. Syst. 29(7), 1816–
1825 (2013)
14. Olma, M., Karpathiotakis, M., Alagiannis, I., Athanassoulis, M., Ailamaki, A.:
Slalom: coasting through raw data via adaptive partitioning and indexing. Proc.
VLDB Endow. 10(10), 1106–1117 (2017)
15. Pimentel, J.F., Murta, L., Braganholo, V., Freire, J.: noWorkflow: a tool for col-
lecting, analyzing, and managing provenance from python scripts. Proc. VLDB
Endow. 10(12), 1841–1844 (2017)
16. Pina, D.B., Neves, L., Paes, A., de Oliveira, D., Mattoso, M.: Análise de
hiperparâmetros em aplicações de aprendizado profundo por meio de dados de
proveniência. In: Anais Principais do XXXIV Simpósio Brasileiro de Banco de
Dados, pp. 223–228. SBC (2019)
17. Silva, V., de Oliveira, D., Valduriez, P., Mattoso, M.: Dfanalyzer: runtime dataflow
analysis of scientific applications using provenance. Proc. VLDB Endow. 11(12),
2082–2085 (2018)
18. The UniProt Consortium: UniProt: the universal protein knowledgebase. Nucleic
Acids Res. 45(D1), D158–D169 (2016)
 Sorting by Reversals and Transpositions
with Proportion Restriction

Klairton Lima Brito1(B) , Alexsandro Oliveira Alexandrino1 ,

Andre Rodrigues Oliveira1 , Ulisses Dias2 , and Zanoni Dias1
1
Institute of Computing, University of Campinas, Campinas, Brazil
{klairton,alexsandro,andrero,zanoni}@ic.unicamp.br
2
School of Technology ,University of Campinas, Limeira, Brazil
[email protected]

Abstract. In the field of comparative genomics, one way of comparing

two genomes is through the analysis of how they distinguish themselves
based on a set of mutations called rearrangement events. When con-
sidering that genomes undergo different types of rearrangements, it can
be assumed that some events are more common than others. To model
this assumption one can assign different weights to different events,
where more common events tend to cost less than others. However,
this approach, called weighted, does not guarantee that the rearrange-
ment assumed to be the most frequent will be also the most frequently
returned by proposed algorithms. To overcome this issue, we investigate
a new problem where we seek the shortest sequence of rearrangement
events able to transform one genome into the other, with a restriction
regarding the proportion between the events returned. Here we consider
two rearrangement events: reversal, that inverts the order and the ori-
entation of the genes inside a segment of the genome, and transposition,
that moves a segment of the genome to another position. We present an
approximation algorithm applicable to any desired proportion, for both
scenarios where the orientation of the genes is known or unknown. We
also show an improved (asymptotic) approximation algorithm for the
case where the gene orientation is known.

Keywords: Rearrangement events · Proportion restriction ·

Approximation algorithm

1 Introduction
When comparing two genomes, one of the main goals is to determine the sequence
of mutations that occurred during the evolutionary process capable of transform-
ing a genome into another. In comparative genomics, we estimate this sequence
through genome rearrangements, evolutionary events (mutations) affecting a
large sequence of the genome.
Two genomes G1 and G2 can be computationally represented as the sequence
of labels assigned to their shared genes (or shared blocks of genes). Labels are
c Springer Nature Switzerland AG 2020
J. C. Setubal and W. M. Silva (Eds.): BSB 2020, LNBI 12558, pp. 117–128, 2020.
https://doi.org/10.1007/978-3-030-65775-8_11
118 K. L. Brito et al.

usually integer numbers. In addition, we associate a positive or negative sign

in each of the numbers, reflecting the orientation of that gene (or block) inside
of the genomes. Assuming that the genomes do not contain duplicated genes,
this representation results in a signed permutation, when the orientation of the
genes is known, and in an unsigned permutation otherwise. One of the genomes
can be seen as the identity permutation, in which the elements are in ascending
order, so problems dealing with genome rearrangements are usually treated as
sorting problems, in which the goal is to transform a given permutation into the
identity.
Two of the most studied genome rearrangements in the literature are the
reversal, that inverts the order and the orientation of the genes inside a seg-
ment of the genome, and transposition, that moves a segment of the genome
to another position. The Sorting by Reversals problem has an exact poly-
nomial algorithm for signed permutations [5] but it is NP-hard for unsigned
permutations [4].
The Sorting by Transpositions problem is NP-hard [3]. When we allow
the use of reversals and transpositions, and assuming that both events occur with
the same frequency (unweighted approach), we have the Sorting by Rever-
sals and Transpositions (SbRT) problem that is NP-hard on signed and
unsigned permutations [7].
In the weighted approach each type of event has an associated cost, and the
goal is to find a sequence of rearrangement events that transforms one genome
into another minimizing the sum of the costs. Oliveira et al. [7] showed that
Sorting by Weighted Reversals and Transpositions (SbWRT) prob-
lem is NP-hard on signed and unsigned permutations when the ratio between
the cost of a transposition and the cost of a reversal is less than or equal to
1.5. Oliveira et al. [8] developed a 1.5-approximation algorithm for SbWRT on
signed permutations considering costs 2 and 3 for reversals and transpositions,
respectively.
The problem with weighted approaches is that they do not guarantee that
lower cost rearrangements, i.e., assumed to be most frequent, will be the most
frequently used by the algorithms. To overcome this issue we propose and inves-
tigate the Sorting by Reversals and Transpositions with Proportion Restriction
problem on signed and unsigned permutations. In this problem, we seek a sorting
sequence with an additional constraint in which the ratio between the number
of reversals and the size of the sequence must be greater than or equal to a
given parameter k ∈ [0..1]. We provide an algorithm that guarantees an approx-
imation for any value of k on signed and unsigned permutations. We also show
an asymptotic algorithm for the signed case with an improved approximation
factor.
This manuscript is organized as follows. Section 2 provides definitions used
throughout the paper. Section 3 presents an approximation algorithm for the
signed and unsigned cases. Section 4 presents an asymptotic approximation algo-
rithm for the signed case with an improved approximation factor. Section 5
concludes the paper.
 Sorting by Reversals and Transpositions with Proportion Restriction 119

2 Basic Definitions
This section formally presents the definitions used in the genome rearrangement
problems. Given two genomes G1 and G2 , each synteny block (common block
of genes between the two genomes) is represented by an integer that also has
a positive or negative sign to indicate its orientation, if known. Therefore, each
genome is a permutation of integers. We assume that one of them is represented
by the identity permutation ιn = (+1 +2 . . . +n) and the other is represented
by a signed (or unsigned) permutation π = (π1 π2 . . . πn ).
We define a rearrangement model M as the set of rearrangement events
allowed to compute the distance. Given a rearrangement model M and a per-
mutation π, the rearrangement distance d(π) is the minimum number of rear-
rangements of M that sorts π (i.e., that transforms π into ι). The goal of the
Sorting by Genome Rearrangements problems consists in finding such distance
and the sequence that reflects it.
In this work, we will assume that M contains both reversals and transposi-
tions. Let us formally define these events.

Definition 1. Given a signed permutation π = (π1 . . . πn ), a reversal ρ(i, j),

with 1 ≤ i ≤ j ≤ n, transforms π in the permutation π · ρ(i, j) = (π1 . . . πi−1
−πj . . . −πi πj+1 . . . πn ).

Definition 2. Given an unsigned permutation π = (π1 . . . πn ), a reversal

ρ(i, j), with 1 ≤ i < j ≤ n, transforms π in the permutation π · ρ(i, j) =
(π1 . . . πi−1 πj . . . πi πj+1 . . . πn ).

Definition 3. Given a permutation π = (π1 . . . πn ), a transposition τ (i, j, k),

with 1 ≤ i < j < k ≤ n + 1, applied to π transforms it in the permutation
π · τ (i, j, k) = (π1 . . . πi−1 πj . . . πk−1 πi . . . πj−1 πk . . . πn ). The effect of a
transposition is the same on signed and unsigned permutations.

The following definition helps us to formally define the problem of sorting by

reversals and transpositions with a constraint on the number of reversals used
in the sorting sequence.

Definition 4. Given a sequence of reversals and transpositions S, let |S| denote

the number of events in S and let |Sρ | denote the number of reversals in S.

Sorting by Reversals and Transpositions with Proportion

Restriction (SbRTwPR)
Input: A permutation π, that can be signed or unsigned, and a rational
number k ∈ [0..1].
Task: Find the shortest sequence S of reversals and transpositions
|S |
that turns π into ι, such that |S|ρ ≥ k.
120 K. L. Brito et al.

Note that when k = 1 the SbRTwPR problem becomes the Sorting by Rever-
sals problem on signed [5] and unsigned [4] permutations. Moreover, when k = 0
we have the Sorting by Reversals and Transpositions problem on signed [10] and
unsigned [9] permutations.
Example 1 shows an optimal solution S for π = (−1 +4 −8 +3 +5 +2 −7 −6)
considering the SbRT and the SbWRT problems (SbWRT using costs 2 for
reversals and 3 for transpositions). Note that half of the operations in S are
reversals and half are transpositions, even using a higher cost for transpositions.
Example 1.
π= (−1 +4 −8 +3 +5 +2 −7 −6)
1
π = π · ρ(1, 5) = (−5 −3 +8 −4 +1 +2 −7 −6)
2 1
π = π · τ (2, 4, 9) = (−5 −4 +1 +2 −7 −6 −3 +8)
3 2
π = π · τ (1, 3, 7) = (+1 +2 −7 −6 −5 −4 −3 +8)
4 3
π = π · ρ(3, 7) = (+1 +2 +3 +4 +5 +6 +7 +8)
S= {ρ(1, 5), τ (2, 4, 9), τ (1, 3, 7), ρ(3, 7)}

Example 2 shows an optimal solution S  for the same signed permutation π

considering the SbRTwPR problem, adopting k = 0.6 (i.e., at least 60% of the
operations in S must be reversals). Compared with Example 1, the sequence S 
has only one more operation than S, while ensuring the minimum proportion of
reversals and using both reversals and transpositions.
Example 2.
π= (−1 +4 −8 +3 +5 +2 −7 −6)
1
π = π · ρ(2, 8) = (−1 +6 +7 −2 −5 −3 +8 −4)
2 1
π = π · ρ(2, 4) = (−1 +2 −7 −6 −5 −3 +8 −4)
3 2
π = π · τ (6, 8, 9) = (−1 +2 −7 −6 −5 −4 −3 +8)
4 3
π = π · ρ(1, 1) = (+1 +2 −7 −6 −5 −4 −3 +8)
5 4
π = π · ρ(3, 7) = (+1 +2 +3 +4 +5 +6 +7 +8)
S = {ρ(2, 8), ρ(2, 4), τ (6, 8, 9), ρ(1, 1), ρ(3, 7)}

In the following, we present breakpoints and the cycle graph, both widely
used to obtain bounds for the distance and to develop algorithms.

2.1 Breakpoints
Given a permutation π = (π1 . . . πn ), we extend π by adding the elements
π0 = 0 and πn+1 = n + 1, with these elements having positive signs when
considering signed permutations. We observe that these elements are not affected
by rearrangement events. From now on, we work on extended permutations.
Definition 5. For an unsigned permutation π, a pair of elements πi and πi+1 ,
with 0 ≤ i ≤ n, is a breakpoint if |πi+1 − πi | = 1.
 Sorting by Reversals and Transpositions with Proportion Restriction 121

The number of breakpoints in a permutation π is denoted by b(π). Given an

operation γ, let Δb(π, γ) = b(π) − b(π · γ), that is, Δb(π, γ) denotes the change
in the number of breakpoints after applying γ to π.

Remark 1. The identity permutation ι is the only permutation with b(π) = 0.

2.2 Cycle Graph

For a signed permutation π, we define the cycle graph G(π) = (V, E), such that
V = {+π0 , −π1 , +π1 , −π2 , +π2 , . . . , −πn , +πn , −πn+1 } and E = Eb ∪ Eg , where
Eb = {(−πi , +πi−1 ) | 1 ≤ i ≤ n + 1} and Eg = {(+(i − 1), −i) | 1 ≤ i ≤ n + 1}.
We say that Eb is the set of black edges and Eg is the set of gray edges.
Note that each vertex is incident to two edges (a gray edge and a black edge)
and, so, there exists a unique decomposition of edges in cycles. The size of a
cycle C ∈ G(π) is the number of black edges in C. A cycle C is trivial if it has
size 1. If C has size less than or equal to 3, then C is called short and, otherwise,
C is called long. The identity permutation ιn is the only one with a cycle graph
containing n + 1 cycles, which are all trivial.
The number of cycles in G(π) is denoted by c(π). Given an operation γ, let
Δc(π, γ) = c(π · γ) − c(π), that is, Δc(π, γ) denotes the change in the number of
cycles after applying γ to π.
The cycle graph G(π) is drawn in a way to highlight characteristics of the
permutation, as shown in Fig. 1. In this representation, we draw the vertices in
a horizontal line, from left to right, following the order +π0 , −π1 , +π1 , . . . , −πn ,
+πn , −πn+1 . The black edges are horizontal lines and the gray edges are arcs.
For 1 ≤ i ≤ n + 1, the black edge (−πi , +πi−1 ) is labeled as i. We represent
a cycle C by the sequence of labels of its black edges following the order they
are traversed, assuming that the first black edge is the one with highest label
(rightmost black edge of C) and it is traversed from right to left. Assuming this
representation, if a black edge is traversed from left to right we add a minus sign
to its label (the first black is always positive since it is traversed from right to
left by convention).
Two black edges of a cycle C are divergent if their labels have different signs,
and convergent otherwise. A cycle C is divergent if at least one pair of black
edges of C are divergent, and it is convergent otherwise.

1 2 3 4 5 6 −7 8
+0 −5 +5 −2 +2 −4 +4 −3 +3 −1 +1 −6 +6 −7 +7 −8

Fig. 1. Cycle Graph for π = (+5 + 2 + 4 + 3 + 1 + 6 − 7). In this cycle graph, we

have the cycles C1 = (5, 3, 4, 1), C2 = (6, 2), and C3 = (8, −7).
122 K. L. Brito et al.

We also classify convergent cycles as oriented and non-oriented. A cycle C =

(c1 , c2 , . . . , ck ) is non-oriented if ci > ci+1 , for all 1 ≤ i < k. Otherwise, we say
that C is oriented.
Two cycles C = (c1 , c2 , . . . , ck ) and D = (d1 , d2 , . . . , dk ) are interleaving if
either |c1 | > |d1 | > |c2 | > |d2 | > . . . > |ck | > |dk | or |d1 | > |c1 | > |d2 | > |c2 | >
. . . > |dk | > |ck |.
Let g1 be a gray edge adjacent to black edges with labels x1 and y1 , such
that |x1 | < |y1 | and let g2 be a gray edge adjacent to black edges with labels x2
and y2 , such that |x2 | < |y2 |. We say that two gray edges g1 and g2 intersect
if |x1 | < |x2 | ≤ |y1 | < |y2 |. Two cycles C and D intersect if an edge from C
intersect with an edge from D.
An open gate is a gray edge from a cycle C that does not intersect with
any other gray edge from C. An open gate g1 from C is closed if another gray
edge (which is not from C) intersects with g1 . All open gates of G(π) must be
closed [8].
In the example of Fig. 1, the cycle C1 = (5, 3, 4, 1) is convergent and oriented,
the cycle C2 = (6, 2) is convergent and non-oriented, and the cycle C3 = (8, −7)
is divergent. The gray edge from C1 adjacent to black edges 1 and 4 intersects
with the gray edge from C2 adjacent to black edges 2 and 6, so the cycles C1
and C2 intersect.

3 Approximation Algorithms
In this section, we present approximation algorithms considering both unsigned
and signed permutations.

3.1 Unsigned Case

Here we present an approximation algorithm with a factor of 3 − k based on

breakpoints for SbRTwPR on unsigned permutations.

Lemma 1 (Kececioglu and Sankoff [6]). For any reversal ρ, Δb(π, ρ) ≤ 2.

Lemma 2 (Walter et al. [10]). For any transposition τ , Δb(π, τ ) ≤ 3.

Lemma 3. Given an instance (π, k) for SbRTwPR on unsigned permutations,

and an optimal sequence of events S, the average number of breakpoints decreased
by an operation in S is less than or equal to 3 − k.

Proof. Since |S| is an optimal sequence for the instance (π, k), we have that at
least |S|k operations present in S are reversals. By Lemmas 1 and 2, we have that
a reversal can remove up to two breakpoints while a transposition can remove
up to three. Let φb(S) denote the average number of breakpoints decreased by
an operation in S, we have that

(2|S|k) + (3|S|(1 − k)) 


φb(S) ≤ = 2k + 3(1 − k) = 3 − k.
|S|
 Sorting by Reversals and Transpositions with Proportion Restriction 123

Theorem 1. Given an instance (π, k) for SbRTwPR on unsigned permuta-

tions, we have that dk (π) ≥ b(π)
3−k .

Proof. Since b(π) breakpoints must be removed in order to turn the permutation
π into ι and, by Lemma 3, up to 3 − k breakpoints are removed per operation
on average, the theorem follows. 


Theorem 2 (Kececioglu and Sankoff [6]). It is possible to turn an unsigned

permutation π into ι using at most b(π) reversals.

Theorem 3. SbRTwPR is approximable by a factor of 3 − k on unsigned per-

mutations.

Proof. By Theorem 2, we can turn any unsigned permutation π into ι using at

|S |
most b(π) reversals. Since we use only reversals, the constraint |S|ρ ≥ k is not
b(π)
violated. By the lower bound showed in Theorem 1, we have b(π) = 3 − k. 

3−k

In order to avoid solutions for the problem consisting exclusively of reversals,

we propose the Algorithm 1. This algorithm guarantees the same approximation
factor for the problem and tends to provide solutions in which the ratio between
the number of reversals and the size of the sorting sequence is close to k.

Algorithm 1: An approximation algorithm for SbRTwPR on unsigned

permutations.
Input: An unsigned permutation π and a value k
Output: A sequence of reversals and transpositions that sorts π
1 Let S ← {}
2 while π = ι do
|Sρ |
3 if |S|+1 ≥ k and there is a transposition τ such that Δb(π, τ ) ≥ 1 then
4 Apply τ in π
5 Append τ to S
6 else
7 Let S  be a sequence of reversals that decreases, on average, one or more
breakpoints per operation [6]
8 Apply S  in π
9 Append S  to S
10 return S

Note that a transposition τ is only applied if two constraints are fulfilled:

|Sρ |
(i) |S|+1 ≥ k, this ensures that the sorting sequence will comply with the main
|S |
restriction of the problem that |S|ρ ≥ k. (ii) Δb(π, τ ) ≥ 1, this constraint ensures
that the sorting sequence will contain a maximum of b(π) operations, since every
reversal sequence removes, on average, one or more breakpoints per operation.
Since Algorithm 1 removes one or more breakpoints by iteration, it guarantees
124 K. L. Brito et al.

that the permutation π will be sorted. In addition, no more than b(π) oper-
ations will be used to sort π, maintaining the approximation factor of 3 − k.
Since each operation (reversal or transposition) can be found in linear time and
|S| ≤ b(π) ≤ n + 1, the running time of Algorithm 1 is O(n2 ).

3.2 Signed Case

Here we present an approximation algorithm with a factor of 3 − 3k
2 based on
the cycle graph for SbRTwPR on signed permutations.

Lemma 4 (Hannenhalli and Pevzner [5]). For any reversal ρ, Δc(π, ρ) ≤ 1.

Lemma 5 (Bafna and Pevzner [1]). For any transposition τ , Δc(π, τ ) ≤ 2.

Lemma 6. Given an instance (π, k) for SbRTwPR on signed permutations,

and an optimal sequence of events S, the average number of cycles increased by
an operation in S is less than or equal to 2 − k.

Proof. Since |S| is an optimal sequence for the instance (π, k), we have that at
least |S|k operations in S sequence are reversals. By Lemmas 4 and 5, we have
that a reversal creates at most one new cycle, while a transposition creates at
most two new cycles. Let φc(S) denote the average number of cycles increased
by an operation in S, we have that:

(1|S|k) + (2|S|(1 − k))

φc(S) ≤ = 1k + 2(1 − k) = 2 − k. 

|S|
Theorem 4. Given an instance (π, k) for SbRTwPR on signed permutations,
we have that dk (π) ≥ n+1−c(π)
2−k .

Proof. Since (n + 1) − c(π) new cycles must be created in order to turn the
permutation π into ι and, by Lemma 6, up to 2 − k new cycles are created per
operation on average, the theorem follows. 


Theorem 5. Given a signed permutation π, there exists a sequence of reversals

S that transforms π into ι such that the average number of cycles increased by
any reversal in S is greater than or equal to 2/3.

Proof. If at any stage G(π) has a divergent cycle C, then there exists a reversal
applied to C that increases the number of cycles by one unit [10]. Otherwise,
G(π) has only convergent cycles, and one of the following is true [8]:

– there exists a long oriented cycle (Fig. 2, Case 1);

– there exists a short cycle C whose open gates are closed by another non-trivial
cycle D (Fig. 2, Case 2);
– there exists a long non-oriented cycle C whose open gates are closed by one
or more non-trivial cycles (Fig. 2, Case 3).
 Sorting by Reversals and Transpositions with Proportion Restriction 125

1 2 3

1 2 3 −4 5

−1 2 3

−1 2 3 4 5
1 −2 3

1 2 3 1 −2 3 4 5

Case 1 Case 2

1 2 3 4 5 6 7

1 −2 3 4 5 6 7

−1 2 3 4 5 6 7

1 2 3 4 5 6 7

Case 3

Fig. 2. Operations applied in each case of Theorem 5.

If G(π) has an oriented long cycle C, then we can apply a reversal on its
black edges in such a way that it turns C into a divergent cycle C  . Since C 
is long, we can apply at least two reversals on C  that increase the number of
cycles by one unit each (Fig. 2, Case 1).
In the other two cases we can turn the cycle C into an oriented cycle C  by
applying one reversal to a cycle D that closes an open gate from C. If C  is short,
we can break it into two trivial cycles with a reversal, and this second reversal
turns D into a divergent cycle D , which guarantees that we can apply a third
reversal to D that increases the number of cycles by one (Fig. 2, Case 2). If
C  is long, then we can apply at least two reversals that increase the number of
cycles by one unit each (Fig. 2, Case 3).
In the three cases above we applied three reversals that increased the number
of cycles by two, and the theorem follows. 


Theorem 6. SbRTwPR is approximable by a factor of 3 − 3k

2 on signed per-
mutations.
126 K. L. Brito et al.

Proof. By Theorem 5, we can turn any signed permutation π into ι using at

|S |
most 3(n+1−c(π))
2 reversals. Since we use only reversals, the constraint |S|ρ ≥ k
is not violated. By the lower bound showed in Theorem 4, we have:
3(n+1−c(π))
3k
2
n+1−c(π)
=3− .
2−k
2 

Note that in order to avoid a solution composed exclusively of reversals,
the approach used in Algorithm 1 can be adapted to be applied in this case as
well. In Sect. 4, we will present an asymptotic approximation algorithm with an
improved approximation factor for the signed case.

4 Asymptotic Approximation for the Signed Case

In this section we show an asymptotic algorithm for SbRTwPR on signed per-
mutations, where k ∈ [0, 1] with an approximation factor of ( 1−
2−k
k ).
3

Definition 6. Let Aρ be an algorithm that sorts a permutation using only signed

reversals and guarantees a ratio of 2/3 of cycles increased by applied reversals
(Theorem 5), and let Aρ (π) represents the sequence of reversals returned by the
algorithm that sorts π.
Now consider Algorithm 2.

Algorithm 2: An approximation algorithm for SbRTwPR on signed per-

mutations.
Input: A signed permutation π and a value k
Output: A sequence of reversals and transpositions that sorts π
1 Let S ← {}
2 while |Aρ (π)| > k(|S| + |Aρ (π)|) do
3 if G(π) has a divergent cycle then
4 Let ρ be a reversal that increases one cycle in G(π)
5 Apply ρ in π
6 Append ρ to S
7 else
8 Let S  be a sequence of at most two transpositions that increases two
cycles in G(π) [2, Theorem 3.4]
9 Apply S  in π
10 Append S  to S
11 Apply Aρ (π) in π
12 Append Aρ (π) to S
13 if |Sρ | < k|S| then
14 Replace the last two transpositions of S with six reversals [8]
15 return S
 Sorting by Reversals and Transpositions with Proportion Restriction 127

Lemma 7. Given a signed permutation π, Algorithm 2 sorts π using at most

(n + 1 − c(π))/(1 − k/3) + 4 operations.

Proof. Let S = (S1 , . . . , S|S| ) be the sorting sequence generated by the algorithm
without considering the substitution of transpositions by reversals applied in line
14. Let S  be the subsequence of operations applied in the while loop of lines
2 to 10. Each operation in S  increases the number of cycles by at least one
unit, and each operation in S \ S  (that is, the operations applied outside the
while loop) increases on average in 2/3 the number of cycles. By the condition
of line 2, we have that |S  | ≥ (1 − k)|S| and, therefore, the average increase in
the number of cycles in S is at least (1 − k)|S||S|+ k|S|2/3 = 1 − k/3. Since these
operations increase at most n + 1 − c(π) cycles, we have that |S| ≤ n +1 −1 − c(π)
k/3 .
In the final sequence, we may increase four operations by replacing the last two
transpositions with six reversals (only if necessary). Therefore, the size of this
sequence is at most n +1 −
1 − c(π)
k/3 + 4.

2−k
Theorem 7. Algorithm 2 is a 1 − k/3 -asymptotic approximation algorithm for
SbRTwPR.

Proof. Since the algorithm only adds transpositions while the condition of line
2 is satisfied and at most two transpositions are added in the sorting sequence
in one iteration, we guarantee that |Sρ | ≥ k by replacing the last two transpo-
sitions by reversals. By Lemma 7 and Theorem 4, the sequence S returned by
Algorithm 2 satisfies |S| ≤ n+1−c(π)
1−k/3 + 4 ≤ 1−k/3
2−k
dk (π) + 4. Therefore, it is a
2−k
1−k/3 -asymptotic approximation algorithm for SbRTwPR. 


5 Conclusion

We investigated the Sorting by Reversals and Transpositions with Proportion

Restriction problem and presented an approximation algorithm with a factor
of 3 − k for unsigned permutations, and an approximation and an asymptotic
approximation algorithm with factors 3 − 3k 2−k
2 and 1− k for signed permutations,
3
respectively.
As future work, we intend to test the proposed algorithms and develop heuris-
tics for the problems. Another interesting research line would be to investigate
the complexity of the problems when 0 < k < 1.

Acknowledgments. This work was supported by the National Council of Techno-

logical and Scientific Development, CNPq (grants 400487/2016-0, 140272/2020-8, and
425340/2016-3), the Coordenação de Aperfeiçoamento de Pessoal de Nı́vel Superior -
Brasil (CAPES) - Finance Code 001, and the São Paulo Research Foundation, FAPESP
(grants 2013/08293-7, 2015/11937-9, 2017/12646-3, and 2019/27331-3).
128 K. L. Brito et al.

References
1. Bafna, V., Pevzner, P.A.: Sorting permutations by transpositions. In: Proceedings
of the Sixth Annual ACM-SIAM Symposium on Discrete Algorithms (SODA 1995),
pp. 614–623. Society for Industrial and Applied Mathematics, Philadelphia (1995)
2. Bafna, V., Pevzner, P.A.: Sorting by transpositions. SIAM J. Discret. Math. 11(2),
224–240 (1998)
3. Bulteau, L., Fertin, G., Rusu, I.: Sorting by transpositions is difficult. SIAM J.
Discret. Math. 26(3), 1148–1180 (2012)
4. Caprara, A.: Sorting permutations by reversals and Eulerian cycle decompositions.
SIAM J. Discret. Math. 12(1), 91–110 (1999)
5. Hannenhalli, S., Pevzner, P.A.: Transforming cabbage into turnip: polynomial algo-
rithm for sorting signed permutations by reversals. J. ACM 46(1), 1–27 (1999)
6. Kececioglu, J.D., Sankoff, D.: Exact and approximation algorithms for sorting by
reversals, with application to genome rearrangement. Algorithmica 13, 180–210
(1995)
7. Oliveira, A.R., Brito, K.L., Dias, U., Dias, Z.: On the complexity of sorting by
reversals and transpositions problems. J. Comput. Biol. 26, 1223–1229 (2019)
8. Oliveira, A.R., Brito, K.L., Dias, Z., Dias, U.: Sorting by weighted reversals and
transpositions. J. Comput. Biol. 26, 420–431 (2019)
9. Rahman, A., Shatabda, S., Hasan, M.: An approximation algorithm for sorting by
reversals and transpositions. J. Discret. Algorithms 6(3), 449–457 (2008)
10. Walter, M.E.M.T., Dias, Z., Meidanis, J.: Reversal and transposition distance of
linear chromosomes. In: Proceedings of the 5th International Symposium on String
Processing and Information Retrieval (SPIRE 1998), pp. 96–102. IEEE Computer
Society, Los Alamitos (1998)
 Heuristics for Breakpoint Graph
Decomposition with Applications
in Genome Rearrangement Problems

Pedro Olı́mpio Pinheiro(B) , Alexsandro Oliveira Alexandrino ,

Andre Rodrigues Oliveira , Cid Carvalho de Souza , and Zanoni Dias

Institute of Computing, University of Campinas, Campinas, Brazil

[email protected],
{alexsandro,andrero,cid,zanoni}@ic.unicamp.br

Abstract. The breakpoint graph of a permutation is a well-known

structure used in genome rearrangement problems. Most studies use the
decomposition of such graph into edge-colored disjoint alternating cycles
to develop algorithms for these problems. The goal of the Breakpoint
Graph Decomposition (BGD) problem is to find a decomposition of the
breakpoint graph with maximum number of cycles. For unsigned per-
mutations, which model genomes without information about gene ori-
entation, the BGD problem is NP-hard. In this work, we developed a
greedy and a Tabu Search algorithm which are compared experimentally
with the approximation algorithm presented by Lin and Jiang [10]. The
experiments revealed that our algorithms find significantly better solu-
tions. Finally, we used our algorithms as part of algorithms for genome
rearrangement problems and the distances calculated in this way have
largely improved.

Keywords: Breakpoint graph · Genome rearrangements · Maximum

cycle decomposition

1 Introduction
The rearrangement distance between two genomes is a problem in comparative
genomics that aims to find a minimum sequence of rearrangements required to
transform one genome into the other.
Genomes in this problem are generally represented as permutations, where
each element of the permutation corresponds to a gene. When the orientation of
the genes is known, we use a plus or minus sign in each element to indicate the
orientation, and we say that the permutation is signed. Otherwise, elements do
not have signs, and we say that the permutation is unsigned.
Since we can model one of the genomes as the identity permutation (i.e.,
permutation (1 2 . . . n) or (+1 + 2 . . . + n)), the problem of transforming one
genome into another by rearrangements is equivalent to that of sorting permu-
tations by rearrangements. We often assume that these permutations have two
extra elements 0 and n + 1 at the beginning and at the end, respectively.
c Springer Nature Switzerland AG 2020
J. C. Setubal and W. M. Silva (Eds.): BSB 2020, LNBI 12558, pp. 129–140, 2020.
https://doi.org/10.1007/978-3-030-65775-8_12
130 P. O. Pinheiro et al.

The most studied rearrangements are the reversal, transposition, and Double
cut-and-join (DCJ) operations. The problems of Sorting by Reversals and Sorting
by DCJs are solvable in polynomial time on signed permutations [8,13], while
they are NP-hard on unsigned permutations [3]. The problems of Sorting by
Transpositions [2] and Sorting by Reversals and Transpositions [11] are also
NP-hard.
The first bounds for these problems used the concept of breakpoints, which
are pairs of elements that are adjacent in the given permutation but not in the
identity permutation. Later, improved bounds and new algorithms were devel-
oped using the breakpoint graph of a permutation [1]. We formally define the
breakpoint graph of a permutation in Sect. 2.
These bounds for the rearrangement distance are based on the number of
cycles in a maximum cardinality decomposition of the breakpoint graph into
edge-colored disjoint cycles. This decomposition is unique on signed permu-
tations or when the model allows only transpositions on unsigned permuta-
tions. When considering reversals on unsigned permutations, Caprara [3] showed
that finding a maximum cardinality decomposition of a breakpoint graph of an
unsigned permutation is NP-hard, and the same is valid for DCJs.
Using the Breakpoint Graph Decomposition as a subproblem, Bafna and
Pevzner [1] presented a 7/4-approximation algorithm for the Sorting by Rever-
sals. This factor was improved by Christie [5] to 1.5 and by Lin and Jiang [10]
to 1.4193 + . Based on a similar strategy, Chen [4] presented a (1.4167 + )-
approximation algorithm for the Sorting by DCJs problem. More recently,
Jiang et al. [9] presented a randomized FPT algorithm for the Sorting by DCJs
with an approximation factor of (4/3 + ).
In this paper, we propose a greedy algorithm and an algorithm based on
the Tabu Search metaheuristic for the Breakpoint Graph Decomposition. We
analyze the performance of these algorithms in practice and compare them with
the algorithm of Lin and Jiang [10]. Furthermore, we present experimental results
of these algorithms applied to the genome rearrangement distance considering a
model with DCJs and a model with reversals and transpositions.
This paper is organized as follows. Section 2 introduces the concepts used
in the algorithms and formalizes the problem. Section 3 presents the heuristics
created for the Breakpoint Graph Decomposition problem. In Sect. 4, we show
the experimental results for the heuristics of Sect. 3. At last, in Sect. 5, we give
our final remarks and discuss directions of future work.

2 Preliminaries
Let π be a permutation (π1 π2 . . . πn ), where πi ∈ {1, . . . , n} and πi = πj
iff i = j, for 1 ≤ i, j ≤ n. The identity permutation ιn is the permutation
(1 2 . . . n), which is the target of the sorting by rearrangements problems.
We extend π by adding the elements π0 = 0 and πn+1 = n + 1. In the next
definitions, we assume that permutations are in its extended form.
We say that (πi , πi+1 ) is a breakpoint if |πi+1 − πi | = 1, for 0 ≤ i ≤ n. The
number of breakpoints in a permutation π is denoted by b(π).
 Algorithms for Breakpoint Graph Decomposition 131

(a)

0 6 4 1 7 3 5 2 8 9

(b)

0 6 4 1 7 3 5 2 8 9

(c)

0 6 4 1 7 3 5 2 8 9

Fig. 1. (a) Breakpoint graph G(π) for π = (0 6 4 1 7 3 5 2 8 9), where b(π) = 8.

(b) Decomposition of G(π) into three cycles C1 = (8, 2, 1, 7), C2 = (5, 3, 4, 6), and
C3 = (2, 5, 4, 1, 0, 6, 7, 3). (c) Decomposition of G(π) into four cycles C1 = (8, 2, 3, 7),
C2 = (2, 5, 4, 1), C3 = (5, 3, 4, 6), and C4 = (7, 1, 0, 6).

Example 1. The permutation π = (0 4 3 5 1 2 6) has breakpoints (0, 4), (3, 5),

(5, 1), and (2, 6). Therefore, b(π) = 4.

The breakpoint graph G(π) = (V, E) of a permutation π is an edge-colored

undirected graph with vertices V = {π0 , π1 , . . . , πn , πn+1 } and edges E = Eb ∪
Eg , where Eb is the set of black edges and Eg is the set of gray edges. For all
0 ≤ i ≤ n, there exists a black edge (πi , πi+1 ) and a gray edge (πi , πi + 1) if the
pair (πi , πi+1 ) is a breakpoint.
The set of black edges Eb connects elements adjacent in π and the set of
gray edges Eg connects elements adjacent in ιn . Elements adjacent in both π
and ιn are not connected in G(π). By convention, we draw the breakpoint graph
by placing each vertex on a horizontal line in the order they appear in the
permutation. Black edges are represented as horizontal lines and gray edges are
represented as arcs. An example is shown in Fig. 1(a).
An alternating cycle C is a cycle such that every pair of consecutive edges
have distinct colors. An alternating cycle C of G(π) is a k-cycle if it has k
black edges. By convention, we list the vertices in an alternating cycle C =
(πx1 , πx2 , . . . , πxm ) assuming that the first vertex is the rightmost element from
π (i.e. x1 ≥ xi , for 2 ≤ i ≤ m) and that (πx1 , πx2 ) is a black edge.
Since each vertex is incident to the same number of black and gray edges,
there exists at least one decomposition of G(π) into edge-disjoint alternating
cycles [3]. In the Breakpoint Graph Decomposition problem, we are interested
in finding a set of edge-disjoint alternating cycles with maximum cardinality.
For a permutation π, we denote by c(π) the number of alternating cycles in an
optimal solution for the Breakpoint Graph Decomposition problem.
132 P. O. Pinheiro et al.

In Figs. 1(b) and 1(c), we show two distinct examples of decompositions

of a breakpoint graph into alternating cycles. The decomposition in Fig. 1(c) is
maximum for the breakpoint graph of the permutation π = (0 6 4 1 7 3 5 2 8 9).

Breakpoint Graph Decomposition (BGD)

Input: A permutation π.
Goal: Find a set of edge-disjoint alternating cycles H = {C1 , C2 , . . .},
such that |H| is maximum.

3 Heuristics for Breakpoint Graph Decomposition

In the following sections, we present two general heuristics for the BGD problem.

3.1 Greedy Algorithm

As shown in Fig. 1(a), each vertex in the breakpoint graph has a maximum of two
black edges and two gray edges. Our greedy algorithm has a subroutine called
bfs cycle that, starting from a given vertex with incident edges, performs a
breadth first search for an alternating cycle, as we explain in detail later in this
section. Depending on which vertex the algorithm chooses to start the search,
different variations could be applied:

– using the leftmost vertex with incident edges, which we call the FIRST app-
roach;
– choosing a vertex at random among those still having incident edges, which
we call the RANDOM approach;
– executing for all vertices still having incident edges, creating a list with the
returned cycles, and, at the end, picking the smallest cycle from this list,
which we call the ALL approach.

The selected cycle is added to the list of cycles and all edges in that cycle
are removed from the breakpoint graph. These steps are repeated until there are
no edges left in the graph, in which case the greedy algorithm stops. Variations
FIRST and RANDOM take O(n2 ) while the variation ALL takes O(n3 ).
As the RANDOM variation is non-deterministic, we also developed a variation
called MAX, that runs the RANDOM k times and returns the decomposition with the
largest number of cycles. The best√trade-off between execution time and solution
√ achieved using k = n n, so the MAX variation has a time complexity
quality was
of O(n3 n).
The bfs cycle subroutine: given a vertex v, we perform a breadth first
search starting at v with the constraint that we explore a new vertex u only if
the explored path from v to u is an alternating path (i.e., consecutive edges have
distinct colors).
When exploring the edges of a vertex u, if there exists an edge (u, v) such
that the path from v to u plus the edge (u, v) forms an alternating cycle, then
 Algorithms for Breakpoint Graph Decomposition 133

we stop the search returning this cycle. Since we do this using a breadth first
search, this subroutine finds the smallest alternating cycle starting with v.
As an example consider the breakpoint graph of Fig. 1(a). The FIRST app-
roach returns the decomposition of Fig. 1(c). It starts at vertex 0, finds the small-
est alternating cycle starting at 0, which is the cycle with vertices (0, 6, 7, 1), and
removes its edges from the graph. Then, it proceeds to find the smallest cycle
starting at vertex 6, which still has black edges. The algorithm finds the cycle
with vertices (6, 4, 3, 5) and removes its edges from the graph. The algorithm
continues until there are no edges in the graph.
Consider the same example but executing the RANDOM approach. Suppose
that the vertices 6, 1, and 0 were chosen in each iteration, in this order. Then,
the algorithm returns the decomposition of Fig. 1(b). In the first iteration, the
algorithm finds the smallest cycle starting at vertex 6, which is the cycle with
vertices (6, 4, 3, 5). In the next iteration, the algorithm finds the smallest cycle
starting at 1, which is the cycle with vertices (1, 7, 8, 2). At this point, the graph
has only one cycle that is chosen in the last iteration.
As a heuristic, our greedy algorithm described above cannot guarantee an
approximation factor. However, it is possible to adapt it in such a way that it
guarantees the same approximation as the Lin and Jiang algorithm (L&J) [10].
Let C be the cycle decomposition returned by L&J. Append to our cycle list
every cycle from C with less than 4 black edges (i.e., the list of cycles that is
enough to guarantee the approximation factor of 1.4193 + ). Remove their edges
from the breakpoint graph and start the breadth first search over this modified
breakpoint graph until no more edges exist. The L&J algorithm requires O(n3 )
time to generate all possible short cycles (i.e., cycles of size 4 and 6) plus O(3m )
to calculate the greatest subset of disjoint short cycles, where m is the number
of short cycles, with  = 0.

3.2 Tabu Search

Tabu Search is a strategy to solve combinatorial optimization problems capa-
ble of using many other methods, such as linear programming and specialized
heuristics, to overcome local optima [6,7].
Glover [6,7] describes the following generic optimization problem to intro-
duce the Tabu Search metaheuristic: given the set X of feasible solutions to an
optimization problem and a objective function f , find an element x ∈ X with
maximum (or minimum) value of f (x).
Given a feasible solution x ∈ X, a movement is an operation s that gen-
erates a new feasible solution s(x) = x . The set of possible movements that
can be applied to a solution x is denoted by S(x). Usually, we are interested in
movements that create better solutions in terms of the objective function.
To avoid local optima, movements that do not increase the objective value
are also executed. However, allowing these movements can trap the search into a
set of solutions without improvements in the objective value. To prevent this sce-
nario, we create a tabu list T with the movements performed in recent iterations
of the search that should not be done again or undone.
134 P. O. Pinheiro et al.

Algorithm 1. Tabu Search Heuristic

1: Choose an initial solution x ∈ X and make x∗ ← x, where x∗ is the best solution
found so far. Start the iteration counter k ← 0 and the tabu list T ← ∅.
2: If S(X) \ T = ∅, then go to step 4. Otherwise, make k ← k + 1 and choose
sk (x) ∈ S(x) \ T such that sk (x) = arg maxs∈S(x)\T {f (s(x))}.
3: Make x ← sk (x). If f (x) > f (x∗ ), then x∗ ← x.
4: If the maximum number of iterations was reached or S(x) \ T = ∅, stop the algo-
rithm and return the solution x∗ . Otherwise, update T and go to step 2.

(a)
... ...
u v w

(b)
... ...
u v w

Fig. 2. (a) Example of a solution for BGD with two cycles C1 (straight edges) and
C2 (dashed edges). These cycles have three vertices (u, v, and w) in common and
they satisfy the conditions to apply the first movement. (b) Three cycles generated by
applying the first movement in the cycles C1 and C2 .

Algorithm 1 describes the Tabu Search heuristic presented by Glover [6].

For the Breakpoint Graph Decomposition problem, the objective function is
to maximize the number of cycles in the solution. The initial solution is given
by one of the algorithms of Sect. 3.1 and the two movements developed are
described next.
The first movement is applied in two cycles C1 and C2 , such that they have
at least three common vertices u, v, and w, transforming them into three cycles.
Given two vertices x and y from the three common vertices u, v, and w, let
1 2
Pxy and Pxy denote the paths in the cycles C1 and C2 , respectively, that go from
x to y and do not include the third common vertex. To apply the first movement,
1 2
the paths Puv and Puv must have lengths with same parity and the first edge
of these two paths must have different colors. The same condition must hold for
1 2 1 2
paths Pvw and Pvw and for the paths Pwu and Pwu .
When these conditions are satisfied, a solution with one more cycle is created
by replacing cycles C1 and C2 with three cycles C1 , C2 , and C3 , where C1 is the
concatenation of Puv1
and Puv2
, C2 is the concatenation of Pvw
1
and Pvw2
, and C3
1 2
is the concatenation of Pwu and Pwu .
Figure 2 shows an example of the first movement being applied to two cycles.
Since this movement always creates a solution with one more cycle, the Tabu
Search algorithms execute it whenever possible.
 Algorithms for Breakpoint Graph Decomposition 135

The second movement is applied in two cycles C1 and C2 , such that they
have at least two common vertices u and v, transforming them into two distinct
cycles.
1
Let Puv and Q1vu be the two distinct paths from u to v and from v to u in
1 2
cycle C1 , respectively, such that Puv begins with a black edge. Let Puv and Q2vu
be the two distinct paths from u to v and from v to u in cycle C2 , respectively,
2
such that Puv begins with a black edge. Note that both Q1vu and Q2vu end with
a gray edge.
1
The second movement can be applied if the length of Puv has the same parity
of the length of Puv . This movement creates two cycles C1 and C2 , such that
2

C1 is the path Puv

1
concatenated with Q2vu and C2 is the path Puv
2
concatenated
1
with Qvu .

(a)
...
u v

(b)
...
u v

Fig. 3. (a) Example of a solution for BGD with two cycles C1 (dashed edges) and
C2 (straight edges). These cycles have two vertices (u and v) in common and they
satisfy the conditions to apply the second movement. (b) Two new cycles generated
by applying the second movement in the cycles C1 and C2 .

Figure 3 shows an example of the second movement applied to two cycles.

When this movement is executed, the pair (u, v) is included to the tabu list T .
Note that this movement creates a new solution with the same number of cycles.
This operation is applied only if the pair (u, v) is not in the tabu list T .
The second movement is useful to continue searching for new solutions after
the algorithm reaches a local optimum. Since, in this case, there is no pair of
cycles that satisfies the conditions of the first movement, the second movement
alters the cycles of the solution possibly enabling the use of the first movement
in the new solution.
We limit the second movement with the tabu list, because, otherwise, the
Tabu Search could enter into an infinite loop, applying the second movement
replacing two cycles C1 and C2 by two new cycles C1 and C2 and, in the next
iteration, reverting the previous operation by applying the second movement to
the new cycles C1 and C2 transforming them back to C1 and C2 .
The complexity of each iteration is O(n3 ). The best trade-off between exe-
cution time and solution quality was achieved using the maximum number of
iterations equals to n. Then the complete search becomes O(n4 ).
136 P. O. Pinheiro et al.

4 Experimental Results
In order to check the efficiency of our heuristics, we implemented Lin and Jiang’s
algorithm (L&J), the greedy heuristic, and the Tabu Search. To compare them,
we generated sets of permutations such that the number of breakpoints is as
large as possible (i.e., n+1 for a permutation of size n). Thus, every breakpoint
graph has n+1 black edges and every vertex has two black edges and two gray
edges, with the exception of the first and the last vertices which have only one
edge of each color. Our sets are separated by permutation size, from 10 to 100
in intervals of 10, with 100 permutations each.
Four different experiments were performed: the first two consist of (i) com-
paring the different versions of the greedy algorithm against the L&J; and (ii)
indicating to the greedy algorithm that it must contain the same short cycles as
those of L&J (so that our greedy algorithm also guarantees the same approxima-
tion factor). The other two experiments consist of running the Tabu Search on
top of the first two tests to check its improvement. Results for variant RANDOM
of the greedy algorithm are the average of 100 executions for each permutation.
Table 1 shows the average number of cycles returned by L&J and the four
variants of the greedy algorithm, namely FIRST, ALL, RANDOM and MAX. We can
see that, except for permutations of size up to 20, all variations of the greedy
algorithm returned decompositions with a greater number of cycles on average
compared to those returned by L&J. Besides, the decompositions obtained by
MAX have on average more cycles than L&J for all permutations of size greater
than or equal to 20. For permutations of size greater than or equal to 60, the
MAX heuristic returned cycle decompositions whose number of cycles are at least
50% greater on average than the cycle decompositions returned by L&J. Results
of Experiment 2 show that by using the same set of short cycles from L&J in
the cycle decomposition, the variations FIRST, ALL, and RANDOM returned cycle
decompositions with more cycles compared to their results on Experiment 1.
However, the MAX variation, which produces the best results on average, returned
cycle decompositions with a smaller number of cycles on average compared to
the results obtained on Experiment 1, for permutations sizes between 20 and 80.
Table 2 shows the average number of cycles returned by Tabu Search using
the output of L&J and the four variants of the greedy algorithm. We can see that
Tabu Search was able to improve the results of all algorithms, especially L&J
that had a great improvement (probably because it returned the lowest average
number of cycles, with a large room for improvement). The MAX variation remains
as the one that returns the greatest number of cycles on average, and the same
behavior of Experiment 2 happened here: on average, the results of Experiment
4 are slightly worse than the results of Experiment 3, for permutations sizes
between 20 and 90.
 Algorithms for Breakpoint Graph Decomposition 137

Table 1. Average number of cycles returned in the Experiments 1 and 2. Experiment

1 consists of the L&J algorithm and all variations of the greedy heuristic. Experiment
2 consists of using the same short cycles returned by L&J in the solution of our greedy
heuristic. Each value represents the average for 100 permutations of size n, which is
indicated in the first column. For the RANDOM approach, we also did the average of 100
executions for each permutation.

n Experiment 1 Experiment 2
L&J FIRST ALL RANDOM MAX FIRST ALL RANDOM MAX
10 4.14 3.99 4.09 3.98 4.14 4.14 4.14 4.14 4.14
20 6.58 6.23 6.63 6.35 6.89 6.75 6.75 6.75 6.75
30 8.01 8.04 8.75 8.27 9.22 8.89 8.90 8.88 8.91
40 9.12 9.98 10.70 10.06 11.39 10.81 10.89 10.84 10.97
50 9.70 11.39 12.75 11.67 13.45 12.61 12.74 12.62 12.97
60 10.20 13.06 14.38 13.20 15.38 14.25 14.47 14.27 14.88
70 10.97 14.45 16.36 14.73 17.13 16.09 16.33 16.02 16.84
80 11.52 15.69 18.16 16.30 19.01 17.77 18.24 17.77 18.95
90 11.73 17.27 19.51 17.60 20.44 19.14 19.63 19.13 20.72
100 12.31 18.35 21.35 19.03 22.17 20.67 21.27 20.71 22.54

4.1 Applications in Genome Rearrangement Distance

Algorithms for the Breakpoint Graph Decomposition problem are often used as
a part of approximation algorithms for genome rearrangement problems consid-
ering unsigned permutations [4,10,12].
Due to the relation between these problems, we can also use the value
b(π) − |H| to evaluate a decomposition H. We recall that b(π) is the number of
breakpoints in a permutation π and c(π) is the number of cycles in an optimal
solution for the BGD problem. Since b(π) is constant for a given permutation π,
searching for a decomposition H of G(π) with maximum cardinality is equivalent
to find a decomposition H of G(π) such that b(π) − |H| is minimum.
The approximation algorithms mentioned in Sect. 1 use the value b(π) − c(π)
to calculate the approximation factor, which is also a lower bound for the reversal
and DCJ distances.

Theorem 1 (Bafna and Pevzner [1]). For any permutation π, the reversal
distance dr (π) ≥ b(π) − c(π).

Theorem 2 (Chen [4] and Yancopoulos et al. [13]). For any permutation
π, the DCJ distance dDCJ (π) ≥ b(π) − c(π). Also, given a decomposition H for
G(π), the DCJ distance dDCJ (π) ≤ b(π) − |H|.

Theorem 3 (Rahman et al. [12]). For any permutation π, the reversal and
transposition distance drt (π) ≥ (b(π) − c(π))/2. Also, given a decomposition H
for G(π), the reversal and transposition distance drt (π) ≤ b(π) − |H|.
138 P. O. Pinheiro et al.

Table 2. Average number of cycles returned by the Tabu Search heuristic using the
cycle decomposition returned by each algorithm as input on Experiments 3 and 4.
Experiment 3 consists of using the Tabu Search heuristic in the solutions of Experiment
1 and Experiment 4 consists of using the Tabu Search heuristic in the solutions of
Experiment 2. Each value represents the average for 100 permutations of size n, which
is indicated in the first column. For the RANDOM approach, we also did the average of
100 executions for each permutation.

n Experiment 3 Experiment 4
L&J FIRST ALL RANDOM MAX FIRST ALL RANDOM MAX
10 4.14 4.09 4.12 4.14 4.14 4.14 4.14 4.14 4.14
20 6.76 6.50 6.80 6.89 6.89 6.77 6.77 6.81 6.81
30 8.87 8.54 9.00 9.22 9.22 9.03 9.02 9.13 9.13
40 10.85 10.50 11.12 11.39 11.40 11.08 11.14 11.33 11.33
50 12.55 12.28 13.06 13.45 13.58 12.99 13.08 13.40 13.42
60 14.39 13.97 14.90 15.40 15.61 14.75 14.85 15.34 15.41
70 16.16 15.72 16.78 17.15 17.48 16.60 16.79 17.35 17.41
80 17.90 17.18 18.67 19.11 19.63 18.39 18.77 19.42 19.54
90 19.24 18.65 20.17 20.58 21.33 19.87 20.24 21.14 21.26
100 20.72 20.29 21.96 22.33 23.27 21.56 21.99 22.94 23.14

We used our algorithms as part of the approximation algorithms for the

genome rearrangement distance to evaluate how the distinct decompositions
affect the resulting sorting sequences. The general idea is that a decomposition
of G(π) in alternating cycles is associated with a signed permutation, and we
can use algorithms for signed permutations once we have found a decomposition
of G(π).
We recall that in sorting problems the goal is to find a sequence that sorts
a permutation π with minimum length. Note that the identity permutation ι is
the only one without breakpoints and, consequently, the breakpoint graph G(ι)
has no edges. In this way, we can interpret sorting a permutation π as removing
all breakpoints and cycles of G(π).
Next, we describe the results of our experiments using algorithms for the
DCJ distance and the reversals and transpositions distance of unsigned permu-
tations. Table 3 presents the results for the DCJ distance and the reversals and
transpositions distance.
For a permutation π and a decomposition H, the DCJ distance is equal to
b(π) − |H| [4,13] and, consequently, the results of this table are similar to the
ones of Experiment 4 of Table 2, since all permutations of size n used in the
experiment have n + 1 breakpoints.
For the experiment with reversals and transpositions distance, we used the
2k-approximation algorithm presented by Rahman et al. [12], where k is the
approximation factor of the algorithm used for the BGD problem. For a per-
mutation π and a decomposition H, this algorithm returns a sorting sequence
S such that (b(π) − |H|)/2 ≤ |S| ≤ b(π) − |H|, where |S| is the length of the
 Algorithms for Breakpoint Graph Decomposition 139

sequence. When considering reversals and transpositions, some characteristics

of the cycles define how easy or difficult it is to sort the permutation. As an
example, when considering reversals, a sequence can remove the two cycles from
Fig. 3(b) using two reversals, but at least three reversals are needed to remove
the two cycles from Fig. 3(a) [8]. For this reason, a decomposition H1 can yield
a sorting sequence with more operations than a sorting sequence using a decom-
position H2 that has fewer cycles than H1 .
All the proposed algorithms of Sect. 3 had significantly better results than
Lin and Jiang’s algorithm. Although the ALL approach results in solutions with
more cycles than the FIRST approach, we can see that, for some values of n, the
FIRST approach yields better results for the distance. Overall, the MAX approach
also had better results.

Table 3. Results for the DCJ distance and the reversals and transpositions (RT) dis-
tance using the original L&J algorithm (Table 1) and the algorithms of the fourth
experiment (Table 2) as a subroutine for the genome rearrangements algorithms. Each
value represents the average for 100 permutations of size n, which is indicated in the
first column. For the RANDOM approach, we also did the average of 100 executions for
each permutation.

n L&J FIRST ALL RANDOM MAX

DCJ RT DCJ RT DCJ RT DCJ RT DCJ RT
10 6.86 6.02 6.86 6.23 6.86 6.23 6.86 6.20 6.86 6.21
20 14.42 12.16 14.23 12.17 14.23 12.15 14.19 12.06 14.19 11.91
30 22.99 18.64 21.97 17.70 21.98 17.79 21.87 17.83 21.87 17.70
40 31.88 25.11 29.92 23.17 29.86 23.22 29.67 23.26 29.67 23.24
50 41.30 32.86 38.01 29.03 37.92 28.89 37.60 28.83 37.58 28.79
60 50.80 39.91 46.25 33.98 46.15 34.26 45.66 34.24 45.59 33.68
70 60.03 46.83 54.40 39.48 54.21 39.44 53.65 39.48 53.59 39.04
80 69.48 54.69 62.61 44.45 62.23 44.56 61.58 44.74 61.46 44.29
90 79.27 61.12 71.13 49.55 70.76 50.10 69.86 50.11 69.74 49.56
100 88.69 68.59 79.44 55.04 79.01 54.91 78.06 55.20 77.86 54.13

5 Conclusion
In this paper, we studied the problem of Breakpoint Graph Decomposition, which
is associated with the genome rearrangement distance on unsigned permutations.
We developed a greedy algorithm and an algorithm based on the Tabu Search
metaheuristic for this problem. In our experiments, the proposed algorithms
yielded better results than the algorithm proposed by Lin and Jiang [10]. The
Tabu Search algorithm, which is given a feasible solution and looks for better
ones through local search, was able to improve the solutions returned by all the
algorithms.
140 P. O. Pinheiro et al.

We also evaluated the solutions in our experiments using them in algorithms

for the rearrangement distance that receive a decomposition of the breakpoint
graph as input. As before, the performance of our algorithms was significantly
better than that of the approximation algorithm of Lin and Jiang.
As future works, we intend to develop algorithms based on the GRASP and
Genetic Algorithms metaheuristics.

Acknowledgments. This work was supported by the National Council of Techno-

logical and Scientific Development, CNPq (425340/2016-3), the Coordenação de Aper-
feiçoamento de Pessoal de Nı́vel Superior - Brasil (CAPES) - Finance Code 001 , and
the São Paulo Research Foundation, FAPESP (grants 2013/08293-7 , 2015/11937-9 ,
2017/12646-3 , 2019/25410-3 , and 2019/27331-3 ).

References
1. Bafna, V., Pevzner, P.A.: Genome rearrangements and sorting by reversals. SIAM
J. Comput. 25(2), 272–289 (1996)
2. Bulteau, L., Fertin, G., Rusu, I.: Sorting by transpositions is difficult. SIAM J.
Discret. Math. 26(3), 1148–1180 (2012)
3. Caprara, A.: Sorting permutations by reversals and Eulerian cycle decompositions.
SIAM J. Discret. Math. 12(1), 91–110 (1999)
4. Chen, X.: On sorting unsigned permutations by double-cut-and-joins. J. Comb.
Optim. 25(3), 339–351 (2013). https://doi.org/10.1007/s10878-010-9369-8
5. Christie, D.A.: A 3/2-approximation algorithm for sorting by reversals. In: Pro-
ceedings of the 9th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA
1998), pp. 244–252. Society for Industrial and Applied Mathematics, Philadelphia
(1998)
6. Glover, F.W.: Tabu search - Part I. INFORMS J. Comput. 1(3), 190–206 (1989)
7. Glover, F.W.: Tabu search - Part II. INFORMS J. Comput. 2(1), 4–32 (1990)
8. Hannenhalli, S., Pevzner, P.A.: Transforming cabbage into turnip: polynomial algo-
rithm for sorting signed permutations by reversals. J. ACM 46(1), 1–27 (1999)
9. Jiang, H., Pu, L., Qingge, L., Sankoff, D., Zhu, B.: A randomized FPT approxima-
tion algorithm for maximum alternating-cycle decomposition with applications. In:
Wang, L., Zhu, D. (eds.) COCOON 2018. LNCS, vol. 10976, pp. 26–38. Springer,
Cham (2018). https://doi.org/10.1007/978-3-319-94776-1 3
10. Lin, G., Jiang, T.: A further improved approximation algorithm for breakpoint
graph decomposition. J. Comb. Optim. 8(2), 183–194 (2004). https://doi.org/10.
1023/B:JOCO.0000031419.12290.2b
11. Oliveira, A.R., Brito, K.L., Dias, U., Dias, Z.: On the complexity of sorting by
reversals and transpositions problems. J. Comput. Biol. 26, 1223–1229 (2019)
12. Rahman, A., Shatabda, S., Hasan, M.: An approximation algorithm for sorting by
reversals and transpositions. J. Discret. Algorithms 6(3), 449–457 (2008)
13. Yancopoulos, S., Attie, O., Friedberg, R.: Efficient sorting of genomic permutations
by translocation inversion and block interchange. Bioinformatics 21(16), 3340–3346
(2005)
 Center Genome with Respect
to the Rank Distance

Priscila Biller1,2 , João Paulo Pereira Zanetti4 , and João Meidanis1,3(B)

1
Institute of Computing, University of Campinas, Campinas, Brazil
[email protected]
2
Department of Mathematics, Simon Fraser University, Burnaby, Canada
3
Scylla Bioinformatics, Campinas, Brazil
4
R&D Seismology and Acoustics, Royal Netherlands Meteorological Institute
(KNMI), De Bilt, Netherlands

Abstract. The rank distance between matrices has been applied to

genome evolution, specifically in the area of genome rearrangements. It
corresponds to looking for the optimal way of transforming one genome
into another by cuts and joins with weight 1 and double-swaps with
weight 2. In this context, the genome median problem, which takes
three genomes A, B, and C and aims to find a genome M such that
d(A, M ) + d(B, M ) + d(C, M ) is minimized, is relevant. This problem
can be stated for any genomic distance, not just the rank distance. In
many cases, the genome median problem is NP-hard, but a number of
approximate methods have been developed.
Here we examine a related problem, the so-called center genome prob-
lem, where we aim to minimize the maximum (instead of the sum) of
pairwise distances between the center genome and the inputs. We show
that, for the rank distance, and for two genomic inputs A and B, it is
not possible to always attain the well-known lower bound d(A, B)/2.
The issue arises when A and B are co-tailed genomes (i.e., genomes with
the same telomeres) with d(A, B) equal to twice an odd number, when
the optimal attainable score is 1 unit larger than the lower bound. In all
other cases, we show that the lower bound is attained.

Keywords: Genome rearrangements · Genome matrices

1 Introduction
The rank distance between matrices has been very successfully used in coding
theory since at least 1985, when Gabidulin published his discoveries in matrix
codes [5]. Recently, applications of the rank distance to genome evolution, specifi-
cally in the area of genome rearrangements, started to emerge [9]. In this context,
the genome median problem, which takes a number of genomes A1 , A2 , . . . , Ak
and aims to find a genome M such that d(A1 , M ) + d(A2 , M ) + . . . + d(Ak , M ) is
minimized, is relevant. This problem can be stated for any genomic distance, not

c Springer Nature Switzerland AG 2020

J. C. Setubal and W. M. Silva (Eds.): BSB 2020, LNBI 12558, pp. 141–149, 2020.
https://doi.org/10.1007/978-3-030-65775-8_13
142 P. Biller et al.

just the rank distance. In many cases, the genome median problem is NP-hard,
but a number of approximate methods have been developed.
With regard to genome medians, much work has been published, especially in
the case of exactly three inputs. This is one of the seminal steps in building phylo-
genetic trees. Finding a genome median is NP-hard for several genome distances,
with the exception of SCJ and breakpoint for multichromosomal genomes. [4,8].
Center genomes, also called closest genomes or minimax genomes, are also
aimed at somehow representing all the inputs, as a sort of average genome.
The center genome problem takes genome inputs A1 , A2 , . . . , Ak and looks
for a genome M minimizing max(d(A1 , M ), d(A2 , M ), . . . , d(Ak , M )). There is
an important difference between using central genomes and median genomes as
subroutines for ancestral reconstruction methods: when just two inputs are used
for the median, the solution will probably be not very relevant, because many
solutions exist, including both input genomes and anything in an optimal path
from one to the other; on the other hand, the center genome, even with just
two inputs, is already restricted enough to be relevant with respect to ancestral
genomes.
For any distance defined as the minimum number of operations, when all
operations have the same weight, clearly the theoretical lower bound for two
genomes is readily achievable: it suffices to start at one of the genomes and
walk towards the other, stopping when the right number of steps have been
performed. However, if an arbitrary number of inputs is allowed, the problem
becomes NP-hard, even for very simple distances such as the SCJ [2].
In contrast, distance measures where operations have distinct weights may
not be able to always attain the lower bound. Here we concentrate on two inputs
and examine the rank distance, which can be defined as the rank of A − B for
genomes (matrices) A and B, but also as the minimum number of cuts, joins,
and double swaps, with weights 1, 1, and 2, respectively, that bring one genome
to the other. Since we have different weights, it is not obvious the lower bound
can be achieved. In fact, we show that it cannot in the case where d(A, B) = 2n
with n odd. In all other cases, the lower bound is achieved.
The rest of this paper is organized as follows. Section 2 contains the defini-
tions used throughout the text. Section 3 presents the results. Finally, Sect. 4
summarizes our work and points to possible continuation of this research.

2 Definitions

We will represent genomes as matrices. For a genome G involving n genes and

therefore 2n gene extremities, we choose an ordering for the extremities (any
ordering is fine), and then define the corresponding genome matrix as follows:
⎧
⎪
⎪ 1 if i = j and extremities i and j are adjacent in G, or
⎨
if i = j and extremity i is a telomere in G
Gij =
⎪
⎪ 0 if i = j and extremities i and j are not adjacent in G, or
⎩
if i = j and extremity i is not a telomere in G
 Center Genome with Respect to the Rank Distance 143

For genomes with just one gene, we have just two extremities. There are only
two genomes: one with an adjacency linking these two extremities, and the other
with just telomeres. Here are some examples of genomes over two genes:
⎡ ⎤ ⎡ ⎤
0100 1000
⎢1 0 0 0⎥ ⎢0 1 0 0⎥
C=⎢ ⎥ ⎢
⎣0 0 1 0⎦,D = ⎣0 0 0 1⎦
⎥

0001 0010

Genome matrices are therefore square matrices of size (2n) × (2n) and have
the following properties:

– They are binary matrices, i.e., have 0’s and 1’s only.
– They are symmetric matrices, that is, they satisfy A = A.
– They are orthogonal matrices, that is, they satisfy A = A−1 .
– They are involutions, that is, they satisfy A2 = I.

It is easy to verify that any two of the last three properties implies the third
one. For binary matrices, being an orthogonal matrix is equivalent to having
just one 1 in each row and in each column. Such binary matrices are called per-
mutation matrices. We can then say that genome matrices are permutation
matrices that are involutions.
Extremities x such that Ax = x are called telomeres of A. A genome with no
telomeres is called circular. Two genomes with exactly the same set of telomeres
are called co-tailed.

3 Results
We recall a lower bound for the score relative to two genomes, and show exactly
the cases where it is possible to achieve such a score. We also show that, in any
case, it is always possible to find a genome within 1 unit of the lower bound.
We start by recalling the notion of intermediate genomes, defined as
genomes that appear in an optimal scenario between two genomes A and B. The
definition depends on A and B, so sometimes we will call them AB-intermediates
for improved clarity. Although initially defined for DCJ [3], the definition works
for any distance.
In addition to being optimal scenario members, intermediate genomes can
be characterized as those for which the triangle inequality becomes an equality.
They are also the medians of two genomes.
Given two genomes A, and B, a center genome for them is a genome M
that minimizes the score sc(M ; A, B), defined as:

sc(M ; A, B) = max(d(A, M ), d(B, M )).

The triangle inequality gives almost immediately a lower bound on the score:
144 P. Biller et al.

Lemma 1. For any three genomes A, B, and M we have:

d(A, B)
sc(M ; A, B) ≥ .
2
Proof. Notice that:
d(A, B) ≤ d(A, M ) + d(B, M ) ≤ 2 max(d(A, M ), d(B, M )) = 2sc(M ; A, B).
From this, the statement easily follows.
In fact, since the score is always an integer, we can strengthen this result and
claim that: 
d(A, B)
sc(M ; A, B) ≥ . (1)
2
It would be tempting to state the following conjecture:
Conjecture 1. For any two genomes A and B over the same genes, there is at
least one genome M over the same genes that satisfies:
d(A, M ) = d(A, B)/2
and
d(B, M ) = d(A, B)/2.
This genome would of course be a center genome, since it would attain the
lower bound established in Eq. 1. However, this is false, as can be seen from the
following example representing genomes that differ by a double swap:
⎡ ⎤ ⎡ ⎤
0100 0010
⎢1 0 0 0⎥ ⎢0 0 0 1⎥
A=⎢ ⎥ ⎢
⎣0 0 0 1⎦,B = ⎣1 0 0 0⎦.
⎥

0010 0100
To compute their distance, let’s subtract B from A:
⎡ ⎤
0 1 −1 0
⎢ 1 0 0 −1 ⎥
A−B =⎢ ⎣ −1 0 0 1 ⎦ .
⎥

0 −1 1 0
This matrix has rank 2. Both A and B are circular genomes, since they do
not have telomeres. Now for a circular genome such as A, the only genomes at
distance 1 from it are the ones obtained by cutting an adjacency, since no extra
adjacencies can be added to A. Genome A has only two adjacencies, so there are
just two genomes at distance 1 from it, namely:
⎡ ⎤ ⎡ ⎤
1000 0100
⎢0 1 0 0⎥ ⎢1 0 0 0⎥
A1 = ⎢ ⎥ ⎢ ⎥
⎣ 0 0 0 1 ⎦ , A2 = ⎣ 0 0 1 0 ⎦ .
0010 0001
However, it can be readily verified that none of these two genomes is at
distance 1 from B. In fact, they are both at distance 3 from B. We conclude
that the center conjecture is not true.
 Center Genome with Respect to the Rank Distance 145

3.1 Co-tailed Genomes

When A and C are co-tailed, we do not always get a center genome satisfying
the lower bound, but we can get within 1 unit of it. This case opens up the
possibility of center genomes that are not intermediates, e.g., in the example of
Sect. 3, the identity matrix is a center genome, but not an intermediate. Let’s
begin by studying properties of intermediate genomes between two co-tailed
ones.
Lemma 2. If A and C are co-tailed genomes and B is an intermediate genome
between A and C, then B is co-tailed with A and C.
Proof. It suffices to show that B is co-tailed with A. Suppose for a moment that
B is not co-tailed with A. Then either A has a telomere that B doesn’t, or B
has a telomere that A doesn’t. The first case is ruled out by Corollary 1 of a
paper by Chindelevitch and Meidanis [1], because a telomere of A would also
be a telomere of C, since they are co-tailed, and would have to be shared by all
AC-intermediate genomes.
So let’s assume that B has a telomere x not shared by A. In this case, at
some point in an optimal operation series going from A to B, there must be a
cut. However, any optimal such series can be extended to a sorting series going
from A to C, since B is intermediate between A and C. However, no cuts can
be present in an optimal scenario linking co-tailed genomes [6]. This shows that
B cannot have telomeres not shared with A and C.
Only double swaps occur in optimal sorting scenarios of co-tailed genomes.
This leads to a parity restriction.
Lemma 3. If A and C are co-tailed genomes, and L = [B0 , B1 , . . . , Bk ] is an
optimal scenario going from A to C, then d(A, C) = 2k.
Proof. According to Lemma 2, all Bi ’s are co-tailed with A, so none of the
operations Bi+1 − Bi can be cuts or joins. Therefore, we have r(Bi+1 − Bi ) = 2
for 0 ≤ i ≤ k − 1. But then
k−1
 k−1

d(A, C) = w(L) = r(Bi+1 − Bi ) = 2 = 2k.
i=0 i=0

It is easy to find center genomes for co-tailed genomes whose distance is a

multiple of 4. However, if their distance is not divisible by 4, we are forced to
take the second best, which is 1 unit off the lower bound.
Lemma 4. If A and C are co-tailed genomes and d(A, C)/2 is even, then there
is a genome B satisfying the center lower bound.
Proof. Let [B0 , B1 , . . . , Bk ] be an optimal scenario going from A to C. We know
that d(A, C) = 2k from Lemma 3. Since k = d(A, C)/2 is even, we can write
k = 2m for some integer m. It is then straightforward to verify that Bm is the
sought AC-intermediate genome satisfying the center lower bound.
146 P. Biller et al.

Lemma 5. If A and C are co-tailed genomes and d(A, C)/2 is odd, then there
is no genome B satisfying the center lower bound.

Proof. If such a genome B existed, then we would have:

d(A, B) = d(B, C) = d(A, C)/2.

This implies that B would be an intermediate genome between A and C.

By Lemma 2, B would be co-tailed with A. But then, by Lemma 3, d(A, B) =
d(A, C)/2 would have to be even, contradicting the hypothesis.

Lemma 6. For any two genomes A and C, there is an intermediate genome B

such that
d(A, C)/2 ≤ d(A, B) ≤ d(A, C)/2 + 1.

Proof. If A = C the result is clear taking B = A. If A = C, let [B0 , B1 , . . . , Bk ]

be an optimal scenario going from A to C and take i as the smallest index such
that d(A, Bi ) ≥ d(A, C)/2. We claim that B = Bi is the sought genome. Notice
that B is an intermediate genome between A and C because it is a member of an
optimal scenario going from A to C. Moreover, the first inequality in the lemma
statement is satisfied because of the choice of i.
For the second equality, notice that, by the minimality of i, we have:

d(A, Bi−1 ) < d(A, C)/2.

Genome Bi−1 exists since A = C implies d(A, C)/2 ≥ 1, so i cannot be

zero. Given that in any scenario the steps have weight 1 or 2, we know that
d(Bi−1 , Bi ) ≤ 2. It follows that

d(A, Bi ) ≤ d(A, Bi−1 ) + d(Bi−1 , Bi ) ≤ d(A, Bi−1 ) + 2 < d(A, C)/2 + 2

or
d(A, Bi ) ≤ d(A, C)/2 + 1,
since both sides are integers.

3.2 Genomes Not Co-tailed

If A and C are not co-tailed, then there are AC-intermediate genomes at any
feasible distance between A and C. To ascertain that, we need a few preliminary
lemmas on operation switch and other properties.

Lemma 7. Let A be a genome, P a cut applicable to A, and Q a double swap

applicable to A + P . Then Q is applicable to A.

Proof. Let Q = W (x, y, z, w). We know that Q is applicable to A + P , which

means that A + P has adjacencies xw and yz. Since P is a cut, which only
removes adjacencies, xw and yz must have been present in A as well, leading to
the conclusion that Q can be applied to A.
 Center Genome with Respect to the Rank Distance 147

An analogous result is valid for joins, saying that joins can be brought back
through double swaps, but we won’t need it now.
Lemma 8. Let A and C be two genomes not co-tailed. Then, for every integer
i such that 0 ≤ i ≤ d(A, C) there is an intermediate genome B between A and
C with d(A, B) = i.
Proof. By induction on d(A, C). The base case is d(A, C) = 1, because A and
C are not co-tailed and hence cannot be equal. The statement is clearly true for
d(A, C) because in this case we only have two possibilities for i, namely, i = 0
or i = 1, and we can take B = A for i = 0 and B = C for i = 1.
Now assume d(A, C) ≥ 2 and consider an integer i such that 0 ≤ i ≤ d(A, C).
Since A and C are not co-tailed, there is either a telomere in A not shared by C
or a telomere in C not shared by A. Without loss of generality, we may assume
that there is a telomere in C not shared by A, otherwise we can just exchange
A and C and i with d(A, C) − i.
Given that there is a telomere in C that is not an A-telomere, destroying
the adjacency of x in A gives us a cut P applicable to A such that A + P is an
intermediate genome between A and C. If A + P is not co-tailed with C, we can
apply the induction hypothesis to A + P and C and get intermediate genomes
at an arbitrary distance j from A + P , provided that 0 ≤ j ≤ d(A + P, C) =
d(A, C) − 1, which will be at distance j + 1 from A. This covers all the distances
we need except 0, for which we can take B = A.
Now if A+P is co-tailed with C, then they are distinct, since d(A, A+P ) = 1
and d(A, C) ≥ 2. Co-tailed genomes can be sorted by double swaps, so there is
a double swap Q applicable to A yielding an intermediate genome A + P + Q
between A+P and C. However, according to Lemma 7, a cut can go forward past
a double swap, which means that Q is applicable to A. The resulting genome,
A + Q, is intermediate between A and C because A + Q + P is just another way
of getting to A + P + Q, which we know is intermediate between A and C. We
can then apply the induction hypothesis to A + Q and C, which are not co-tailed
since A + Q is co-tailed with A, obtaining intermediate genomes at distances i
from A for 2 ≤ i ≤ d(A, C). For i = 0 we have A, and for i = 1 we have A + P .
This completes the induction step and the proof of our lemma.

3.3 Main Result

Theorem 1. Let A and C be arbitrary genome matrices over the same genes.
Then:
1. If A and C are not co-tailed, then there is a genome matrix B such that:

d(A, C)
d(A, B) =
2
and  
d(A, C)
d(B, C) = .
2
148 P. Biller et al.

2. If A and C are co-tailed and d(A, C) is a multiple of 4, then there is a genome

matrix B such that:
d(A, C)
d(A, B) =
2
and
d(A, C)
d(B, C) = .
2
3. If A and C are co-tailed and d(A, C) is not a multiple of 4, then there is no
genome matrix B such that:

d(A, C)
d(A, B) =
2
and
d(A, C)
d(B, C) = .
2
However, there is a genome matrix B such that:

d(A, C)
d(A, B) = +1
2
and
d(A, C)
d(B, C) = − 1.
2
Proof. Part 1 is a consequence of Lemma 8, since 0 ≤ d(A, C)/2 ≤ d(A, C).
Part 2 is a consequence of Lemma 4. Part 3 is a consequence of Lemmas 5 and 6.

4 Conclusions

In this paper we showed that center genomes do not always attain the theoreti-
cal lower bound in the case of two genomes, with respect to the rank distance.
In spite of that, their are easy to calculate, and provide an attractive alterna-
tive to the median in ancestral genome reconstruction, even in the two-input
version, which is already more restrictive than its median counterpart. Given
that computing a median is NP-hard for the majority of relevant distances, its
replacement by a center solution would bring a significant gain.
Nevertheless, it would be interesting to extend this analysis to three inputs,
and determine what happens there. Probably the arbitrary input version is NP-
hard, as similar problems with simpler distances have already been proved NP-
hard [2,7]. In addition, considering genomes with unequal gene content would
also be worthwhile.

Acknowledgements. We thank funding agency FAPESP (Brazil) for financial sup-

port (Grant numbers 2012/13865-7, 2012/14104-0, and 2018/00031-7). PB would also
like to acknowledge the Canada 150 Research Chair program.
 Center Genome with Respect to the Rank Distance 149

References
1. Chindelevitch, L., Zanetti, J.P.P., Meidanis, J.: On the rank-distance median of 3
permutations. BMC Bioinform. 19(Suppl 6), 142 (2018). https://doi.org/10.1186/
s12859-018-2131-4
2. Cunha, L.F.I., Feijão, P., dos Santos, V.F., Kowada, L.A., de Figueiredo, C.M.H.:
On the computational complexity of closest genome problems. Discret. Appl. Math.
274, 26–34 (2020)
3. Feijão, P.: Reconstruction of ancestral gene orders using intermediate genomes.
BMC Bioinform. 16(Suppl 14), S3 (2015). https://doi.org/10.1186/1471-2105-16-
S14-S3
4. Feijão, P., Meidanis, J.: SCJ: a breakpoint-like distance that simplifies several rear-
rangement problems. Trans. Comput. Biol. Bioinform. 8, 1318–1329 (2011)
5. Gabidulin, E.M.: Theory of codes with maximum rank distance. Probl. Peredachi
Inf. 21(1), 3–16 (1985)
6. Meidanis, J., Biller, P., Zanetti, J.P.P.: A matrix-based theory for genome rearrange-
ments. Technical Report, IC-18-10. Institute of Computing, University of Campinas,
August 2018
7. Popov, V.: Multiple genome rearrangement by swaps and by element duplications.
Theor. Comput. Sci. 385(1–3), 115–126 (2007)
8. Tannier, E., Zheng, C., Sankoff, D.: Multichromosomal median and halving problems
under different genomic distances. BMC Bioinform. 10(1), 120 (2009). https://doi.
org/10.1186/1471-2105-10-120
9. Zanetti, J.P.P., Biller, P., Meidanis, J.: Median approximations for genomes modeled
as matrices. Bull. Math. Biol. 78(4), 786–814 (2016). A preliminary version appeared
on the Proceedings of the Workshop on Algorithms for Bioinformatics (WABI) 2013
 ImTeNet: Image-Text Classification
Network for Abnormality Detection
and Automatic Reporting on
Musculoskeletal Radiographs

Leodécio Braz , Vinicius Teixeira , Helio Pedrini(B) , and Zanoni Dias

Institute of Computing, University of Campinas, Campinas, Brazil

{leodeciobraz,viniciusteixeira}@liv.ic.unicamp.br,
{helio,zanoni}@ic.unicamp.br

Abstract. Deep learning techniques have been increasingly applied to

provide more accurate results in the classification of medical images and
in the classification and generation of report texts. The main objective of
this paper is to investigate the influence of fusing several features of het-
erogeneous modalities to improve musculoskeletal abnormality detection
in comparison with the individual results of image and text classifica-
tion. In this work, we propose a novel image-text classification frame-
work, named ImTeNet, to learn relevant features from image and text
information for binary classification of musculoskeletal radiography. Ini-
tially, we use a caption generator model to artificially create textual data
for a dataset lacking text information. Then, we apply the ImTeNet, a
multi-modal information model that consists of two distinct networks,
DenseNet-169 and BERT, to perform image and text classification tasks
respectively, and a fusion module that receives a concatenation of fea-
ture vectors extracted from both. To evaluate our proposed approach, we
used the Musculoskeletal Radiographs (MURA) dataset and compare the
results obtained with image and text classification scheme individually.

Keywords: Deep learning · Musculoskeletal abnormalities · X-ray

1 Introduction
Musculoskeletal disorders represent a major health problem that affects a large
part of the population. The X-ray images are one of the most commonly acces-
sible radiological examinations used to detect and locate abnormalities in radio-
graphic studies. The process of interpreting image examinations is a typically
complex task. Specialist doctors are usually responsible for conducting the inter-
pretation of these types of information [8]. The first step is to read and analyze
the images in order to have a knowledge base to write a report on what is in the
image and provide a diagnosis of normal or abnormal, for example.
Determining whether a radiographic study is normal or abnormal is a chal-
lenging problem. If a study is interpreted as normal, the possibility of disease is
c Springer Nature Switzerland AG 2020
J. C. Setubal and W. M. Silva (Eds.): BSB 2020, LNBI 12558, pp. 150–161, 2020.
https://doi.org/10.1007/978-3-030-65775-8_14
 ImTeNet: Image-Text Classification Network 151

(a) Humerus (b) Hand

Fig. 1. Original images from MURA dataset and corresponding heatmaps.

excluded, which may eliminate the need for the patient to undergo other tests
or procedures [13]. In Fig. 1, we present two examples of images from MURA
dataset and fracture location using activation maps.
To assist in the diagnostic process, many automated computational meth-
ods, such as Computer-Aided Diagnosis (CAD) have been explored [10]. With
advances in techniques such as natural language processing (NLP) and computer
vision (CV), various deep learning methods have been developed to automati-
cally interpret medical images and reports in order to assist clinicians who are
subjected to these tasks on a daily basis.
More recently, transformers have shown success in labeling radiological
images [5,16]. However, using these methods, a large amount of resources are
required to annotate the data manually to obtain a higher classification score.
Several methods of deep learning have been developed for the task of classifi-
cation, localization and interpretation of radiology images. Recent advances in
deep convolutional neural network (DCNN) architectures have improved the per-
formance of CAD systems, which support health experts [14]. The advancement
in hardware and software development has made it possible that the amount of
medical data collected per patient has increased considerably [12].
In this work, we propose an image-text classification network, called
ImTeNet, for the automatic detection and notification of fractures on muscu-
loskeletal radiographs. Our method uses information extracted from images and
artificially generated captions to obtain a classification label. Initially, we use a
caption generator model to create textual data for a dataset without text infor-
mation. Then, we apply our multi-modal information model, which consists of
two distinct networks, DenseNet-169 and BERT, to perform image and text clas-
sification tasks respectively, and a fusion module that receives a concatenation
of feature vectors extracted from both.
Our main contributions are summarized as follows: (i) the proposed method
is applied with a clinical objective for diagnostic abnormality recognition, (ii) dis-
tinct classifiers based on deep learning are trained with radiographs and texts for
abnormality classification and (iii) experimental results on the MURA dataset
show that the combination of image and text features can increase the classifi-
cation results.
The text is organized as follows. Section 2 describes some relevant approaches
related to the topic under investigation. Section 3 presents the proposed image-
152 L. Braz et al.

text classification network for musculoskeletal abnormality detection. Section 4

describes the dataset used in the experiments and some implementation details.
Section 5 describes and evaluates the experimental results. Section 6 concludes
the paper with some final considerations.

2 Related Work
Deep learning methods are the state of the art for classification tasks in the med-
ical field. Moreover, Deep Convolutional Neural Networks (DCNNs) are widely
used mainly for image domains [3,13,21]. Many Natural Language Processing
(NLP) methods have been developed to extract structured labels from free-text
medical reports [1,12]. Our work is closely related to approaches that explore
the use of DCNNs on medical images, with the aim of extracting relevant infor-
mation, as well as approaches that explore textual medical information and the
combination of both.
Wang et al. [21] explored the use of many pre-trained DCNN models to
perform a multi-label classification of thoracic disease on chest X-ray images.
Chen et al. [3] proposed a dual asymmetric feature learning network called
DualCheXNet to explore the complementarity and cooperation between the
two networks to learn discriminative features. Rajpurkar et al. [13] proposed
the Musculoskeletal Radiographs (MURA) dataset and explored the use of a
DenseNet-169 [7] to perform a binary classification task.
Smit et al. [16] proposed a method, called CheXbert, for radiology report
labeling, combining existing report labelers with hand-annotations to obtain
accurate results. They used a BERT [4] model and obtained state-of-the-art
results for report labeling on the chest X-ray datasets. Pelka et al. [12] proposed
an approach to combine automatically generated keywords with radiographs,
presenting a method to allow multi-modal image representations by fusing tex-
tual information with radiographic images to perform body part and abnormality
recognition on MURA dataset. In contrast to these works, our approach explores
the cooperation between image and text networks to learn the classification of
abnormalities in a complementary and accurate way.

3 Proposed Method
In this section, we present details of the proposed method. We provide some
notations used throughout this paper, as well as describe the different tasks
performed and how we combine each one to produce our proposed architecture.

3.1 Caption Generation

Since text representations are not available on the MURA [13] dataset, we
used a caption generator model, trained on the Radiology Objects in COntext
(ROCO) [11] dataset, to automatically generate medical reports for the images
on MURA dataset.
 ImTeNet: Image-Text Classification Network 153

The ROCO dataset contains 81,825 radiology images including several medi-
cal imaging modalities, such as X-ray, Computer Tomography (CT), Ultrasound,
Mammography (UM), Magnetic Resonance Imaging (MRI) and various other
modalities. All images on ROCO have a corresponding caption information.
For this caption generation task, inspired by Pelka et al. [12], we propose an
encoder-decoder framework [8,20]. As an encoder, we use a Convolutional Neural
Network (CNN) model, called ResNet-152 [6], pretrained on ImageNet [15]. The
encoder extracts the feature vector from an input image and this feature vector
is linearly transformed to be the input to the decoder network. The decoder is a
Long Short-Term Memory (LSTM) network, which receives as input the feature
vector from the encoder and produces the image caption as output.

<start> x-ray showing <end> MURA dataset

ROCO dataset
Softmax Softmax Softmax Softmax
FC layer

Linear

LSTM

LSTM

LSTM

LSTM
DCNN ...

Wemb Wemb Wemb

<start> x-ray wrist

Train caption generator (CG)

<start> x-ray showing the fracture... <end>

<start> postoperative radiograph... <end> Caption
MURA labels + <start> x-ray showing the right wrist... <end>
generator
...

TCL ICL

BERT DenseNet-169 + CAM

FCL

Linear(6656,256)

Linear(256,128)

Linear(128,1)

Fig. 2. Overview of the proposed method workflow. A DenseNet-169 network and

a BERT model are used to learn from image and text representation, respectively,
given an image and its corresponding caption. The output features of each network are
concatenated in the fusion classifier and the outputs of each classifier are averaged to
provide the final output prediction.

Figure 2 shows the caption generation model (CG) scheme. We trained the
model using a corpus of paired image and captions from the ROCO dataset1 .
After the model training step, we construct a dataset, called MURA caption,
executing the caption generator for each image on MURA. At the end of this
task, we have a combination of an original image and label from MURA and an
1
No additional datasets were used for training.
154 L. Braz et al.

automatically generated caption associated with it. Figure 3 shows examples of

these captions for four radiographs from the MURA training set.

Set: Shoulder Set: Wrist Set: Shoulder Set: Humerus

Label: Normal Label: Normal Label: Abnormal Label: Abnormal
Caption: the right Caption: X-ray Caption: X-ray of Caption: X-ray
shoulder is normal showing the of the the shoulder show- showing the fracture
right wrist ing the acromioclav- of the right humerus
icular joint

Fig. 3. Examples of generated captions. All of this information is present on the

MURA caption dataset. The caption generator model was trained using all images
from the ROCO dataset.

3.2 Image-Text Classification Framework

Our focus in this work is to develop and evaluate an approach based on deep
learning that, using multi-modal information, can make the detection of abnor-
malities in musculoskeletal radiograph samples more accurate. The intuition
behind our proposed method is that the combination of visual and text informa-
tion will benefit the classification task, improving its results. Figure 2 illustrates
the main steps of our architecture. After the caption generation task, three steps
are applied: (i) Image Classification Level (ICL), (ii) Text Classification Level
(TCL), and (iii) Fusion Classification Level (FCL).
1. ICL Step: At this step, we trained a DCNN model, called DenseNet-
169 [7], pre-trained on ImageNet [15], to perform an image classification. We
also used this model as a feature extractor. In this model, we added an atten-
tion module called Class Activation Mapping (CAM) [22], which is employed
to indicate the discriminative region detected by the DCNN model to iden-
tify the correct class. After this attention module, we applied an Average-Max
(AVG-MAX) pooling layer to reduce computational complexity and extract low
(average) and high (max) level features from the neighborhood. The AVG-MAX
pooling is the concatenation of the average pooling and max pooling results.
2. TCL Step: In this step, we proposed a Natural Language Processing
(NLP) approach to extract structured labels from free-text image captions. We
fine-tuned the BERT [4] model to perform a text classification. The BERT model
architecture is based on a multilayer bidirectional transformer [18], and pre-
trained in an unsupervised way in two tasks: masked language model and next
 ImTeNet: Image-Text Classification Network 155

sentence prediction [2]. Our proposed method follows the same architecture as
BERT. Each image caption text is tokenized, and the maximum number of tokens
in each input sequence is limited to 64. The final-layer hidden state is then fed as
input to each of the BERT linear heads. We changed the BERT output dimen-
sion to 1 to cover our binary classification problem. We also used this model
as a feature extractor and then applied an Average-Max (AVG-MAX) pooling
operation in hidden state layers.
3. FCL Step: At this step, the output features from ICL and TCL models
are concatenated into a single feature vector used as input to the FCL. In the
fusion classifier, the M input features are fed into three dense layers. As a loss
function, we used the Binary Cross Entropy Loss (BCEL).

4 Experimental Setup

In this section, we present details about the dataset used in our experiments, as
well as some implementation details.

4.1 Musculoskeletal Radiographs (MURA) for Abnormality

Classification

Musculoskeletal Radiographs (MURA) is a large dataset of bone X-rays [13].

MURA dataset consists of 14,863 musculoskeletal studies of the upper extrem-
ity, from 12,173 patients, containing a total of 40,561 multi-view radiographic
images. All radiographs presented on MURA belong to one of the seven standard
upper extremities (elbow, finger, forearm, hand, humerus, shoulder and wrist)
and were labeled as normal or abnormal by board-certified radiologists from the
Stanford Hospital [13]. Figure 4 shows radiographs representing each of the seven
anatomy classes.

Wrist Shoulder Humerus Elbow Finger Hand Forearm

Fig. 4. Examples of radiographs present on MURA dataset. The images ‘Forearm’,

‘Humerus’ and ‘Shoulder’ belong to the abnormality positive class, whereas the images
‘Elbow’, ‘Wrist’, ‘Finger’ and ‘Hand’ belong to the abnormality negative class. All
images were randomly chosen from the MURA training set.

The original dataset was split into training set(11,184 patients, 13,457 stud-
ies, 36,808 images), validation set (783 patients, 1,199 studies, 3,197 images),
and test set (206 patients, 207 studies, 556 images).
156 L. Braz et al.

For the purpose of comparing our results and evaluating our proposed
method, as the test set is not public accessible, the validation set was adopted
as a test set. In addition, we split stratified the original training set into 80%
and 20% to generate new training and validation sets, respectively.

4.2 Implementation Details

We implemented the proposed image-text classification framework with the
PyTorch deep learning toolbox on a GTX Titan V 12 GB GPU.
For the caption generation, we used the Adam [9] optimizer with a batch
size of 64, learning rate of lr = 1e−3, with a vocabulary size of 16,380. We
employed the same hyper-parameters as those based on the work described by
Vinyals et al. [20]. For the image-text classification, in the image level step, we
initially resized each original image to 384 × 384 pixels and randomly cropped
it to 320 × 320 pixels. In the training stage, we applied a random horizontal
flip. We used the Adam optimizer with a mini-batch size of 8 and learning rate
of lr = 1e−4, reducing by a factor of 1e−1 when the validation loss reaches a
plateau.
In the text level step, the BERT model was trained using a Mean-Square
loss, as we changed the default number of labels to 1 and Adam optimization
with a learning rate of lr = 2e − 5, as used by Devlin et al. [4] for fine-tuning
tasks. Finally, in the fusion level step, we received the feature vectors from image
and text levels and concatenated them, resulting in a single feature vector of size
6,656. For the training fusion model, we used the same hyper-parameter as the
image level.

5 Results and Discussion

Our proposed model takes one or more views for a study case as input. In
each view, our method predicts the probability of abnormality. We compute
the abnormality classification for the study by taking the majority abnormality
classification occurrence output by network for each image.2 The model makes
the binary prediction of abnormal if the probability of abnormality for the image
is greater than 0.5.
To assess the effectiveness of the entire proposed method, we compare the
results of the image, text and fusion classifier level individually and the proposed
ImTeNet architecture, which performs a majority vote of labels of the three
aforementioned classifiers. In Table 1, we report the performance of only image
and text classification in the DenseNet-169 and BERT models, respectively. In
addition, we report the results obtained with the proposed fusion classifier and
with our final prediction approach.
For ‘Elbow’ and ‘Wrist’ studies, the performance of ImTeNet is lower than
the fusion classifier individually and both performance rates are higher than
2
If the normal and abnormal classification occurrences are equal, we perform an
arithmetic mean of the probabilities.
 ImTeNet: Image-Text Classification Network 157

Table 1. We compare the performance of the final predictions of DenseNet-169, BERT,

Fusion and our ImTeNet, applying a voting method among the three proposed clas-
sifiers, with Balanced Accuracy and Cohen’s kappa statistic. We highlight the best
(bold) and the second best (underline) metric rates in each type of the studies and in
aggregate.

DenseNet-169 BERT Fusion ImTeNet

Balanced Kappa Balanced Kappa Balanced Kappa Balanced Kappa
accuracy accuracy accuracy accuracy
Elbow 0.8537 0.7214 0.6322 0.2562 0.8731 0.7512 0.8634 0.7364
Finger 0.8222 0.6521 0.6747 0.3446 0.8300 0.6651 0.8294 0.6647
Forearm 0.8292 0.6652 0.5354 0.0705 0.8471 0.6974 0.8526 0.7114
Hand 0.7680 0.5722 0.4578 -0.0882 0.7835 0.5935 0.7707 0.5745
Humerus 0.9038 0.8074 0.5868 0.1730 0.9039 0.8075 0.9113 0.8222
Shoulder 0.7724 0.5456 0.5473 0.0934 0.7737 0.5467 0.7840 0.5673
Wrist 0.8584 0.7390 0.5997 0.1973 0.8806 0.7772 0.8738 0.7672
Average 0.8296 0.6718 0.5762 0.1495 0.8417 0.6912 0.8407 0.6920

the results of DenseNet-169 and BERT individual. For ‘Finger’ study, the per-
formance of ImTeNet is comparable to the fusion performance, which presents
the best results. The performance of the ImTeNet presents the best results for
‘Shoulder’, ‘Humerus’ and ‘Forearm’ studies. In the latter, in particular, our
ImTeNet provides the greatest gain compared to the DenseNet-169 model. For
‘Hand’ study, the BERT model presents a negative Kappa, indicating that there
is no agreement on its results and the performance of ImTeNet is comparable to
the performance of DenseNet-169 individually, with no relevant gain, the fusion
classifier individually presents the best results. The results of the BERT model,
as expected due the caption format, show the worst results in all studies.
Compared to the baseline results of DenseNet-169, the proposed method
presented an overall performance gain, even with artificially generated text data
where most contain texts with poor quality or lacking relevant information. In
general, our model was able to extract and combine discriminative features from
each proposed branch (ICL and TCL). Our method obtained an average balanced
accuracy and Cohen’s kappa statistic score of 0.8407 and 0.6920, respectively,
with a significant difference of 0.0138 and 0.0202 for the DenseNet-169 individu-
ally. We also evaluate our proposed ImTeNet and compare with another existing
method for musculoskeletal abnormality classification [12] on MURA dataset,
under the same ruled test set, defined in Sect. 4.1.
To allow a fair comparison, we report the evaluation results of each method
with the accuracy scores, as shown in Table 2. The ‘Visual’ column is related
to the image classification task, the ‘Text’ column is associated with the text
classification task, whereas the last column is related to the proposed method.
Compared to this baseline, our proposed ImTeNet yields a high performance for
musculoskeletal abnormality classification on the MURA dataset. Our method
has an accuracy score of 0.8511 with a difference of 0.0356 for the best result
method by Pelka et al. [12].
158 L. Braz et al.

Table 2. Comparison against other approaches on the MURA dataset. We compute

the mean accuracy score for each method. The best result is highlighted (bold).

Method Visual Text Proposed

Pelka et al. [12] 0.7985 0.5424 0.8155
ImTeNet 0.8418 0.5807 0.8511

5.1 Analysis

We analyzed specific examples from the test set and their prediction on each
level, as illustrated in Fig. 5. In addition, we applied the BERTViz [19] tool to
visualize the BERT model attention produced in each input caption.
In the first two examples, all levels were able to correctly label the input
image and text. The BERT model correctly detected relevant information and
was able to discard spurious information present in both captions. The DenseNet-
169 model could also identify relevant regions in the image. On the third example,
which contains poor textual information with absolutely no semantic informa-
tion, the BERT model incorrectly labeled the abnormality as negative, while the
DenseNet-169 labeled the abnormality as positive with a high confidence. It is
also possible to observe a well-defined highlighted area of the radiograph that
was most important in making the prediction. In the last example, while the
DenseNet-169 model was unable to detect the presence of abnormalities in the
image, the BERT model was able to distinguish relevant information from the
caption and correctly label the abnormality as positive. All features and pre-
dictions performed by DenseNet-169 and BERT model were considered in the
fusion level classifier, which could increase the classification results.
Notwithstanding, our study has some limitations. First, our approach relies
on the existence of a good caption generator model. Second, and very related to
the first, our proposed generator model is designed to deal with musculoskeletal
radiograph texts, however, MURA does not have this type of data. We tried
to provide this information by artificially creating textual data, however, when
observing examples of generated captions, many samples do not have relevant
information, such as the caption of the third image in Fig. 5, in which we expected
a text containing information such as “x-ray”, “implant” and “humerus”, how-
ever, <start> the of the the of the the of the <end> was obtained. Third, the
average length of our caption texts is lower than other datasets usually per-
formed with the BERT model [17]. We conjecture that longer text information
can increase the BERT results.
 ImTeNet: Image-Text Classification Network 159

Fig. 5. Original image, DenseNet-169 heatmap and artificially generated caption from
test set. A check mark indicates the correct label prediction for each level. The
BERTViz tool was used to visualize the BERT attention maps for each caption.

6 Conclusions and Future Work

In this study, we proposed an automatic caption generator for musculoskeletal
radiograph images and a method for automatically combining these artificially
generated captions with radiographs for accurate musculoskeletal abnormality
classification. A DCNN model was trained in musculoskeletal radiograph images.
In parallel, the BERT model was trained in automatically generated captions
associated with each musculoskeletal radiograph image. Finally, we proposed a
data fusion approach, that is carried out by combining features from different
heterogeneous modalities, to concatenate features extracted from DCNN and
BERT models. This process allowed for an enriched multimodal data represen-
tation used as input to a fusion module. This multi-modal data representation
presented the highest prediction results.
To create a caption generation model, image-caption pairs from the Radi-
ology Objects in COntext (ROCO) dataset were adopted to train our encoder-
decoder model framework. We use this caption generation model to generate text
160 L. Braz et al.

representations for the Musculosketal Radiograph (MURA) dataset. Through

the proposed multimodal data representation method (ImTeNet), we outper-
formed the results achieved only with image and text features individually with
DenseNet-169 and BERT, respectively.
The proposed work can be further improved through the exploration of bet-
ter caption generator models trained with specific musculosketal text represen-
tations to avoid text quality problems, as mentioned in Sect. 5.1. In addition,
other fusion methods to combine features of different heterogeneous modalities
could be explored. Since there are several tasks and sources in medicine, the pro-
posed work has great potential to be applied in other medical areas by combining
metadata with other imaging modalities.

Acknowledments. The authors would like to thank FAPESP (grants #2015/11937-9,

#2017/12646-3, #2017/16246-0, #2017/12646-3 and #2019/20875-8), CNPq (grants
#304380/2018-0 and #309330/2018-1) and CAPES for their financial support.

References
1. Annarumma, M., Withey, S.J., Bakewell, R.J., Pesce, E., Goh, V., Montana, G.:
Automated triaging of adult chest radiographs with deep artificial neural networks.
Radiology 291(1), 196–202 (2019)
2. Beltagy, I., Lo, K., Cohan, A.: SciBERT: A Pretrained Language Model for Scien-
tific Text. arXiv preprint arXiv:1903.10676 (2019)
3. Chen, B., Li, J., Guo, X., Lu, G.: DualCheXNet: dual asymmetric feature learning
for thoracic disease classification in chest X-rays. Biomed. Signal Process. Control
53, 101554 (2019)
4. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: pre-training of
deep bidirectional transformers for language understanding. arXiv preprint
arXiv:1810.04805 (2018)
5. Drozdov, I., Forbes, D., Szubert, B., Hall, M., Carlin, C., Lowe, D.J.: Supervised
and unsupervised language modelling in chest X-ray radiological reports. PLoS
ONE 15(3), e0229963 (2020)
6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition.
In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016)
7. Huang, G., Liu, Z., van der Maaten, L., Weinberger, K.Q.: Densely connected
convolutional networks. In: IEEE Conference on Computer Vision and Pattern
Recognition (CVPR) (2017)
8. Jing, B., Xie, P., Xing, E.P.: On the automatic generation of medical imaging
reports. In: 56th Annual Meeting of the Association for Computational Linguistics
- Proceedings of the Conference (Long Papers), vol. 1, pp. 2577–2586 (2018)
9. Kingma, D.P., Ba, J.: Adam: a method for stochastic optimization, pp. 1–15. arXiv
preprint arXiv:1412.6980 (2014)
10. Kooi, T., et al.: Large scale deep learning for computer aided detection of mam-
mographic lesions. Med. Image Anal. 35, 303–312 (2017)
11. Pelka, O., Koitka, S., Rückert, J., Nensa, F., Friedrich, C.M.: Radiology objects
in context (ROCO): a multimodal image dataset. In: Stoyanov, D., et al. (eds.)
LABELS/CVII/STENT -2018. LNCS, vol. 11043, pp. 180–189. Springer, Cham
(2018). https://doi.org/10.1007/978-3-030-01364-6 20
 ImTeNet: Image-Text Classification Network 161

12. Pelka, O., Nensa, F., Friedrich, C.M.: Branding - fusion of meta data and mus-
culoskeletal radiographs for multi-modal diagnostic recognition. In: International
Conference on Computer Vision Workshop (ICCV), pp. 467–475 (2019)
13. Rajpurkar, P., et al.: MURA: large dataset for abnormality detection in muscu-
loskeletal radiographs. arXiv preprint arXiv:1712.06957 (2017)
14. Ranjan, E., Paul, S., Kapoor, S., Kar, A., Sethuraman, R., Sheet, D.: Jointly
learning convolutional representations to compress radiological images and clas-
sify thoracic diseases in the compressed domain. In: 11th Indian Conference on
Computer Vision, Graphics and Image Processing, pp. 1–8 (2018)
15. Russakovsky, O., et al.: ImageNet large scale visual recognition challenge. Int. J.
Comput. Vis. 115(3), 211–252 (2015). https://doi.org/10.1007/s11263-015-0816-y
16. Smit, A., Jain, S., Rajpurkar, P., Pareek, A., Ng, A.Y., Lungren, M.P.: CheXbert:
combining automatic labelers and expert annotations for accurate radiology report
labeling using BERT. arXiv preprint arXiv:2004.09167 (2020)
17. Sun, C., Qiu, X., Xu, Y., Huang, X.: How to fine-tune BERT for text classification?
In: Sun, M., Huang, X., Ji, H., Liu, Z., Liu, Y. (eds.) CCL 2019. LNCS (LNAI),
vol. 11856, pp. 194–206. Springer, Cham (2019). https://doi.org/10.1007/978-3-
030-32381-3 16
18. Vaswani, A., et al.: Attention is all you need. In: Guyon, I., et al. (eds.) Advances
in Neural Information Processing Systems 30, pp. 5998–6008 (2017)
19. Vig, J.: A Multiscale visualization of attention in the transformer model, pp. 1–6.
arXiv preprint arXiv:1906.05714 (2019)
20. Vinyals, O., Toshev, A., Bengio, S., Erhan, D.: Show and tell: a neural image cap-
tion generator. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit. 7(12),
3156–3164 (2015)
21. Wang, X., Peng, Y., Lu, L., Lu, Z., Bagheri, M., Summers, R.M.: ChestX-ray8:
hospital-scale chest X-ray database and benchmarks on weakly-supervised clas-
sification and localization of common thorax diseases. In: IEEE Conference on
Computer Vision and Pattern Recognition (CVPR) (2017)
22. Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., Torralba, A.: Learning deep fea-
tures for discriminative localization. In: IEEE Conference on Computer Vision and
Pattern Recognition (CVPR) (2016)
 A Scientometric Overview of Bioinformatics
Tools in the Pseudomonas Putida Genome Study

Alex Batista Trentin(B) , Beatriz Thomas Metzner , Glecio Oliveira Barros ,

and Deborah Catharine de Assis Leite

Federal University of Technology – Paraná, Dois Vizinhos, Brazil

[email protected]

Abstract. Pseudomonas putida is a microorganism widely used in environmental

science due to its high degradation power of recalcitrant compounds. This study
aimed to perform a scientometric analysis of the global panorama of publications
on the P. putida genome, in addition to the bioinformatics tools mainly used.
The growth of publications on the P. putida genome is continuous until 2020,
being France, Spain and USA the main countries with publications on the sub-
ject. Illumina was the main sequencing platform used and, in this set of articles,
120 genomes were sequenced, 106 complete and 14 drafts. The main assembly
software was SPAdes, comprising 23.3% of the articles, and NCBI PGAP was the
main genome annotation tool, in 25% of the documents. Thus, this study allowed
the visualization of the main bioinformatics tools used for the analysis of the P.
putida genome, besides presenting the advance of the research on this subject, and
also supporting next studies with P. putida.

Keywords: Bioinformatic tools · Scientometric analysis · Pseudomonas putida

1 Introduction

The planet is fighting against all types of environmental pollution, with the soil being
one of the places most affected by environmental degradation. There is, therefore, a need
to control soil pollution in order to maintain fertility as well as productivity [1, 2]. In
this context, the development of technologies to remedy these degraded environments
becomes indispensable [3].
Pseudomonas genera, belonging to the Gammaproteobacteria class, have been
extensively studied over time due to their great potential for degradation and biotrans-
formation of xenobiotic and recalcitrant compounds, and are likely to be used in vari-
ous biotechnological processes of environmental recovery [4]. Pseudomonas putida is
an ubiquitous bacterium, mainly in soil, classified as chemoorganotrophic, capable of
metabolizing large carbon chains, as well as several recalcitrant pollutants [5].
From the sequencing of the genome, as well as from bioinformatics analyses, it is
possible to know the proteins and metabolic pathways of Pseudomonas putida applied
to environmental remediation and degradation of xenobiotic compounds [6]. Thus,

© Springer Nature Switzerland AG 2020

J. C. Setubal and W. M. Silva (Eds.): BSB 2020, LNBI 12558, pp. 162–167, 2020.
https://doi.org/10.1007/978-3-030-65775-8_15
 A Scientometric Overview of Bioinformatics Tools 163

using molecular level analysis it is possible to study new techniques and environmental
applications for this organism [7].
This work aimed to perform a systematic analysis, by means of the scientomet-
ric methodology, of the global publications on the genome of Pseudomonas putida,
analyzing mainly the bioinformatics tools used for the assembly and annotation of the
sequenced genomes.

2 Methods
The searches for performing the scientometric analysis were performed in the Web
of Science database using the terms “Genome” AND “Pseudomonas putida” AND
“soil”. Initially 159 documents were found on the researched terms, however, a manually
filtering was performed in order to find only those documents that reported the sequencing
process of the microorganism. After the filtering, 29 papers were left, being 100% journal
articles, containing the sequencing of the genome.
After filtering, the data was extracted to Microsoft Excel and Citespace software in
order to produce graphs of data and connections on: main countries with publications
on the subject, knowledge areas applied, keywords most used, number of publications
and citation per year. Also, all the documents were read, in order to extract the quantity
of sequenced genomes, the sequencing platforms, besides the genome annotation and
assembly softwares, these data were tabulated and had graphs generated in Excel.

3 Results and Discussion

The first publication studying the genomic sequencing of Pseudomonas putida bac-
terium associated with soil, within the database, was in 2002, followed by a gap of 11
years, only starting the publications again in 2014, however, this time with a continuous
growth of publications until 2020. The rate of citations on the subject grew continu-
ously from 2002, totaling 1129 citations until August 2020, as shown in the Fig. 1. The
growth in genomic studies of this soil-related organism is due to the extremely versatile
metabolism of P. putida, it’s great capacity of adaptation in several environments, the
resistance to physical-chemical stresses, and the genes associated with degradation of
recalcitrant compounds. This way, studies using this organism in biological remediation
processes have maintained constant growth [8, 9]. Also, the top 5 of the main countries
that published on the subject are: Spain, Japan, USA, Germany and France. The number
of publications by country each year is also shown in Fig. 1.

3.1 Knowledge Areas

From the analysis performed at CiteSpace Software it is possible to see that Micro-
biology, Environmental Sciences and Biotechnology are the main areas of knowledge
associated with the theme, fact based on the search direction of the research, aiming
at the application of P. putida in soils, due to the high diversity of metabolisms of this
microorganism associated with bioremediation [10]. The Fig. 2 presents the network of
164 A. B. Trentin et al.

Fig. 1. Number of publications by country and year, and the number of citations per year

Fig. 2. Network of connections over knowledge areas. Yellow circles represent the centrality.
(Color figure online)

connections related to the areas of study, and the higher the font of the letter, the more
frequently this area is used in these researches.
Also, the yellow circle represents the centrality, a factor that indicates the amount
of connections that the theme performs [11], thus, it is visualized that Microbiology is
the only area that presents a significant centrality, due to the fact that this set of data
approaches the use of a microorganism, therefore, even the most specific researches are
still interconnected with microbiology, making microbiology coherent as a central area
of knowledge [12].

3.2 Main Journals

In this data set 12 journals published the articles, most of them related to microbiol-
ogy, environmental science or genomics. The main journal was Microbiology Resource
 A Scientometric Overview of Bioinformatics Tools 165

Announcements, with 9 articles published, 31% of the 29 documents. Table 1 presents

all journals, along with the number of documents and the impact factor.

Table 1. Journals and number of published articles

Journal Papers Impact factor

Microbiology Resource Announcements 9 0.89
Environmental Microbiology 3 4.93
Environmental Microbiology Reports 3 2.97
Amb Express 2 2.49
Applied And Environmental Microbiology 2 4.016
Current Microbiology 2 1.73
Environmental Science And Pollution Research 2 3.30
Plos One 2 2.74
Biotechnology And Bioengineering 1 4.00
Microbial Ecology 1 3.86
Scientific Reports 1 3.99
Standards In Genomic Sciences 1 1.44

3.3 Sequencing Data

In total, 117 genomes of Pseudomonas putida were sequenced in the 26 documents,
106 of them complete and 14 drafts. Four sequencing platforms were used for these
works, and some of these documents used more than one platform, the main one being
Illumina, in 17 papers; followed by Roche 54 GS FLX, in 5 papers; Pac Bio, in 4 of the
26 documents; and Ion Torrent, in 1 document. Figure 3A and B presents the percentages
of whole/draft genome and the main platforms.

Fig. 3. (A) Percentage of 120 whole/draft genome. (B) Percentage of the main sequencing
platforms used.
166 A. B. Trentin et al.

3.4 Assembly

In total, 9 genome assembly softwares were used, the most used being SPAdes, one of the
main bacterial genome assembly softwares [13]. Other software such as CLC Genomics
Workbench, Newbler and HGAP were used, as shown in Fig. 4. However, many papers
(6 out of 29) did not inform the assembly software.

Fig. 4. Genome assembly softwares used

3.5 Annotation

In the same way as in the assembly, most of the articles did not inform the genome
annotation software (7 of the 26 documents). There were 5 softwares used in the 26
papers, the main one being the NCBI Prokaryotic Genome Annotation Pipeline, a tool
in constant modification and evolution for the annotation process [14]. Figure 5 shows
the software used along with the percentage within the 26 papers.

Fig. 5. Genome annotation softwares used in the data set

 A Scientometric Overview of Bioinformatics Tools 167

4 Conclusions
Pseudomonas putida is a widely studied organism with an extreme potential in the area
of bioremediation of affected environments. In this context, the genomic study of this
microorganism is of extreme relevance, since from the analyses at molecular level it is
possible to know the main metabolic pathways, as well as the genes and proteins used
by P. putida to perform the degradation processes of recalcitrant compounds.
In this way, the bioinformatics tools are able to analyze these molecular data and
allow the development of new technologies using this microorganism, besides allowing
the dissemination of genomic data. Thus, this systematic study allowed the visualization
of the main bioinformatics tools used for the analysis of the P. putida genome, besides
presenting the advance of the research on the subject, helping possible future studies
with P. putida with application in soil environmental science and other environments.

References
1. Ashraf, A.M., Maah, J.M., Yusoff, I.: Soil contamination, risk assessment and remediation,
vol. 1, no. 1, pp. 3–5. Intech (2014)
2. Ranga, P., Sharma, D., Saharan, S.B.: Bioremediation of azo dye and textile effluents using
Pseudomonas putida MTCC 244. Asian J. Microbiol. Biotechnol. Environ. Sci. 22(2), 88–94
(2019)
3. Horemans, B., Breugelmans, P., Saeys, W., Springael, D.: A soil-bacterium compatibility
model as a decision-making tool for soil bioremediation. Environ. Sci. Technol. 51(3), 1605–
1615 (2017)
4. Loh, K.C., Cao, B.: Paradigm in biodegradation using Pseudomonas putida—A review of
proteomics studies. Enzyme Microbial Technol. 43(1), 1–12 (2008)
5. Tsirinirindravo, H.L., et al.: Bioremediation of soils polluted by petroleum hydrocarbons by
Pseudomonas putida. Int. J. Innov. Eng. Sci. Res. 2(5), 9–18 (2018)
6. Iyer, R., Iken, B., Damania, A., Krieger, J.: Whole genome analysis of six organophosphate-
degrading rhizobacteria reveals putative agrochemical degradation enzymes with broad
substrate specificity. Environ. Sci. Pollut. Res. 25(14), 13660–13675 (2018)
7. Nelson, K.E., et al.: Complete genome sequence and comparative analysis of the metabolically
versatile Pseudomonas putida KT2440. Environ. Microbiol. 4(12), 799–808 (2002)
8. Crovadore, J., et al.: Whole-genome sequence of Pseudomonas putida strain 1312, a potential
biostimulant developed for agriculture. Microbiol. Resour. Announc. 10(7), 1–2 (2018)
9. Weimer, A., et al.: Industrial biotechnology of Pseudomonas putida: advances and prospects.
Appl. Microbiol. Biotechnol. 104(1), 7745–7766 (2020)
10. Abyar, H., et al.: The role of Pseudomonas putida in bioremediation of naphthalene and
copper. World J. Fish Marine Sci. 3(5), 444–449 (2011)
11. Ping, Q., He, J., Chen, C.: How many ways to use CiteSpace? A study of user interactive
events over 14 months. J. Assoc. Inf. Sci. Technol. 68(5), 1234–1256 (2017)
12. Chen, C.: Manual do CiteSpace, 1st edn. Drexel University, Philadelphia (2014)
13. Prjibelski, A., et al.: Using SPAdes de novo assembler. Curr. Prot. Bioinform. 70(1), 1–29
(2020)
14. Tatusova, T., et al.: NCBI prokaryotic genome annotation pipeline. Nucl. Acids Res. Adv.
1(1), 1–11 (2016)
 Polysome-seq as a Measure
of Translational Profile
from Deoxyhypusine Synthase Mutant
in Saccharomyces cerevisiae

Fernanda Manaia Demarqui(B) , Ana Carolina Silva Paiva ,

Mariana Marchi Santoni , Tatiana Faria Watanabe,
Sandro Roberto Valentini , and Cleslei Fernando Zanelli

Department of Biological Sciences, School of Pharmaceutical Sciences,

São Paulo State University (UNESP), Araraquara – SP 14800-903, Brazil
[email protected]

Abstract. The profile of proteins observed in a cell is characterized by

the control of gene expression, which has several regulation points acting
individually or in concert, such as epigenetic, transcriptional, transla-
tional, post-transcriptional or post-translational modification. Copulat-
ing the total mRNA data and mRNAs actively translated can facilitate
the identification of the key regulatory points of gene expression. Here,
we analyze the transcriptional and translational profiles of the deoxy-
hypusine synthase mutant dys1-1 in yeast. This enzyme is involved in
the post-translational modification of translation factor eIF5A, which
has an important role in the elongation translational process. This work
presents gene expression data from the total mRNA levels and the
polysomally-loaded mRNAs for the Saccharomyces cerevisiae DYS1 and
dys1-1 strains, based on RNA-seq and Polysome-seq. Our results showed
that for this mutant, most of the changes in the transcripts forwarded for
translation are due to transcriptional control; and, to solve translation
problems, cell responds with positive regulation of ribosome biogenesis.
Besides, polysome-seq as a tool to study translation profiles is useful to
understand gene expression changes.

Keywords: eIF5A · Gene regulation · Ribosome biogenesis

1 Introduction
Protein synthesis consists of decoding the messenger RNA. This process is cat-
alyzed by ribosomes and mediated by translation factors. The regulation of the
repertoire of proteins expressed in a cell is determined by the selective con-
trol of gene expression by several cellular mechanisms, such as epigenetic, tran-
scriptional, translational, post-transcriptional or post-translational modification
[4,18,21].
c Springer Nature Switzerland AG 2020
J. C. Setubal and W. M. Silva (Eds.): BSB 2020, LNBI 12558, pp. 168–179, 2020.
https://doi.org/10.1007/978-3-030-65775-8_16
 Polysome-seq as a Measure of Translational Profile from dys1-1 Mutant 169

The eukaryotic translation elongation factor 5A (eIF5A - ortholog elongation

factor P (EF-P) of bacteria) is a highly conserved protein in eukaryotes and
archaea [5,7,19]. In addition, eIF5A is essential for cell viability in all tested
eukaryotes [3,20].
eIF5A undergoes a post-translational modification which leads to hypusine
biosynthesis, called hypusination. This process is irreversible and involves two
enzymatic steps. In the first one, a deoxyhypusine synthase catalyzes the modifi-
cation of a specific lysine residue (K51 in Saccharomyces cerevisiae) to a hypusine
in a spermidine-dependent manner. In the second one, it occurs a hydroxyla-
tion by deoxyhypusine hydroxylase with molecular oxygen as the source. Both
enzymes are also evolutionarily conserved [1,15]. Hypusinated eIF5A is described
to aid in the efficiency of peptide binding of motifs that tend to induce ribo-
somes stalling and also assists with translational termination [22]. In this study,
by measuring the total mRNAs of cells (transcriptome) and the polysomally-
loaded mRNAs (translatome) for the yeast deoxyhypusine synthase mutant dys1-
1 and its wild-type counterpart [9], we obtained a picture of overall relationship
between the two changes for the majority of genes. Polysome-seq can explain
the regulation of post-transcriptional gene expression, as a reliable measure for
a translational profiling study, showing the mRNA recruited for translation. We
show that the majority of statistically significant differences at RNA-seq level
correspond to similar differences at Polysome-seq level, suggesting that, in most
transcripts for this mutant, changes in translation are due to a transcriptional
control and ribosome biogenesis is the main response to translational problems.

2 Materials and Methods

2.1 Strain and Growth Conditions

Saccharomyces cerevisiae strains SVL613 (MATa leu2 trp1 ura3 his3 dys1::HIS3
[DYS1/TRP1/CEN - pSV520]) and SVL614 (MATa leu2 trp1 ura3 his3
dys1::HIS3 [dys1 W75R T118A A147T /TRP1/CEN - pSV730]), DYS1 and
dys1-1, respectively, were used to RNA highthroughput experiments. Cells were
grown under previously described conditions [9].

2.2 Polysome Profilling

For the polysome profiling assay, cell extracts from DYS1 and dys1-1 strains
were prepared as described in [9]. Briefly, the cell cultures were grown to mid-
log phase (OD600 nm = 0.6) and cross-linked with 1% formaldehyde for 1 h
in ice bath. 15 A260 nm units of cell lysates were layered onto 10–50% (w/w)
sucrose gradients and centrifuged for 3 h (39.000 rpm at 4 ◦ C in Beckman SW41-
Ti rotor). The absorbance at 254 nm of gradient fractionation was continuously
measured. Fractions corresponding to mRNA populations bound by 3 ribosomes
were pooled and stored at −80 ◦ C for future RNA isolation.
170 F. M. Demarqui et al.

2.3 RNA Isolation

For total RNA isolation, DYS1 and dys1-1 strains were grown in exponen-
tial phase an OD600 0.6. Cultures were centrifuged and cell pellets were
stored at −80 ◦ C. Cell lysis was conducted with zymolyase and total RNA was
extracted using the RNeasy mini kit (cat. number 74104, Qiagen). The polysome-
associated RNA from pooled fractions was extracted using TRIzol R Reagent,
following the manufacturer’s protocol (cat. n 15596026, ThermoFisher Scien-
tific). Both total RNA and polysome-associated RNA were quantified using a
NanoDrop 2000 Spectrophotometer (ThermoFisher) and the integrity was veri-
fied by electrophoresis gel on 2100 Bioanalyzer equipment (Agilent, Santa Clara,
CA), using a High Sensitivity Total RNA Analysis Chip.

2.4 Library Preparation and Sequencing

Library preparation and sequencing (RNA-seq) for total and polysome-
associated RNA were conducted by Life Sciences Core Facility (LaCTAD) from
State University of Campinas (UNICAMP). Three biological replicates for tran-
scriptome analysis (RNA-seq of total RNA) or translatome analysis (RNA-seq
of polysome-associated RNA) from DYS1 and dys1-1 strains were carried out
according to the manufacturer’s guidelines for TrueSeq kit (catalog number RS-
1222001, Illumina) by selection of mRNA by poly-A tail. These 12 libraries were
sequenced for 51 cycles paired-end on a Illumina HiSeq 2500 platform.

2.5 RNA-seq Data Analysis

The public server (usegalaxy.org/) was used to process the highthroughput
data. FASTQ files had their quality checked by the FastQC tool (Galaxy Ver-
sion 0.72). TrimGalore! (Galaxy Tool Version: 0.4.3.1 + galaxy1) was used to
remove reads with Phred quality score <25 and adapter strings. Files were
mapped against a S. cerevisiae non-coding RNA (ncRNA) sequence file (down-
loads.yeastgenome.org/sequence/S288C reference/rna/archive/rna coding R64-
1-1 20110203.fasta.gz), by Bowtie software (Galaxy Tool Version: 1.1.2) with
the parameters –v 2 –y –a –m 1 –best –strata –S –p 4. The mapping and quan-
tification of reads was performed by Stringtie software (Galaxy Tool Version:
1.3.4) with standard parameters. Only genes in which the median read count of
the three replicates was larger than 10 in all conditions (dys1-1 and DYS1 strain,
for RNA-seq and for Ribo-Seq from polysome-profiling) were kept. The filtered
table of counts contained data for 5.334 genes. Count of reads was converted
into RPM (reads per million).

2.6 Ribo-Seq and Protein Abundance Comparative Analysis

We used the table of counts converted in log2 RPM to compare the relative
abundance of total or polysome-bound mRNAs in wild-type strain between two
published ribosome profiling data: RPM normalized data from Ribo-Seq [23] and
Ribo-Seq from polysome-profiling [10]; and protein abundance estimation [6].
 Polysome-seq as a Measure of Translational Profile from dys1-1 Mutant 171

2.7 Differential Expression Analysis

Non-normalized RNA-seq count tables were used as input in anota2seq (ver.

1.2.0; datatype = “RNA-seq”, normalize = TRUE, transformation = “TMM-
log2”) and normalized using Trimmed Mean of M-values (TMM). Changes
in translational efficiencies were assessed using the anota2seqAnalyze function.
We applied eanota2seqSelSigGenes function to identify differentially expressed
genes, separately for RNA-seq and polysome-profiling RNA-seq data and anal-
ysis of partial variance for identification of gene expression modes from both
profiles. Significance was determined using an adjusted p-value limit of 0.05.

2.8 Enrichment of Gene Ontology and Enrichment Analysis

of Transcription Factors

For the regulatory gene groups, we performed gene ontology (GO) analysis with
terms of biological process to determine whether specific biological functions
were enriched using Yeastmine database [8]. Fisher’s exact test was used to test
for statistically significant differences, and the Holm-Bonferroni correction test
procedure to adjust for the effects of multiple tests [2]. GO terms were considered
significant when FDR <0.05. Gene lists obtained via the statistical differential
from transcriptome profile were submitted to the PSCAN (v1.5, http://159.149.
160.88/pscan/) online tool.

3 Results and Discussion

3.1 RNA-seq and Polysome-Seq Experiments in DYS1 and dys1-1

Strains

We conducted transcriptional and translational profiling (Fig. 1A) for S. cere-

visiae dys1-1 strain and its wild-type counterpart. The number of RNA-seq reads
mapping to a gene was used to quantify the relative abundance of the transcript,
whereas the Polysome-seq provided a quantification of the translatome (Table 1).

Table 1. Number of mapped reads for each sample

Profile DYS11 DYS12 DYS13 dys1-11 dys1-12 dys1-13

Transcriptional 55801619 34587137 116149587 30329854 45292306 51380070
Translational 1214644 1221375 1070191 6150284 1936203 6720246

After filtering out non-expressed genes (see Methods), the table of read counts
per gene contained data for 5,334 S. cerevisiae annotated ORFs. Both transcrip-
tional and translational profiles results were highly reproducible among biological
replicates for each strain (Fig. 1B and 1C) (Table 2 and 3).
172 F. M. Demarqui et al.

Fig. 1. (A) Experimental approaches for studying the transcribed and recruited
mRNAs for translation. Transcriptional profile: the total RNA is extracted, the mRNAs
are separated and subjected to large-scale sequencing. Translational profile: extracts
are separated by ultracentrifugation through sucrose gradient which is then fraction-
ated while its absorbance is continuously monitored at 254 nm (A254), allowing the
separation of free RNA, the 40S and 60S ribosomal subunits, the 80S monosomes and
the polysomes. The RNA is isolated from individualized gradient fractions and pooled
for further large-scale analysis. (B) Principal Component Analysis indicating the dis-
tribution of replicates in the plan. Three biological replicates independent of the DYS1
and dys1-1 strains are represented in the distribution graphs along two main compo-
nents, from the normalized RPM values of the genes sequenced by RNA-seq of each
profile. (C) Linear correlation between replicates of log2 RPM values of genes sequenced
by RNA-seq. The linear correlation of the log2 RPM values of experimental replicates
for the transcriptional profile varied between 0.94 and 0.98 whereas for the translational
profile this value varied between 0.98 and 0.99.
 Polysome-seq as a Measure of Translational Profile from dys1-1 Mutant 173

Table 2. Pearsons correlation values for log2 RPM values from transcriptional profile
for each replicate

DYS11 DYS12 DYS13 dys1-11 dys1-12 dys1-13

DYS12 0.973 1
DYS13 0.983 0.992 1
dys1-11 0.940 0.929 0.935 1
dys1-12 0.935 0.916 0.922 0.995 1
dys1-13 0.925 0.906 0.914 0.986 0.990 1

Table 3. Pearsons correlation values for log2 RPM values from translational profile for
each replicate

DYS11 DYS12 DYS13 dys1-11 dys1-12 dys1-13

DYS12 0.973 1
DYS13 0.968 0.973 1
dys1-11 0.845 0.853 0.847 1
dys1-12 0.845 0.855 0.854 0.986 1
dys1-13 0.854 0.867 0.865 0.989 0.988 1

3.2 Polysome-seq as a Measure for Translational Profile

One technique aimed for studying the composition of mRNAs recruited for trans-
lation by large-scale analysis is the polysome profiling, which segregates mRNAs
associated with polysomes from ribosome-free mRNAs, associated with RNA-seq
(Fig. 1A). In addition to Polysome-seq, Ribo-seq methodology, or ribosome pro-
filing, is based on the sequencing of ribosome-protected fragment (RPF) mRNAs
[12]. We observed high Pearson correlations with the log2 RPM wild-type data
from this study to ribosome profiling wild-type data available in the literature
[10,23] and (Fig. 2A and 2B).
Next, we compared the wild-type strain quantification of gene expression by
RNA-seq and Polysome-seq to published proteomic data [6]. The correlation and
coefficient of determination from translatome (Polysome-seq) to the proteome
normalized abundances (Fig. 2C) was higher than the transcriptome measure-
ments (Fig. 2D), indicating that this former quantification of gene expression
provides a more accurate picture of protein abundance, since translation is reg-
ulated by (1) translation rate, (2) translation rate modulation, (3) modulation
of a protein’s half-life, (4) protein synthesis delay, (5) protein transport [17,18].
So Polysome-seq allows a better understanding of regulatory mechanisms that
involves post-transcriptional gene expression programs [11,13], as regulation via
tuning transcript levels alone [16], resulting in a profile of selected mRNAs
recruited for translation.
174 F. M. Demarqui et al.

3.3 Yeast Hypusination Mutant dys1-1 Responds Transcriptionally

for Gene Regulation

We first calculated the gene expression level fold change (FC) between the two
strains using RNA-seq and Polysome-seq data separately and we observed similar
numbers of differentially expressed genes (DEGs) for both profiles - 2432 and
2826 DEGs for transcriptional and translational level, respectively - (Fig. 3A and
3B), however, Polysome-seq data had a higher variance than RNA-Seq data for
the significant log2 FC distribution (Fig. 3C), a consistent result for a mutant
involved with a translational factor.
To establish the relationship between mRNA and polysome-associated
mRNA changes when comparing DYS1 and dys1-1, we categorized DEGs into
gene expression modes by computing analysis of partial variance with transcrip-
tome and translatome (Fig. 3D): (1) Homodirectional DEGs, significantly change
in both profiles in a concordant way, indicating a transcriptional regulation; (2)
Polysome-only DEGs, up or down polysome-associated mRNA with no signifi-
cant changes in mRNA levels, a result of translation regulatory mode; (3) Tran-
scriptome only DEGs, differences in mRNA levels not followed by a significant
change in polysome-associated mRNA, a result of buffering regulatory mode; (4)
Antidirectional DEGs, significantly change in both profiles but antidirectional
ways. Most DEGs (67%) showed a coupled significant change, i. e., genes with sig-
nificant homodirectional change in both the transcriptome and the translatome
(Fig. 3E). This result is in accordance with the fact that under stress conditions,
differential expressed proteins correlated strongly with the corresponding mRNA
level, indicating that transcriptional control seems to be the major driver behind
changes in protein levels [14].
Transcriptionally regulated genes were significantly enriched for Gene Ontol-
ogy (GO) biological process terms as “maturation of SSU-rRNA” (GO:0030490),
“transposition” (GO:0032196), “RNA modification” (GO:0009451) (Table 4)
and Transcription Factors (TF) as Tod6, Dot6 and Stb3 (Table 5). Additionally,
BUD27, the gene that encodes a protein which impacts the homeostasis of the
ribosome biogenesis process by regulating the activity of the three RNA poly-
merases [17], is classified as an homodirectional gene and upregulated in both
profiles. Taking together, these results revealed a cell response to ribosome bio-
genesis, a high-energy consumption process that requires stringent regulation to
ensure proper ribosome production to deal with cell growth and protein synthesis
in different environmental and metabolic situations [17].
The results of this study illustrate the use of Polysome-seq as a measurement
of mRNAs recruited for translation. We identified for a deoxyhypusine synthase
mutant dys1-1, a protein involved in translation, a pattern of gene expression
control that is transcription dependent and upregulation of ribosome synthesis
is one of the cell responses to translation impairment.
 Polysome-seq as a Measure of Translational Profile from dys1-1 Mutant 175

Fig. 2. Polysome-seq correlates satisfactorily to Ribo-seq data and is a good predictor

of protein abundance. (A) Correlation between the translational profile (log2 RPM)
of this study and the translational profile of obtained by Ribo-seq (log2 RPM) [23].
(B) Correlation between the translational profile of this study (log2 RPM) and the
translational profile of obtained by a combination of polysomal profile followed by
Ribo-seq (log2 RPM) [6]. C) Distribution between protein abundance (molecules per
cell) and the translational profile (log2 RPM) of this study. D) Distribution between
protein abundance (molecules per cell) and the transcriptional profile (log2 RPM) of
this study. Protein abundance data are indicated in molecules per cell according to [6].
176 F. M. Demarqui et al.

Fig. 3. Volcano plot of the distribution of the transcripts differentially expressed in the
transcriptional profile (A) and translational profile (B). The values of −log1 0 p-value
were plotted according to the differencial expression between DYS1 and dys1-1 (log2 fold
change). Downregulated genes are highlighted in blue (left), upregulated genes, in orange
(right); dashed horizontal line indicates an adjusted p-value of 0.05. (C) Distribution of
gene expression fold change (FC) values. FC was calculated as the ratio between the num-
ber of reads in dys1-1 and DYS1 strains. We took the average number of reads per gene
among the replicates. (D) Scheme of differential expression analysis between the tran-
scriptional and translational profile of the dys1-1 mutant. Genes classified as differentially
expressed were called transcriptome only (blue), polysome only (orange), antidirectional
(purple) - significantly opposite variations between transcriptional and translational pro-
files - and homodirectional (green) - variations significantly converging between both pro-
files. (E) Distribution of the log2 fold change of the transcriptional and translational pro-
file. Genes showing statistical differences between dys1-1 and DYS1 were simultaneously
compared in the two profiles. Categories are defined in 3D. (Color figure online)
 Polysome-seq as a Measure of Translational Profile from dys1-1 Mutant 177

Table 4. Gene Ontology (GO) analysis of transcriptionally regulated mRNAs from

dys1-1 mutant as determined by anota2seq

Analysis by Anota2seq - “Homodirectional”

Term ID Description log1 0 p-value Dispensability N genes
GO:0030490 Maturation of SSU-rRNA −3.000.000 0.00 53
GO:0032196 Transposition −53.792 0.03 42
GO:0032197 Transposition, −34.237 0.12 40
RNA-mediated
GO:0006278 RNA-dependent DNA −76.021 0.24 39
biosynthetic process
GO:0090305 Nucleic acid phosphodiester −61.694 0.28 98
bond hydrolysis
GO:0001510 RNA methylation −45.888 0.29 24
GO:0006396 RNA processing −64.134 0.37 175
GO:0009451 RNA modification −27.911 0.40 43
GO:0034660 ncRNA metabolic process −28.050 0.48 166
GO:0000966 RNA 5’-end processing −62.684 0.48 28

Table 5. Transcriptional factor (TF) enrichment analysis of differentially expressed

genes in the transcriptional profile from dys1-1 mutant as determined by anota2seq

MatrixID Matrix name p-value

MA 0350.1 TOD6 1,92E−23
MA 0351.1 DOT6 1,72E−21
MA 0390.1 STB3 1,14E−05
MA 0378.1 SFP1 4,49E−03
MA 0398.1 SUM1 4,64E−02
MA 0345.1 NHP6A 1,12E−01
MA 0346.1 NHP6B 2,73E−01
MA 0386.1 SPT15 0.000184904
MA 0418.1 YAP6 0.000859765
MA 0435.1 YPR015C 0.00121852

Acknowledgement. This study was financially supported by grant #10/50044-6,

São Paulo Research Foundation (FAPESP). This study was financed in part by the
Coordenação de Aperfeiçoamento de Pessoal de Nı́vel Superior - Brasil (CAPES) -
Finance Code 001.
178 F. M. Demarqui et al.

References
1. de Almeida, O.P., et al.: Hypusine modification of the ribosome-binding protein
eIF5A, a target for new anti-inflammatory drugs: understanding the action of the
inhibitor GC7 on a murine macrophage cell line. Curr. pharm. Des. 20(2), 284–92
(2014)
2. Benjamini, Y., Yekutieli, D.: The control of the false discovery rate in multiple test-
ing under dependency. Ann. Stat. (2001). https://doi.org/10.1214/aos/1013699998
3. Buskirk, A.R., Green, R.: Ribosome pausing, arrest and rescue in bacteria and
eukaryotes. Philos. Trans. R. Soc. Lond. Ser. B Biol. Sci. 372, 20160183 (2017).
https://doi.org/10.1098/rstb.2016.0183
4. Chassé, H., Boulben, S., Costache, V., Cormier, P., Morales, J.: Analysis of transla-
tion using polysome profiling. Nucleic Acids Res. (2017). https://doi.org/10.1093/
nar/gkw907
5. Chen, K.Y., Liu, A.Y.: Biochemistry and function of hypusine formation on eukary-
otic initiation factor 5A. NeuroSignals (1997). https://doi.org/10.1159/000109115
6. Csárdi, G., Franks, A., Choi, D.S., Airoldi, E.M., Drummond, D.A.: Accounting
for experimental noise reveals that mRNA levels, amplified by post-transcriptional
processes, largely determine steady-state protein levels in yeast. PLoS Genet.
11(5), e1005206 (2015). https://doi.org/10.1371/journal.pgen.1005206
7. Dever, T.E., Ivanov, I.P.: Roles of polyamines in translation. J. Biol.
Chem. 293(48), 18719–18729 (2018). https://doi.org/10.1074/jbc.TM118.003338.
http://www.jbc.org/
8. Engel, S.R., et al.: The reference genome sequence of saccharomyces cerevisiae:
then and now. G3: Genes Genomes Genetics 4(3), 389–398 (2014). https://doi.
org/10.1534/g3.113.008995
9. Galvão, F.C., Rossi, D., Silveira, W.D.S., Valentini, S.R., Zanelli, C.F.: The deoxy-
hypusine synthase mutant dys1-1 reveals the association of eIF5A and Asc1 with
cell wall integrity. Plos One 8(4), e60140 (2013). https://doi.org/10.1371/journal.
pone.0060140
10. Heyer, E.E., Moore, M.J.: Redefining the translational status of 80S monosomes.
Cell 164(4), 757–769 (2016). https://doi.org/10.1016/j.cell.2016.01.003
11. Ingolia, N.T.: Ribosome profiling: new views of translation, from single codons
to genome scale. Nat. Rev. Genet. 15, 205–213 (2014). https://doi.org/10.1038/
nrg3645
12. Ingolia, N.T., Ghaemmaghami, S., Newman, J.R., Weissman, J.S.: Genome-wide
analysis in vivo of translation with nucleotide resolution using ribosome profiling.
Science 324(5924), 218–223 (2009). https://doi.org/10.1126/science.1168978
13. Jin, H.Y., Xiao, C.: An integrated polysome profiling and ribosome profiling
method to investigate in vivo translatome. In: Methods in Molecular Biology, vol.
1712, pp. 1–18. Humana Press Inc. (2018). https://doi.org/10.1007/978-1-4939-
7514-31
14. Lahtvee, P.J., et al.: Absolute quantification of protein and mRNA abun-
dances demonstrate variability in gene-specific translation efficiency in yeast.
Cell Syst. 4(5), 495–504 (2017). https://doi.org/10.1016/J.CELS.2017.03.003.
https://www.sciencedirect.com/science/article/pii/S2405471217300881#mmc4
15. Landau, G., Bercovich, Z., Park, M.H., Kahana, C.: The role of polyamines in
supporting growth of mammalian cells is mediated through their requirement for
translation initiation and elongation. J. Biol. Chem. 285(17), 12474–12481 (2010).
https://doi.org/10.1074/jbc.M110.106419
 Polysome-seq as a Measure of Translational Profile from dys1-1 Mutant 179

16. Liu, Y., Beyer, A., Aebersold, R.: On the dependency of cellular protein levels on
mRNA abundance. Cell 165, 535–550 (2016). https://doi.org/10.1016/j.cell.2016.
03.014
17. Martı́nez-Fernández, V., et al.: Prefoldin-like Bud27 influences the transcription of
ribosomal components and ribosome biogenesis in Saccharomyces cerevisiae. RNA
(2020). https://doi.org/10.1261/rna.075507.120
18. Piccirillo, C.A., Bjur, E., Topisirovic, I., Sonenberg, N., Larsson, O.:
Translational control of immune responses: from transcripts to trans-
latomes. Nat. Immunol. 15(6), 503–511 (2014). https://doi.org/10.1038/ni.2891.
http://www.nature.com/doifinder/10.1038/ni.2891
19. Rossi, D., Kuroshu, R., Zanelli, C.F., Valentini, S.R.: eIF5A and EF-P: two unique
translation factors are now traveling the same road. Wiley Interdiscip. Rev.: RNA
5(2), 209–222 (2014). https://doi.org/10.1002/wrna.1211
20. Schnier, J., Schwelberger, H.G., Smit-McBride, Z., Kang, H.A., Hershey, J.W.:
Translation initiation factor 5A and its hypusine modification are essential for cell
viability in the yeast Saccharomyces cerevisiae. Mol. Cell. Biol. (1991). https://
doi.org/10.1128/MCB.11.6.3105
21. Schuller, A.P., Green, R.: Roadblocks and resolutions in eukaryotic translation.
Nat. Rev. Mol. Cell Biol. 19(8), 526–541 (2018). https://doi.org/10.1038/s41580-
018-0011-4. http://www.nature.com/articles/s41580-018-0011-4
22. Schuller, A.P., Wu, C.C.C., Dever, T.E., Buskirk, A.R., Green, R.: eif5a functions
globally in translation elongation and termination. Mol. Cell 66(2), 194–205 (2017).
https://doi.org/10.1016/j.molcel.2017.03.003
23. Sen, N.D., Zhou, F., Ingolia, N.T., Hinnebusch, A.G.: Genome-wide analysis of
translational efficiency reveals distinct but overlapping functions of yeast DEAD-
box RNA helicases Ded1 and eIF4A. Genome Res. 25(8), 1196–1205 (2015).
https://doi.org/10.1101/gr.191601.115
 Anti-CD3 Stimulated T Cell
Transcriptome Reveals Novel ncRNAs
and Correlates with a Suppressive Profile

Manuela M. do Almo1,2 , Isabel G. Sousa1 , Waldeyr Mendes Cordeiro da

Silva1,3 , Thomas Gatter4 , Peter F. Stadler4,5,6,7,8 , Steve Hoffmann4,9,10 ,
Andrea Q. Maranhão1,11 , and Marcelo Brigido1,11(B)
1
Department of Cell Biology, Institute of Biological Science,
University of Brası́lia, Brası́lia, Brazil
[email protected]
2
Molecular Pathology Graduation Program, Medicine Faculty,
University of Brası́lia, Brası́lia, Brazil
3
NEPBio - Federal Institute of Goiás, Formosa, Brazil
4
Bioinformatics Group, Department of Computer Science;
Interdisciplinary Center for Bioinformatics,
University of Leipzig, Leipzig, Germany
5
Institute for Theoretical Chemistry,
University of Vienna, Vienna, Austria
6
Facultad de Ciencias,
Universidad Nacional de Colombia, Bogotá, Colombia
7
Max Planck Institute for Mathematics in the Sciences, Leipzig, Germany
8
Santa Fe Institute, Santa Fe, USA
9
Computational Biology Group, Leibniz Institute on Ageing -
Fritz Lipmann Institute (FLI), Jena, Germany
10
Friedrich-Schiller-University Jena, Jena, Germany
11
Instituto de Investigação em Imunologia,
Instituto Nacional de Ciência e Tecnologia (iii-INCT), São Paulo, Brazil

Abstract. T lymphocytes are key players in immunity. Anti-CD3 anti-

bodies activate T cells and promote a suppressive phenotype in vivo.
Although the pharmacological use of these antibodies is widely studied,
the underlying mechanisms are still poorly understood. Here we describe
the response of the non-coding RNA transcriptome of T cells after in vitro
stimulation of peripheral blood mononuclear cells (PBMC) by anti-CD3
and demonstrate that several novel long non-coding RNA are associated
with antibody treatment. Regulated long intergenic non-coding RNAs
were associated with lymphocyte activation and signaling pathways. In
particular the lncRNA transcripts WFDC21P and GAPLINC are regu-
lated in stimulated T cell.

Keywords: Anti-CD3 · lncRNA · Gene regulation · Alternative

splicing · Transcriptome

c Springer Nature Switzerland AG 2020

J. C. Setubal and W. M. Silva (Eds.): BSB 2020, LNBI 12558, pp. 180–191, 2020.
https://doi.org/10.1007/978-3-030-65775-8_17
 Anti-CD3 Stimulated T Cell Transcriptome Reveals Novel ncRNAs 181

1 Introduction
Targeting T lymphocytes was among the first monoclonal antibodies (mAb) asso-
ciated immunotherapies. Anti-CD3 therapy was used to control T cell activity,
suppress the immune response, and substitute the polyclonal anti-thymocyte
antibody preparation, previously used for graft rejection [17]. Muronomab, a
monoclonal antibody specific for the human CD3 antigen, was the first mAb
used in clinical studies, but its use was abolished due to its overall toxicity
[21]. Nowadays, novel and less toxic CD3 specific antibodies have reemerged as
promising therapeutics for controlling autoimmune diseases.
To better understand the human T cell reprogramming after anti-CD3 treat-
ment, we previously investigated the protein-coding (PTC) genes [25] regulated
ex vivo in a PBMC milieu. Based on a new transcript prediction algorithm,
we reannotated the non-coding transcriptome of human T cells treated with
anti-CD3 antibodies to unveil differentially expressed lncRNA (DEL) that may
be involved in CD3 targeted antibody therapy. We observed several novel non-
coding transcripts along with previously annotated ones, and we discuss their
possible participation in T cell fate and the suppressive phenotype.

2 Materials and Methods

2.1 Sample Preparation and Sequencing

We extend here analysis of the RNA-seq data GEO database (GSE112899) orig-
inally described in [25]. In brief, Ficoll-Paque purified volunteer human PBMC
were cultured with or without anti-CD3 antibodies. We used three antibody
preparations: an anti-CD3 Monoclonal Antibody - OKT3, and two Recombi-
nant Antibodies derived from OKT3, and produced in a human IgG1 scFvFc
format: a humanized (FvFcR) and a chimeric antibody (FvFcM). After 72 h in
culture, CD3+ cells were enriched by negative selection. The RNA-seq data were
produced with an Illumina HiSeq in 2 × 150 bp paired-end mode [25]. All human
blood experiments were performed in accordance with the Ethics Committee of
the University of Brasilia guidelines, which approved the study protocol (CAAE:
32874614.4.0000.0030).

2.2 Genome Mapping and Transcript Prediction

All sequenced reads produced by Illumina were analyzed for quality control
using FASTQC [1]. The reads were filtered using BBDuk [3] at k = 31 to a ref-
erence of ribosomal kmers provided by the developers. Adapters were trimmed
by cutadapt [15], and reads were then aligned to the HG19/GRCh37 Human
Genome using HISAT2 [13] at standard settings. Ryūtō [7] was run on the align-
ments to predict individual transcripts for each set. Transcript predictions were
joined using the TACO meta assembler [16]. Results were compared to the Gen-
code V19 annotation to identify novel transcripts. Salmon [18] was used to realign
182 M. M. do Almo et al.

the filtered reads for each sample to the full set of GENCODE. A count matrix
was created from transcripts together with the additional predicted novel tran-
scripts. DESeq2 [14] was used to identify differentially expressed genes (DEGs)
on all replicates, applying a significance threshold for the adjusted p-values of
0.05.

2.3 Data Mining and Ontology Classification

The annotated transcriptome and DEGs were explored using in-house awk
scripts and BEDTools [19] commands and visualized with Integrative Genomics
Viewer (IGV) [22] and Ensembl browser (ENSEMBL) [12]. Venn diagrams were
designed using UGENT tools1 . The protein-coding (PTC) transcripts (Ensembl
transcript type) were excluded from analysis, and only non-coding transcripts
from non-protein-coding and protein-coding genes were considered. Only tran-
scripts with base mean above zero were considered. The overlaps between dif-
ferent transcriptomes were analyzed using bedtools intersect [19]. Non-coding
potentials were calculated using Portrait [2] using a cutoff of 60%.

2.4 cDNA Synthesis Transcript Testing

The cDNA was synthesized with the SuperScript IV Reverse Transcriptase
kit (Invitrogen, Carlsbad, CA, USA) from the total RNA extracted using
miRNeasy R
Mini Kit (Qiagen, Valencia, CA, USA) [25] from T lymphocytes
from three donors, treated or not with anti-CD3 antibodies, following the manu-
facturer’s guidelines. Primers were designed for the predicted transcripts lnc-DC
and GAPLINC, that were amplified by PCR with Taq Platinum DNA Poly-
merase (Invitrogen, Carlsbad, CA, USA) under the following conditions: 35
cycles where the DNA was denatured at 94 ◦ C for 30 s, annealed at 52 ◦ C for 30 s
and elongated at 72 ◦ C for 1 min. The amplification cycles were performed in the
SimpliAmp thermocycler (Applied Biosystems). The PCR products were ligated
into the pGEM-T Easy Vector (Promega, Madison, Wisconsin) and transformed
into XL1Blue competent cells. The cloning was confirmed by plasmid enzymatic
digestion with Nco I and Not I. Plasmids were sequenced by the Sanger method.

2.5 Gene Expression Analysis by qPCR Assays

The quantitative PCR assays were performed in triplicate with total RNA iso-
lated from T cells utilized for cDNA synthesis using RT2 First Strand Kit (Qia-
gen, Valencia, CA, USA). As previously described [25], expression genes were
quantified using RT2 qPCR SYBR Green/ROX Master-Mix (Qiagen, Valencia,
CA, USA) following the manufacturer’s instructions. The reference gene Beta-2
microglobulin (B2M) was used as the endogenous control. The assays were per-
formed using the ABI Step One Plus Real-Time PCR System (Applied Biosys-
tems). The 2-ΔΔCt method was used to calculate the levels of the lncRNA
1
bioinformatics.psb.ugent.be/webtools/Venn.
 Anti-CD3 Stimulated T Cell Transcriptome Reveals Novel ncRNAs 183

transcripts (fold change) and for analysis of the obtained data, the RT2 Profiler
PCR Array Data Analysis (SABiosciences Frederick, MD, USA) software was
used. Real-Time qPCR p-values were calculated based on Student’s t-test using
RT2 Profiler PCR Array Data Analysis software.

3 Results

The transcriptome reconstructed from the union of all mapped reads comprises
174,649 transcripts of which about a third are protein-coding, nearly half (44.1%)
are correspond to known non-coding Ensembl/VEGA transcripts, and 20% pre-
viously unannotated transcripts (Fig. 1). The distribution of Ensembl/VEGA
transcript types among the known ncRNA is summarized in Fig. 1-A. The most
abundant types are transcripts with retained introns (IR) and Processed Tran-
scripts (PT) that are not classified as belonging to one of the lncRNAs classes.

Fig. 1. Several non-protein coding transcripts are differentially expressed in anti-CD3

stimulated cells. (A) Gene type distribution of the observed ncRNA in all experimen-
tal samples. The distribution of non-protein-coding transcripts is detailed in the right;
Volcano plot of DEL transcripts regulated by (B) OKT3; (C) FvFcR; and (D) FvFcM.
Each group is compared to the non-treated sample; (E) Venn diagram comparing sta-
tistically significative DEL in each experimental group and their intersections.
184 M. M. do Almo et al.

Long intergenic ncRNA (LincRNA), Antisense RNA (AS), and pseudogenes

were also significantly observed. The three antibody treatments resulted in a
total of 726 differentially expressed lncRNA (DELs). OKT3 induced a larger
number of ncRNA compared to FvFcR and FvFcM, as observed in the volcano
plots (Fig. 1-D). The Venn diagram (Fig. 1-E) summarizes the DEL set seen in
each treatment and their intersections. The number of statistically significant,
differentially expressed transcripts after each antibody treatment category is
summarized in Table 1 for the five major categories of lncRNA. Anti-CD3 anti-
bodies regulate only a small fraction of total observed lncRNA, ranging from
0,9% of retained intron transcripts (IR) after OKT3 stimulation to 0,02% of
pseudogenes after FvFcM stimulation.

Table 1. Differentially expressed lncRNA after anti-CD3 treatment

Transcript type∗ OKT3 FvFcR FvFcM Total transcripts#

Up Down Up Down Up Down
Intron Retained 90 132 76 36 30 18 22,360
Unclassified Processed Transcript 85 46 69 10 27 13 21,074
Antisense 11 13 8 1 2 6 6,737
LincRNA 17 19 15 4 3 1 6,552
Pseudogene& 4 4 4 3 0 3 5,696
∗ Transcript type classification after ENSEMBL.
& Pseudogene includes: pseudogene, processed pseudogene, transcribed unprocessed pseudogene, unprocessed
pseudogene and polymorphic pseudogene type.
# Total number of transcripts in the T cell transcriptome.

3.1 LincRNAs

LincRNAs are known to be involved in critical processes in the differentiation

of T cells. The whole T cell transcriptome reveals 6,552 LincRNA, and a small
proportion of them are regulated by anti-CD3 stimulation. Some of them are
known to be expressed T cells, such as NEAT1, a nuclear long ncRNA associ-
ated with Th2 [10] and Th17 differentiation [23]. NEAT1 is down-regulated after
OKT3 and FvFcR treatment, but only OKT3 induces a statistically significant
three-fold reduction in transcript level. Linc00861 was found to be expressed in
several CD4 and CD8 cells. OKT3 stimulation also diminished linc00861 with
high confidence (p < 10−13 ), while repression by FvFcR shows a barely signifi-
cant adjusted p-value (<4 × 10−3 ).
Three lincRNA were observed to be activated in all antibodies treatment,
AC017002.1, LINC00339, and LINC01132. Only the first two were previously
found in T cells [5]. AC017002.1 was mostly but not exclusively detected in mem-
ory Treg. This lincRNA is in close genomic proximity to BCL2L11, a proapop-
totic gene involved in the T cell negative selection in thymus associated with T
cell activation by high-affinity antigens [8].
 Anti-CD3 Stimulated T Cell Transcriptome Reveals Novel ncRNAs 185

3.2 Novel Predicted Transcripts

Among the reconstructed transcripts, about a fifth (35,149) remained unanno-

tated transcripts (TU) (Fig. 1-A). Most of these TU overlapped with protein-
coding and non-protein coding genes (Fig. 2-A), and thus most likely constitute
previously undescribed isoforms of known genes. However, 575 TUs do not over-
lap any annotated gene and are novel genes candidate. The TU set was further
tested for their coding potential that suggested that at least 413 (72%) of them
are novel lncRNA transcripts (Fig. 2-B). These transcripts contained between 1
and 15 exons, with a modal exon number of 2 per gene (Fig. 2-C). Among these
putative novel transcripts revealed by RNA-seq, 77 are regulated by anti-CD3
stimulation, 72 by OKT3, 19 by FvFcR, and 13 by FvFcM (Fig. 2-D). All three
antibodies regulated eight of them.
Interestingly, some genomic regions accumulate novel TU with CD4 and CD8
specific transcripts suggesting a regulatory role for such loci. An example is the
TU35249, which appears close to KLF3-AS1 and KLF3 loci (data not shown).
All three transcripts are repressed during anti-CD3 treatment. TU35249 also
seems to be coregulated with the antisense RNA and overlaps a CD4 and a CD8
specific transcript previously reported in [11,20].

3.3 Testing Novel LncRNA Isoforms

Two of the computationally predicted TU were tested experimentally for differ-

ential expression after T cell stimulation with anti-CD3. These transcripts were
cloned and sequenced, and their expression levels were quantified by qPCR. The
data on T cell RNA-seq suggested transcriptional activity in the locus WFDC21P
on Chromosome 17. This locus is described as an ancient pseudogene with coding
capacity in mammals, including primates, but not in humans [6]. In the genus
Homo, this locus was reported to code for a lincRNA, Lnc-DC, found to regulate
STAT3 activity in dendritic cells (DC) [27]. We observed several transcripts in
this locus, and two novel isoforms TU20859 and TU20860 were found repressed
in OKT3 stimulated T cell (padj < 0.05). These isoforms differ from the Lnc-
DC due to the use of novel exons at the 5’ end (Fig. 3). Transcriptional activity
was validated in T cell RNA by qPCR analyses that corroborate the RNA-seq
data to demonstrate the presence of WFDC21P transcripts in resting T cell,
besides their repression after anti-CD3 treatment. Moreover, sequence analyses
of cDNA amplicon are compatible with the TU20859 and TU20861 transcripts,
but no cDNA clone was found consistent with the presence of TU20860, the
transcript Lnc-DC (Genbank NR 030732.1). Though, the WFDC21P transcript
that is repressed after anti-CD3 stimulation may be a novel transcript distinct
from the previously described Lnc-DC [27].
GAPLINC is a lincRNA that has been described as a marker for gastric ade-
nocarcinoma [9] and was not reported to be expressed in lymph nodes [5]. The
data on T-cell RNA-seq presented here suggested the presence of GAPLINC
transcripts in the untreated cells and significant reduction after anti-CD3 treat-
ment. All four reference transcripts of GAPLINC were observed, but no signifi-
186 M. M. do Almo et al.

cant differential expression was observed, except for the GAPLINC-2042 , barely
significantly repressed by OKT3 with a padj = 0.0135. Along with them, two
novel isoforms, TU21901 and TU21904, were detected. TU20901 was the most
abundant transcript predicted in this locus and is a DEL, repressed as a result of
OKT3, FvFcR, and FvFcM treatment (padj value of 0.0001, 0.0487, and 0.0502,
respectively). The qPCR quantification suggested that all antibodies induced
certain repression, especially for a particular primer pair, which detects all vari-
ants except for GAPLINC-201 (Fig. 4).
The analysis of cDNA clones obtained from PCR for exons 1 to 4 and 3 to
4 yielded sequences that confirmed the presence of predicted transcripts. Three
independent clones could unequivocally validate TU21901 with the same exon-

Fig. 2. A set of unannotated predicted transcripts may correspond to novel lncRNA.

(A) a large fraction of the predicted transcriptome could be machine annotated (blue),
from the unannotated transcripts most overlaps known gene loci (salmon), except for a
small fraction (red) of them. (B) Non-coding probability of the unannotated nonover-
lapping transcripts. (C) Histogram of exon content of the unannotated nonoverlapping
transcripts. The number of transcripts is quoted following the number of predicted
exons. (D) Venn diagram showing that part of the unannotated nonoverlapping tran-
scripts is regulated by the anti-CD3 treatment. In blue are transcripts regulated by
OKT3, red, FvFcR and green, FvFcM. Yellow eclipse reflects the total number of
unannotated nonoverlapping transcripts. (Color figure online)

2
ENST00000581442.1.
 Anti-CD3 Stimulated T Cell Transcriptome Reveals Novel ncRNAs 187

exon junction. Two clones showed the unique junction of GAPLINC-204, where
exon 2 uses an alternative donor splice site compared to TU21901. Three other
predicted clones showed an exon-exon junction that is shared either by TU21904
or by the previously reported transcript GAPLINC-202. Therefore, the data
showed that along with GAPLINC-202 and GAPLINC-204, at least the novel
regulated transcript TU21901 could be found in non-stimulated T cells.

Fig. 3. The WFDC21P transcript is depicted to reveals the Lnc-DC along with
TU20859, TU20860, and TU20861 intron-exon structure. (A) The transcripts in the
opposite strand of chromosome 17 were rotated to facilitate visualization. Primers used
to check for transcripts are marked in red and green. Quantitative expression by qPCR
assay of lncRNAs was performed with total RNA extracted from T cells stimulated
with anti-CD3 antibodies. The results are expressed as the fold change relative to
unstimulated T cells (n = 5; p < 0.05). (B) Expression of transcripts detected with the
primer pair for the first and third exon of Lnc-DC (red). (C) Expression of transcripts
detected with the primer pair for the second and third exon of Lnc-DC (green). (Color
figure online)

4 Discussion
The stimulation of T lymphocytes with anti-CD3 antibodies induces a change
in transcriptional landscape. In this new study, we reanalyzed the data on the
188 M. M. do Almo et al.

anti-CD3 induced T cells [25] to unveil the lncRNA transcriptome. We produced

a new read mapping of the reads and focused on the reconstruction of splicing
isoforms. The number of annotated non-protein-coding transcript observed in
the cell transcriptome was close to that in the hg19 human genome assembly
[12], suggesting a good coverage of the total universe of lncRNAs. The focus of
the research presented here, however, was not the complete lncRNA set, but the
differentially expressed transcriptome; trying to figure out the changes in genetic
programming that is achieved after antibody stimulation. Due to limitation of
our model system, we are not able to pinpoint specific T cell subpopulations.
Nevertheless, considering DEL observed in the whole T cell mixture in a PBMC
context, we speculate that a particular T cell population is imposed (or polar-
ized), either by expansion or differentiation.

Fig. 4. Transcriptional activity of the GAPLINC locus. (A) Novel transcripts are
depicted along with annotated transcripts. In the top, the genomic view of transcripts
with exon-intron structured. Primers used to check for transcripts are marked in green
and magenta. Quantitative expression by qPCR assay of lncRNAs was performed with
total RNA extracted from T cells stimulated with anti-CD3 antibodies. The results
are expressed as the fold change relative to unstimulated T cells (n = 5; p < 0.05). (B)
qPCR with a primer to the junction of the first and second exon and another for the
third exon of GAPLINC-204 (red), detecting all transcripts except GAPLINC 201. (C)
Expression of transcripts detected with a primer pair for the third and fourth exon of
GAPLINC-202/TU21904 (green). (Color figure online)
 Anti-CD3 Stimulated T Cell Transcriptome Reveals Novel ncRNAs 189

LincRNA expression seems to be more cell type-specific than protein-coding

genes [4]. After that, we inspect the ncRNA transcriptome for lincRNA DEL,
which could support discrete changes in cell differentiation, explaining the
upraise of a suppressive phenotype. Among the annotated lincRNA, only three
(AC017002.1, LINC00339, LINC01132) were consistently activated with the
three anti-CD3 molecules. AC017002.1 was associated with memory CD4 cells,
and its upregulation may reflect the expansion of memory cells, a reported effect
of commercial anti-CD3 [26].
The antibody molecule format has a significant impact on the expression
profile. OKT3 is far the most effective in regulating lncRNA. It chimeric for-
mat FvFcM and the humanized antibody fragment FvFcR neverthess display
a similar, although generally less intense response. The former, a mouse mAb,
has a much stronger mitogenic response than FvFc format [24,25]. Despite the
differences in the regulated gene set, several of DEL are consistently regulated
after all antibody molecules. Yet, other regulated lncRNA is antibody specific,
such as FvFcM specific DEL. The chimeric antibody is the only antibody that
significantly regulates TU13951. Maybe variation in the antibody’s paratope and
the Fc component affect the strength and quality of TCR signaling, and further
engineering may improve regulatory bias reducing the inflammatory response.

5 Conclusion

We investigate the lncRNA transcriptome of T lymphocyte cells to uncover the

changes incurred by anti-CD3 immunotherapy, which could reflect in a suppres-
sive phenotype. Several lncRNAs and known lincRNAs were observed, and its
role in the reversal of inflammation may be associated with the induction of a
regulatory phenotype.
The successful release of novel anti-CD3 therapeutics will readdress the inves-
tigation of the novel suppressive and tolerogenic effect of these immune phar-
maceuticals in humans. The selective immunoregulation of the anti-CD3 treat-
ment observed in clinics may become the basis of novel therapy for autoimmune
disease. In this sense, the development of clinical-grade markers could help this
development. The data on lncRNA revealed in this work may not only contribute
to novel markers to follow immunoregulation on the whole T cell context but
also contribute to potential non-coding RNA regulation. Association of anti-CD3
data may yield new disease markers and treatment targets among the differen-
tially expressed lncRNA. Beyond simplifying therapeutics monitoring, biological
relevant ncRNA could become a pharmacological target for future therapies. As
a result, a new generation of more powerful pharmaceuticals to immunosuppress
and control the immune response.

Acknowledgments. We are thankful to CAPES and CNPq for scholarship funding

and to FAPDF for financial support of this project. We are also grateful to Prof.
Concepta M. McManus for English correction.
190 M. M. do Almo et al.

References
1. Andrews, S., et al.: FastQC: a quality control tool for high throughput sequence
data (2010). https://www.bioinformatics.babraham.ac.uk/projects/fastqc/
2. Arrial, R.T., Togawa, R.C., Brigido, M.M.: Screening non-coding RNAs in tran-
scriptomes from neglected species using PORTRAIT: case study of the pathogenic
fungus Paracoccidioides brasiliensis. BMC Bioinform. 10(1), 239 (2009). https://
doi.org/10.1186/1471-2105-10-239
3. BBTools: BBDuk. http://jgi.doe.gov/data-and-tools/bb-tools/
4. Cabili, M.N., et al.: Integrative annotation of human large intergenic noncoding
RNAs reveals global properties and specific subclasses. Genes Dev. 25(18), 1915–
1927 (2011)
5. GTEx Consortium et al.: Genetic effects on gene expression across human tissues.
Nature 550(7675), 204–213 (2017)
6. Dijkstra, J.M., Ballingall, K.T.: Non-human lnc-DC orthologs encode Wdnm1-like
protein. F1000Research 3, 160 (2014)
7. Gatter, T., Stadler, P.F.: Ryūtō: network-flow based transcriptome reconstruction.
BMC Bioinform. 20(1), 190 (2019). https://doi.org/10.1186/s12859-019-2786-5
8. Hojo, M.A., et al.: Identification of a genomic enhancer that enforces proper apop-
tosis induction in thymic negative selection. Nat. Commun. 10(1), 1–15 (2019)
9. Hu, Y., et al.: Long noncoding RNA GAPLINC regulates CD44-dependent cell
invasiveness and associates with poor prognosis of gastric cancer. Cancer Res.
74(23), 6890–6902 (2014)
10. Huang, S., et al.: NEAT1 regulates Th2 cell development by targeting STAT6 for
degradation. Cell Cycle 18(3), 312–319 (2019)
11. Hudson, W.H., et al.: Expression of novel long noncoding RNAs defines virus-
specific effector and memory CD8+ T cells. Nat. Commun. 10(1), 1–11 (2019)
12. Hunt, S.E., et al.: Ensembl variation resources. Database 2018 (2018)
13. Kim, D., Langmead, B., Salzberg, S.L.: HISAT: a fast spliced aligner with low
memory requirements. Nat. Methods 12(4), 357–360 (2015). https://doi.org/10.
1038/nmeth.3317
14. Love, M.I., Huber, W., Anders, S.: Moderated estimation of fold change and dis-
persion for RNA-seq data with DESeq2. Genome Biol. 15(12), 550 (2014). https://
doi.org/10.1186/s13059-014-0550-8
15. Martin, M.: Cutadapt removes adapter sequences from high-throughput sequencing
reads. EMBnet. J. 17(1), 10–12 (2011)
16. Niknafs, Y.S., et al.: Taco produces robust multisample transcriptome assemblies
from RNA-seq. Nat. Methods 14(1), 68–70 (2017)
17. Norman, D.J., et al.: The use of OKT3 in cadaveric renal transplantation for
rejection that is unresponsive to conventional anti-rejection therapy. Am. J. Kidney
Dis. 11(2), 90–93 (1988). https://doi.org/10.1016/S0272-6386(88)80186-0
18. Patro, R., et al.: Salmon provides fast and bias-aware quantification of transcript
expression. Nat. Methods 14(4), 417–419 (2017)
19. Quinlan, A.R.: BEDTools: the Swiss-army tool for genome feature analysis. Curr.
Protocols Bioinform. 47(1), 11–12 (2014)
20. Ranzani, V., et al.: The long intergenic noncoding RNA landscape of human lym-
phocytes highlights the regulation of T cell differentiation by linc-MAF-4. Nat.
Immunol. 16(3), 318 (2015)
21. Reichert, J.M.: Marketed therapeutic antibodies compendium. MAbs 4(3), 413–
415 (2012). https://doi.org/10.4161/mabs.19931
 Anti-CD3 Stimulated T Cell Transcriptome Reveals Novel ncRNAs 191

22. Robinson, J.T., et al.: Integrative genomics viewer. Nat. Biotechnol. 29(1), 24–26
(2011)
23. Shui, X., et al.: Knockdown of lncRNA NEAT1 inhibits Th17/CD4+ T cell dif-
ferentiation through reducing the STAT3 protein level. J. Cell. Physiol. 234(12),
22477–22484 (2019)
24. Silva, H.M., et al.: Novel humanized anti-CD3 antibodies induce a predominantly
immunoregulatory profile in human peripheral blood mononuclear cells. Immunol.
Lett. 125(2), 129–136 (2009)
25. Sousa, I.G., et al.: Gene expression profile of human T cells following a single
stimulation of peripheral blood mononuclear cells with anti-CD3 antibodies. BMC
Genomics 20(1), 593 (2019). https://doi.org/10.1186/s12864-019-5967-8
26. Tooley, J.E., et al.: Changes in T-cell subsets identify responders to FcR-
nonbinding anti-CD3 mAb (teplizumab) in patients with type 1 diabetes. Eur.
J. Immunol. 46(1), 230–241 (2016)
27. Wang, P., et al.: The STAT3-binding long noncoding RNA lnc-DC controls human
dendritic cell differentiation. Science 344(6181), 310–313 (2014)
 A Simplified Complex Network-Based
Approach to mRNA and ncRNA
Transcript Classification

Murilo Montanini Breve and Fabrı́cio Martins Lopes(B)

Departamento Acadêmico de Computação (DACOM),

Universidade Tecnológica Federal do Paraná (UTFPR) Câmpus Cornélio Procópio,
Av. Alberto Carazzai,
1640, Cornélio Procópio, PR CEP 86300-000, Brazil
[email protected], [email protected]

Abstract. Bioinformatics is an interdisciplinary area that presents sev-

eral important computational challenges. These challenges are usually
related to the large volume of biological data generated and that needs
to be analyzed for information discovery. An important challenge is the
need to distinguish mRNAs and ncRNAs in an efficient and assertive
way. The correct identification of these transcripts is due to the exis-
tence of thousands of non-coding transcripts, whose function and mean-
ing are not known, as well as the challenge to understand the expres-
sion and regulation of genetic information. On the other hand, the com-
plex network theory has been successfully applied in many real-world
problems in different contexts. Therefore, this work presents a simplified
and efficient complex network-based approach for the classification of
mRNA and ncRNA sequences. Experiments were performed to evaluate
the proposed approach considering a dataset with six different species
and with important methods in the literature such as CPC, CPC2 and
PLEK. The results indicated the assertiveness of the proposed approach
achieving average accuracy rates higher than 98% in the classification of
mRNA and ncRNA considering all compared species. Besides, the pro-
posed approach presents fewer variations on its results when compared
to competitor methods, indicating its robustness and suitability for the
classification of transcripts.

Keywords: RNA classification · Complex networks · Feature

extraction · Bioinformatics · Pattern recognition

1 Introduction
After more than 150 years of the discovery of nucleic acids, the interest in the
study of these molecules has been growing over the years [1]. Since the sequenc-
ing of bacteriophage φX174 in 1977 [2], a high amount of organisms have been
sequenced and stored in databases. The advances in the development of sequenc-
ing technologies has led to the generation of a huge volume of biological data.
c Springer Nature Switzerland AG 2020
J. C. Setubal and W. M. Silva (Eds.): BSB 2020, LNBI 12558, pp. 192–203, 2020.
https://doi.org/10.1007/978-3-030-65775-8_18
 A Simplified Complex Network-Based Approach to mRNA and ncRNA 193

Thus, it became essential that computational tools were developed to analyze

these data. This need led to the emergence of a new interdisciplinary research
field: Bioinformatics, which had its beginning in the 1970’s, defined as “the
study of informatic processes in biotic systems”[3]. Since then, bioinformatics
has become an important interdisciplinary research field.
Nowadays, the high-throughput sequencing methods, such as RNA-seq,
allows the generation of a massive volume of transcritomic data [4]. Besides
the quantification of transcriptomic sequences [5], it is important to assign an
annotation to these sequences (reads), e.g. their functionality [6].
Two classes of transcripts have received a lot of attention due to their func-
tionality in organisms. The first is the mRNAs that carry information for protein
synthesis. The second are the non-coding RNAs (ncRNAs), for which the func-
tionality of only a few is known. In fact, the functionality of the vast majority
of ncRNA remains unknown [7].
Therefore, it is important to identify the classes of RNA in order to better
understand the mechanisms of action and their functionality and thus contribute
to effective analysis and annotation of these sequences. In particular, ncRNAs are
very heterogeneous in terms of length, conformation and cellular function. The
ncRNA can be organized into long non-coding RNA (lncRNA) with sequences
>200 nucleotides and small non-coding RNA (sncRNA) [8,9]. In addition, recent
findings indicate that lncRNAs has an important participation in many biological
processes, such as transcriptional regulation [10], complex diseases [11] including
cancer [12,13], and also analyzed regarding preserved regions in their structure
due to the evolution of living beings [14].
In this context, some methods have been proposed in the literature with
the objective of classifying transcripts, such as: Coding Potential Calculator
(CPC) [15] and its updated version CPC2 [16], PLEK [17] and BiologicAl
Sequences NETwork (BASiNET) [18]. However, the CPC and CPC2 methods
use biological information about the composition of sequences such as open read-
ing frame (ORF) features and also sequence alignment scores from UniProt
protein database. PLEK is a predictor of long non-coding RNAs and messen-
ger RNAs based on an improved k-mer scheme that considers the nucleotide
frequency (k-mer) as feature taking into account a sliding window to count the
k-mer ranging from 1 to 5, not considering sequence alignment or features related
to the structure of the molecule, such as the position or adjacency between the
nucleotides. BASiNET presents a methodology for feature extraction without
considering any biological information a priori about the sequence. This method
performs a mapping from biological sequences to networks (vertices and edges)
and topological measurements are extracted from these networks through the
complex networks theory[19,20].
This work proposes a method for the classification of coding (mRNA) and
non-coding (ncRNA) sequences, based on the BASiNET method, thus consider-
ing the mapping of biological sequences in networks and extracting topological
measures, however improving the learning process and the feature selection lead-
ing to more efficient and accurate classification. The proposed method was evalu-
194 M. M. Breve and F. M. Lopes

ated and compared with the methodologies CPC, CPC2 and PLEK, considering
CPC2 dataset[16] with six different species in which methodology obtained ade-
quate results and with greater accuracy than the competing methods considering
all the species evaluated.

2 Graphs and Complex Networks

The study of graphs can be seen as a subarea of mathematics, having as function

the analysis of the relations between the objects of a certain context [21]. For
this purpose structures called graphs are used, in short, a graph can be seen as a
set of nodes connected by edges, which represent connections between the nodes.
For example, if a sequence composed of the ‘ACG’ nucleotides is considered, and
the neighborhood of the nucleotides is used as a criterion for its connection, this
sequence can be mapped in a graph such as the one shown in Fig. 1.

A C G
Fig. 1. Graph generated from the sequence ‘ACG’, considering the nucleotides as
nodes and its neighborhood as edges.

If the frequency of neighborhood of a given nucleotide is considered, the

graph may contain weights on its edges, maintaining the unique occurrence of
the nodes and traversing the sequence for the identification of neighborhoods.
Besides, instead of considering a single nucleotide as a graph node, it can be
considered k-mers, with different values of k. Figure 2 shows an example of a
graph generated from the sequence ‘GCACCGGCCG’ considering k = 2.

1
AC CG

1 2

GC

Fig. 2. Graph generated from the sequence ‘GCACCGGCCG’, considering k-mers,

with k = 2 as nodes and its neighborhood as edges.

Complex networks can be seen as graphs with non-trivial topologies [22].

These complex networks are present in several areas of knowledge, having been
 A Simplified Complex Network-Based Approach to mRNA and ncRNA 195

successfully applied in the representation of many real-world relationships in

different areas [18,23–32].
An important aspect of the complex networks is their dynamics, i.e. how the
topological structure of networks changes by some factor such as time, threshold,
perturbation, etc. leading to a dynamics for networks [19,20].
The theory of complex networks presents a variety of topological measure-
ments that can be used to characterize and to represent the topologies of these
networks [19].

3 Materials and Methods

3.1 Materials

This work adopts a dataset in order to validate the proposed method as well as
to compare its results with the competitor methods. The adopted dataset was
obtained from CPC2 [16], which contains six species of organisms: Arabidopsis
thaliana, human, zebrafish, fruitfly, mouse and worm. The CPC2 dataset presents
transcripts (mRNA), small non-coding transcripts (sncRNAs) and long non-
coding transcripts (lncRNAs). In this work the sncRNA and lncRNA sequences
were grouped in a ncRNA subset. Redundant sequences were removed from the
dataset. The adopted dataset, the number of samples per class and species are
presented in Table 1.

Table 1. Description of the number of samples per class of the dataset adopted in
this work.

Species Class of RNAs Number of samples

Arabidopsis mRNA 15931
ncRNA 3853
Human mRNA 6142
ncRNA 12015
Zebrafish mRNA 2344
ncRNA 1528
Fruitfly mRNA 3680
ncRNA 3556
Mouse mRNA 10638
ncRNA 12251
Worm mRNA 3551
ncRNA 9313
196 M. M. Breve and F. M. Lopes

3.2 Methods
The first step is to map the sequences in complex networks. Therefore, the
sequences were organized in FASTA files, and the messenger RNA and non-
coding RNA sequences were separated, that is, they were previously classified in
a supervised learning model.
For the mapping of sequences in complex networks, two parameters were
adopted: k and step. The step parameter defines the distance traveled in the
sequence to define the neighborhood of each iteration. The k parameter defines
the amount of nucleotides for forming the k-mer. Figure 3 shows an example
in which the network was generated from the sequence ‘ACCGATG’ with the
parameters step = 1 and k = 3, as defined by the BASiNET [18] method.

ATG
1

1
ACC CCG GAT
1
1

CGA

Fig. 3. Graph generated from the sequence ‘ACCGATG’, considering k-mers, with
k = 3 as nodes and its neighborhood with step = 1 as edges.

The second step is to consider a feature selection approach in order to reduce

noise and improve network representativeness leading to reduced complexity and
contributing to the classification step. The feature selection consists of identi-
fying the exclusive edges for each type of sequences: mRNA and ncRNA, so
that before performing the feature extraction (topological measurements), filter
the edges avoiding the appearance of edges which are repeated in both classes.
Figure 4 presents an overview of the proposed feature selection approach. It is
important to note that the network’s adjacency matrix is transformed into a
binary matrix by filtering the edges according to a percentage of exclusivity
(defined as parameter), leading to the identification of the unique edges between
the RNA classes and providing the filtering in a simplified way by a subtraction
of matrices.
To identify the exclusivity parameter that produces an adequate selection,
the adopted dataset was analyzed and the exclusivity rates were identified in an
iterative way, producing the plot shown in Fig. 5. It is possible to observe that
 A Simplified Complex Network-Based Approach to mRNA and ncRNA 197

mRNA ncRNA
(Sequence) (Sequence)

Network Network
(mRNA) (ncRNA)

Adjacency Adjacency
matrix maxtrix
(mRNA) (ncRNA)

Binary Binary
Adjacency Adjacency
Matrix Matrix
(% of ex- (% of ex-
clusivity) clusivity)
(mRNA) (ncRNA)

Subtraction
(Matrices)

Filter
(Exclusive
Edges)

Fig. 4. Overview of the proposed feature selection approach.

1000

750
Quantity of Edges

500 Exclusive
Not exclusive

250

0
0

20

40

60

(%)

Fig. 5. Quantity of exclusive and non-exclusive edges identified in the dataset in both
classes of RNA.
198 M. M. Breve and F. M. Lopes

the higher point on the curve of the exclusive edges is between 40% and 60%,
and that the saturation point of the exclusive curve is between 40% and 50%.
Thus, the exclusivity parameter adopted in this work was 45% for the feature
selection step.
Figure 6 presents an overview of the feature selection step. It is possible to
notice that (a) presents the exclusive edges, (b) presents the original network
and (c) the network filtered by considering only the exclusive edges, i.e. the
edges that are not present in the filter (exclusive edge) are removed. This feature
selection reduces the complexity of the BASiNET method [18]. More specifically,
BASiNET considers all the network edges and applies a threshold iteratively to
remove less frequent edges (with lower weights). Thus, the topological measure-
ments are extracted at each iteration at different levels of network resolution.
As a consequence, the proposed feature selection approach eliminates the need
to apply the threshold and leads to an improved and simplified approach.

Fig. 6. Overview of the feature selection applied on networks.

After the feature selection step, the next step is to extract topological mea-
surements from the networks. This work adopts some complex network measure-
ments commonly used in the proposed approach, such as: assortativity, medium
degree, maximum degree, minimum degree, average centrality of intermediation,
cluster coefficient, average short path length, average standard deviation, fre-
quency of motifs with size 3 and frequency of motifs with size 4 [20]. However,
each measurement has a different value range, thus a Min-Max is applied and
the measurements values are rescaled in range [0,1].
 A Simplified Complex Network-Based Approach to mRNA and ncRNA 199

The last step consists of classifying the sequences based on their topological
features extracted from their respective networks. For this step, the supervised
learning was adopted taking into account the Random Forest [33] classifier. The
R project [34] was adopted and the rfUtilities [35] package was also considered
for 10-fold cross-validation.

4 Results and Discussion

To evaluate the proposed approach, the adopted dataset (Sect. 3.1) was consid-
ered in the same way for all the methods. The proposed approach was performed
considering the following parameters values: edge exclusivity = 45%, step = 1
and k = 3, which were presented and their values justified in Sect. 3.2.
Table 2 presents the accuracy rates of classifications regarding the mRNAs
and ncRNAs using the 10-fold cross validation. It is possible to verify the ade-
quacy of the proposed approach for the correct identification between ncRNA
and mRNA sequences with a high accuracy rate achieving accuracy rates higher
than 98,7% for all species.

Table 2. Accuracy rates in the classification of mRNA and ncRNA sequences using
the proposed method for different species.

Species Accuracy
mRNA ncRNA
Human 6142 (100%) 12005 (99.9%)
Fruitfly 3665 (99,5%) 3551 (99,7%)
Mouse 10638 (100%) 12251 (100%)
Zebrafish 2313 (98,7%) 1528 (100%)
Arabidopsis 15930 (100%) 3834 (100%)
Worm 3523 (99,7%) 9312 (100%)

Figure 7 presents the average accuracy of the adopted methods for the clas-
sification of the mRNA and ncRNA sequences, considering each species of the
dataset. It can be noted that the CPC and PLEK present greater variations
in their results. While the proposed approach and the CPC2 show more stable
behaviors, the proposed approach presents superior results when compared to all
competitor methods, indicating the suitability in the classification of the mRNA
and ncRNA sequences.
200 M. M. Breve and F. M. Lopes

PLEK
CPC
100 CPC2
Proposed approach

95
(%)

90

85
Fr

A
H

W
Ze

ra
um

ou

or
br
itfl

bi

m
se

afi
an

do
y

sh

ps
is
Fig. 7. Comparison of the accuracy of the methods considering the species separately.

Table 3. Comparison of classifiers by class and average accuracy.

PLEK CPC CPC2 Proposed

mRNA 79.9% 99,50% 95,20% 99,65%
ncRNA 92.31% 87,30% 97,14% 99,92%
Average accuracy 90.75% 93,20% 96,1% 99,78%

Table 3 presents the average results considering all the species available in the
dataset. It can be noted that the classification of mRNA and ncRNA sequences
by the proposed approach presents results superior to competing methods, again
indicating their adequacy in the classification of transcripts.
The results indicated a high accuracy in the identification of RNA sequences
by the proposed approach. The average results obtained both considering the
species individually and when grouped indicate the robustness of the proposed
approach, with small variations and superior accuracies when compared with
competitor methods. Therefore, the feature selection approach by filtering the
exclusive edges proved to be adequate for the correct identification of the fea-
tures, reducing the complexity of the classification and with a high accuracy rate
for transcripts identification.

5 Conclusion
Complex network theory has been successfully applied in modeling various real-
world problems, in particular in bioinformatics. This work applies the theory of
 A Simplified Complex Network-Based Approach to mRNA and ncRNA 201

complex networks by adopting the mapping of transcripts in networks and per-

forming the extraction of topological measurements from these networks. These
measurements are used as features vector for the classification of mRNA and
ncRNA sequences. More specifically, this work is based on the BASiNET method
[18], and proposes a feature selection approach by filtering exclusive edges of the
networks, leading to a reduction in the complexity and obtaining robust and
adequate results.
Experiments were performed to evaluate the proposed approach considering
the CPC2 dataset [16] with mRNA and ncRNA sequences with six different
species. The experiments were performed comparing the classification results
with important methods in the literature such as CPC [15], CPC2 [16] and
PLEK [17]. The results were obtained considering the 10-fold cross-validation.
In general, the proposed approach achieved average accuracy rates above 98%
in the classification of mRNA and ncRNA considering the all compared species,
indicating superior results when compared with competitor methods. Besides,
the proposed approach presents less variations on its results when compared to
competitor methods, indicating its robustness and suitability for the classifica-
tion of transcripts.
As a future work, a further study considering the proposed approach can
lead to a better understanding of the structure of non-coding RNA. In addition,
the proposed approach can be applied to biological sequences considering other
contexts and of different types of biological sequences.

Acknowledgments. This work was financed in part by the Coordenação de Aper-

feiçoamento de Pessoal de Nı́vel Superior - Brasil (CAPES), Conselho Nacional de
Desenvolvimento Cientı́fico e Tecnológico - CNPq (Grant number 406099 /2016-2) and
the Fundação Araucária e do Governo do Estado do Paraná/SETI (Grant number
035/2019).

References
1. Dahm, R.: Friedrich Miescher and the discovery of DNA. Dev. Biol. 278(2), 274–
288 (2005)
2. Sanger, F., et al.: Nucleotide sequence of bacteriophage φ X174 DNA. Nature
265(5596), 687–695 (1977)
3. Hogeweg, P.: The roots of bioinformatics in theoretical biology. PLOS Comput.
Biol. 7(3), 1–5 (2011)
4. Wang, Z., Gerstein, M., Snyder, M.: RNA-Seq: a revolutionary tool for transcrip-
tomics. Nat. Rev. Genet. 10(1), 57–63 (2009)
5. Costa-Silva, J., Domingues, D., Lopes, F.M.: RNA-Seq differential expression anal-
ysis: an extended review and a software tool. PLOS One 12(12), 1–18 (2017)
6. Garber, M., Grabherr, M.G., Guttman, M., Trapnell, C.: Computational methods
for transcriptome annotation and quantification using RNA-Seq. Nat. Methods
8(6), 469–477 (2011)
7. Panwar, B., Arora, A., Raghava, G.P.S.: Prediction and classification of ncRNAs
using structural information. BMC Genomics 15(1), 1–13 (2014)
202 M. M. Breve and F. M. Lopes

8. Wang, K.C., Chang, H.W.: Molecular mechanisms of long noncoding RNAs. Mol.
Cell 43(6), 904–914 (2011)
9. Peng, Y., Li, J., Zhu, L.: Chapter 8 - cancer and non-coding RNAs. In: Ferguson,
B.S. (ed.)Nutritional Epigenomics, volume 14 of Translational Epigenetics, pp.
119–132. Academic Press (2019)
10. Long, Y., Wang, X., Youmans, D.T., Cech, T.R.: How do lncRNAs regulate tran-
scription? Sci. Adv. 3(9), p. eaao2110 (2017)
11. Chen, X., Yan, G.-Y.: Novel human lncRNA-disease association inference based on
lncRNA expression profiles. Bioinformatics 29(20), 2617–2624 (2013)
12. Huarte, M.: The emerging role of lncRNAs in cancer. Nat. Med. 21(11), 1253–1261
(2015)
13. Peng, W.-X., Koirala, P., Mo, Y.-Y.: LncRNA-mediated regulation of cell signaling
in cancer. Oncogene 36(41), 5661–5667 (2017)
14. Kung, J.T., Colognori, D., Lee, J.T.: Long noncoding RNAs: past, present, and
future. Genetics 193(3), 651–669 (2013)
15. Kong, L., et al.: CPC: assess the protein-coding potential of transcripts using
sequence features and support vector machine. Nucleic Acids Res. 35(suppl 2),
W345–W349 (2007)
16. Kang, Y.J., et al.: Cpc2 a fast and accurate coding potential calculator based on
sequence intrinsic features. Nucleic Acids Res. 45(W1), W12–W16 (2017)
17. Li, A., Zhang, J., Zhou, Z.: PLEK: a tool for predicting long non-coding RNAs
and messenger RNAs based on an improved k-mer scheme. BMC Bioinformatics
15(1), 311 (2014)
18. Ito, E.A., Katahira, I., Vicente, F.F.D.R., Pereira, L.F.P., Lopes, F.M.: BASiNET
- biological sequences network: a case study on coding and non-coding RNAs iden-
tification. Nucleic Acids Res. 46(16), e96–e96 (2018)
19. Boccaletti, S., Latora, V., Moreno, Y., Chavez, M., Hwang, D.-U.: Complex net-
works: structure and dynamics. Phys. Rep. 424(4), 175–308 (2006)
20. Costa, L.D.F., Rodrigues, F.A., Travieso, G., Villas Boas, P.R.: Characterization
of complex networks: a survey of measurements. Adv. Phys. 56(1), 167–242 (2007)
21. Diestel, R.: Graph Theory, 3rd edn. Springer-Verlag, Heidelberg (2005)
22. Barabási, A.L.: Linked: How Everything Is Connected to Everything Else and
What It Means. Plume, New York (2003)
23. Watts, D.J., Strogatz, S.H.: Collective dynamics of small-world networks. Nature
393, 440–442 (1998)
24. Barabási, A.-L., Albert, R.: Emergence of scaling in random networks. Science
286(5439), 509–512 (1999)
25. Newman, M.E.J.: The structure and function of complex networks. SIAM Rev.
45(2), 167–256 (2003)
26. Lopes, F.M., Cesar, R.M., da F. Costa, L.: AGN simulation and validation model.
In: Bazzan, A.L.C., Craven, M., Martins, N.F. (eds.) BSB 2008. LNCS, vol.
5167, pp. 169–173. Springer, Heidelberg (2008). https://doi.org/10.1007/978-3-
540-85557-6 17
27. Yi Ming Zou: Modeling and analyzing complex biological networks incooperating
experimental information on both network topology and stable states. Bioinfor-
matics 26(16), 2037–2041 (2010)
28. Lopes, F.M., Cesar Jr, R.M., Costa, L.D.F.: Gene expression complex networks:
synthesis, identification, and analysis. J. Comput. Biol. 18(10), 1353–1367 (2011)
29. Lopes, F.M., Martins Jr., D.C., Barrera, J., Cesar Jr., R.M.: A feature selection
technique for inference of graphs from their known topological properties: revealing
scale-free gene regulatory networks. Inf. Sci. 272, 1–15 (2014)
 A Simplified Complex Network-Based Approach to mRNA and ncRNA 203

30. da Rocha Vicente, F.F., Lopes, F.M.: SFFS-SW: a feature selection algorithm
exploring the small-world properties of GNs. In: Comin, M., Käll, L., Marchiori, E.,
Ngom, A., Rajapakse, J. (eds.) PRIB 2014. LNCS, vol. 8626, pp. 60–71. Springer,
Cham (2014). https://doi.org/10.1007/978-3-319-09192-1 6
31. de Lima, G.V.L., Castilho, T.R., Bugatti, P.H., Saito, P.T.M., Lopes, F.M.: A com-
plex network-based approach to the analysis and classification of images. CIARP
2015. LNCS, vol. 9423, pp. 322–330. Springer, Cham (2015). https://doi.org/10.
1007/978-3-319-25751-8 39
32. de Lima, G.V., Saito, P.T., Lopes, F.M., Bugatti, P.H.: Classification of texture
based on bag-of-visual-words through complex networks. Expert Syst. Appl. 133,
215–224 (2019)
33. Liaw, A., Wiener, M.: Classification and regression by randomforest. R News 2(3),
18–22 (2002)
34. R Core Team. R: A Language and Environment for Statistical Computing. R Foun-
dation for Statistical Computing, Vienna, Austria (2014)
35. Evans, J.S., Murphy, M.A.: rfUtilities. R package version 2.1-3 (2018)
 A Systems Biology Driven Approach to Map
the EP300 Interactors Using Comprehensive
Protein Interaction Network

Shivananda Kandagalla1(B) , Maria Grishina1 , Vladimir Potemkin1(B) ,

Sharath Belenahalli Shekarappa2 , and Pavan Gollapalli3
1 Laboratory of Computational Modeling of Drugs, Higher Medical and Biological School,
South Ural State University, Chelyabinsk 454080, Russia
[email protected], [email protected]
2 Department of PG Studies and Research in Biotechnology and Bioinformatics, Kuvempu
University, JnanaSahyadri, Shankaraghatta, Shivamogga 577451, Karnataka, India
3 Central Research Lab, K.S. Medical Academy, Nitte University (Deemed To Be University),

Deralakatte, Mangalore 575018, Karnataka, India

Abstract. EP300 is one of the putative tumor-suppressor genes and is

mutated/deleted, under expressed/overexpressed in several types of cancer. The
role of EP300 and its interactions during cancer is crucial to explore its reprogram-
ming events that lead to malignant phenotype and acquisition of drug resistance.
In this context, all the experimentally valid EP300 interactors were collected from
the primary protein-protein interaction (PPI) databases and followed by tracing
their subcellular location using the UniProtKB database. Further, all the EP300
interactors were categorized based on their subcellular location and functionally
annotated with the DAVID gene ontology tool. Subsequently, the interactome
of EP300 with its interactors was constructed and identified TP53, CREBBP,
JUN, HDAC1, CTNNB1, MYC, PCNA, HDAC2, FOS, and KAT2B as the top
first neighbors of EP300. Together, the present analysis gives a comprehensive
overview on EP300 interactors located in different subcellular locations.

Keywords: EP300 · Interactome · Cytoscape

1 Introduction
EP300(p300) is a ubiquitously expressed transcriptional coactivator and a member of the
EP300/CBP family of type 3 major lysine (K) acetyltransferases (KAT3), present in all
mammals and found in many multicellular organisms, such as flies, worms, and plants.
In humans, 31 exons in chromosome 22 (locus 22q13) codes for the p300 gene, and gene
size spans approximately 90 kb. Overexpression and inappropriate activation of EP300
are associated with malignancy, tumor size, poor differentiation, tumor progression, and
poor prognosis [1–3]. Increased expression of EP300 has been observed in advanced
human malignancies, such as liver, prostate cancers, primary human breast cancers,
etc., [4].Recent reports highlight EP300 as a central regulator of angiogenesis, hypoxia,

© Springer Nature Switzerland AG 2020

J. C. Setubal and W. M. Silva (Eds.): BSB 2020, LNBI 12558, pp. 204–214, 2020.
https://doi.org/10.1007/978-3-030-65775-8_19
 A Systems Biology Driven Approach to Map the EP300 Interactors 205

and EMT pathway in esophageal squamous carcinoma [5]. The increased expression of
cancer stem cell markers, tumorsphere formation was observed in EP300-depleted cells
and diminished in EP300-overexpressing cells [6]. Apart from cancer, EP300 is a key
player in Rubinstein − Taybi syndrome (RTS or RSTS) disease [7].
EP300 shares high sequence homology with CBP (CREBBR or KAT3A), and less
with other acetyltransferases [8]. Both proteins have almost 86% amino acid residue iden-
tity in the catalytic domain and significant sequence homology was found in several types
of protein-protein interacting motifs, and other non-catalytic domains [9]. In EP300, the
acetyltransferase domain spans from residues 1284 to 1673, and IBiD (Interferon Bind-
ing Domain) located at the C-terminal side. The IBiD contains an NCBD (Nuclear Coac-
tivator Binding Domain) and glutamine-rich domain, followed by a proline-containing
PxP motif. There are three cysteine/histidine-rich domains (C/H) like C/H1, C/H2 (which
is part of the catalytic domain), and C/H3. The C/H1 and C/H3 domains contain transcrip-
tional adaptor zinc fingers (TAZ1 and TAZ2), and additionally, C/H3 domain contains a
ZZ zinc finger. The C/H2 domain contains a plant homeodomain (PHD) and the domains
such as interferon binding homology domain (IHD), KIX domain, and bromodomain is
located between the C/H1 and C/H2 domains [10, 11].
EP300 functions as acetyltransferase by facilitating transcription through acetylation
of histones, transcription factors, sequence-specific DNA binding factors, and basal tran-
scriptional machinery. During intracellular or extracellular signaling the cell must turn
different subsets of genes to regulate different cellular functions accomplished by acety-
lation of histone proteins during transcription. Most of the cellular signaling pathway
such as cAMP signaling pathway, HIF-1 signaling pathway, FoxO signaling pathway,
cell cycle, Wnt signaling pathway, Notch signaling pathway, TGF-beta signaling path-
way, adherens junction signaling, Jak-STAT signaling pathway, DNA damage pathways,
and other pathways use EP300 as downstream effector protein [12, 13]. EP300 responds
to those signaling pathways differently, which mainly depends on the cell environment
and its phosphorylation state. Various proteins such as PKC, cyclin E/CDK-2, CaMKIV,
IKK, and AKT phosphorylate EP300 at different sites which ultimately impact on its
acetyltransferase activity. Along with, self-modification (auto-acetylation) of EP300
also influences on the acetyltransferase activity. EP300 contain methylation sites near
the KIX domain and lysine SUMOylation site near the bromodomain. EP300 also has
acetylation site (17 lysine residues) in the regulatory loop of acetyltransferase domain
and their acetylation is essential for its acetyltransferase activity, and for binding with
other proteins. In addition, EP300 through protein interacting domains binds to the dif-
ferent proteins and thereby it regulates wide variety of signaling pathways [14, 15]. All
these reports clearly show that EP300 regulates signaling pathways by interacting with
multiple proteins and targeting these interactions during disease conditions could be a
good solution. In this concern, all the experimentally valid datasets of EP300 interac-
tors were collected from primary protein interaction databases, followed by tracing their
subcellular locations and functional annotations. Finally, the interactome of EP300 with
its interactors was developed and first-degree interactors were identified.
206 S. Kandagalla et al.

2 Materials and Method

2.1 Collection of EP300 Interactors

The experimentally detected proteins having the interaction with EP300 were extracted
from the public databases such as, IntAct [16], BioGRID [17], APID [18], PINA [19],
Mentha [20], HitPredict [21], WiKi-Pi [22], PIPs [23], PPI-finder [24] and PrePPI [25].
Non-human interactors of EP300 were excluded from the study. Using the UniProt
Knowledge base (UniProtKB) Id mapping, the gene symbols and protein symbols were
identified [26].

2.2 Protein Class and Subcellular Location Analysis

The subcellular location of EP300 interactors was explored using the UniProtKB
database based on the record “Subcellular location”. UniProtKB database act as a cen-
tral hub in identifying functional information of proteins with accurate annotations,
and also it includes widely accepted biological ontologies, classifications and cross-
references, and clear indications of the quality of annotation in the form of evidence
attribution of experimental and computational data (https://www.uniprot.org/help/uni
protkb). The PANTHER classification system was used to identify the protein classes
of EP300 interactors. The PANTHER (Protein Analysis THrough Evolutionary Rela-
tionships) database contains comprehensive information on the evolution and function
of protein-coding genes from 104 completely sequenced genomes. PANTHER classi-
fication tools allow users to classify new protein sequences and to analyze gene lists
obtained from large scale genomics experiments [27, 28].

2.3 Functional Annotation and Pathway Enrichment Analysis

The EP300 interactors located in the different subcellular location were functionally
annotated with gene ontology (GO) terms in the PANTHER database and the pathway
enrichment analysis was performed in the DAVID database (The Database for Annota-
tion, Visualization, and Integrated Discovery) against PANTHER and KEGG pathways
with a p-value < 0.05 [29].

2.4 Construction of EP300 Interactome

The primary protein interaction data of EP300 interactors were extracted from STRING
database v10.5 [31] with a high confidence score of 0.9. The interactions in the STRING
database are derived from different sources: text mining, experiments, co-expression,
neighborhood, gene fusion, and co-occurrence. The high confidence interaction of score
above 0.9 indicates, all the interactions are validated in all the above-mentioned sources.
The low confidence (score 0.7) interactions were considered for N4BP2, MSTO1, MYB,
HOXD10, and KLF16 interactors. The interactome of EP300 with its interactors was
constructed using Cytoscape 3.4.0 [30] based on the subcellular location and the first-
degree interactors of EP300 were identified from the core interactome.
 A Systems Biology Driven Approach to Map the EP300 Interactors 207

3 Results and Discussion

A total of 854 predicted or experimentally validated EP300 interactors were obtained
from public databases as illustrated in the methods section. The predicted EP300 interac-
tors in many of the above primary databases are mainly from indirect clues such as data
mining, Bayesian prediction, structural information, and more information can be found
in [32]. It is difficult to maintain the accuracy of predicted results from the indirect clue
and Zhang QC et al. [25] reported that the interactions from indirect methods are often
more indicative of functional associations between two proteins than of direct physi-
cal interactions. Hence, the interactors only with experimental evidence were selected
to maintain accuracy and other computational predictions without experimental valida-
tions were excluded for the analysis. A total of 540 EP300 interactors were included for
further analysis and a complete list is provided in the supplementary file [37].

3.1 Analysis of Subcellular Location and Protein Class

The subcellular location analysis shows that EP300 interactors were located in different
cell locations, and further based on location we categorized EP300 interactors into three
broad classes: (i) cytoplasm; (ii) nucleus; (iii) both in cytoplasm and nucleus. Among 540
EP300 interactors, 202 interactors present in both cytoplasm and nucleus, 72 interactors
present in the cytoplasm, 263 interactors in the nucleus, and the remaining interactors
subcellular location is not available in the UniprotKB database (excluded for further anal-
ysis). Cytoplasm location includes apical cell membrane, cell membrane, cytoskeleton,
focal adhesion, mitochondrion outer mem-brane, etc., and nucleus location is found to
include chromosome, centromere, nucleus matrix, PML body, etc. Further analysis of
the protein class of these EP300 interactors shows that they are mainly associated with
the protein class nucleic acid binding and transcription factors. Other enriched top pro-
tein classes are transferase, hydrolase, enzyme modulator, receptor, etc., are shown in
subsequent figures (Fig. 1D; 2D; 3D). All these results give a comprehensive overview
of EP300 interactors protein classes and their subcellular locations.

3.2 Functional Enrichment Analysis of EP300 Interactors

The functional enrichment analysis of EP300 interactors present in the different subcel-
lular locations was done separately. The analysis shows that EP300 interactors present
in the cytoplasm enriched in various biological and molecular functions. As shown in
Fig. 1A, the interactors present in the cytoplasm are mainly participating in response to
the extracellular stimulus, positive regulation of apoptosis, positive regulation of pro-
grammed cell death, positive regulation of cell death, and regulation of apoptosis. Further,
these EP300 interactors are mainly enriched in the molecular function of ribonucleotide
binding, nucleoside binding, and ATP binding process (Fig. 1B). The EP300 interactors
present in the nucleus mainly participate in the regulation of transcription, regulation
of transcription (DNA-dependent), regulation of transcription from RNA polymerase II
promoter, and regulation of RNA metabolic process(Fig. 2A). DNA binding, transcrip-
tion regulator activity, and transcription factor activity are the top enriched GO molecular
function are shown in Fig. 2B.
208 S. Kandagalla et al.

Finally, the analysis of EP300 interactors present in both cytoplasm and nucleus
revealed that most of the interactors are participate in the regulation of transcription, pos-
itive regulation of macromolecule metabolic process, and regulation of RNA metabolic
process (Fig. 3B). The transcriptional regulator activity, transcription factor binding,
transcriptional activator activity, and transcription factor activity are the top enriched
molecular function (Fig. 3B). Together, this analysis provides the functional significance
of EP300 interactors present in the different subcellular locations. Interactors present
in the cytoplasm were mainly involved in the biological process such as extracellular
stimulus, positive regulation of apoptosis, positive regulation of programmed cell death,
positive regulation of cell death, etc. Whereas interactors present in both cytoplasm and
nucleus were engaged in almost similar biological processes.

Fig. 1. Functional enrichment of EP300 interactors present in the Cytoplasm: Top annotated
EP300 interactors involved in A) Biological Process, B) Molecular function, C) Pathways (KEGG
and PANTHER), D) PANTHER protein class.

3.3 Pathway Enrichment Analysis of EP300 Interactors

The EP300 interactors present in the cytoplasm were enriched during pathogenic
Escherichia coli infection and ubiquitin-mediated proteolysis based on KEGG pathway
analysis. The PANTHER pathway results suggest that EP300 interactors are mainly asso-
ciated with the apoptosis signaling pathway (Fig. 1C). The EP300 interactors present in
nucleus were enriched in the cell cycle, DNA replication in KEGG pathways and the
p53, p53 pathway feedback loop 2, oxidative response, and Wnt signaling are the top
PANTHER enriched pathways (Fig. 2C).
 A Systems Biology Driven Approach to Map the EP300 Interactors 209

Fig. 2. Functional enrichment of EP300 interactors present in the Nucleus: Top annotated EP300
interactors involved in A) Biological Process, B) Molecular function, C) Pathways (KEGG and
PANTHER), D) PANTHER protein class.

Finally, EP300 interactors present in both cytoplasm and nucleus are enriched during
pathways in cancer, chronic myeloid leukemia, prostate cancer, acute myeloid leukemia,
pancreatic cancer, cell cycle, and ErbB signaling pathway are among the top enriched
KEGG pathways. The PDGF signaling pathway, JAK/STAT signaling pathway, B cell
activation, T cell activation, p53 pathway, EGF receptor signaling pathway, p53 pathway
feedback loops 2 and TGF-beta signaling pathway are among the top enriched PAN-
THER pathways (Fig. 3C). Collectively these analysis provides the details on EP300
interactors associated pathways. From the results, it can be seen that apart from normal
pathways, EP300 interactors also enriched in associated disease related pathways such
as cancer, infection, etc. Further analysis of EP300 interactors associated with disease
related pathways gives broad insights on the role of EP300 and also, it provides a new
avenue in developing new drugs.

3.4 EP300 Interactome and Identification of First-Degree Nodes

The interactome of EP300 with its interactors was constructed to check the influence
of EP300 based on the analysis of the first-degree nodes. First, the network of EP300
interactors present in the nucleus was constructed and the network consists of 1165
nodes and 6046 edges (Fig. 4A). Further first-degree nodes of EP300 were identified
and these nodes have direct contact with EP300 and any alteration in these nodes changes
the signaling pattern. A total of 135 nodes form the direct connection with EP300 and
the top nodes based on the degree are TP53, CREBBP, HIST2H2BE, HDAC1, JUN,
HIST2H2AC, H2AFZ, and MYC. Next, interactome of EP300 interactors present in the
cytoplasm were constructed and the network has 635 nodes with 2331 edges (Fig. 4A).
Further identified first-degree nodes of the EP300 in the network and with 18 nodes
210 S. Kandagalla et al.

Fig. 3. Functional enrichment of EP300 interactors present in the both Cytoplasm and Nucleus:
Top annotated EP300 interactors involved in A) Biological Process, B) Molecular function, C)
Pathways (KEGG and PANTHER), D) PANTHER protein class.

EP300 has a direct connection. The UBA52, NR3C1, GRIP1, SREBF1, and JUN are
the top nodes based on the degree. Further, the network of EP300 interactors present in
the cytoplasm and nucleus were constructed and the network has 1214 nodes with 5443
edges (Fig. 4A). Total 89 nodes have a direct connection with EP300 and among TP53,
JUN, AKT1, CREBBP, HDAC2, HDAC1, and MYC are top nodes based on degree
(complete list is provided in the supplementary file [37]).
Altogether, the final interactome consists of 2388 nodes with 12577 edges. Among
2388 nodes Ep300 form the direct interaction with only 186 nodes (Fig. 4C) and among
TP53, CREBBP, JUN, HDAC1, CTNNB1, MYC, PCNA, HDAC2, FOS and KAT2B are
the top interactors. Previously several reports show the significance of EP300 interactions
with TP53, CREBBP, JUN, CTNNB1, and MYC in several pathophysiological condi-
tions [33–36], and still, their role is not clearly understood. Further, in vitro validation
of these interactors is required to understand the role of EP300 in different cancer condi-
tions and which ultimately helps in developing the novel inhibitor also, these interactors
act as potential biomarkers.
 A Systems Biology Driven Approach to Map the EP300 Interactors 211

Fig. 4. Interactome of EP300 with it interactors, A) The PPI network of EP300 interactors present
in nucleus are colored in red, green nodes corresponds to cytoplasm, and grey node corresponds
to both nucleus and cytoplasm. B) Venn diagram showing number of EP300 interacting partners
present in different subcellular location. C) First neighbors of EP300 in the interactome.

4 Conclusion
The evaluation of EP300 interactors present in different subcellular locations provides
a broad sense to the role of EP300 in complex disease or cellular events. The functional
and pathways enrichment analysis of EP300 interactors clearly shows their involvement
in several pathological conditions and mainly in cancer. Among the EP300 interactors,
TP53, CREBBP, JUN, HDAC1, CTNNB1, MYC, PCNA, HDAC2, FOS, and KAT2B
are the top first-degree nodes, and these interactors are the key players with which EP300
interact and perform its functions. Further, in vitro validation of these interactors with
EP300 is required in different cancer conditions. Altogether, the present analysis gives
the complete overview on EP300 interactors presents different subcellular locations.

Acknowledgement. The work was supported by Act 211 Government of the Russian Federation,
contract 02.A03.21.0011 and by the Ministry of Science and Higher Education of Russia (Grant
FENU-2020–0019).
212 S. Kandagalla et al.

Conflict of Interest. All authors declared that they have no competing interest.

References
1. Karamouzis, M.V., Konstantinopoulos, P.A., Papavassiliou, A.G.: Roles of CREB-binding
protein (CBP)/p300 in respiratory epithelium tumorigenesis. Cell Res. 17, 324–332 (2007).
https://doi.org/10.1038/cr.2007.10
2. Dutto, I., Scalera, C., Prosperi, E.: CREBBP and p300 lysine acetyl transferases in the DNA
damage response. Cell. Mol. Life Sci. 75(8), 1325–1338 (2017). https://doi.org/10.1007/s00
018-017-2717-4
3. Mees, S.T., Mardin, W.A., Wendel, C., et al.: EP300-A miRNA-regulated metas-tasis sup-
pressor gene in ductal adenocarcinomas of the pancreas. Int. J. Cancer 126, 114–124 (2010).
https://doi.org/10.1002/ijc.24695
4. Yang, H., Pinello, C.E., Luo, J., et al.: Small-molecule inhibitors of acetyltrans-ferase p300
identified by high-throughput screening are potent anticancer agents. Mol. Cancer Ther. 12,
610–620 (2013). https://doi.org/10.1158/1535-7163.MCT-12-0930
5. Bi, Y., Zhang, L., et al.: EP300 promotes tumor development and correlates with poor
prognosis in esophageal squamous carcinoma. Oncotarget 9(1), s376–s392 (2018)
6. Asaduzzaman, M., et al.: Tumour suppressor EP300, a modulator of paclitaxel resistance and
stemness, is downregulated in metaplastic breast cancer. Breast Cancer Res. Treat. 163(3),
461–474 (2017). https://doi.org/10.1007/s10549-017-4202-z
7. Babu, A., et al.: Chemical and genetic rescue of an ep300 knockdown model for Rubinstein
Taybi Syndrome in zebrafish. Biochim. Biophys. Acta - Mol. Basis Dis. 1864, 1203–1215
(2018). https://doi.org/10.1016/j.bbadis.2018.01.029.
8. Gayther, S.A., Batley, S.J., Linger, L., et al.: Mutations truncating the EP300 acetylase in
human cancers. Nat. Genet. 24, 300–303 (2000). https://doi.org/10.1038/73536
9. Chan, H.M., La Thangue, N.B.: p300/CBP proteins: HATs for transcriptional bridges and
scaffolds. J. Cell. Sci. 114, 2363–2373 (2001)
10. Ogryzko, V.V., Schiltz, R.L., Russanova, V., et al.: The transcriptional coactiva-tors p300 and
CBP are histone acetyltransferases. Cell 87, 953–959 (1996)
11. Yang, X.J., Seto, E.: Lysine acetylation: codified crosstalk with other post-translational
modifications. Mol. Cell. 31, 449–461 (2008). https://doi.org/10.1016/j.molcel.2008.07.002
12. Vo, N., Goodman, R.H.: CREB-binding protein and p300 in transcriptional regulation. J. Biol.
Chem. 276, 13505–13508 (2001). https://doi.org/10.1074/jbc.R000025200https://doi.org/10.
1074/jbc.R000025200
13. Bedford, D.C., Brindle, P.K.: Is histone acetylation the most important physio-logical function
for CBP and p300? Aging (Albany NY) 4, 247–55 (2012). https://doi.org/10.18632/aging.
100453
14. Attar, N., Kurdistani, S.K.: Exploitation of EP300 and CREBBP lysine acetyltransferases by
cancer. Cold Spring Harb. Perspect. Med. 7, a026534 (2017). https://doi.org/10.1101/cshper
spect.a026534
15. Dancy, B.M., Cole, P.A.: Protein lysine acetylation by p300/CBP. Chem. Rev. 115, 2419–2452
(2015). https://doi.org/10.1021/cr500452k
16. Hermjakob, H., Montecchi-Palazzi, L., Lewington, C., et al.: IntAct: an open source molecular
interaction database. Nucleic Acids Res. 32, D452–D455 (2004). https://doi.org/10.1093/nar/
gkh052
17. Stark, C., Breitkreutz, B.J., Reguly, T., et al.: BioGRID: a general repository for interaction
datasets. Nucleic Acids Res. 34, D535–D539 (2006). https://doi.org/10.1093/nar/gkj109
 A Systems Biology Driven Approach to Map the EP300 Interactors 213

18. Alonso-López, D., Gutiérrez, M.A., Lopes, K.P., et al.: APID interactomes: providing
proteome-based interactomes with controlled quality for multiple spe-cies and derived
networks. Nucleic Acids Res. 44, W529–W535 (2016). https://doi.org/10.1093/nar/gkw363
19. Cowley, M.J., Pinese, M., Kassahn, K.S., et al.: PINA v2.0: mining interactome modules.
Nucleic Acids Res. 40, D862–D865 (2012). https://doi.org/10.1093/nar/gkr967
20. Calderone, A., Castagnoli, L., Cesareni, G.: mentha: aresource for browsing integrated
protein-interaction networks. Nat. Methods 10, 690–691 (2013). https://doi.org/10.1038/
nmeth.2561
21. Patil, A., Nakai, K., Nakamura, H.: HitPredict: a database of quality assessed protein-protein
interactions in nine species. Nucleic Acids Res. 39, D744–D749 (2011). https://doi.org/10.
1093/nar/gkq897
22. Orii, N., Ganapathiraju, M.K.: Wiki-Pi: a web-server of annotated human protein-protein
interactions to aid in discovery of protein function. PLoS ONE 7, e49029 (2012). https://doi.
org/10.1371/journal.pone.0049029
23. McDowall, M.D., Scott, M.S., Barton, G.J.: PIPs: human protein-protein interac-tion pre-
diction database. Nucleic Acids Res. 37, D651–D656 (2009). https://doi.org/10.1093/nar/
gkn870
24. He, M., Wang, Y., Li, W.: PPI Finder: a mining tool for human protein-protein interactions.
PLoS ONE 4, e4554 (2009). https://doi.org/10.1371/journal.pone.0004554
25. Zhang, Q.C., Petrey, D., Garzón, J.I., et al.: PrePPI: a structure-informed data-base of protein-
protein interactions. Nucleic Acids Res. 41, D828–D833 (2013). https://doi.org/10.1093/nar/
gks1231
26. Bateman, A., Martin, M.J., O’Donovan, C., et al.: UniProt: the universal protein knowl-
edgebase. Nucleic Acids Res. 45, D158–D169 (2017). https://doi.org/10.1093/nar/gkw
1099
27. Thomas, P.D., Campbell, M.J., Kejariwal, A., et al.: PANTHER: a library of pro-tein families
and subfamilies indexed by function. Genome Res 13, 2129–2141 (2003). https://doi.org/10.
1101/gr.772403
28. Mi, H., Dong, Q., Muruganujan, A., et al.: PANTHER version 7: improved phy-logenetic
trees, orthologs and collaboration with the gene ontology consortium. Nucleic Acids Res. 38,
D204–D210 (2010). https://doi.org/10.1093/nar/gkp1019
29. Huang, D.W., Sherman, B.T., Tan, Q., et al.: The DAVID gene functional classification tool:
a novel biological module-centric algorithm to functionally analyze large gene lists. Genome
Biol. 8, R183 (2007). https://doi.org/10.1186/gb-2007-8-9-r183
30. Shannon, P., Markiel, A., Ozier, O., et al.: Cytoscape: a software environment for integrated
models of biomolecular interaction networks. Genome Res. 13, 2498–2504 (2003). https://
doi.org/10.1101/gr.1239303
31. Szklarczyk, D., Morris, J.H., Cook, H., et al.: The STRING database in 2017: quality-
controlled protein-protein association networks, made broadly accessible. Nucleic Acids Res.
45, D362–D368 (2017). https://doi.org/10.1093/nar/gkw937
32. Kong, F.Y., et al.: Bioinformatics analysis of the proteins interacting with LASP-1 and their
association with HBV-related hepatocellular carcinoma. Sci. Rep. 7, 1–5 (2017). https://doi.
org/10.1038/srep44017
33. Kaypee, S., et al.: Mutant and wild-type tumor suppressor p53 induces p300 autoacetylation.
iScience 4, 260–272 (2018). https://doi.org/10.1016/j.isci.2018.06.002
34. Attar, N., Kurdistani, S.K.: Exploitation of EP300 and CREBBP lysine acetyltransferases by
cancer. Cold Spring Harb. Perspect. Med. 7 (2017). https://doi.org/10.1101/cshperspect.a02
6534.
214 S. Kandagalla et al.

35. Yu, W., et al.: Cellular Physiology and Biochemistry Cellular Physiology and Biochemistry
Yu et al.: B-catenin cooperates with CREB binding protein β-catenin cooperates with CREB
binding protein to promote the growth of tumor cells cellular physiology and biochemistry
cellular physiology and biochemistry. Cell Physiol. Biochem. 44, 467–478 (2017). https://
doi.org/10.1159/000485013.
36. Wang, Y.N., Chen, Y.J., Chang, W.C.: Activation of extracellular signal-regulated kinase
signaling by epidermal growth factor mediates c-Jun activation and p300 recruitment in ker-
atin 16 gene expression. Mol. Pharmacol. 69, 85–98 (2006). https://doi.org/10.1124/mol.105.
016220
37. Kandagalla, S., Grishina, M., Potemkin, V., Shekarappa, S.B., Gollapalli, P.: A systems biol-
ogy driven approach to map the EP300 interactors using comprehensive protein interaction
network (2020). https://doi.org/10.5281/ZENODO.4112838
 Analyzing Switch Regions of Human Rab10
by Molecular Dynamics Simulations

Levy Bueno Alves1(B) , William O. Castillo-Ordoñez2 , and Silvana Giuliatti1

1 Department of Genetics, University of São Paulo, Ribeirão Preto, Brazil
[email protected]
2 Department of Biology, University of Cauca, Popayán, Colombia

Abstract. Rab10 is a small GTPase that regulates cellular processes by alternat-

ing between its GDP-bound inactive and the GTP-bound active states. Studies
have shown that functional deficiencies in the Rab10 pathways are implicated in
ciliophaties, gliobastomas and neurodegenerative diseases. Thus, the modulation
of Rab10 activity may represent an interesting strategy in drug discovery. In order
to identify potential Rab10 inhibitors for the treatment of Alzheimer’s disease, we
studied the mobility of the switch1-interswitch-switch2 surface to understand the
active “ON” and inactive “OFF” states of this enzyme. Even today, no in silico
study on Rab10 linked to GTP and GDP has been carried out. We used molecular
dynamics simulations to investigate the atomic movements of the Rab10 switch
regions associated with these nucleotides. We found noticeable differences in the
local flexibility of switch 1 when Rab10 was linked to GDP. However, the heuristic
method used was not able to successfully differentiate the flexibility of switch 2
region. We hypothesized that the flexibility of the switch 1 region can be used as
an indicator of in silico studies that search potential competitive inhibitors based
on nucleotides against Rab10. Furthermore, the present study can be useful for
research that involves the description on-to-off process of other target proteins.

Keywords: Small GTPases · Structural flexibility · In silico

1 Introduction

Rab10 is a small monomeric enzyme that belongs to the Rab GTPases family [1]. It is
responsible for regulating intracellular traffic in various pathways of different cellular
sublocations, having roles in the endoplasmic reticulum, trans-Golgi network, endo-
somes, lysosomes and primary cilium [2]. Functional deficiencies in the Rab10 pathways
are implicated in ciliopathies [3], glioblastomas [4] and neurodegenerative diseases [5].
Studies have shown that Rab10 has a relevant role in Alzheimer disease (AD), helping in
the process of the amyloid precursor protein (APP) and in the production of Aβ through
intracellular vesicle transport [6]. Such evidence paves the way for the application of
new strategies for targeting drugs in the treatment of AD. Thus, modulation of Rab10
activity may represent an alternative to reduce the proportion of neurotoxic Aβ, making
it a potential therapeutic target for the prevention and treatment of AD [5].

© Springer Nature Switzerland AG 2020

J. C. Setubal and W. M. Silva (Eds.): BSB 2020, LNBI 12558, pp. 215–220, 2020.
https://doi.org/10.1007/978-3-030-65775-8_20
216 L. B. Alves et al.

Rab GTPases regulate cellular processes by alternating the nucleotides GTP and
GDP. When linked to GTP, its switch 1, interswitch and switch 2 regions interacts with a
series of effector proteins promote downstream signaling events. On the other hand, the
hydrolysis of GTP results in conformational changes in the G domain of these enzymes,
inactivating them [7]. The differences between the conformations of the G domain linked
to GDP and GTP suggest that after the hydrolysis of GTP the switch 1 and switch 2
regions show a high degree of flexibility and disorder. In contrast, such regions are
stabilized in the active state, which favors Rab10 to be recognized by effector proteins
[8].
Therefore, the present study aimed to detail the structural flexibility of Rab10 and
its switch regions, considering simulations of 200 ns of molecular dynamics (MD). In
the past, 10 ns MD simulations have been done to investigate the internal movements of
Rab5a of wild and mutant type [9]. However, even today, no in silico study on Rab10
linked to GTP and GDP has been carried out. Due to the unavailability of Rab10 crys-
tallographic models associated with GTP and GDP, we used the molecular docking
technique to form complexes with such nucleotides. Thus, it was possible to analyze
atomic movements using classical mechanics and verify whether the heuristic method
used was able to describe the active “ON” and inactive “OFF” state of this enzyme.
The results discussed here may be useful for MD studies that aim to identify potential
competitive inhibitors based on nucleotides against Rab10.

2 Methodology
2.1 Molecular Docking

Rab10 (PDB ID: 5SZJ) [10], GDP and GTP structures were downloaded from the Protein
Data Bank (PDB) [11]. Modeller software v9.23 [12] was used to fill the missing atoms
of Rab10. The addition of hydrogen in each structure, considering the protonation state
of the atoms at physiological pH, was performed using the Open Babel 3.0.0 software
[13]. The Autodock Vina 1.1.2 software [14] was used to docking the nucleotides at the
active site of Rab10. The grid box with a size of 22 Å3 , was defined by the average of
the Cartesian coordinates of the co-crystallized GNP nucleotide, being 34.692, 26.252
and –46.027 for the x, y and z axes, respectively. The GNP compound was redocked to
validate the docking study. The poses of each nucleotide were chosen by means of the
lowest binding energy and the highest number of intermolecular bonds. The interactions
between the ligands and receptor were calculated using the Maestro 12.3 interface [15].

2.2 Molecular Dynamics

The GROMACS package version 2019.3 [16] was used in the MD simulations of com-
plexes with GDP and GTP. The force field used was CHARMM36 [17]. The ligand
parameters were obtained by the CGenFF server [18]. The complexes were centralized
in cubic boxes, where the distance between the solute and the edge was 14 Å. The
molecules were solvated with TIP3P water molecules and neutralized by adding the
appropriate number of Na + Cl- ions considering the ionic concentration of 0.15 M. The
 Analyzing Switch Regions of Human Rab10 217

energy minimization was performed using the steepest descent method with a maximum
force of 1000 kJ.mol−1 .nm−1 . After minimization, the systems were equilibrated in two
stages: a canonical NVT set (number of particles, volume and temperature) followed
by an isothermal-isobaric NPT set (number of particles, pressure and temperature). The
NVT equilibrium was performed with a constant temperature of 300 K for 500 ps. The
NPT equilibrium was performed with a constant pressure of 1 bar and a constant tem-
perature of 300 K for 500 ps. The production step was carried out at 300 K for 200 ns
and the trajectories were saved every 10 ps. The tools of the root mean square deviation
(RMSD), root mean square fluctuation (RMSF), radius of gyration (Rg ) and solvent
accessible surface area (SASA) were used for the trajectory analysis.

3 Results and Discussion

3.1 Molecular Docking Study

The lowest energy values for each test were grouped and their molecular interactions
were analyzed. The most promising poses of each ligand are described in Table 1. The
comparison between the co-crystallized GNP ligand pose and all docking poses indicated
RMSD ≤ 0.60 Å. These values are lower than the tolerance level of 2.0 Å, indicating
that the docking protocol has been validated. The GDP and GTP nucleotides were suc-
cessfully docking at the active site of Rab10. These complexes presented intermolecular
bonds in common: 2 saline bridges are formed with residue K22 and another with D125;
6 hydrogen bonds involve residues G21, K22, T23, C24, D125 and K154; and 2 π stack-
ing interactions are found in residues F34. Moreover, the presence of γ-phosphate in
GTP guarantees four more bonds of hydrogen with residues G19, T41, G67 and D125.
These binding modes are consistent with the interactions found in the crystal of the
Rab10 structure associated with GNP [10].

Table 1. Score by the Autodock Vina and the number of interactions calculated by Maestro.

Ligand Score Salt Bridge HBonds Stacking

(Kcal.mol−1 )
GDP – 10,7 3 6 2
GTP – 11,5 3 10 2

3.2 Molecular Dynamics Study

The RMSD is a crucial parameter to analyze the stability of biomolecular simulations

along the trajectories of MD. Based in our findings, we observed that the stability of the
trajectories of the two systems is only achieved after 100 ns of simulation. The results
of the RMSD of these systems are shown in Table 2. considering all residues of the
enzyme and those present in the switch regions. In these results, the switch 1 region
218 L. B. Alves et al.

showed significant differences when Rab10 was associated with the tested nucleotides.
Here, the Rab10_GDP system showed greater fluctuations compared to Rab10_GTP.
This is explained by the absence of γ-phosphate in GDP, which makes the switch 1
region more flexible due to the lack of stabilizing bonds between the enzyme and this
nucleotide. In contrast, the results found for switch 2 did not reflect the nature of the
structural flexibility of Rab10. High RMSD values were expected in switch 2 when Rab10
was associated with GDP, indicating the high fluctuations resulting from disordered
movements. However, in this region the Rab10_GDP system showed a lowest RMSD
value. In the case of interswitch, although the RMSD has been higher in the Rab10_GTP
system, the variations in flexibility were not significant.

Table 2. Analysis of RMSD (nm) for the entire enzyme, switch 1, interswitch and switch 2.

Systems Entire enzyme Switch 1 Interswitch Switch 2

Res. M1–P175 Res. D31–I44 Res. D45–T65 Res. A66–A82
Rab10_GDP 0.34 ± 0.07 0.49 ± 0.08 0,33 ± 0.06 0,47 ± 0.07
Rab10_GTP 0.29 ± 0.10 0.26 ± 0.09 0,27 ± 0.05 0,49 ± 0.10

Figure 1 shows the results of the RMSF, SASA and Rg analyzes of the enzyme. The
RMSF allows analyzing the amino acid residues that contributed most to the fluctuations
during the simulation. As it can be seen in Fig. 1A, the residues composes the switch 1
region showed greater fluctuations when Rab10 is linked to GDP and lesser fluctuations
when linked to GTP. Although the Rab10_GDP system had predominantly higher peaks
than Rab10_GTP in the interswitch region, the difference in RMSF values between the
two systems was subtle. In relation to the switch 2 region, the RMSF values reflected
greater fluctuations when Rab10 is associated with GTP, which is erroneous and does
not represent Rab10’s biological behavior.
The values of the radius of gyration of switch 1 (see Fig. 1B) confirm that this
region has more disordered movements when Rab10 is associated with GDP. Here, the
Rab10_GDP system showed an average of 0.94 ± 0.03 nm, while Rab10_GTP 0.91 ±
0.01 nm. Constant values of Rg indicate structures folded in a stable way, this indicates
that the switch 1 region of Rab10 has greater flexibility when linked to GDP. The analysis
of the interswitch radius of gyration (see Fig. 1C) showed that for the two systems studied,
the flexibility is stable, where the calculated mean of the interswitch region was 1.07
± 0.01 nm for all systems. In switch 2 (see Fig. 1D), the Rab10_GTP system showed
greater disorder in fold movements (0.79 ± 0.03 nm), while Rab10_GDP had better
stability, with Rg of 0.81 ± 0.01 nm.
The SASA analysis allowed to quantify the molecular surface and describe the con-
tact between Rab10 and the solvent. The systematic increase in SASA indicates the
destabilization of the biomolecule, which can expose its hydrophobic regions to the sol-
vent [19]. Figure 1E shows that the SASA of the Rab10_GDP system has predominantly
higher peaks than Rab10_GTP. When Rab10 is linked to GDP, the average of SASA
was 105.92 ± 2.40 nm2 ; when connected to GTP, it was 101.91 ± 2.86 nm2 . Thus, we
 Analyzing Switch Regions of Human Rab10 219

1.50
RMSF (nm)
1.00 S1 Int S2
0.50

0.00
0 25 50 75 100 125 150 175
A Residue number

1.20 1.20 1.20

1.10 1.10 1.10
Rg (nm)

Rg (nm)

Rg (nm)
1.00 1.00 1.00
0.90 0.90 0.90
0.80 0.80 0.80
0.70 0.70 0.70
0 100 200 0 100 200 0 100 200
B Time (ns) C Time (ns) D Time (ns)
120
SASA (nm²)

110

100

90
0 25 50 75 100 125 150 175 200
E Time (ns)
Fig. 1. Analysis of the trajectories obtained in the MD simulation: the grey line represents the
Rab10_GTP system and the black line Rab10_GDP system. (A) RMSF of the entire enzyme:
Switch 1 (S1) region is defined by positions 31–44, while interwitch (In) and Switch 2 (S2), 45–
65, 66–82, respectively. (B) Rg of switch 1. (C) Rg of Interswitch. (D) Rg of switch 2. (E) SASA
of the entire enzyme.

can infer that the disordered movements of switch 1 and the absence of γ-phosphate
contribute to the increase in SASA of Rab10.

4 Conclusions

In short, the MD simulations used in this study were able to obtain notable differences in
the switch 1 region of Rab10, enabling the identification of its active “ON” and inactive
“OFF” states. However, the classical mechanics method was unable to accurately predict
the disordered movements of the switch 2 region. We hypothesized that the flexibility
of the switch 1 sensitive region can be used as an indicator of in silico studies that
search potential competitive inhibitors based on nucleotides against Rab10. Our findings
suggest that the in silico study of the flexibility of sensitive regions involved in the on-
to-off mechanism of other protein targets may be useful in the discovery of potential
drug candidates.
220 L. B. Alves et al.

References
1. Yan, T., Wang, L., Gao. J., et al.: Rab10 Phosphorylation is a Prominent Pathological Feature
in Alzheimer’s Disease. J. Alzheimer Dis. 63(1), 157–165 (2018)
2. Chua, C.E.L., Tang, B.L.: Rab10 – a traffic controller in multiple cellular pathways and
locations. J. Cellular Phys. 233(9), 6483–6494 (2018)
3. Ordónez, A.J.L., Fernández, B., Fdez, E., et al.: RAB8, RAB10 and RILPL1 contribute to
both LRRK2 kinase–mediated centrosomal cohesion and ciliogenesis deficits. Human Mol.
Genetycs 28(21), 3552–3568 (2019)
4. Shen, G., Mao, Y., Su, Z., et al.: PSMB8-AS1 activated by ELK1 promotes cell proliferation
in glioma via regulating miR-574-5p/RAB10. Biomed. Pharmacother. 122(1), 109658 (2020)
5. Ridge, P.G., Karch, C.M., Hsu, S., et al.: Linkage, whole genome sequence, and biological
data implicate variants in RAB10 in Alzheimer’s disease resilience. Genome Med. 9(1), 100
(2017)
6. Tavana, J.P., Rosene, M., Jensen, N.O., et al.: RAB10: an Alzheimer’s disease resilience locus
and potential drug target. Clin. Interv. Aging 14(1), 73–79 (2019)
7. Good, R.G., Müller, M.P., Wu, Y.: Mechanisms of action of Rab proteins, key regulators of
intracellular vesicular transport. Bio. Chem. 398(5–6), 565–575 (2017)
8. Pylypenko, O., Hammich, H., Yu, I., et al.: Rab GTPases and their interacting protein partners:
Structural insights into Rab functional diversity. Small GTPases 9(1–2), 22–48 (2018)
9. Wang, J., Chou, K.: Insight into the molecular switch mechanism of human Rab5a from
molecular dynamics simulations. Bio. Biophys. Res. Commun. 390(3), 608–612 (2009)
10. Rai, A., Oprisko, A., Campos, G., et al.: bMERB domains are bivalent Rab8 family effectors
evolved by gene duplication. eLife 5(1), e186475 (2016)
11. Berman, H.M., Westbrook, J., Feng, et al.: The Protein Data Bank. Nucleic Acids Res. 28(1),
235–242 (2000)
12. Sali, A., Blundell, T.L.: Comparative protein modelling by satisfaction of spatial restraints.
J. Mol. Biol. 234(1), 779–815 (1993)
13. O’Boyle, N.M., Banck, M., James, C.A., et al.: Open Babel: An open chemical toolbox. J.
Cheminformatics 3(1), 33 (2011)
14. Trott, M., Olson, A.J.: AutoDock Vina: Improving the speed and accuracy of docking with
a new scoring function, efficient optimization, and multithreading. J. Comput. Chem. 31(1),
455–461 (2010)
15. Schrödinger Release 2020–3.: Maestro. New York NY (2020)
16. Spoel, D.V.D., Lindahl, E., Hess, B., et al.: GROMACS: Fast, flexible, and free. J. Comput.
Chem. 26(1), 1701–1718 (2005)
17. Huang, J., MacKerell Jr., A.D.: CHARMM36 all-atom additive protein force field: Validation
based on comparison to NMR data. J. Comput. Chem. 34(1), 2135–2145 (2013)
18. Vanommeslaeghe, K., MacKerell Jr., A.D.: Automation of the CHARMM General Force Field
(CGenFF) I: Bond Perception and Atom Typing. J. Chem. Inf. Model. 52(12), 3144–3154
(2012)
19. Paul, M., Panda, M.K., Thatoi, E.H.: Developing Hispolon-based novel anticancer thera-
peutics against human (NF-κβ) using in silico approach of modelling, docking and protein
dynamics. J. Biomol. Struct. Dyn. 37(15), 3947–3967 (2019)
 Importance of Meta-analysis in Studies
Involving Plant Responses to Climate
Change in Brazil

Janaina da Silva Fortirer1 , Adriana Grandis1 ,

Camila de Toledo Castanho2 , and Marcos Silveira Buckeridge1(B)
1
Laboratory of Physiology and Ecology of Plants (Lafieco), Department of Botany,
Biosciences Institute, University of São Paulo, São Paulo, Brazil
[email protected]
2
Departamento de Ciências Ambientais,
Universidade Federal de São Paulo - UNIFESP, São Paulo, Brazil

Abstract. Meta-analysis synthesizes individual research results on the

same subject and provides information that indicates bottlenecks in
research. Due to the massive production of data, integrative analyzes
are necessary, presenting more consistent views of biological phenom-
ena. Broad themes such as plants’ response to climate change have been
the subject of meta-analyses since 1996, as there is a global concern
about the effect of elevated CO2 on plants and forests. We propose using
meta-analysis to compile existing data, including studies related to the
effects of high CO2 on Brazilian biomes’ vegetation. For that, we found
36 articles on the theme after a systematic review. Physiological param-
eters such as photosynthesis, leaf area, and non-structural carbohydrates
are essential to understand the plant’s responses to elevated CO2 using
meta-analysis. However, these parameters are not present in a consider-
able portion of the literature, decreasing the statistical power of meta-
analytical strategies. The meta-analysis of plants’ biological responses is
usually performed with several species, although there are also studies
with single species. The use of many species increases the variance of
the effects, highlighting the need for multilevel modeling to consider the
dependence among data on the same species. We discuss how to carry
out studies considering the variables needed in future meta-analyses to
contribute to better data integration relevant to national reports. In
this way, we expect that meta-analytical strategies could be essential for
national decision-making and complement global analyses such as those
made by the IPCC.

Keywords: Elevated CO2 · High temperature · Plant physiology ·

Tropical species

Supported by Capes. Fellowship: 88882.461730/2019-01.

c Springer Nature Switzerland AG 2020
J. C. Setubal and W. M. Silva (Eds.): BSB 2020, LNBI 12558, pp. 221–234, 2020.
https://doi.org/10.1007/978-3-030-65775-8_21
222 J. da S. Fortirer et al.

1 Introduction
The number of scientific publications is increasing exponentially, and there is a
need for revisions that compile data produced to guide several study areas [34].
For many years these reviews were carried out by narrative reviews [12]. How-
ever, the narrative review is subjective and often not reproducible, as it can be
skewed depending on the point of view and preferences of each author [36]. To
answer this demand, the reviews have been applying the meta-analysis method-
ology that incorporates a previous systematic review to extract information,
following the studies’ inclusion criteria, where all the steps are documented [28].
Meta-analysis is a set of statistical methods that quantitatively compares the
results of different studies that address a common issue [10,22]. Meta-analysis is
the grandmother of ‘big data’ and ‘open science’. The implementation of meta-
analytic techniques was the first effort to collect and synthesize pre-existing data
to determine patterns, make predictions, and make evidence-based decisions[11].
Furthermore, qualitative data presentation that indicates gaps in current knowl-
edge and new research needs to be performed [33]. The meta-analysis is the
analysis of the analyzes [5]. After Glass [10] first used the term “meta-analysis”
in 1976, the method has been widely applied and developed in the areas of
medicine and sociology [41]. In the 1990 s, meta-analysis has been used in ecol-
ogy and evolutionary biology [33] and is not yet widely known in the biological
sciences [33].
Several studies show terrestrial ecosystems’ responses to climate change,
mainly due to increased atmospheric CO2 . The concentration of carbon diox-
ide (CO2 ) in the atmosphere has increased from ∼ 280 ppm (parts per million)
to ∼ 410 ppm from the industrial revolution to the present [19]. This increase
in atmospheric CO2 is due to fossil fuels, forest burning, and land-use changes
[18,19]. The projection for the 2100 s of the Intergovernmental Panel on Climate
Change (IPCC) is that the CO2 levels will increase to 1300 ppm [18]. How-
ever, CO2 is not only one of the leading gases responsible for the greenhouse
effect (GHG) [19,38], but it is also an essential component for photosynthesis,
leading to growth and higher productivity of the ecosystem [2,23]. Plant’s dry
mass consists of 40% carbon fixed by photosynthesis [25]. The CO2 concentra-
tion increases lead to a rise in temperature, and together they induce drastic
changes in the terrestrial ecosystem, such as changing the pattern of rainfall in
certain regions [23], or they can also cause tree mortality due to water losses
or the forest burning [29]. Thus, most studies performed in this century try to
understand how plants can respond to the increase in CO2 . Indeed, several indi-
vidual publications have focused on the three variables that mostly affected the
climate: CO2 , temperature, and water stress.
Meta-analyses on climate change have been primarily applied in studies with
temperate climate species and have proven to be a valuable tool in this field
[2,6,47]. Curtis et al. [6] used the meta-analysis to summarize more than 500
studies on the high CO2 effects. They conclude that there is a 28% increase in
tree biomass allocation. Wand et al. [47] showed that biomass increased by 33%
and 44% when submitted to high CO2 in plants holding C3 and C4 photosynthe-
 Meta-analysis on Plants and Climate Change in Brazil 223

sis, respectively. In another meta-analysis covering 120 individual studies, the

main effects on the plant physiology under high CO2 in the Free Air Carbon
Enrichment (FACE) system were described [2]. However, as mentioned above,
these studies were performed on plants from the temperate climate, and it is
necessary to include studies from tropical climate [2]. Tropical and subtropi-
cal forests contribute significantly to carbon assimilation and storage on the
planet. However, until 2009, research on tropical species represented less than
10% of the studies [24]. Despite representing 13% of the Earth’s surface, tropi-
cal forests resemble, in addition to their enormous genetic diversity, the largest
carbon reservoir (52%) in the world (340 Gt/C biomass) [24]. The Brazilian ter-
ritory sits on a larger tropical area in South America. Displaying six different
biomes (Amazon, Cerrado, Atlantic Forest, Caatinga, Pampas, and Pantanal),
the neotropical region is often considered a hotspot of biodiversity [1]. Due to
the tropical species richness found in Brazil, studies involving the effects of an
increase of CO2 on plants’ growth and development in these biomes have been
performed [3,35], contributing to the knowledge of the tropical plant responses
to climate change. However, there is still no meta-analysis study that synthesizes
the total effect of high CO2 on Brazil’s tropical plants from these published data.
To address this problem, this work had the following objectives: 1) to approach
the importance of the meta-analysis methodology with a systematic review for
use in studies on climate change; 2) Provide an overview of Brazilian research on
the effect of high CO2 on tropical plants, and 3) discuss future studies on which
species and biomes should be prioritized and some parameters that could better
explain the variation in the high CO2 effects on plants. This way of compiling
the data may, in the future, provide a scientific basis for the adoption of the best
public policy and a reliable database about the response of Brazilian forests to
climate change.

2 Methodology
2.1 How to Get the Data to Perform the Meta-analysis
The steps for performing meta-analysis follow the scientific research procedures:
problem, question, hypothesis, method, data collection, and data analysis [27].
Data collection in the meta-analysis studies is carried out employing a systematic
review, which selects primary studies and, in this case, experimental studies on
plant responses to the increase in CO2 [28]. A systematic review is necessary
to identify and describe all the steps performed in selecting studies and the
extraction of data so that the general result will be reproducible [28](Fig. 1A). In
this work, we describe the history of meta-analysis using studies on the effect of
elevated CO2 on plants. These studies were obtained from the search in the Web
of Science. The keywords used in the search were: “meta-analysis” OR “meta-
analytic” AND “CO2” AND “plant,” on August 18, 2020. To verify the primary
studies on the effect of increasing CO2 in plants in Brazil, a search was carried
out in three databases (Web of Science, Scielo, and Brazilian Digital Library of
Theses and Dissertations). The words used in this search were: “elevated CO2”
224 J. da S. Fortirer et al.

OR “increas* CO2” OR “CO2 enrichment” OR “rising CO2” OR “high* CO2”.

This search was conducted in March 2019. The criteria of inclusion of the works
found in these searches were as follows: (I) studies conducted in Brazil with the
effect of elevated CO2 on plant physiological responses; (II) experiments with
Free Air Carbon Enrichment (FACE) or Open-Top Chamber (OTC) systems;
(III) studies that added: sample size, mean and standard deviation/error of the
control and treatment group (Fig. 1A).
After completing all stages of the systematic review, the full text of Brazil-
ian primary studies included in the meta-analysis was screened to extract the
following data: species name, sample number, average and standard devia-
tion/standard error for the control group (ambient CO2 ), and treatment (ele-
vated CO2 ) for each variable (see Fig. 3), experimental system (FACE or OTC)
and experiment time. The moderating variables contribute to part of the vari-
ance observed in a meta-analysis. The moderator analysis can be conducted to
determine the heterogeneity sources and the extent to which it contributes to the
observed variability in effect sizes among studies [36]. In this sense, the moderat-
ing variables verified were the life habit (tree or herbaceous) and the photosyn-
thetic pathway (with C3 or C4 metabolism) to check whether these moderating
variables influence the plants’ magnitude responses to elevated CO2 .

2.2 How to Apply a Meta-analysis

The first stage to perform a meta-analysis is determining the type of effect size
[27] (Fig. 1B). Effect size is the basic unit of meta-analysis. It makes possible the
standardization of individual studies’ results by providing an average estimate
of the effect of elevated CO2 compared to ambient CO2 [12,14], and the effect
size’s in logarithmic scale is generally used [27]. The choice of the effect size used
for the dataset extracted from the primary studies raises the necessity to check
whether there is dependence on the data. Research in biology may include more
than one analysis in the same experiment, and there are also different studies
for the same species or studies that investigate several species. When these data
types are included in the meta-analysis, the study requires a hierarchical model
approach because it allows dependence on the data set [33]. After adjusting to
the models, it is necessary to perform the heterogeneity test to determine the
variation among studies, which is not attributable to the sample variance [33].
The publication bias test should be applied in the meta-analysis. The publica-
tion bias occurs when the published studies are not representative of the totality
of the studies performed. For example, significant results that confirm the expec-
tations of the research are more likely to be published than non-significant results
[13]. The methods commonly used to assess publication bias are funnel charts
[39] and the Egger test [9]. Finally, it is recommended to perform the sensitivity
test to check the data’s consistency and possible outliers. For the present study,
we used “metafor” package for analysis [46] and ggplot2 package for the graphics
[48], both in the R version 3.6.0 program [42].
 Meta-analysis on Plants and Climate Change in Brazil 225

3 Results and Discussion

3.1 History of Climate Change Conferences and the Impact
on the Meta-analytical Production on the Effect of Elevated
CO2 on Plants

Figure 2 presents historical highlights of the evolution of global awareness about

climate change (Fig. 2A). These historical steps are contrasted with the evolu-
tion of the scientific publication regarding the effects of elevated CO2 and tem-
perature on plants (Fig. 2B). These steps in climate change awareness history
(Fig. 2A) highlight society, scientists, and the government’s concern with climate
change. These and many other conferences (Conference of the Parties - COPs)
and the Intergovernmental Panel on Climate Change (IPCC) reports have been
gradually changing political decision-making worldwide. Nevertheless, in order
to design effective public policies to mitigate and adapt to the effects of climate
change, there is still an urgent need to produce more scientific knowledge about
the impact of climate change worldwide.
Due to concerns about air pollution and acid rain, in 1972, the first major
United Nations Organization (ONU) conference on the environment was held in
Stockholm [37] (Fig. 2A). After 18 years, in 1990, the IPCC published the first
report on the increase in GHG concentrations on Earth’s atmosphere [16]. After
that, Brazil started to participate actively in these discussions on climate and
environment, and in 1992 it held the United Nations Conference on Environ-
ment and Development, called Rio 92 [32,37]. In 1995, the 1st Conference of the
Parties (COP-1) was held in Germany, and the IPCC Second Report [17] was
also published.
COP-3 was held in 1997 in Kyoto - Japan. It generated targets for a 5% reduc-
tion in Greenhouse Gas emissions (GHG) [43]. GHG is one of the main factors
that increase the effect of climate change [18] (Fig. 2A). Around this period,
the first individual publications relating plants and climate change appeared in
the literature, and consequently, the employment of meta-analyzes to assess the
common effect among all experiments became possible (Fig. 2B). Between 1996
(the first meta-analysis on this theme) and 2006, four meta-analytical publica-
tions per year were published (Fig. 2B). By 2007 twice as many meta-analysis
studies were produced on the effect of elevated CO2 on plants. This increase
of meta-analytical reviews is concomitant with the fourth IPCC report. This
is consistent with the previously mentioned need for large-scale data analysis
studies to guide global reports such as the IPCC ones.
The fourth IPCC report, released in 2007, brought a strong message confirm-
ing that the increase in CO2 results from human activities. One of the report’s
key messages was that global warming could be irreversible if drastic measures
were not implemented. Some ways of mitigating and adapting to climate scenar-
ios were suggested [19]. That year the IPCC won the Nobel Peace Prize for its
achievement.
The GHG emissions, mainly CO2 , and the increase in temperature were
reported in previous reports. However, only in 2009 at COP-15 in Copenhagen,
226 J. da S. Fortirer et al.

Fig. 1. (A) Flowchart showing the systematic review steps modified from Liberati et
al. [28]. Numbers in parentheses represent the data obtained in the systematic review
carried out in this work for publications on the CO2 increases and its physiological
responses in plants using individual Brazilian plant studies. After performing the sys-
tematic review and data compilation, (B) flowchart describing the steps to perform a
meta-analysis. Based on Lei et al. [27].

it was proposed that there would be a 2 ◦ C increase in global temperature [44].

This prediction was only possible because the individual studies were carried out
very rapidly between 2007 and 2009, resulting in meta-analytical articles that
compiled many studies and showed how the temperature could affect plants and
forests’ productivity (Fig. 2B).
Analyzing the meta-analytical articles (Fig. 2B), one year before and in the
years following the IPCC reports (Fig. 2A), the number of reviews adopting this
approach increased dramatically. One possible explanation is the adjustment of
the timing of scientific literature publications on climate change and the pub-
lication of the IPCC reports. While an IPCC report is being prepared – what
can take a few years – the scientific community is consulted and becomes aware
of its ongoing work. As the IPCC sets limit dates for inclusion in the report’s
citations, many publications tend to be released before the IPCC deadline. As
a result, many publications became available after the report’s launch, increas-
ing the amount of data available and turning new meta-analyses more viable.
This relationship among chronology of events and world reports with scientific
knowledge production shows the importance of both to the development of new
technologies to adapt or mitigate the consequences of climate change. Due to
the large number of scientific data generated, the projections have become more
realistic and can be calculated using more complex mathematical models. In
the fifth report from IPCC, it was estimated that from 2100, atmospheric CO2
 Meta-analysis on Plants and Climate Change in Brazil 227

Fig. 2. (A) World history events ranging from meetings to IPCC reports that addressed
global climate change. (B) Distribution over years of published manuscripts (from 1996
to 2019) of meta-analyses performed worldwide that address plant responses to high
CO2 (236 studies).

would reach an atmospheric concentration of about 1.300 ppm [18]. Moreover,

the global temperature is likely to reach 1.5 ◦ C around 2030 if there is no action
to reduce GHG emissions [20,45].
From 1996 to 2019, we found 236 articles that performed meta-analyses
related to high CO2 effects on plants (Fig. 2B). The pioneering study that used
the meta-analysis to assess elevated CO2 responses on plants was conducted by
Curtis [7] (Fig. 1B). In 2005, Ainsworth & Long [2] published a meta-analysis
228 J. da S. Fortirer et al.

with 120 works showing physiological aspects and some species’ productivity.
This work was a milestone for studies researching climate change, with 3.157
citations (until October 20, 2020), demonstrating the power of quantitative syn-
thesis of meta-analysis.
The meta-analytic studies addressed some ecological processes and relation-
ships, including the effect of elevated CO2 on photosynthesis and plant respira-
tion, growth, competition, and interaction among plants, productivity, exchange
and conductance of gases in the leaves, soil respiration, carbon, and nitrogen
accumulation in the soil, and seed production [7,30,47]. This shows that the
meta-analysis has been widely used in research containing experiments with ele-
vated CO2 in plants and comes with each new publication trying to couple the
high CO2 with other factors such as temperature and water conditions.
Alternatively, most meta-analyses, such as Ainsworth & Long [2], only
include experiments performed with plants from temperate climate regions. This
points to the need for studies involving tropical plant responses to climate change
[2,21]. The importance of carrying out studies on species in a tropical climate
is highlighted by the fact that from the 236 meta-analyses found in this work,
only three studies on experiments with tropical species carried out in Brazil were
included among the primary studies reviewed [4,31,40].

3.2 Integrative Analysis of Variables that Are Contemplated

in Brazilian Studies Eligible for a Meta-analysis on Plant
Responses to Climate Change
The systematic review carried out in the three databases returned 1.625 studies
(Fig. 1A). After reading titles and abstracts, we excluded 837 studies because
they were not experimental studies or were not conducted in Brazil. An evalua-
tion of the texts was performed for the 788 eligible articles, and 728 articles were
excluded for not meeting at least one of the following criteria: research conducted
in Brazil on the elevated CO2 effects on physiological responses in plants and
experiments with FACE or OTC. In the next stage of the systematic review, 60
studies were included in the qualitative synthesis, but of these, only 36 displayed
the required data for meta-analysis, which are sample size, mean, and standard
deviation/error of the control and treatment groups (Fig. 1A). The small number
of studies included in the meta-analysis demonstrated that the primary studies
do not report all the information necessary to perform a meta-analysis. From the
studies included in the quantitative synthesis, the information described in the
methodology was extracted, and then the steps shown in Fig. 1B were followed.
Of the 36 studies performed in Brazil, 27 species were contemplated, 13 clas-
sified as trees, and 14 herbaceous species. Regarding the photosynthetic pathway,
20 species were classified with the C3 photosynthetic pathway, and seven species
were from plants displaying the C4 photosynthetic pathway. Furthermore, there
were relatively few studies with C4 plants compared to C3 (Fig. 3B). Despite
some studies with maize [26], the elevation of CO2 did not have any effect, con-
firming the hypothesis that species displaying C4 photosynthesis type do not
respond or respond very weakly to CO2 elevation. However, when sugarcane,
 Meta-analysis on Plants and Climate Change in Brazil 229

also a tropical C4 grass, was subjected to elevated CO2 , an increase in photo-

synthesis and biomass was observed [8]. These results show that there seem to
exist alternative physiological mechanisms by which C4 species could respond to
elevated CO2 .
Species categorization, according to their habit, can provide information
about the magnitude of the high CO2 effects. In this sense, whether plants are
trees or herbs and the photosynthesis metabolism (C3 or C4) seem to directly
influence the size of the high CO2 effect in the plants studied (Fig. 3B). However,
when categorizing the different species, the studies number becomes smaller, as
shown in Fig. 3, in which the herbs category C4 is left with only three observa-
tions (Fig. 3C), reducing the power of statistical analysis.

Fig. 3. (A) Number of times each variable was used within the 36 studies included
in this meta-analysis. (B) Size of the high CO2 effects on plant growth. (C) Size of
the effect of high CO2 on plant photosynthesis. Data represent means ±95% CI for
each functional group analyzed. Values that do not overlap to zero have a significant
difference (α < 0.05). The numbers of observations are presented in parentheses.

Life’s habit (trees or herbs) influences the magnitude of high CO2 responses
on plant biomass. From 1,625 studies included for the meta-analysis, 36 observa-
tions were obtained to estimate the effect of biomass (growth) (Figs. 1 and 3A).
This higher growth variable number is due to the measurement facility during
experiments since a single scale can be used to obtain the variable data. This
representativeness of studies enables a more detailed analysis of the effect of high
230 J. da S. Fortirer et al.

Fig. 4. Physiological variables more commonly measured in experiments with plants

submitted to climate changes (CO2 , water stress, and temperature). The variables more
frequently analyzed in Brazilian studies are in the yellow boxes (continuous line box),
and variables in the red box (dashed line box) did not have enough studies (number
of studies less than five) to perform a more consistent meta-analysis, needing to be
measured in future studies. A = net CO2 assimilation rate, Jmax = maximum electron
transport rate, Vcmax = maximum carboxylation of Rubisco, Ci:Ca = intercellular and
environmental CO2 rate, Rd = dark respiration, E = leaf transpiration, gs = stomatal
conductance, WUE = water use efficiency, C:N = carbon and nitrogen ratio, C =
carbon, and N = nitrogen. (Color figure online)

CO2 on plants. However, if the photosynthesis variable, which represents how

much the plant assimilates carbon, is observed, only 22 studies display this type
of variable (Fig. 3A). A smaller number of observations reported photosynthe-
sis measurements, which leads to a more complex level of plant responses. It is
 Meta-analysis on Plants and Climate Change in Brazil 231

important to note that to measure photosynthesis, it is necessary to use expensive

equipment to be operated by a highly qualified, well-trained professional. Alto-
gether, these limitations may explain the lower number of publications reporting
this kind of data in the literature.
A meta-analysis’s reliability depends on having several individual studies and
the relatively high number of variables within them. In the search for Brazilian
studies, there was a diversity of variables reported. However, most variables
were not common to all studies found (Fig. 3A). In this sense, it was necessary
to standardize the number of variables that could be evaluated. The most stud-
ied variables in the studies were growth (plant biomass - 36), followed by soluble
sugars (27), and A (photosynthetic assimilation - 22). The other variables were
reported with a very low frequency in the selected studies (Fig. 3A). Few obser-
vations for each variable can influence the meta-analysis’s statistical power since
the higher the sample number, the greater the consistency of the result [15].
Our data show the diversity of variables that should be taken into account in
the future for meta-analytical studies of plants in elevated CO2 and particularly
in studies of plant species in Brazilian biomes. Furthermore, standardization of
the variables used in experiments with single species would allow future meta-
analyzes to be more reliable, affording more consistent conclusions about how
the effect of elevated CO2 influences the plants’ physiology.
This work suggests that some of the physiological parameters presented in
Fig. 4, which contain the main variables that would bring more strength to
Brazilian studies’ future meta-analysis on plants’ high CO2 effects. The variables
that are within the red (dashed line box) in Fig. 4 are those that have relatively
little data. The variables in yellow boxes (continuous line box) are those that
have a significant amount of data for a meta-analytical study. However, there
are insufficient individual studies to perform a more detailed meta-analysis to
verify moderators’ influence on the magnitude of high CO2 effects in tropical
species.
Regarding the responses of tropical plant species to elevated CO2 , our sys-
tematic review, followed by classification criteria, generated 36 studies suitable
for meta-analysis. Our results indicate significant differences in plant responses
regarding habit (trees or herbs) and the photosynthesis type. However, as the
biological implications were out of this chapter’s scope, the more profound explo-
ration of the meta-analysis performed with the Brazilian studies will be reported
and discussed elsewhere.

4 Conclusion

In this study, we surveyed the history of global meta-analyses of the effects of

CO2 and temperature on plants. The purpose of the present work was not a
complete biological analysis of the results of our meta-analysis but rather to
bring a perspective on how important meta-analytical approaches are in the
area of Climate Change. We also discussed the limitations inherent in the meta-
analytical approach.
232 J. da S. Fortirer et al.

We found a close correlation (likely involving the release of the IPCC reports)
between the rise in meta-analysis concerning the responses of plants to elevated
CO2 and the socio-political events that determined sustainable development
goals on the planet during the last 50 years.
Given the importance of Brazilian biomes and agriculture for the planet and
the importance of land use for the impacts of global climate change, we conclude
that it is urgently needed that more studies would be performed with species
from the neotropics. A meta-analysis will undoubtedly be crucial to understand
such impacts and for decision making. However, most studies would have to
bring a plethora of variables that can afford the use of this valuable statistical
method.

References
1. Aguiar, S., Santos, I.D.S., Arêdes, N., Silva, S.: Redes-bioma: Informação e comu-
nicação para ação sociopolı́tica em ecorregiões. Ambiente Soc. 19(3), 233–252
(2016)
2. Ainsworth, E.A., Long, S.P.: What have we learned from 15 years of free-air
co2 enrichment (face)? a meta-analytic review of the responses of photosynthe-
sis, canopy properties and plant production to rising co2. New Phytol. 165(2),
351–372 (2005). https://doi.org/10.1111/j.1469-8137.2004.01224.x
3. Arenque, B.C., Grandis, A., Pocius, O., de Souza, A.P., Buckeridge, M.S.:
Responses of Senna reticulata, a legume tree from the amazonian floodplains, to
elevated atmospheric CO2 concentration and waterlogging. Trees 28(4), 1021–1034
(2014). https://doi.org/10.1007/s00468-014-1015-0
4. Cleveland, C.C., et al.: Relationships among net primary productivity, nutrients
and climate in tropical rain forest: a pan-tropical analysis. Ecol. Lett. 14(9), 939–
947 (2011). https://doi.org/10.1111/j.1461-0248.2011.01658.x
5. Cook, T.D.: The potential and limitations of secondary evaluations. In: Analysis
and Responsibility, Educational Evaluation (1974)
6. Curtis, P.S., Wang, X.: A meta-analysis of elevated CO2 effects on woody plant
mass, form, and physiology. Oecologia 113(3), 299–313 (1998). https://doi.org/10.
1007/s004420050381
7. Curtis, P.: A meta-analysis of leaf gas exchange and nitrogen in trees grown under
elevated carbon dioxide. Plant, Cell Environ. 19(2), 127–137 (1996). https://doi.
org/10.1111/j.1365-3040.1996.tb00234.x
8. De Souza, A.P., et al.: Elevated CO2 increases photosynthesis, biomass and pro-
ductivity, and modifies gene expression in sugarcane. Plant, Cell Environ. 31(8),
1116–1127 (2008). https://doi.org/10.1111/j.1365-3040.2008.01822.x
9. Egger, M., Smith, G.D., Schneider, M., Minder, C.: Bias in meta-analysis detected
by a simple, graphical test. Bmj 315(7109), 629–634 (1997). https://doi.org/10.
1136/bmj.315.7109.629
10. Glass, G.V.: Primary, secondary, and meta-analysis of research. Educ. Res. 5(10),
3–8 (1976). https://doi.org/10.3102/0013189X005010003
11. Gurevitch, J., Koricheva, J., Nakagawa, S., Stewart, G.: Meta-analysis and the
science of research synthesis. Nature 555(7695), 175–182 (2018). https://doi.org/
10.1038/nature25753
12. Harrison, F.: Getting started with meta-analysis. Methods Ecol. Evol. 2(1), 1–10
(2011). https://doi.org/10.1111/j.2041-210X.2010.00056.x
 Meta-analysis on Plants and Climate Change in Brazil 233

13. Haworth, M., Hoshika, Y., Killi, D.: Has the impact of rising CO2 on plants been
exaggerated by meta-analysis of free air CO2 enrichment studies? Front. Plant Sci.
7, 1153 (2016). https://doi.org/10.3389/fpls.2016.01153
14. Hedges, L.V., Gurevitch, J., Curtis, P.S.: The meta-analysis of response ratios in
experimental ecology. Ecology 80(4), 1150–1156 (1999). https://doi.org/10.1890/
0012-9658(1999)080[1150:TMAORR]2.0.CO;2
15. Hedges, L.V., Pigott, T.D.: The power of statistical tests for moderators in meta-
analysis. Psychol. Methods 9(4), 426 (2004). https://doi.org/10.1037/1082-989X.
9.4.426
16. IPCC: Climate change: The ipcc 1990 and 1992 assessments (1990). https://www.
ipcc.ch/report/climate-change-the-ipcc-1990-and-1992-assessments/. Accessed 25
Aug 2020
17. IPCC: Sar climate change 1995: Synthesis report (1995). https://www.ipcc.ch/
site/assets/uploads/2018/05/2nd-assessment-en-1.pdf. Accessed 25 Aug 2020
18. IPCC: Tar climate change 2001: Synthesis report (2001). https://www.ipcc.ch/
report/ar3/syr/. Accessed 25 Aug 2020
19. IPCC: Ar4 climate change 2007: Synthesis report (2007). https://www.ipcc.ch/
report/ar4/syr/. Accessed 25 Aug 2020
20. IPCC: Global warming of 1.5 ◦ c: Special report (2018). https://www.ipcc.ch/sr15/.
Accessed 25 Aug 2020
21. Jones, A.G., Scullion, J., Ostle, N., Levy, P.E., Gwynn-Jones, D.: Completing the
face of elevated CO2 research. Environ. Int. 73, 252–258 (2014). https://doi.org/
10.1016/j.envint.2014.07.021
22. Koricheva, J., Gurevitch, J., Mengersen, K.: Handbook of meta-analysis in ecology
and evolution. Princeton University Press, New Jersey (2013)
23. Körner, C.: Plant CO2 responses: an issue of definition, time and resource supply.
New Phytol. 172(3), 393–411 (2006). https://doi.org/10.1111/j.1469-8137.2006.
01886.x
24. Körner, C.: Responses of humid tropical trees to rising CO2 . Annu. Rev. Ecol. Evol.
Syst. 40, 61–79 (2009). https://doi.org/10.1146/annurev.ecolsys.110308.120217
25. Lambers, H., Chapin III, F.S., Pons, T.L.: Plant Physiological Ecology. Springer,
New York (2008)
26. Leakey, A.D., et al.: Photosynthesis, productivity, and yield of maize are not
affected by open-air elevation of CO2 concentration in the absence of drought.
Plant Physiol. 140(2), 779–790 (2006). https://doi.org/10.1104/pp.105.073957
27. Lei, X., Peng, C., Tian, D., Sun, J.: Meta-analysis and its application in global
change research. Chin. Sci. Bull. 52(3), 289–302 (2007). https://doi.org/10.1007/
s11434-007-0046-y
28. Liberati, A., et al.: The Prisma statement for reporting systematic reviews and
meta-analyses of studies that evaluate health care interventions: explanation and
elaboration. J. Clin. Epidemiol. 62(10), e1–e34 (2009). https://doi.org/10.1016/j.
jclinepi.2009.06.006
29. Lovejoy, T.E., Nobre, C.: Amazon tipping point: last chance for action (2019).
https://doi.org/10.1126/sciadv.aba2949
30. Luo, Y., Hui, D., Zhang, D.: Elevated CO2 stimulates net accumulations of carbon
and nitrogen in land ecosystems: a meta-analysis. Ecology 87(1), 53–63 (2006).
https://doi.org/10.1890/04-1724
31. Moles, A.T., et al.: Which is a better predictor of plant traits: temperature or
precipitation? J. Veg. Sci. 25(5), 1167–1180 (2014). https://doi.org/10.1111/jvs.
12190
234 J. da S. Fortirer et al.

32. MRE: Ministério das relações exteriores do brasil (2019). http://www.itamaraty.

gov.br/pt-BR. Accessed 25 Aug 2020
33. Nakagawa, S., Noble, D.W., Senior, A.M., Lagisz, M.: Meta-evaluation of meta-
analysis: ten appraisal questions for biologists. BMC biology 15(1), 1–14 (2017).
https://doi.org/10.1186/s12915-017-0357-7
34. Nakagawa, S., et al.: Research weaving: visualizing the future of research synthesis.
Trends Ecol. Evol. 34(3), 224–238 (2019). https://doi.org/10.1016/j.tree.2018.11.
007
35. Palacios, C., Grandis, A., Carvalho, V., Salatino, A., Buckeridge, M.: Isolated and
combined effects of elevated CO2 and high temperature on the whole-plant biomass
and the chemical composition of soybean seeds. Food Chem. 275, 610–617 (2019).
https://doi.org/10.1016/j.foodchem.2018.09.052
36. Quintana, D.S.: From pre-registration to publication: a non-technical primer for
conducting a meta-analysis to synthesize correlational data. Front. Psychol. 6, 1549
(2015). https://doi.org/10.3389/fpsyg.2015.01549
37. Ribeiro, W.C.: A ordem ambiental internacional. Editora Contexto (2001)
38. Ruddiman, W.F.: The anthropogenic greenhouse era began thousands of years
ago. Climatic Change 61(3), 261–293 (2003). https://doi.org/10.1023/B:CLIM.
0000004577.17928.fa
39. Sterne, J.A., Egger, M.: Funnel plots for detecting bias in meta-analysis: guidelines
on choice of axis. J. Clin. Epidemiol. 54(10), 1046–1055 (2001). https://doi.org/
10.1016/S0895-4356(01)00377-8
40. Stevens, N., Lehmann, C.E., Murphy, B.P., Durigan, G.: Savanna woody encroach-
ment is widespread across three continents. Glob. Change Biol. 23(1), 235–244
(2017). https://doi.org/10.1111/gcb.13409
41. Sutton, A.J., Higgins, J.P.: Recent developments in meta-analysis. Stat. Med.
27(5), 625–650 (2008). https://doi.org/10.1002/sim.2934
42. Team, R.C., et al.: R: A language and environment for statistical computing (2013)
43. UNFCCC: Kyoto protocol to the united nations framework convention on climat-
echange (1997). http://unfccc.int/resource/docs/convkp/kpeng.pdf. Accessed 25
Aug 2020
44. UNFCCC: Statements on behalf of the group of g77 and china (2009). https://
unfccc.int/resource/docs/2009/cop15/eng/11a01.pdf. Accessed 25 Aug 2020
45. UNFCCC: Adoption of the paris agreement. united nations framework conven-
tion on climate change (2015). http://unfccc.int/paris agreement/items/9485.php.
Accessed 25 Aug 2020
46. Viechtbauer, W.: Conducting meta-analyses in r with the metafor package. J. Stat.
Softw. 36(3), 1–48 (2010). https://doi.org/10.18637/jss.v036.i03
47. Wand, S.J., Midgley, G.F., Jones, M.H., Curtis, P.S.: Responses of wild c4 and
c3 grass (poaceae) species to elevated atmospheric CO2 concentration: a meta-
analytic test of current theories and perceptions. Glob. Change Biol. 5(6), 723–741
(1999). https://doi.org/10.1046/j.1365-2486.1999.00265.x
48. Wickham, H.: Elegant Graphics for Data Analysis (ggplot2). Springer, New York
(2009)
 A Brief History of Bioinformatics Told by Data
Visualization

Diego Mariano1(B) , Mívian Ferreira2 , Bruno L. Sousa2 ,

Lucianna H. Santos1 , and Raquel C. de Melo-Minardi1
1 LBS - Laboratory of Bioinformatics and Systems, Department of Computer Science,
Universidade Federal de Minas Gerais - UFMG, Belo Horizonte, Brazil
[email protected]
2 LLP - Laboratory of Programming Languages, Department of Computer Science,

Universidade Federal de Minas Gerais - UFMG, Belo Horizonte, Brazil

Abstract. Bioinformatics is an interdisciplinary research field that aims to ana-

lyze biological data through computational approaches. In the last years, the evo-
lution of technological resources has provided a tidal wave of biological data. Con-
sequently, an unprecedented amount of studies using bioinformatics approaches
have been released, increasing peer-reviewed published papers. Here, we tell a
brief history of bioinformatics based on literature data analysis and visualization.
We collected abstracts and other metadata from papers published from 1998 to
2019 in four leading bioinformatics journals: (i) Oxford Bioinformatics; (ii) BMC
Bioinformatics; (iii) Briefings in Bioinformatics; and (iv) PLoS Computational
Biology. Our results show an increase in publication number and international
collaborations. We also observed an increase in publications by Chinese authors.
Latin America continues to have a low percentage of global scientific bioinformat-
ics production. However, Brazil excels in this region, being responsible for almost
half of Latin America papers published. Our results also point out the recent trend
of using Python as the programming language for bioinformatics applications, fol-
lowed by Perl, Java, and R. We hope these data visualizations can provide insights
to understand the recent changes and evolution in the bioinformatics field. The
developed interactive visualizations are available at http://bioinfo.dcc.ufmg.br/his
tory/.

Keywords: Bioinformatics · Computational biology · Data visualization

1 Introduction
Bioinformatics is an interdisciplinary research field whose principle is using models
and algorithms to analyze biological data and solve biologically related problems [1].
Bioinformatics’ roots are in the early 1960s when computers, used for military purposes,
became available for universities and research institutes. At that time, researchers began
to use computers to try answering fundamental questions in life sciences [2].
Margaret Dayhoff was a pioneer in bioinformatics studies at that time. She proposed
the use of mathematical approaches for analyzing amino acid frequencies and mutation

© Springer Nature Switzerland AG 2020

J. C. Setubal and W. M. Silva (Eds.): BSB 2020, LNBI 12558, pp. 235–246, 2020.
https://doi.org/10.1007/978-3-030-65775-8_22
236 D. Mariano et al.

probabilities in biological sequences. Since the late 1950s, experimental approaches have
allowed the sequencing of small protein structures, such as insulin [3]. This culminated
in the creation of the first database of amino acid sequences and structures, the so-called
“Atlas of Protein Sequence and Structure” [4]. Dayhoff and collaborators also proposed
computational methods for sequence comparisons to detect homologous proteins using
a substitution matrix called PAM (Percent Accepted Mutation), which contributed to the
rising of the molecular evolution field [2].
Also, Needleman and Wunsch proposed a dynamic programming method for
sequence alignment that is the base for the state-of-the-art methods used nowadays
[5]. By the end of the 1960s, we observed the rising of structural bioinformatics, for
example, with the construction of a three-dimensional model of a Cytochrome c protein
[6]. At the beginning of the 1980s, the first methods for phylogenetic tree inference
based on DNA sequences and the maximum likelihood were proposed [7]. Moreover,
BLAST, a tool for local sequence alignments, was proposed in the 1990s mainly due to
the increase of sequences availability [8].
However, three main milestones led to the modern bioinformatics field were: (i) the
DNA sequencing methods, (ii) the genome projects, and (iii) the rise of supercomputers
and the Internet [2]. DNA sequencing methods have existed since the 1970s, such as the
Sanger chain-termination sequencing method [9]. However, these methods were slow
and expensive. For instance, the human genome project (HGP) was an international
research effort to map all of the genes and sequences of the human genome. This project
started at the beginning of the 1990s but was only officially completed in 2003. The HGP
gave us a genetic blueprint of the human being, but the sequencing costs were high.
The game changed when ingenious strategies were used in combination with com-
putational approaches. Shotgun sequencing was introduced in the 1970s [10]. In 1995, a
similar strategy was used to obtain a complete nucleotide sequence of the Haemophilus
influenzae bacterium [11] and, in 2010, to get the complete genome of the fly Drosophila
melanogaster, a eukaryote with a sequence length of ~ 120 Mb [12]. All these develop-
ments led to the emergence of Next-Generation Sequencing (NGS) platforms [13]. These
technologies are characterized by high-throughput, with reduced processing time, and
costs each time lower. This culminated in the diffusion of genome projects, even in small
or middle research laboratories. Later, other high-throughput technologies appeared,
such as microarray and RNA-Seq, used to analyze gene expression, and cryo-electron
microscopy, which sped up the resolution of 3D macromolecular structures.
Thus, bioinformatics went from a tool for biological analysis to an interdisciplinary
research field, mainly focused on the development of new models, algorithms, tools,
and new types of analysis based on computational approaches to deal with biological
data and get knowledge from them. In the last years, these evolutions of technological
resources originated a tidal wave of data [14]. Consequently, an unprecedented amount
of studies using bioinformatics approaches have been released, increasing peer-reviewed
published papers in the field.
Here we tell the recent history of Bioinformatics based on metadata collected from
scientific papers. We know that the history of bioinformatics has been told before in some
publications [1, 15, 16]. However, we aimed to understand the recent state-of-the-art of
the bioinformatics research, visualize changes motivated by technological evolution, and
 A Brief History of Bioinformatics Told by Data Visualization 237

detect trends in published studies. We also performed an exploratory analysis of data

collected from the bioinformatics literature and constructed visualizations to get insights
into the bioinformatics field’s brief history.

2 Materials and Methods

2.1 Data Collection
We empirically selected four journals for performing theses analyses: (i) Oxford Bioin-
formatics (ISSN 1367-4803); (ii) BMC Bioinformatics (ISSN 1471-2105); (iii) Brief-
ings in Bioinformatics (ISSN 1467-5463); and (iv) PLoS Computational Biology (ISSN
1553-7358). We collected data using the PubMed API [17, 18] and in-house scripts
coded using Python. Firstly, we collected a list of ID entries for all papers available at
PubMed until October 7th , 2019, for the four journals. Then, we collected an XML file
with PMID, date of publication, journal data, title, abstract, authors, and their affiliations
for each entry. Citations were collected using Scopus Search API (https://dev.elsevier.
com/).

2.2 Data Visualizations

Exploratory data analysis was performed using Python and R scripts. The web tool
was developed using CodeIgniter Web framework (https://codeigniter.com), DataTables
(https://datatables.net), jQuery (https://jquery.com), and other frameworks. Data were
stored in a structured database using MySQL (https://www.mysql.com). Interactive visu-
alizations were constructed using D3.js (https://d3js.org) and Flourish Data Visualization
& Storytelling tool (https://flourish.studio). Static visualizations were constructed using
RStudio (https://rstudio.com) and were refined using Adobe Photoshop CC 2019.

3 Results and Discussion

We collected 13,798 items from Oxford Bioinformatics, 1,418 items from Briefings in
Bioinformatics, 9,265 items from BMC Bioinformatics, and 6,623 items from PLoS
Computational Biology. In the publication timeline analysis for the last 22 years, we can
observe the increase in the number of papers being published (Fig. 1A).
We can notice an expressive growth around 2007-2009 and the emergence of impor-
tant open access journals in this period (BMC Bioinformatics and PLoS Computational
Biology).
We also observed a substantial increase in international collaborations in publications
(Fig. 1B), which agrees with the global trend, as corroborated by an analysis performed
for other general science journals, such as Nature [19]. In 2019, the number of publi-
cations showing co-authors with different nationalities corresponded to approximately
25%. In contrast, the proportion of single authors’ publications had a substantial drop.
It was also an evident rise in the numbers of papers published by Chinese authors
and a reduction of the United Kingdom proportion in global production. Latin America
has a low participation in global scientific production. However, Brazil leads the bioin-
formatics scientific production in the region, being responsible for almost half of Latin
America publications (Fig. 1C-D).
238 D. Mariano et al.

Fig. 1. (A) Publications by journal (1998–2019). (B) The proportion of international collabora-
tions. “Domestic” corresponds to collaborations between researchers of the same institute or the
same country, “Multinational” from different countries, and “Single author” to individual papers.
(C) Nationality estimated based on author affiliation declarations (hence, this may not represent
the real nationality). (C) The amount and (D) the proportion. United Kingdom’s (UK) data was
not included in the European Union (EU) group to ease the comprehension (even, before Brexit).

3.1 The History that the Keywords in Papers Tell Us

We analyzed the keywords reported in the papers to establish what are the major research
topics addressed in bioinformatics publications. We obtained 7,167 unique keywords
reported in the papers. From them, eleven were reported in the top five positions from
1998 to 2019 (Fig. 2).
Since 2011, only a restricted set of keywords has scored the top five list: humans,
computational biology, algorithms, models, and software. Interestingly, the keyword
“humans” that leads the top five groups since 2011, was less cited in the previous years,
except in 2000–2001 when it appeared in the fifth position. Even though most of the
bioinformatics research is more focused on developing new algorithms, models, and
software, a fact well illustrated by the other four keywords in the rank, it was expected
to find papers using keywords describing application areas for these approaches. Indeed,
we can observe that “humans” appear in the top five list in the years closer to the first
 A Brief History of Bioinformatics Told by Data Visualization 239

Fig. 2. Top five keywords listed in bioinformatics papers from 1998 to 2019.

announcements of the human genome draft (2001). We hypothesized that, in the fol-
lowing years, the bioinformatics research was characterized by considerable efforts to
establish the architecture for data processing and storage, collection and organization of
data, and development of desktop and web tools. Since many publications have accom-
plished these objectives, this provided scaffolds for more applicable studies focused on
human data.
The top five ranking list only told us the main and general topics. To tell more
about recent bioinformatics content history, it is necessary to analyze the importance
of some keywords and correlate them with historical facts. To obtain these insights, we
constructed word clouds with keywords with at least 50 occurrences by year (available
on the website). Analyzing the word cloud, we can detect when some topics turned
into trends or reduced, comparing the word size changes according to different periods.
Although this analysis is quite limited to show specific changes in the keyword use (since
it is hard to compare and find many words together through different word clouds), this
could be used to identify targets for further analysis. Hence, based on this analysis, we
raised some questions: what topic is the target for a higher number of studies: genome,
transcriptome, or proteome? Are there keywords that are less used nowadays than in
the past? What is related to the recent increasing in molecular dynamics publications?
Why “artificial intelligence” was very reported in the 2000s, but fewer today? To try to
answer these questions, we constructed several visualizations and discussed what they
depict.

3.2 Omics
A Bioinformatics’ fundamental consists of the use of computational approaches to ana-
lyze data related to the central dogma of molecular biology, i.e., the process in which
information flow from DNA to RNA to protein. Thus, genome, transcriptome, and pro-
teome have been considered the main topics in this area, compounding the “Omics”
240 D. Mariano et al.

study fields, such as genomics, transcriptomics, and proteomics (and new study areas
such as metagenomics, metabolomics, and so on).
A priori, we inquired which of these topics is the target of a higher number of studies.
In Fig. 3A, we plotted the percentage of reports of each keyword. We can observe the
sovereignty of genome studies, only defied by proteome studies by a short period between
2007 and 2008. The first occurrence of transcriptome keyword was in 2010. Since then,
the proportion of transcriptome studies is slightly increasing, while the proportion of
proteome studies has considerably decreased. Metagenome and metabolome are recent
study areas, and hence, still have few occurrences. We also included metagenome and
metabolome keywords in this plot to illustrate the interest in recent “Omics” approaches.

Fig. 3. Keywords analysis. (A) Percentage of citations of “Omics” keywords from 1998 to 2019.
(B) Use of the keywords “Database Management Systems”, “User-Computer Interface”, “Pro-
gramming Languages”, and “Sequence Alignment”, “Protein Structure”, and “Systems Biology”
from 1998 to 2019. (C) Use of “Artificial Intelligence” and other correlated keywords from 1998
to 2019.
 A Brief History of Bioinformatics Told by Data Visualization 241

3.3 Keywords in Disuse

We also observed a peculiar behavior in the usage of some keywords, such as: “Database
Management Systems”, “User-Computer Interface”, “Programming Languages”, and
“Sequence Alignment” (Fig. 3B). These keywords correspond to essential topics in
bioinformatics. So, why have they been less used nowadays?
Our hypothesis relies on the release of the HGP and the blast of the genome projects.
At first, researchers did not know how to deal with high amounts of data. Besides, the
scientific community interest was more focused on the development of novel models and
algorithms to solve relevant problems than on data management. This naturally reflected
in publications aiming to establish better ways to model, store, and access these data.
Also, the high number of software executed by command lines created a gap of user-
friendly tools that could be explored by publications. Nowadays, these necessities also
persist, but because of the high number of studies, some topics present tools widely con-
solidated, such as “sequence alignment”. Although this topic is essential, nowadays, the
use of sequence alignment is considered trivial, with standard tools already established
and few novel inventions.
We also observed a slight reduction in the use of “protein structure” and “systems
biology” keywords. However, this could change in the next years due to the evolution
in the Cryogenic electron microscopy methods used for determining protein structures
and the increase of computational power that will allow simulations of larger systems.
Additionally, the interest in programming languages initially increased, impelled by
projects that aimed to develop libraries for analyzing biological data, such as packages
derived from the projects Bioconductor [20–24], and later Biopython [25, 26]. When a
large number of packages were obtained, research topics began to focus, for example,
on applications of use.

3.4 Artificial Intelligence and the Influence of the Pop Culture in Science
We analyzed the use of the keyword “Artificial Intelligence” and other related keywords.
We observed that this keyword’s use increased from 2001 until 2008, when reduced
(Fig. 3C). Then, publications using “Artificial Intelligence” were made, but the authors
started to use more specific descriptions, such as “neural networks”, “support vector
machine”, “machine learning”, and “data mining”.
It is interesting to report that artificial intelligence is studied since the middle of the
20th century, as well as the other listed topics. Hence, what could explain the expressive
increase in the use of this topic as a keyword and its subsequent reduction?
A peculiar explanation for this phenomenon is that in 2001 was released the famous
movie “A.I. Artificial Intelligence” (directed by Steven Spielberg). The film’s popularity
may have influenced the keyword “artificial intelligence” in academic works. Later, this
keyword’s use decreased, and more specific descriptions of the A.I. techniques used were
adopted. This suggests that pop culture could influence how studies are disseminated.

3.5 A Promising Future for Molecular Dynamics

Classical and quantum molecular dynamics are a set of techniques to simulate the
dynamic behavior, movements, and interactions of molecules in a system across periods.
242 D. Mariano et al.

They have been used, for example, to perform computational predictions of cancer drug
resistance [27], to understand allosteric immune escape pathways in the HIV-1 envelope
glycoprotein [28], and to simulate the action of enzymes used in biofuel production [29,
30]. Although these methods are known to have high computational cost requirements,
since 2011, the number of citations of “molecular dynamics simulation” has increased
substantially (Fig. 4).

Fig. 4. (A) Use of “molecular dynamics” as a keyword (above) compared to (B) the number of
transistors in NVIDIA GPUs (below). Also, it is important to highlight that molecular dynam-
ics researchers prefer to publish their experiments in specialized journals, such as the Journal
of Biomolecular Structure and Dynamics. Source: Adapted from https://vintage3d.org/dbn.php.
Accessed on September 16th 2020.

We searched for connections in the number of reports with the number of transistors in
NVIDIA graphic cards from 1998 until 2020. We observed that the number of citations of
molecular dynamics simulation seems to correlate to the evolution of Graphic Processing
Unities (GPUs). The evolution of GPUs is impelled by the game industry because of the
necessity of more realistic games. However, researchers have adapted the GPU usage to
process force fields calculus used in molecular dynamics simulations. Thus, molecular
dynamics scientists have successfully used gamer graphic cards in research applications,
which led to better results than those from CPU only supercomputers from a few years
ago.
 A Brief History of Bioinformatics Told by Data Visualization 243

Based on this analysis, we foresee a promising future for molecular dynamics

research making good use of GPUs evolution. Figure 4 shows that the new generation
of graphic cards brought a huge increase in GPUs’ number of transistors. The number
of transistors in graphic cards can be related to their performance. The more quantity
of transistors, the more the capacity of data processing. For instance, the graphic card
GeForce256 SDR (released in 1999) presented 23,000,000 transistors. A new GeForce
RTX 3090 (released in 2020) has 28,000,000,000 transistors (~ 1,200 × more). This is
possible due to the reduction in the size of the components. GeForce256 SDR used the
chip NV10 with lithography of 220 nm. On the other hand, RTX 3090 uses the chip
GA102 with lithography of only 8 nm. More powerful graphic cards will increase the
number and the size of studies using these techniques.

3.6 Popular Programming Languages for Bioinformatics

Our results also point out a recent trend of using Python as the programming language
for bioinformatics research and applications, followed by Perl (perhaps the first bioinfor-
matic programming language), Java, and R. A recent large-scale analysis study of code
from bioinformatics projects available at GitHub suggested that the languages Python,
R, and Java are more succinct than lower-level languages [31]. This allows develop-
ers to construct advanced programs written in fewer code lines, which can contribute
to the popularity of these languages for bioinformatics. Besides, initiatives for teach-
ing Python to deal with biological data have contributed to expanding this language by
bioinformaticians [32, 33].

3.7 Limitations of This Study

This study’s major limitation is that it could include only a few of the interesting findings
and correlations in the main text because of space restrictions. We expect to have solved
this by developing an interactive web tool and turn it publicly available for users to
explore data.
Furthermore, this research is restricted by the journals selected and the period used for
data gathering. Frequent updates of data collected can be a challenge due to restrictions
of the data sources APIs. Also, the keyword analysis topics were defined by an empirical
analysis of the word clouds. In the future, we intend to improve these analyses using
context-dependent string search engines, such as the natural language processing (NLP)
API from Google Cloud.
In the popular programming language analysis, we found a lower number of results
once we used text mining to detect occurrences in the abstract (many authors did not
declare the programming language used). However, we expect to get a realistic sample
of programming languages used. We would like to emphasize that it is not possible to
define the “best programming language” for bioinformatics research applications. The
choice of a programming language depends on the developer’s familiarity with their
syntax, the context of use, the requirements, and the algorithms’ computational costs
to be implemented. Besides, we consider the availability of packages and libraries for
biological data analysis, and developers’ community support is an essential item to be
observed.
244 D. Mariano et al.

Last, we have tried to establish which authors contributed more to bioinformatics

based on the number of publications and citations. However, this analysis was limited
due to the challenges in disambiguating names. This could be solved by the adoption
of unique codes for identifying scientists, such as ORCID. Unfortunately, these data
were not available for most of the databases consulted. Even so, we constructed lists of
authors that most collaborated in studies published in the main bioinformatics journals
(stratified by publication number, citations, and countries) and made them available on
the website. Also, PubMed API returned a limited number of citations in documents
indexed there. We used the Scopus API to improve this analysis.

4 Conclusion
In this paper, we presented an overview of bioinformatics publications in the last 22
years based on four high-impact bioinformatics journals. We consider that this paper’s
main scientific contribution is to present the reports of the state-of-the-art of publications
in bioinformatics (past and present), presenting our predictions based on what the data
showed us. We show an increasingly collaborative world: data show an increase in
the bioinformatics’ scientific production with the increasing participation of several
countries. However, there is still much to improve (see Latin America’s low participation
in global scientific production). The keyword analysis showed changes in the major
topics addressed in bioinformatics papers. We also hypothesized a growth in molecular
dynamic simulation publications due to the correlation between published works and
recent evolutions in GPUs. The most cited programming languages analysis showed
Python, Perl, Java, and R as the most popular programming languages for bioinformatics.
Additionally, we proposed a web tool for exploratory data analysis as supplementary
material. Here, we present only an overview of what the data showed us, focusing on
some details that caught our attention. Our readers can interact with the data in the web
tool, obtain insights, and maybe reach conclusions that we may not even have imagined
when writing this article. We also want to encourage (and provoke) our readers to think
about bioinformatics’ perspectives as a science (and their specific areas of activity).
The scientist’s role is to observe, question, and propose solutions that lead to society’s
improvement. For scientists to fulfill their roles, they should be able to adapt to changes
and know the history and discuss the future is a fundamental step to this. We hope
that these data visualizations can provide insights and raise more thought-provoking
discussions about bioinformatics evolution, trends, and perspectives. We provided the
data sets and other interactive visualizations at http://bioinfo.dcc.ufmg.br/history/.

Acknowledgments. The authors thank the funding agencies: CAPES, FAPEMIG, and CNPq.
This study was financed in part by the Coordenação de Aperfeiçoamento de Pessoal de Nível Supe-
rior - Brasil (CAPES) - Finance Code 001. Project grant number 51/2013 - 23038.004007/2014-82.
 A Brief History of Bioinformatics Told by Data Visualization 245

References
1. Akalin, P.K.: Introduction to bioinformatics. Mol. Nutr. Food Res. 50, 610–619 (2006)
2. Hagen, J.B.: The origins of bioinformatics. Nat. Rev. Genet. 1, 231–236 (2000)
3. Moore, S., Spackman, D.H., Stein, W.H.: Automatic recording apparatus for use in the
chromatography of amino acids, pp. 1107–1115 (1958)
4. Dayhoff, M.O.: Atlas of protein sequence and structure. National Biomedical Research
Foundation (1972)
5. Needleman, S.B., Wunsch, C.D.: A general method applicable to the search for similarities
in the amino acid sequence of two proteins. J. Molecular Biol. 48, 443–453 (1970)
6. Levinthal, C.: Molecular model-building by computer. Sci. Am. 214, 42–52 (1966)
7. Felsenstein, J.: Evolutionary trees from DNA sequences: a maximum likelihood approach. J.
Mol. Evol. 17, 368–376 (1981)
8. Stephen, F., Altschu, P., Warren, G., Webb, M., Eugene, W., Myers, David, J.L.: Basic Local
Alignment Search Tool (1990)
9. Sanger, F., Nicklen, S., Coulson, A.R.: DNA sequencing with chain-terminating inhibitors.
Proc. National Acad. Sci. 74, 5463–5467 (1977)
10. Staden, R.: A strategy of DNA sequencing employing computer programs. Nucleic Acids
Res. 6, 2601–2610 (1979)
11. Fleischmann, R.D., Adams, M.D., White, O., Clayton, R.A., Kirkness, E.F., Kerlavage, A.R.,
et al.: Whole-genome random sequencing and assembly of Haemophilus influenzae Rd.
Science 269, 496–512 (1995)
12. Adams, M.D., Celniker, S.E., Holt, R.A., Evans, C.A., Gocayne, J.D., Amanatides, P.G., et al.:
The genome sequence of drosophila melanogaster. Science 287, 2185–2195 (2000)
13. Mariano, D.C.B., Pereira, F.L., Aguiar, E.L., Oliveira, L.C., Benevides, L., Guimarães, L.C.,
et al.: SIMBA: a web tool for managing bacterial genome assembly generated by Ion PGM
sequencing technology. BMC Bioinform. 17(Suppl 18), 456 (2016)
14. It’s sink or swim as a tidal wave of data approaches. Nature, 399, 517 (1999)
15. Gauthier, J., Vincent, A.T., Charette, S.J., Derome, N.: A brief history of bioinformatics.
Brief. Bioinform. 20, 1981–1996 (2019)
16. Hogeweg, P.: The Roots of Bioinformatics in Theoretical Biology. PLoS Comput. Biol. 7,
e1002021 (2011)
17. Canese, K., Weis, S.: PubMed: The Bibliographic Database. National Center for Biotechnol-
ogy Information (US) (2013). https://www.ncbi.nlm.nih.gov/books/NBK153385/. Accessed
14 Sep 2020
18. NCBI Resource Coordinators: Database resources of the national center for biotechnology
information. Nucleic Acids Res. 46, D8–D13 (2018)
19. Monastersky, R., Noorden, R.V.: 150 years of nature: a data graphic charts our evolution.
Nature 575, 22–23 (2019)
20. Carey, V.J., Gentry, J., Whalen, E., Gentleman, R.: Network structures and algorithms in
Bioconductor. Bioinformatics 21, 135–136 (2005)
21. Chen, H., Lau, M.C., Wong, M.T., Newell, E.W., Poidinger, M., Chen, J.: Cytofkit: a biocon-
ductor package for an integrated mass cytometry data analysis pipeline. PLoS Comput. Biol.
12, e1005112 (2016)
22. Fournier, F., Joly Beauparlant, C., Paradis, R., Droit, A.: rTANDEM, an R/Bioconductor
package for MS/MS protein identification. Bioinformatics 30, 2233–2234 (2014)
23. Gådin, J.R., van’t Hooft, F.M., Eriksson, P., Folkersen, L.: AllelicImbalance: an R/ biocon-
ductor package for detecting, managing, and visualizing allele expression imbalance data
from RNA sequencing. BMC Bioinform. 16, 194 (2015)
246 D. Mariano et al.

24. Gentleman, R.C., Carey, V.J., Bates, D.M., Bolstad, B., Dettling, M., Dudoit, S., et al.: Biocon-
ductor: open software development for computational biology and bioinformatics. Genome
Biol. 5, R80 (2004)
25. Talevich, E., Invergo, B.M., Cock, P.J., Chapman, B.A.: Bio.Phylo: a unified toolkit for
processing, analyzing and visualizing phylogenetic trees in Biopython. BMC Bioinform. 13,
209 (2012)
26. Cock, P.J.A., Antao, T., Chang, J.T., Chapman, B.A., Cox, C.J., Dalke, A., et al.: Biopy-
thon: freely available Python tools for computational molecular biology and bioinformatics.
Bioinformatics 25, 1422–1423 (2009)
27. Sun, X., Hu, B.: Mathematical modeling and computational prediction of cancer drug
resistance. Brief. Bioinform. 19, 1382–1399 (2018)
28. Sethi, A., Tian, J., Derdeyn, C.A., Korber, B., Gnanakaran, S.: A mechanistic understanding
of allosteric immune escape pathways in the HIV-1 envelope glycoprotein. PLoS Comput.
Biol. 9, e1003046 (2013)
29. Costa, L.S.C., Mariano, D.C.B., Rocha, R.E.O., Kraml, J., da Silveira, C.H., Liedl, K.R., et al.:
Molecular dynamics gives new insights into the glucose tolerance and inhibition mechanisms
on β-glucosidases. Molecules 24, 3215 (2019)
30. Lima, L.H.F., de Fernandez-Quintéro, M., Rocha, R.E.O., Mariano, D.C.B., Melo-Minardi,
R.C., de Liedl, K.R.: Conformational flexibility correlates with glucose tolerance for point
mutations in β-glucosidases – a computational study. J. Biomolecular Structure Dyn. 1–20
(2020)
31. Russell, P.H., Johnson, R.L., Ananthan, S., Harnke, B., Carlson, N.E.: A large-scale analysis
of bioinformatics code on GitHub. PLoS ONE 13, e0205898 (2018)
32. Ekmekci, B., McAnany, C.E., Mura, C.: An Introduction to Programming for Bioscientists:
A Python-Based Primer. PLoS Comput. Biol. 12, e1004867 (2016)
33. Mariano, D., Martins, P., Helene Santos, L., de Melo- Minardi, R.C.: Introducing Program-
ming Skills for Life Science Students. Biochemistry and Molecular Biology Education (2019).
https://doi.org/10.1002/bmb.21230
 Computational Simulations
for Cyclizations Catalyzed by Plant
Monoterpene Synthases

Waldeyr Mendes Cordeiro da Silva1,2,3(B) , Daniela P. de Andrade1 ,

Jakob L. Andersen4 , Maria Emı́lia M. T. Walter2 , Marcelo Brigido2 ,
Peter F. Stadler3,4,5,7,8 , and Christoph Flamm6
1
NEPBio - Federal Institute of Goiás, Formosa, Brazil
[email protected]
2
University of Brası́lia, Brası́lia, DF, Brazil
3
Department of Computer Science, Bioinformatics Group, Interdisciplinary Center
for Bioinformatics, University of Leipzig, Leipzig, Germany
4
Department of Mathematics and Computer Science, University of Southern
Denmark, Odense, Denmark
5
German Centre for Integrative Biodiversity Research (iDiv) Halle-Jena-Leipzig,
Leipzig, Germany
6
Institute for Theoretical Chemistry, University of Vienna, Vienna, Austria
7
Max Planck Institute for Mathematics in the Sciences, Leipzig, Germany
8
Santa Fe Institute, Santa Fe, USA

Abstract. Metabolic pathways collectively define the biochemical reper-

tory on an organism exposing the steps of its production. In silico meta-
bolic pathways have been reconstructed using a wide range of computa-
tional methods. The reconstructed metabolic pathways can vary in some
aspects, among which, in the context of this work, it is relevant to remark
the data structure and granularity of biochemical details. Inferring chem-
ical reactions using graph grammar rules is a method that exposes the ini-
tial, intermediates, and final products by modeling pathways over graphs
and hypergraphs. Plant monoterpenes are volatile compounds with appli-
cations in industry, biotechnology, and medicine. They also play a vital
ecological role. The last enzymatic reaction in the plant monoterpenes pro-
duction chain is plentiful of possibilities due to the promiscuous nature of
the terpene synthases (TPS), a case in which the application of inferring
chemical reactions using graph grammar rules is suitable. In this work, we
designed graph grammar rules that express cyclization reactions catalyzed
by plant monoterpene synthases. As a result, it was generated a reachable
chemical space of potential plant monoterpene blend, which can be compu-
tationally exploitable. In addition, these graph grammar rules were added
to the 2Path, and a graphical interface was provided to aid the simulation
code outlining.

Keywords: Plant · Monoterpene · Biosynthesis · Computational ·

Simulation
c Springer Nature Switzerland AG 2020
J. C. Setubal and W. M. Silva (Eds.): BSB 2020, LNBI 12558, pp. 247–258, 2020.
https://doi.org/10.1007/978-3-030-65775-8_23
248 W. M. C. da Silva et al.

1 Introduction
Terpenes are the oldest and various plant natural products as essential oils,
and some of them, as volatile compounds, play a crucial role in response to
herbivores, interaction with other plants, and attracting pollinators [33,36]. The
benefits of terpenes properties for humans spread through their use as flavors
in agricultural and industrial products, or as fragrances in foods and cosmetics,
plus pharmaceuticals and biofuels [2,39].
The wide range of terpenes and its “chemodiversity” is expected as a char-
acteristic of life, taking into account the considerable biodiversity of plants and
their interactions with other organisms [16,36]. This chemodiversity is related
to the chemical mechanisms catalyzed by Terpene synthases/cyclases (TPS)
influencing the terpenes’ variety. This variety may be related to their biolog-
ical function by adjusting the mixture and amount of terpenes to the specificity
of the target, both in communication relations and concerning protection against
numerous predators, parasites and competitors [16,32,36].
Terpenes are named according to the number of C5 isoprenoid units incor-
porated into their carbon skeletons as mono- (C10 ), sesqui- (C15 ), di- (C20 ),
sester- (C25 ), tri- (C30 ) and sesquarterpenes (C35 ) [37]. The isoprenoid units can
be Isopenthenyl Diphosphate (IP P ) or its allylic isomer Dimethylallyl Diphos-
phate (DM ADP ) that are condensed by prenyltransferases to produce larger
prenyl diphosphates, such as the monoterpene precursor Geranyl Diphosphate
(GDP ), which is the minimum-length cyclization substrate terpene biosynthe-
sis [10]. Commonly, after the GP P diphosphate loss and a C1 − C6 bond forma-
tion, the monoterpene formation proceeds through the α-terpinyl cation (Fig. 1)
by means of a cascade of reactions that include C − C bonds, Wagner-Meerwein
rearrangements, allyl- and methyl-shifts caused by conformational changes of
intermediate cations, and carbocation capture by water and hydride [33].
There are two classes of TPSs: Class I and Class II, defined by catalytically
essential amino acid motifs [7,27]. TPSs I convert linear, all-trans, isoprenoids,
geranyl (C10)-, farnesyl (C15)-, or geranylgeranyl (C20)-diphosphate into numer-
ous varieties of monoterpenes, sesquiterpenes, and diterpenes. The TPSs I bind
their substrate by coordinating a trinuclear divalent metal ion catalytic site
(generally a M g 2+ ), consisting of a central cavity formed by mostly antiparallel
α-helices. This catalytic site has an aspartate-rich DDxxD/E motif, and often
another N SE/DT E motif in the C-terminal portion [24]. TPSs Class II act
by triggering GGPP protonation, which results in successive carbocations and
cyclizations to form, for example, copalyl-diphosphate (CP P ) [27]. In the Class
II TPSs, the DxDD motif (distinct from the TPS I DDxxD/E motif) catalyzes
the reaction, also using a M g 2+ cofactor to assist substrate binding and posi-
tioning [15]. The terpenes diversity also can be influenced by nucleotide changes
in the alleles of TPS genes [26]. In plants, the production of terpenes can be
compartmentalized (Fig. 2), and monoterpenes, for example, can be produced in
specialized structures: plastids [31].
Enzyme function prediction is particularly challenging when dealing with
TPS because of their capability to produce numerous carbon skeletons by
 Computational Simulations for Cyclizations Catalyzed by PMS 249

Fig. 1. Examples in the blend of monoterpenes produced by plants through Geranyl

Diphosphate (GP P ) cyclizations.

Fig. 2. Plant compartmentalization of terpenes biosynthesis.

250 W. M. C. da Silva et al.

catalyzing complex carbocation rearrangements. Thus, beyond the primary

sequence homology, metadata, such as environmental or experimental condi-
tions and cheminformatics approaches could help annotate the TPS accurately.
Considering that the TPS annotation may be enhanced by exploring the feasible
terpene blend produced by cyclization of their precursors, this work presents a
computational approach to generate it by simulating carbocations of GPP cat-
alyzed by in vitro or in vivo patterns of plant monoterpene synthases that have
already been described in the literature.
The work is organized as follows. Section 2 introduces the method, explain-
ing how the chemical network is produced, its data structure, and traversing.
Section 3 discusses the results and compares this approach with others, high-
lighting the contributions of our results for molecular biology and cheminfor-
matics. Finally, Sect. 4 presents the conclusion and an outline of the next steps
to consolidate the approach as a tool for the scientific community.

2 Method

Degenhardt et al. [11] described reaction mechanisms for plant mono- and sesqui-
terpene synthases identifying several monoterpenes as products of these chem-
ical rearrangements. Based on Degenhardt et al. [11] review and other sources
of plant enzymatic GP P cyclizations [6,13,17,18,36], we extended the approach
presented by Silva et al. [33] for plant sesquiterpenes biosynthesis by including
monoterpenes.
The method formally models the chemical reactions on a mechanistic level,
building pathways assembled as a chemical network. The chemical network is
abstracted as a directed multi-hypergraph, where the vertices correspond to
molecules and hyperedges to reactions. Each vertex represents a molecule, which
is abstracted by an undirected graph, where atoms are vertices, and the bonds
are edges. In other words, each vertex of the chemical network is composed of
an undirected graph, which represents a molecule, and the chemical reactions
on these molecules are modeled as graph transformations. The accumulation of
graph transformations following the provided rules compose a network obtained
in a given number of iterations that exposes the initial, intermediate, and final
compounds.
Rule-based graph transformations can be described formally by graph gram-
mars that generalize the much more commonly used term-rewriting systems.
Each rule describes a specific class of chemical reactions such as a ring closure
from C1 to C6 atom, or an allyl-shift. Each rule has a pattern L of atoms and
bonds that need to be present in the educts for the corresponding reaction. The
matched part of educts is then transformed as specified by the rule.
We used the double pushout (DPO) formalism [29] for graph rewriting
because it is particularly suitable to model chemistry: it ensures reversibility
of transformation and supports well-defined atom maps [3]. Here, each rule has
l r
the form p = (L ←−K − → R) where L is the left graph, R is the right graph,
and K is a context graph. Graph morphisms l and r describe the embedding
 Computational Simulations for Cyclizations Catalyzed by PMS 251

of the context into the L and R by connecting these graphs by l : K → L and

r : K → R. If a rule p is applied to a graph G, it is mandatory that L “matches” a
part of G. The existence of another graph morphism (m : L → G) captures this,
and together with the rule p and the matching morphism m, define the trans-
p,m
formation of the substrate G to the product H, written as G ==⇒ H. Figure 3
shows an example of a rule for quenching by water, where the molecules (H2 O
and α-terpynol), the rule, and the matching morphism are exposed according to
the DPO graph transformation.

Fig. 3. Example of rule for quenching by water, and its application to the molecules
H2 O and α-terpynol.

MedØlDatschgerl [4], or MØD for short, is a software package for chemically

inspired graph transformation that can target a set of initial molecules and
generate the reaction networks automatically, applying rule-based graph trans-
formations. Some of the rules from Silva et al. [33] were revised to compose a
more comprehensive new set of rules able to generating mono- and sesquiter-
penes. A total of 17 graph transformation rules were manually designed in GML
format [21] to represent the literature-based chemical mechanisms that lead to
the compounds shown in the Fig. 1.
These rules transform the initial set of molecules (GP P and H2 O), match-
ing them in a first iteration, and generating a new set of chemically feasible
molecules. Then, in a new iteration, the set of rules is again applied to the new
set of compounds, generating a third set of compounds. The same process can be
repeated for an arbitrary number of iterations. It is possible to customize both
the set of rules and the number of iterations for each simulation. We have built
a Web interface to facilitate this task, which allows us to generate a simulation
252 W. M. C. da Silva et al.

source code file according to the chosen parameters. A graphical explanation of

the proposed method is shown in Fig. 4 and the graphical representaion of the
rules is shown in Fig. 5.

Fig. 4. Method summary with examples of abstractions/codes, and their applications

for the chemical network generation based on graph transformations.
 Computational Simulations for Cyclizations Catalyzed by PMS 253

(a) 1,3 hydrid shift. (b) H + loss.

(c) Capture by H2 O. (d) GPP diphosphate loss.

(f) Diphosphate capture by geranyl

cation.
(e) Alternative GPP diphosphate loss.

(g) LPP diphosphate loss. (h) Alternative LPP diphosphate loss.

(i) 1-6 ring closure. (j) 3-7 ring closure.

(k) 2-7 ring closure. (l) Wagner-Meerwein 1,2 alcyl shift

(m) 1-8 cyclization (n) 2-6 closure

(p) Diphosphate capture by bornyl

Cation.
(o) 5-7 closure

(q) Allylic charge shift

Fig. 5. Graphical representation of the 17 chemical inspired built graph transformation

rules for the GP P cyclizations.
254 W. M. C. da Silva et al.

3 Results and Discussion

This work provides material and method for computer simulations capable of
reproducing the blend of monoterpenes produced by plants using chemical mech-
anisms reported in the literature. Based on the presented seventeen graph gram-
mar rules, which embedded the enzymatic cyclizations of Geranyl Diphosphate
(GP P ) described in [6,11,13,17,18,36], the simulations were able to generate
a massive universe of monoterpenes as final products, including those shown in
Fig. 1. Along with them, the simulations also produced all the cyclization steps,
explicitly exposing the final and intermediates compounds, forming an exten-
sive chemical network of potential monoterpene products. The entire chemical
network, a hypergraph, can be computationally accessed, processed for further
analysis, and conveniently exported in a P DF report.
Combining the graph transformation rules and the number of iterations
allows exploring a vast blend of feasible monoterpenes, exposing all the enzy-
matic mechanisms of cyclization of its precursor, the GP P . Notably, in this
chemical space, there are three groups of monoterpenes: i) those that are found in
nature and are known and deposited in scientific databases such as CHEBI [19],
PubChem [25], and KEGG [23]; ii) those that are found in nature but remain
unknown; iii) those that are physically feasible but probably will not be found
in nature, maybe due to the high energy cost of their production. Scientific
advances in technologies such as mass spectrometry combined with gas chro-
matography have allowed allocating more and more monoterpenes in the first
group, the known ones. For all these groups, the results presented here allow us
to clarify the chemical mechanisms of their biosynthesis.
Along with this, the analysis of omic data has increasingly provided evidence
based on the primary sequence of amino acid residues for functional protein
annotation. Like other enzymes, the TPS sequence of amino acid residues influ-
ences its three-dimensional structure [28], and mutations in these amino acids
can affect their structure, and hence their function, causing changes in the effi-
ciency, specificity, or concentration of their products [8,10,40]. Also, different
blends of terpenes are produced under different scenarios atop their primary
amino acid sequence. [5]. These scenarios can be, for example, the availability
of GP P , the subcellular compartmentalization, the pH, developmental stage,
and organ, abiotic and biotic stress, and fitness costs [1,2,5,20,38]. The depen-
dency on conditions beyond sequence homology makes it challenging to annotate
TPS accurately. Thus, the results presented here could compose a computational
annotation system that considers the following elements: the biological sequence
of the TPS, a blend of feasible terpenes produced by it, and a weighted subset
of this blend produced by a particular TPS under a particular scenario.
Although simulations can be customized, source code writing may be non-
trivial for non-computational researchers. Looking to find a way to make the
simulation environment more friendly, we created a Web interface that relieves
this condition, making customization of the simulation more intuitive and less
laborious as some drag and drop and clicks are enough. Such an interface (Fig. 6,
 Computational Simulations for Cyclizations Catalyzed by PMS 255

combined with a Docker Image with all the environment ready, is available in
the GitHub1 .

Fig. 6. On-line interface to build the simulation code.

3.1 Comparison with Related Works

There are some important computational metabolic pathway prediction initia-
tives that consider intermediate transformations in chemical reactions such as
RetroRules [14], Biotransformer [12], Isegawa et al. [22], and Tian et al. [37].
Although providing comparable results, these approaches differ by using distinct
methods and data structures to represent the molecules and their transforma-
tions. RetroRules [14] and Biotransformer [12] uses chemical reaction descrip-
tions encoded by SMARTS [35] and SMIRKS [34]. Isegawa et al. [22] used
AFIR [30] method to computationally predict pathways for terpene formation.
Tian et al. [37] presented a computational approach to generate all possible car-
bocations of monoterpene synthases defining and organizing the product chem-
ical space.
This work has similar objectives and results to RetroRules [14], Biotrans-
former [12], and Tian et al. [37], but quite a different methodology, presenting
itself as a versatile alternative. Chow et al. [9] used Tian et al. [37] approach to
characterize a sesquiterpene synthase from Streptomyces clavuligerus, which cor-
roborates the idea that the chemical networks can help the functional assignment
of TPS.

4 Conclusion and Future Work

This work presents a way to generate and explore the monoterpenes biosynthesis
and their diversity. It is done through a computationally tractable data structure
1
https://github.com/waldeyr/2PathTerpenes.
256 W. M. C. da Silva et al.

representing a chemical network that exposes the inputs, intermediates, and final
compounds at an atom-bond level. The results presented here, allied to biolog-
ical scenarios and the primary sequences of TPS, can enhance their functional
annotation. Moreover, plant metabolic engineering could take advantage of this
approach to help designing genes that produce monoterpenes’ desired assortment
for ecological and industrial purposes. These monoterpene results are added to
the previous results of sesquiterpenes [33] to compose a terpene biosynthesis
simulation system called 2Path. Also, it brings a Web interface that makes writ-
ing the simulations more intuitive and less laborious as some drag and drop,
and clicks are enough. All the work is available at https://github.com/waldeyr/
2PathTerpenes.

Acknowledgments. This research was funded in part by CAPES through a sandwich

scholarship to W.M.C.d.S. It was additionally supported in part by the Independent
Research Fund Denmark, Natural Sciences, grant DFF-7014-00041. M.E.M.T.W. and
M.B. has been continuously supported by productivity fellowship from CNPq.

References
1. Abbas, F., Ke, Y., Yu, R., Yue, Y., Amanullah, S., Jahangir, M.M., Fan, Y.:
Volatile terpenoids: multiple functions, biosynthesis, modulation and manipula-
tion by genetic engineering. Planta 246(5), 803–816 (2017). https://doi.org/10.
1007/s00425-017-2749-x
2. Aharoni, A., Jongsma, M.A., Bouwmeester, H.J.: Volatile science? metabolic engi-
neering of terpenoids in plants. Trends Plant Sci. 10(12), 594–602 (2005)
3. Andersen, J.L., Flamm, C., Merkle, D., Stadler, P.F.: Inferring chemical reaction
patterns using graph grammar rule composition. J. Syst. Chem. 4, 4 (2013)
4. Andersen, J.L., Flamm, C., Merkle, D., Stadler, P.F.: A Software Package for
Chemically Inspired Graph Transformation. In: Echahed, R., Minas, M. (eds.)
ICGT 2016. LNCS, vol. 9761, pp. 73–88. Springer, Cham (2016). https://doi.org/
10.1007/978-3-319-40530-8 5
5. Block, A.K., Vaughan, M.M., Schmelz, E.A., Christensen, S.A.: Biosynthesis and
function of terpenoid defense compounds in maize (zea mays). Planta 249(1), 21–
30 (2019)
6. Bohlmann, J., Steele, C.L., Croteau, R.: Monoterpene synthases from grand fir
(Abies grandis): cDNA isolation, characterization, and functional expression of
myrcene synthase, (-)-(4S)- limonene synthase, and (-)-(1S,5S)-pinene synthase. J.
Biol. Chemistry 272(35), 21784–21792 (1997)
7. Chen, F., Tholl, D., Bohlmann, J., Pichersky, E.: The family of terpene synthases
in plants: a mid-size family of genes for specialized metabolism that is highly
diversified throughout the kingdom. Plant J. 66(1), 212–229 (2011)
8. Chen, H., Li, G., Köllner, T.G., Jia, Q., Gershenzon, J., Chen, F.: Positive dar-
winian selection is a driving force for the diversification of terpenoid biosynthesis
in the genus oryza. BMC Plant Biol. 14(1), 1–12 (2014)
9. Chow, J.Y., et al.: Computational-guided discovery and characterization of a ses-
quiterpene synthase from Streptomyces clavuligerus. Proceedings of the National
Academy of Sciences of the United States of America 112(18), 5661–6 (2015)
 Computational Simulations for Cyclizations Catalyzed by PMS 257

10. Christianson, D.W.: Structural and chemical biology of terpenoid cyclases. Chem-
ical Rev. 117(17), 11570–11648 (2017)
11. Degenhardt, J., Köllner, T.G., Gershenzon, J.: Monoterpene and sesquiterpene
synthases and the origin of terpene skeletal diversity in plants. Phytochemistry
70(15–16), 1621–1637 (2009)
12. Djoumbou-Feunang, Y., Fiamoncini, J., Gil-de-la-Fuente, A., Greiner, R., Manach,
C., Wishart, D.S.: BioTransformer: a comprehensive computational tool for small
molecule metabolism prediction and metabolite identification. Journal of Chemin-
formatics 11(1), 1–25 (2019). https://doi.org/10.1186/s13321-018-0324-5
13. Dong, L., Jongedijk, E., Bouwmeester, H., Van Der Krol, A.: Monoterpene biosyn-
thesis potential of plant subcellular compartments. New Phytologist 209(2), 679–
690 (2016)
14. Duigou, T., du Lac, M., Carbonell, P., Faulon, J.L.: Retrorules: a database of
reaction rules for engineering biology. Nucleic Acids Res. 47(D1), D1229–D1235
(2018)
15. Gao, Y., Honzatko, R.B., Peters, R.J.: Terpenoid synthase structures: a so far
incomplete view of complex catalysis. Nat. Prod. Rep. 29(10), 1153 (2012)
16. Gershenzon, J., Dudareva, N.: The function of terpene natural products in the
natural world. Nat. Chem. Biol. 3(7), 408–414 (2007)
17. Godard, K.A., White, R., Bohlmann, J.: Monoterpene-induced molecular responses
in Arabidopsis thaliana. Phytochemistry 69(9), 1838–1849 (2008)
18. Gutensohn, M., et al.: Cytosolic monoterpene biosynthesis is supported by plastid-
generated geranyl diphosphate substrate in transgenic tomato fruits. Plant J.
75(3), 351–363 (2013)
19. Hastings, J., et al.: Chebi in 2016: Improved services and an expanding collection
of metabolites. Nucleic Acids Res. 44(D1), D1214–D1219 (2016)
20. Heinig, U., Gutensohn, M., Dudareva, N., Aharoni, A.: The challenges of cellular
compartmentalization in plant metabolic engineering. Current Opinion Biotechnol.
24(2), 239–246 (2013)
21. Himsolt, M.: Gml: a portable graph file format. Html page under http://www.fmi.
uni-passau.de/graphlet/gml/gml-tr.html, Universität Passau (1997)
22. Isegawa, M., Maeda, S., Tantillo, D.J., Morokuma, K.: Predicting pathways for
terpene formation from first principles-routes to known and new sesquiterpenes.
Chem. Sci. 5(4), 1555–1560 (2014)
23. Kanehisa, M., et al.: The kegg database. In: Novartis Foundation Symposium, pp.
91–100. Wiley Online Library (2002)
24. Kempinski, C., Jiang, Z., Bell, S., Chappell, J.: Metabolic engineering of higher
plants and algae for isoprenoid production. In: Schrader, J., Bohlmann, J. (eds.)
Biotechnology of Isoprenoids. ABE, vol. 148, pp. 161–199. Springer, Cham (2015).
https://doi.org/10.1007/10 2014 290
25. Kim, S., et al.: Pubchem 2019 update: improved access to chemical data. Nucleic
Acids Res. 47(D1), D1102–D1109 (2019)
26. Köllner Schnee, C., Gershenzon, J., Degenhardt, J.: The Variability of Sesquiter-
penes Emitted from Two Zea mays\nCultivars Is Controlled by Allelic Variation
of Two Terpene\nSynthase Genes Encoding Stereoselective Multiple\nProduct
Enzymes. The Plant Cell 16(May), 1115–1131 (2004)
27. Liu, W., et al.: Structure, function and inhibition of ent-kaurene synthase from
bradyrhizobium japonicum. Sci. Rep. 4, 612 (2014)
28. Liu, W., et al.: Structure, function and inhibition of ent-kaurene synthase from
bradyrhizobium japonicum. Sci. Rep. 4, 6214 (2014)
258 W. M. C. da Silva et al.

29. Löwe, M.: Algebraic approach to single-pushout graph transformation. Theor.

Comput. Sci. 109(1), 181–224 (1993)
30. Maeda, S., et al.: Artificial Force Induced Reaction (AFIR) method for exploring
quantum chemical potential energy surfaces. Chemical Record 16(5), 2232–2248
(2016)
31. Nagegowda, D.A.: Plant volatile terpenoid metabolism: biosynthetic genes, tran-
scriptional regulation and subcellular compartmentation. FEBS Letters 584(14),
2965–2973 (2010)
32. O’maille, P.E., et al.: Quantitative exploration of the catalytic landscape separating
divergent plant sesquiterpene synthases. Nat. Chemical Biol. 4(10), 617–623 (2008)
33. da Silva, W.M.C., et al.: Exploring plant sesquiterpene diversity by generating
chemical networks. Processes 7(4), 240 (2019)
34. Systems, D.C.I.: A reaction transform language. http://daylight.com/dayhtml/
doc/theory/theory.smirks.html. Accessed 30 Jan 2019
35. Systems, D.C.I.: SMARTS - a language for describing molecular patterns. http://
www.daylight.com/dayhtml/doc/theory/theory.smarts.html (2008). Accessed 30
Jan 2019
36. Talapatra, S.K., Talapatra, B.: Biosynthesis of terpenoids: the oldest natural prod-
ucts. Chemistry of Plant Natural Products, pp. 317–344. Springer, Heidelberg
(2015). https://doi.org/10.1007/978-3-642-45410-3 5
37. Tian, B., Poulter, C.D., Jacobson, M.P.: Defining the product chemical space of
monoterpenoid synthases. PLoS Comput. Biol. 12(8), 1–13 (2016)
38. Vattekkatte, A., Garms, S., Brandt, W., Boland, W.: Enhanced structural diversity
in terpenoid biosynthesis: enzymes, substrates and cofactors. Organic Biomolecular
Chemistry 16(3), 348–362 (2018)
39. Zhang, Y., Nielsen, J., Liu, Z.: Engineering yeast metabolism for production of
terpenoids for use as perfume ingredients, pharmaceuticals and biofuels. FEMS
Yeast Res. 17(8), fox080 (2017)
40. Zhuang, X., et al.: Dynamic evolution of herbivore-induced sesquiterpene biosyn-
thesis in sorghum and related grass crops. The Plant J. 69(1), 70–80 (2012)
 Oncogenic Signaling Pathways
in Mucopolysaccharidoses

Gerda Cristal Villalba Silva1,2,3 , Luis Dias Ferreira Soares3,5 ,

and Ursula Matte1,2,3,4(B)
1 Postgraduate Program in Genetics and Molecular Biology, UFRGS,
Porto Alegre 91501970, Brazil
[email protected]
2 Gene Therapy Center, HCPA, Porto Alegre 90035903, Brazil
3 Bioinformatics Core, HCPA, Porto Alegre 90035903, Brazil
4 Department of Genetics, UFRGS, Porto Alegre 91501970, Brazil
5 Graduation Program on Biotechnology/Bioinformatics, UFRGS,

Porto Alegre 91501-970, Brazil

Abstract. Cancer cells depend on several signaling pathways and organelles, such
as the lysosomes. Defects in the activity of lysosomal hydrolases involved in gly-
cosaminoglycan degradation lead to a group of lysosomal storage diseases called
Mucopolysaccharidoses (MPS). In MPS, secondary cell disturbance affects path-
ways common to cancer. This work aims to identify oncogenic pathways related
to cancer in the different MPS datasets available in public databases and compare
the ontologies across the different types of MPS. For this, we used 12 expres-
sion datasets of 6 types of MPS. Statistical analysis was based on hypergeometric
distribution followed by FDR correction. We found several enriched pathways
across the 12 MPS studies, among being 57.65% were KEGG pathways, 32.5%
of GO Biological Process, 2.5% GO Celular Component, and 7.35% GO Molec-
ular Function. Hippo signaling pathway and MAPK signaling pathway appear
in all datasets. Proteoglycans in cancer, Rap1 signaling pathway, and Cytokine-
mediated signaling pathway appears in 11 of 12 datasets. The lysosome partici-
pates in several biological processes, like autophagy, cell adhesion and migration,
and antigen presentation. These processes also may affect in several types of can-
cer and Lysosomal Storage Diseases. Studying the tumor ontology signature in
lysosomal disorders may help understand lysosomal storage diseases and cancer’s
underlying mechanisms. This may help amplify therapeutic approaches for both
types of diseases.

Keywords: Cancer pathways · Gene ontology · Lysosomal storage diseases

1 Introduction
Several metabolic pathways are deranged in cancer cells. The proliferation ability of
tumors depends on a cascade of signaling pathways in several cancer cells’ organelles,
such as the lysosomes [1]. Lysosomes are cellular compartments responsible, among

© Springer Nature Switzerland AG 2020

J. C. Setubal and W. M. Silva (Eds.): BSB 2020, LNBI 12558, pp. 259–264, 2020.
https://doi.org/10.1007/978-3-030-65775-8_24
260 G. C. V. Silva et al.

other functions, for the degradation of macromolecules through acid hydrolases con-
tained within them. Defects in these enzymes culminate in the lysosomal accumulation
of intermediate metabolites or macromolecules, known as lysosomal storage diseases
[2]. Lysosomal Storage Diseases (LSD) are a group of more than 50 rare metabolic
diseases, among which we can highlight the Mucopolysaccharidoses (MPS). In MPS,
secondary cell disturbance affects pathways common to cancer.
This work aims to identify oncogenic pathways related to cancer in the different
datasets of MPS available in public databases and to compare the ontologies across the
different types of MPS.

2 Methods
Gene expression analysis considered 12 datasets available at GEO (https://www.ncbi.
nlm.nih.gov/geo), from six different MPS types. For RNA-seq data, we used edgeR, and
for microarray data, we used R packages according to the experiment’s platform. Fur-
thermore, the data present in this work are available in the MPSBase (https://www.ufrgs.
br/mpsbase/). Statistical analysis was based on hypergeometric distribution followed by
FDR correction. We perform the enrichment analysis in Cytoscape, with Bingo and
ClueGo plugins. We search the child terms with QuickGo. We selected 12 datasets,
being 2 of MPS I; 1 from MPS II; 1 from MPS IIIA; 3 MPS IIIB; 1 MPS VI; and
4 from MPS VII. These datasets comprise an RNA-seq data of Illumina HiSeq 2500
platform of human iPSC-derived Neuronal Stem Cell (MPS I, GSE111906); Agilent-
021193 Canine (V2) microarray of Ascending Aorta, Descending Aorta and Carotid
Aorta (MPS I, GSE78889); AB SOLiD 3 Plus System (Mus musculus) of Brain sam-
ples (MPS II, GSE95224); Agilent-028005 SurePrint G3 Mouse GE 8×60K Microarray
of Brain and Blood samples (MPS IIIA, GSE97759); Agilent-012694 Whole Mouse
Genome G4122A of Lateral entorhinal cortex and Medial entorhinal cortex (MPS IIIB,
GSE15758); Affymetrix Human Exon 1.0 ST Array of iPSC-derived Neuronal Stem
Cell (MPS IIIB, GSE23075); Affymetrix Human Exon 1.0 ST Array of HeLa depleting
NAGLU (MPS IIIB, GSE32154); Affymetrix Mouse Gene 1.0 ST Array of ARSB null
mouse hepatic cells (MPS VI, GSE77689); Illumina Mouse-8 Expression BeadChip of
Descending aorta (MPS VII, GSE30657); Affymetrix Mouse Genome 430A 2.0 Array
of six brain regions (MPS VII, GSE34071); Affymetrix Mouse Exon 1.0 ST Array of
iPS embryo-derived ES cells with controls derived from B6 Blu ES cells and Mouse
embryonic fibroblast (MPS VII, GSE36017); and Affymetrix Mouse Genome 430A 2.0
Array of hippocampus (MPS VII, GSE76283).

3 Results
We found 680 oncogenic enriched ontologies across the 12 MPS studies, among being
392 were KEGG pathways (57.65%), 221 GO Biological Process (32.5%), 17 GO Celu-
lar Component (2.5%), and 50 of GO Molecular Function (7.35%). Hippo signaling
pathway and MAPK signaling pathway appears in all datasets. Proteoglycans in cancer,
Rap1 signaling pathway, and Cytokine-mediated signaling pathway appears in 11 of 12
datasets (see Fig. 1).
 Oncogenic Signaling Pathways in Mucopolysaccharidoses 261

(a)

(b)

(c)

(d)

Fig. 1. Top Gene ontology of oncogenic terms of MPS. (a) KEGG pathways; (b) GO Biological
Process; (c) Molecular Function; (d) Cellular Component.
262 G. C. V. Silva et al.

Considering only the MPS types, the ontologies Axon guidance, Focal adhesion,
Hippo signaling pathway, MAPK signaling pathway, Metabolic pathways, Pathways in
cancer, PI3K-Akt signaling pathway, Proteoglycans in cancer, Rap1 signaling pathway,
and Ras signaling pathway are present in all the MPS types found in the GEO. The
following Table 1 gives a summary of the most frequent oncogenic ontologies according
to the MPS type.

Table 1. Prevalent oncogenic enriched pathways of datasets analyzed. In bold, the ontology
appears in all MPS types.

Term Ontology MPS MPS MPS MPS MPS MPS

I II IIIA IIIB VI VII
Apelin signaling pathway KEGG X X X X
Apoptosis KEGG X X X X X
Autophagy KEGG X X X X
Axon guidance KEGG X X X X X X
Calcium signaling pathway KEGG X X X X X
cAMP signaling pathway KEGG X X X X X
Focal adhesion KEGG X X X X X X
FoxO signaling pathway KEGG X X X X X
Hepatocellular carcinoma KEGG X X X X
Hippo signaling pathway KEGG X X X X X X
MAPK signaling pathway KEGG X X X X X X
Metabolic pathways KEGG X X X X X X
mTOR signaling pathway KEGG X X X X X
Oxytocin signaling pathway KEGG X X X X X
Pathways in cancer KEGG X X X X X X
PI3K-Akt signaling pathway KEGG X X X X X X
Proteoglycans in cancer KEGG X X X X X X
Rap1 signaling pathway KEGG X X X X X X
Ras signaling pathway KEGG X X X X X X
Wnt signaling pathway KEGG X X X X X X
Chemokine-mediated signaling pathway GO_BP X X X X X
Cytokine-mediated signaling pathway GO_BP X X X X X
EGFR signaling pathway GO_BP X X X X X
Slit-Robo signaling complex GO_CC X X X X
p53 binding GO_MF X X X X
 Oncogenic Signaling Pathways in Mucopolysaccharidoses 263

The dataset with the most enriched pathways is GSE32154 (MPS IIIB, Homo sapi-
ens) with 90 ontologies (see Fig. 2). The dataset with the most enriched KEGG terms is
GSE30657 (MPS VII, Mus musculus) with 60 KEGG terms. The GSE32154 (MPS IIIB,
Homo sapiens) have the most GO Biological Process enriched terms, with 45 terms.
For GO Cellular Component, the dataset with more enriched terms in this category
is GSE32154 (MPS IIIB, Homo sapiens) with 5 terms. Lastly, in the GO Molecular
Function, the GSE32154 is the dataset with more enriched terms (10 terms found).

Fig. 2. Number of enriched cancer ontologies across the MPS datasets.

4 Discussion

Oncogenic activation can lead to the destabilization of lysosomal membranes and an

increase of lysosomal hydrolases into the cytosol, where they can contribute to the
demise of the cancer cell [3].
Axon guidance and Wnt signaling pathway are related to the neurological impairment
found in several MPS types [4]. Alterations in autophagy are frequently found in MPS
patients with neurodegenerative symptoms [4]. In cancer, autophagy are related to cancer
initiation, proliferation, and survival [5].
The ontologies Cell cycle, Hippo, Notch, PI-3-Kinase/Akt, RAS, TGFβ signaling,
P53 and β-catenin/WNT signaling pathway are considered canonical oncogenic path-
ways. Unfortunately, 89% of tumors found in the TCGA consortium had at least one
driver alteration in these pathways, and 57% percent of the tumors had at least one alter-
ation potentially targetable by currently available treatments [6]. We hypothesize that
these signaling pathways are altered because glycosaminoglycans play an essential role
264 G. C. V. Silva et al.

in the composition of the extracellular matrix [7], helping to regulate processes such
as metabolic signaling, apoptosis, cell migration, adhesion, and antigen presentation, in
both cancer and MPS.

5 Concluding Remarks
The available public data is essential for amplified the multi-omic knowledge of complex
and rare diseases. Bioinformatic approaches, such as gene enrichment analysis, may
help us understand the complexity of processes deranged in several diseases. Studying
the tumor ontology signature in lysosomal disorders may help understand lysosomal
storage diseases and cancer’s underlying mechanisms. This may help amplify therapeutic
approaches for both types of diseases.

References
1. Cairns, R.A., Harris, I.S., Mak, T.W.: Regulation of cancer cell metabolism. Nat. Rev. Cancer
11(2), 85–95 (2011). https://doi.org/10.1038/nrc2981
2. Matte, U., Pasqualim, G.: Lysosome: the story beyond the storage. J. Inborn Errors Metab.
Screen. 4, e160044 (2016). https://doi.org/10.1177/2326409816679431
3. Kallunki, T., Olsen, O.D., Jäättelä, M.: Cancer-associated lysosomal changes: friends or foes?
Oncogene 32(16), 1995–2004 (2012). https://doi.org/10.1038/onc.2012.292
4. Fiorenza, M.T., Moro, E., Erickson, R.P.: The pathogenesis of lysosomal storage disorders:
beyond the engorgement of lysosomes to abnormal development and neuroinflammation. Hum.
Mol. Genet. 27(R2), R119–R129 (2018). https://doi.org/10.1093/hmg/ddy155
5. Martinez-Carreres, L., Nasrallah, A., Fajas, L.: Cancer: linking powerhouses to suicidal bags.
Front. Oncol. 7, 204 (2017). https://doi.org/10.3389/fonc.2017.00204
6. Sanchez-Vega, F., et al.: Oncogenic signaling pathways in the cancer genome atlas. Cell 173(2),
321–337 (2018). https://doi.org/10.1016/j.cell.2018.03.035
7. Davidson, S.M., Vander Heiden, M.G.: Critical functions of the lysosome in cancer biology.
Ann. Rev. Pharmacol. Toxicol. 57(1), 481–507 (2017). https://doi.org/10.1146/annurev-pha
rmtox-010715-103101
 Natural Products as Potential Inhibitors
for SARS-CoV-2 Papain-Like Protease:
An in Silico Study

Jesus Alvarado-Huayhuaz1 , Fabian Jimenez2 , Gerson Cordova-Serrano3 ,

Ihosvany Camps4 , and Wilmar Puma-Zamora1,5(B)
1
Laboratorio de Investigación en Biopolı́meros y Metalofármacos, Universidad
Nacional de Ingenierı́a, Lima, Peru
[email protected]
2
Department of Technological Operations, Alluriam HealthCare Inc.,
Orlando, FL, USA
3
Escuela de Farmacia y Bioquı́mica, Universidad Marı́a Auxiliadora, Lima, Peru
4
Laboratorio de Modelagem Computacional, Universidade Federal de Alfenas,
Minas Gerais, Brazil
5
Escuela de Farmacia y Bioquı́mica, Universidad Nacional de San Antonio Abad del
Cusco, Cusco, Peru
[email protected]

Abstract. In December 2019 in Wuhan, China, the first case of a patient

infected with a new coronavirus was reported, later called SARS-CoV-
2 by the World Health Organization, and so far, there is no approved
drug or vaccine against this new virus. Considering this, previous studies
identified several essential SARS-CoV-2 viral proteins, among these to
the papain-like protease (PLpro ), an enzyme with an important role in
viral spread and evading the host immune response, making it an attrac-
tive drug target. For this reason, using a library of 213,038 structures of
natural products and virtual screening, we identified 10 molecules with
high affinity for residues of the SARS-CoV-2 PLpro allosteric site, which
could be tested in vitro against the virus or as lead-compounds to develop
inhibitors that are more effective.

Keywords: SARS-CoV-2 · Natural products · Virtual screening

1 Introduction
SARS-CoV-2 is an RNA virus whose genome has 10 open reading frames that
encode various structural and non-structural proteins [23]. Among these pro-
teins, one of the most studied together with the main protease and the protein
spike, is the papain-like protease (PLpro ) which is necessary to facilitate the
spread of the virus and help to evade the host immune response [7]. One of
the most outstanding aspects of SARS-CoV-2 PLpro is its role as an ubiquitin-
specific protease, especially the ubiquitin-like protein ISG15, which is a known
c Springer Nature Switzerland AG 2020
J. C. Setubal and W. M. Silva (Eds.): BSB 2020, LNBI 12558, pp. 265–270, 2020.
https://doi.org/10.1007/978-3-030-65775-8_25
266 J. Alvarado-Huayhuaz et al.

regulatory agent of human immune response pathways. In this way, it is proposed

that blocking SARS-CoV-2 PLpro could stop the viral infection and improve the
immune response [18].
Since to date there is no officially an effective vaccine or antiviral against
this coronavirus [2], we highlight the importance of rapid search techniques like
virtual screening (VS) and molecular docking to identify possible inhibitors of
this enzyme, using the evaluation of molecular interactions in its active and
allosteric sites. In this work, we focus on natural products (NPs) because in
these organic molecules we can find a large surface area and multiple functional
groups capable of interacting by hydrogen bonds or van der Waals forces.

2 Methods
We carried out two stages. First, a VS with AutoDock Vina (referred to as
Vina from here on) [19] and 213,038 chemical structures of NPs collected in the
freely available Universal Natural Product Database - In Silico MS/MS Database
(UNPD-ISDB) [1].
To ensure the reliability of the VS results, the second stage consisted of
molecular docking divided into three sub-stages with Glide [5] from the suite
Schrödinger. The best hits of the VS were docked with standard precision (Glide-
SP), extra precision (Glide-XP), and then with the Induced Fit Docking (IFD)
protocol, that allows flexibility to the residues from the pocket. Finally, the
binding free energy (ΔG b ) of the best 10 complexes from IFD protocol was cal-
culated using the physics-based MM-GBSA (Molecular Mechanics-Generalized
Born Surface Area) method with the Prime [10] module of the suite Schrödinger.

Preparation of the Receptor. From Protein Data Bank we retrieve the struc-
ture of the SARS-CoV-2 PLpro with PDB ID 6W9C (resolution 2.7 Å). For VS,
all water molecules and ions were removed and polar hydrogens were added to
the receptor using AutoDock Tools v1.5.6 [14]. For molecular docking, the pro-
tease was prepared by removing unnecessary water molecules and ions with the
Protein Preparation Wizard [13] module and considering the protonation states
of the residues at pH 7.4 with Epik [17].

Ligand Preparation. The structures of NPs in InChI strings were obtained

from the UNPD-ISDB. 3D coordinates were assigned to them and then they were
geometrically minimized with 20 thousand steps using the Open Babel software
(v3.0.0) [16]. For VS and molecular docking, we add appropriate hydrogens for
pH 7.4.

Validation of Docking Protocol. For the validation we used the crystallized

structure of SARS-CoV PLpro (PDB ID 3E9S) to check the suitability and accu-
racy of Vina and Glide as appropriate docking tools. The protein and its native
ligand were prepared under the same conditions detailed above. The box size
 Natural Products Against SARS-CoV-2 Papain-Like Protease 267

and the coordinates of the centroid of the binding site were calculated in the
same way as for VS and molecular docking stages.

Virtual Screening. Ligands and protein in PDBQT format were used as input
files. The centroid of the binding site (x = −50.1, y = 15.1 and z = 34.4) and
the optimal box size (40 × 40 × 40 Å), which cover the active and allosteric
sites, were obtained with the POCASA server [25] and the eBoxSize script [4]
respectively. Vina was used with an exhaustiveness level of 24 in the Quinde 1
supercomputer.

Molecular Docking and MM-GBSA. The best 1000 ligands were selected
from VS according to their binding affinity values. These were docked with Glide-
SP in the SARS-CoV-2 PLpro active and allosteric sites. From the results with
Glide-SP, the best 100 hits were selected and docked with Glide-XP. From the
last step with Glide-XP, the best 10 hits were selected and again they were
docked with the IFD protocol, allowing flexibility to the residues close to the
active and allosteric sites. For these steps, the OPLS 2005 force field was used,
and the same conditions of binding site coordinates and box size from the VS
were used. Finally, the ΔG b of the best ligand-receptor complexes from IFD was
estimated, according to Eq. 1, with the Prime module using an implicit solvation
model VSGB and the OPLS 2005 force field. In Eq. 1, G complex , G protein , and
G ligand are the free energies of the complex, protein, and ligand, respectively.

ΔGb = Gcomplex − (Gprotein + Gligand ) (1)

3 Results and Discussion

Validation of Docking Protocol. Using the Open Babel obrms tool, the
RMSD was calculated between the native co-crystallized ligand (TTT; 5-amino-
2-methyl-N-[(1R)-1-naphthalen-1-ylethyl] benzamide) of the SARS-CoV PLpro
and its best re-docked conformations obtained with Vina (0.47 Å), Glide-SP
(0.57 Å) and Glide-XP (0.81 Å). Since RMSD values are less than 2.0 Å with
respect to the native ligand, it is considered that the protocols used in the present
study are reliable [21].

Virtual Screening. According to Vina, 88,848 NPs obtained a binding affin-

ity between −6 and −7 kcal/mol, which represents 41.7% approximately of all
ligands (Fig. 1). The 1000 NPs selected for the following molecular docking tests
had a binding affinity between −8 to −13 kcal/mol.

Molecular Docking and MM-GBSA. We identify 10 molecules with the

best ΔG b values. Hydrogen bonds (H-bonds) interactions were found with many
residues from the subsites S3 and S4 (underlined in Table 1). These molecules
268 J. Alvarado-Huayhuaz et al.

〈-1,0] 2 Frequency

Binding affinity ranges (kcal/mol)

〈-2,-1] 15
〈-3,-2] 176
〈-4,-3] 2,496
〈-5,-4] 16,819
〈-6,-5] 44,234
〈-7,-6] 88,848
〈-8,-7] 49,858
〈-9,-8] 9,772
〈-10,-9] 725
〈-11,-10] 71
〈-12,-11] 20
〈-13,-12] 2
0 20,000 40,000 60,000 80,000 100,000
Number of ligands

Fig. 1. Frequency of ligands according to binding affinity, calculated by Vina.

could block the enzyme activity through non-covalent interactions (between 6 to

9 H-bonds and lipophilic interactions) with many residues of the SARS-CoV-2
PLpro allosteric site, inducing the closure of the loop [3]. These hits are relatively
large glycosides and tannins with previously reported biological properties.

Table 1. ΔG b values, interacting residues, classification and biological source of the

10 molecules with the best affinity for SARS-CoV-2 PLpro

ISDB IDΔG b (kcal/mol)H-bonds interactions Type of compounds Biological source

208044 −131.251 K157, E161, D164, R166, E167, Steroidal glycoside M. pubescens [22]
Y264, Y268, Q269, T301
139907 −129.517 C155, N156, D164, R166, E167, Steroidal glycoside O. japonicus [12]
Y273, D302
113050 −127.701 C155, K157, E161, Y171, P248, Steroidal glycoside A. filicinus [26]
G266, Q269
066371 −121.304 E161, R166, Y171, E203, P248, Hydrolyzable tanninQ. phillyraeoides [15]
Y264, Y273, T301
141146 −118.245 K157, L162, Y171, P248, G266, Terpene glycoside P. quinquefolium [20]
N267, Y268, Q269
058326 −105.371 C155, N156, E167, Y171, P248, Steroidal glycoside A. asphodeloides [11]
Y264, G266, Y273, T301
150945 −86.977 E161, L162, D164, E167, Y264, Steroidal glycoside A. amurensis [9]
Y273, T301
035549 −75.780 K157, E161, L162, D164, E167 G266, Steroidal glycoside D. inarticulata [24]
N267, Q269
171225 −73.354 E161, L162, E167, S245, Y273, D302. Hydrolyzable tanninL. speciosa [6]
*Pi-cation: R166
122929 −56.530 E161, L162, D164, R166, Y273, T301Hydrolyzable tanninJ. nigra [8]

4 Conclusion

In this study, we identified 10 molecules that can interact favorably with residues
of the allosteric site of PLpro , being considered stable complexes according to
their ΔG b values. This search was done using a library of 213,038 structures
of NPs collected in the UNPD-ISDB, and techniques like virtual screening and
molecular docking. These molecules are mainly glycosides and tannins, and their
 Natural Products Against SARS-CoV-2 Papain-Like Protease 269

molecular structures are relatively large, presenting between 6 to 9 H-bonds and

lipophilic interactions with residues of SARS-CoV-2 PLpro allosteric site. How-
ever, still are needed prediction of pharmacokinetic properties and molecular
dynamics simulations to evaluate the possibility that these molecules can be
used for in vitro PLpro inhibition assays, or lead optimization processes.

Acknowledgments. The team gives special thanks to Siembra Empresa Pública

(Imbabura, Ecuador) for their support with Quinde 1 supercomputer. Also thanks to
Dr. Meng Guo and Dr. Li Wang (National Supercomputing Center in Jinan, China),
for their support with the implementation of Vina and helpful discussions, and to Dr.
Ana Valderrama (Universidad Nacional de Ingenierı́a, Perú) for her personal financial
contribution to the project. Finally, remembering that #SinCienciaNoHayFuturo.

References
1. Allard, P.M., et al.: Integration of molecular networking and In-Silico MS/MS frag-
mentation for natural products dereplication. Anal. Chem. 88, 3317–3323 (2016)
2. Arnold, C.: Race for a vaccine. New Sci. 245, 44–47 (2020)
3. Estrada, E.: COVID-19 and SARS-CoV-2. Modeling the present, looking at the
future. Phys. Rep. 869, 1–51 (2020)
4. Feinstein, W.P., Brylinski, M.: Calculating an optimal box size for ligand dock-
ing and virtual screening against experimental and predicted binding pockets. J.
Cheminform. 7(1), 1–10 (2015). https://doi.org/10.1186/s13321-015-0067-5
5. Friesner, R.A., et al.: Extra precision glide: docking and scoring incorporating a
model of hydrophobic enclosure for protein-ligand complexes. J. Med. Chem. 49,
6177–6196 (2006)
6. Guo, S., et al.: The anti-diabetic effect of eight Lagerstroemia speciosa leaf extracts
based on the contents of ellagitannins and ellagic acid derivatives. Food Funct. 11,
1560–1571 (2020)
7. Harcourt, B.H., et al.: Identification of severe acute respiratory syndrome coron-
avirus replicase products and characterization of papain-like protease activity. J.
Virol. 78, 13600–13612 (2004)
8. Ho, K.V., et al.: Identifying antibacterial compounds in black walnuts (Juglans
nigra) using a metabolomics approach. Metabolites 8, 58 (2018)
9. Ishii, T., Okino, T., Mino, Y., Tamiya, H., Matsuda, F.: Plant-growth regulators
from common starfish (Asterias amurensis Lütken) waste. Plant Growth Regul.
52, 131–139 (2007)
10. Jacobson, M.P., Friesner, R.A., Xiang, Z., Honig, B.: On the role of the crystal
environment in determining protein side-chain conformations. J. Mol. Biol. 320,
597–608 (2002)
11. Ji, D., Huang, Z.y., Fei, C.h., Xue, W.w., Lu, T.l.: Comprehensive profiling and
characterization of chemical constituents of rhizome of Anemarrhena asphodeloides
Bge. J. Chromatogr. B 1060, 355–366 (2017)
12. Kang, Z.Y., Zhang, M.J., Wang, J.X., Liu, J.X., Duan, C.L., Yu, D.Q.: Two new
furostanol saponins from the fibrous root of Ophiopogon japonicus. J. Asian Nat.
Prod. Res. 15, 1230–1236 (2013)
13. Madhavi Sastry, G., Adzhigirey, M., Day, T., Annabhimoju, R., Sherman, W.:
Protein and ligand preparation: parameters, protocols, and influence on virtual
screening enrichments. J. Comput. Aided. Mol. Des. 27, 221–234 (2013)
270 J. Alvarado-Huayhuaz et al.

14. Morris, G.M., et al.: AutoDock4 and AutoDockTools4: automated docking with
selective receptor flexibility. J. Comput. Chem. 30, 2785–2791 (2009)
15. Nonaka, G.i., Nakayama, S., Ishioka, I.N.: Tannins and related compounds.
LXXXIII. Isolation And Structures of Hydrolyzable Tannins, Phillyraeoidins A-
E From Quercus Phillyraeoides. Chem. Pharm. Bull. 37, 2030–2036 (1989)
16. O’Boyle, N.M., Banck, M., James, C.A., Morley, C., Vandermeersch, T., Hutchison,
G.R.: Open babel: an open chemical toolbox. J. Cheminform. 3, 33 (2011)
17. Shelley, J.C., Cholleti, A., Frye, L.L., Greenwood, J.R., Timlin, M.R., Uchimaya,
M.: Epik: a software program for pK a prediction and protonation state generation
for drug-like molecules. J. Comput. Aided. Mol. Des. 21, 681–691 (2007)
18. Shin, D., et al.: Papain-like protease regulates SARS-CoV-2 viral spread and innate
immunity. Nature (2020)
19. Trott, O., Olson, A.J.: AutoDock Vina: Improving the speed and accuracy of dock-
ing with a new scoring function, efficient optimization, and multithreading. J.
Comput. Chem. 31, 455–461 (2009)
20. Tung, N.H., Shoyama, Y.: Eastern blotting analysis and isolation of two new
dammarane-type saponins from american ginseng. Chem. Pharm. Bull. 60, 1329–
1333 (2012)
21. Wang, R., Lu, Y., Wang, S.: Comparative evaluation of 11 scoring functions for
molecular docking. J. Med. Chem. 46, 2287–2303 (2003)
22. Zhao, W., Xu, R., Qin, G., Vaisar, T., Lee, M.S.: Saponins from mussaenda
pubescens. Phytochemistry 42, 1131–1134 (1996)
23. Wu, F., et al.: A new coronavirus associated with human respiratory disease in
China. Nature 579, 265–269 (2020)
24. Yang, W.L., Tian, J., Peng, S.L., Guan, J.F., Ding, L.S.: Chemical constituents of
Diuranthera inarticulata. Yao Xue Xue Bao 36, 590–4 (2001)
25. Yu, J., Zhou, Y., Tanaka, I., Yao, M.: Roll: a new algorithm for the detection of
protein pockets and cavities with a rolling probe sphere. Bioinformatics 26, 46–52
(2009)
26. Zhou, L., Cheng, Z., Chen, D.: Simultaneous determination of six steroidal saponins
and one ecdysone in Asparagus filicinus using high performance liquid chromatog-
raphy coupled with evaporative light scattering detection. Acta Pharm. Sin. B 2,
267–273 (2012)
 Author Index

Alexandrino, Alexsandro Oliveira 117, 129 do Almo, Manuela M. 180

Almeida, Nalvo F. 1 Dorte, André 57
Alvarado-Huayhuaz, Jesus 265 dos Santos Adan, Wenny Camilla 93
Alves, Levy Bueno 215
Andersen, Jakob L. 247 Entenza, Julio Omar Prieto 25
Araújo, Carlos Leonardo 38 Evangelista, Adriane Feijo 81
Araújo, Fabrício 13 Evangelista, Renato Feijo 81

Barros, Glecio Oliveira 162 Ferraz, Matheus Vitor Ferreira 93

Biller, Priscila 141 Ferreira, Mívian 235
Blanco, Iago Rodrigues 38 Flamm, Christoph 247
Braga, Marcus 13 Folador, Adriana Carneiro 38
Braz, Leodécio 150 Fortirer, Janaina da Silva 221
Breve, Murilo Montanini 192
Brigido, Marcelo 180, 247
Gatter, Thomas 180
Brito, Klairton Lima 117
Giuliatti, Silvana 215
Buckeridge, Marcos Silveira 221
Gollapalli, Pavan 204
Grandis, Adriana 221
Camps, Ihosvany 265 Grishina, Maria 204
Castillo-Ordoñez, William O. 215
Cavalcante, Renata Lilian Dantas 50 Haeusler, Edward Hermann 25
Cereda, Luis 57 Hoffmann, Steve 180
Cordova-Serrano, Gerson 265 Holanda, Maristela 1
Cutigi, Jorge Francisco 81
Jimenez, Fabian 265
da Silva Segundo, Leodécio Braz 69
da Silva, Waldeyr Mendes Cordeiro 180, Kandagalla, Shivananda 204
247
de Andrade, Daniela P. 247 Lifschitz, Sérgio 1, 25
de Armas, Elvismary Molina 1 Lins, Roberto Dias 93
de Assis Leite, Deborah Catharine 162 Lopes, Bruno 105
de Carvalho, Andre C. P. L. F. 81 Lopes, Fabrício Martins 192
de Melo-Minardi, Raquel C. 235 Lopes, Maurício 57
de Oliveira Lage Ferreira, Cynthia 81
de Oliveira, Daniel 1, 105 Maranhão, Andrea Q. 180
de Souza, Cid Carvalho 129 Mariano, Diego 235
de Souza, Jorge Estefano Santana 50 Matte, Ursula 259
de Toledo Castanho, Camila 221 Mattoso, Marta 105
Demarqui, Fernanda Manaia 168 Meidanis, João 141
Dias, Luiz Gustavo 105 Metzner, Beatriz Thomas 162
Dias, Ulisses 117 Miranda, Fábio 13
Dias, Zanoni 57, 69, 117, 129, 150 Miyazaki, Luiz 57
272 Author Index

Oliveira, Andre Rodrigues 117, 129 Shekarappa, Sharath Belenahalli 204

Oliveira, Gabriel 57 Silva, Artur 13
Ortega, Jose Miguel 50 Silva, Gerda Cristal Villalba 259
Simao, Adenilso 81
Padilha, Rafael 57 Soares, Luis Dias Ferreira 259
Paiva, Ana Carolina Silva 168 Sousa, Bruno L. 235
Pedrini, Helio 150 Sousa, Isabel G. 180
Pinheiro, Kenny 13 Stadler, Peter F. 180, 247
Pinheiro, Pedro Olímpio 129
Potemkin, Vladimir 204 Teixeira, Vinicius 150
Puma-Zamora, Wilmar 265 Trentin, Alex Batista 162

Ramos, Rodrigo Henrique 81 Valentini, Sandro Roberto 168

Ramos, Rommel 13
Walter, Maria Emília M. T. 247
Sakamoto, Tetsu 50 Watanabe, Tatiana Faria 168
Santoni, Mariana Marchi 168
Santos, Lucianna H. 235 Zanelli, Cleslei Fernando 168
Sato, Guilherme Tiaki Sassai 69 Zanetti, João Paulo Pereira 141