Methods in

Molecular Biology 1375

Pietro Hiram Guzzi Editor

Microarray
Data Analysis
Methods and Applications
Second Edition
METHODS IN MOLECULAR BIOLOGY

Series Editor
John M. Walker
School of Life and Medical Sciences
University of Hertfordshire
Hatfield, Hertfordshire, AL10 9AB, UK

For further volumes:

http://www.springer.com/series/7651
Microarray Data Analysis

Methods and Applications

Second Edition

Edited by

Pietro Hiram Guzzi

Department of Surgical and Medical Sciences,
University “Magna Græcia” of Catanzaro, Catanzaro, Italy
Editor
Pietro Hiram Guzzi
Department of Surgical and Medical Sciences
University “Magna Græcia” of Catanzaro
Catanzaro, Italy

ISSN 1064-3745 ISSN 1940-6029 (electronic)

Methods in Molecular Biology
ISBN 978-1-4939-3172-9 ISBN 978-1-4939-3173-6 (eBook)
DOI 10.1007/978-1-4939-3173-6
Library of Congress Control Number: 2016932899

Springer New York Heidelberg Dordrecht London

# Springer Science+Business Media New York 2007, 2016
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,
express or implied, with respect to the material contained herein or for any errors or omissions that may have been made.

Printed on acid-free paper

Humana Press is a brand of Springer

Springer Science+Business Media LLC New York is part of Springer Science+Business Media (www.springer.com)
Preface

The development of novel technological platforms in molecular biology has given a large
input to research and in particular has caused a big development of bioinformatics to
support storage, management, and analysis of a large amount of data about different
aspects of the omic world. We here in particular focus on two main techniques for studying
the activity of transcriptome, i.e., the set of molecules that play a role in the complex
mechanism of protein synthesis. Such a study focuses on the role of mRNA, i.e., coding
fragments of messenger RNA, and miRNA, i.e., small fragments of noncoding RNA. This
study has been conducted through two main technological platforms: microarray and
miRNA-microarray. More recently, the advent of next-generation sequencing techniques
is gaining a prominent role. Despite this, classical microarray studies are still alive since
there are a considerable number of published papers related to the generation and analysis
of microarray data.
The flow of information in this field starts from technological platforms that produce
different data. Examples of such platforms are microarray for studying the expression of
messenger RNA (mRNA) and microRNA (miRNA); genomic microarrays for studying
copy number variations (CNV) or single-nucleotide polymorphisms (SNP); novel micro-
arrays for studying noncoding RNAs (e.g., miRNA); and genomic arrays for pharmacoge-
nomics.
Classical studies focused on the individuation of the role of a single class of molecules
into a specific disease. Therefore they contained the analysis of a single class of data. More
recently, the biological assumption that different molecules (e.g., miRNA, mRNA, or
Transcription Factors) are strongly correlated has determined the rise of a novel discipline,
often referred to as computational systems biology or network systems biology. In such
discipline computer science, bioinformatics, and mathematical modeling play a synergistic
role in the interpretation of large data sets belonging to different data sources. Conse-
quently, a big attention has been paid to the development of integrated methods of
analysis, often based on distributed or high-performance architectures (e.g., Cloud) or
on semantic-based approaches, for extracting biologically relevant knowledge from data. In
parallel, a growing number of biological and medical papers have demonstrated the real
application of these methodologies.
This book is intended to cover main aspects of this area, and it covers a large area, from
the description of methodologies for data analysis to the real application. The intended
audience is students or researchers that need to learn main topics of research as well as
practitioners that need to have a look on applications. The structure of the presentation of
all the chapters makes it adapt even for the use in bioinformatics courses.
The book is composed of 15 chapters. It starts by presenting main concepts related to
data analysis. Wu and Gantier present main methodologies for preprocessing of microarray
data in Chapter 1. Cristiano and Veltri present a survey of miRNA Data analysis in
Chapter 2 while Calabrese and Cannataro discuss the rise of Cloud-based approaches in
Chapter 3. Chapter 4 by Lopez Kleine et al. presents the application of data mining
techniques for data analysis and in Chapter 5 Deveci et al. focus on the use of biclustering
to query different datasets. In Chapter 6 Chang and Lin discuss a web-based tool to
analyze the evolution of miRNA clusters. Roy et al. present in Chapter 7 the application

v
vi Preface

of biclustering to mine patterns of co-regulated genes. Chapters 8 and 9 present the use of
ontologies; in particular, Ovaska discusses the use of csbl.go tool while Agapito and Milano
survey main existing tools for semantic similarity analysis of microarray data. Wang et al. in
Chapter 10 introduce the integration of microarray and proteomic data. Chapter 11 by
Koumakis et al. discusses the relevance of Gene Regulatory Network Inference, while
Chapter 12 by Roy and Guzzi focuses on the assessment of Gene Regulatory Network
methods. The remaining chapters present some relevant applications in different medical
fields. Chapter 13 by Gan et al. is related to the analysis of Mouse data for metabolomics
studies. Chapter 14 by Di Martino et al. surveys the functional analysis of microRNA data
in multiple myeloma that is currently a big research area. Chapter 15 by Bhawe and Aghi
presents the application of microarray data analysis in glioblastomas. Finally, Chapter 16
discusses the analysis of microRNA data in cardiogenesis.

Catanzaro, Italy Pietro Hiram Guzzi

Contents

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
Contributors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix

Normalization of Affymetrix miRNA Microarrays

for the Analysis of Cancer Samples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Di Wu and Michael P. Gantier
Methods and Techniques for miRNA Data Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
Francesca Cristiano and Pierangelo Veltri
Bioinformatics and Microarray Data Analysis on the Cloud . . . . . . . . . . . . . . . . . . . . . 25
Barbara Calabrese and Mario Cannataro
Classification and Clustering on Microarray Data for Gene Functional
Prediction Using R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
Liliana López Kleine, Rosa Montaño, and Francisco Torres-Avilés
Querying Co-regulated Genes on Diverse Gene Expression Datasets
Via Biclustering. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
Mehmet Deveci, Onur Küçüktunç, Kemal Eren, Doruk Bozdağ,
Kamer Kaya, and Umit € V. Çatalyürek
MetaMirClust: Discovery and Exploration of Evolutionarily Conserved
miRNA Clusters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
Wen-Ching Chan and Wen-chang Lin
Analysis of Gene Expression Patterns Using Biclustering . . . . . . . . . . . . . . . . . . . . . . . 91
Swarup Roy, Dhruba K. Bhattacharyya, and Jugal K. Kalita
Using Semantic Similarities and csbl.go for Analyzing Microarray Data. . . . . . . . . . . 105
Kristian Ovaska
Ontology-Based Analysis of Microarray Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
Agapito Giuseppe and Marianna Milano
Integrated Analysis of Transcriptomic and Proteomic Datasets Reveals
Information on Protein Expressivity and Factors Affecting Translational
Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
Jiangxin Wang, Gang Wu, Lei Chen, and Weiwen Zhang
Integrating Microarray Data and GRNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
L. Koumakis, G. Potamias, M. Tsiknakis, M. Zervakis, and V. Moustakis
Biological Network Inference from Microarray Data, Current Solutions,
and Assessments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
Swarup Roy and Pietro Hiram Guzzi
A Protocol to Collect Specific Mouse Skeletal Muscles
for Metabolomics Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
Zhuohui Gan, Zhenxing Fu, Jennifer C. Stowe, Frank L. Powell,
and Andrew D. McCulloch
Functional Analysis of microRNA in Multiple Myeloma . . . . . . . . . . . . . . . . . . . . . . . . 181
Maria Teresa Di Martino, Nicola Amodio, Pierfrancesco Tassone,
and Pierosandro Tagliaferri

vii
viii Contents

Microarray Analysis in Glioblastomas. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195

Kaumudi M. Bhawe and Manish K. Aghi
Analysis of microRNA Microarrays in Cardiogenesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
Diego Franco, Fernando Bonet, Francisco Hernandez-Torres,
Estefania Lozano-Velasco, Francisco J. Esteban, and Amelia E. Aranega
Erratum to: Classification and Clustering on Microarray Data for Gene
Functional Prediction Using R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
Liliana López Kleine, Rosa Montaño, and Francisco Torres-Avilés

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
Contributors

MANISH K. AGHI  Graduate Division of Biomedical Sciences (BMS),

Department of Neurosurgery and Brain Tumor Research Center,
University of California at San Francisco (UCSF), San Francisco, CA, USA
NICOLA AMODIO  Department of Experimental and Clinical Medicine,
T. Campanella Cancer Center, Magna Graecia University
and Medical Oncology Unit, Catanzaro, Italy
AMELIA E. ARANEGA  Cardiovascular Development Group, Department of Experimental
Biology, University of Jaén, Jaen, Spain
DHRUBA K. BHATTACHARYYA  Tezpur University, Napaam, India
KAUMUDI M. BHAWE  Graduate Division of Biomedical Sciences (BMS),
Department of Neurosurgery and Brain Tumor Research Center,
University of California at San Francisco (UCSF), San Francisco, CA, USA
FERNANDO BONET  Cardiovascular Development Group, Department of Experimental
Biology, University of Jaén, Jaen, Spain
DORUK BOZDAĞ  Biomedical Informatics, The Ohio State University, Columbus, OH, USA
BARBARA CALABRESE  Department of Medical and Surgical Sciences, University
Magna Graecia of Catanzaro, Catanzaro, Italy
MARIO CANNATARO  Department of Medical and Surgical Sciences, University
Magna Graecia of Catanzaro, Catanzaro, Italy
€
UMIT V. ÇATALYÜREK  Biomedical Informatics, Department of Electrical
and Computer Engineering, The Ohio State University, Columbus, OH, USA
WEN-CHING CHAN  Kaohsiung Chang Gung Memorial Hospital, Kaohsiung, Taiwan,
People’s Republic of China; Institute of Biomedical Sciences, Academia Sinica, Taipei,
Taiwan, People’s Republic of China
LEI CHEN  Laboratory of Synthetic Microbiology, School of Chemical Engineering
and Technology, Tianjin University, Tianjin, People’s Republic of China; Key Laboratory
of Systems Bioengineering, Ministry of Education of China, Tianjin, People’s Republic
of China; Collaborative Innovation Center of Chemical Science and Engineering,
Tianjin, People’s Republic of China
FRANCESCA CRISTIANO  Bioinformatic Bioinformatics Laboratory, Department of Surgical
and Medical Sciences, University Magna Græcia of Catanzaro, Catanzaro, Italy
MEHMET DEVECI  Computer Science and Engineering, The Ohio State University,
Columbus, OH, USA
KEMAL EREN  Computer Science and Engineering, The Ohio State University, Columbus,
OH, USA
FRANCISCO J. ESTEBAN  System Biology Group, Department of Experimental Biology,
University of Jaén, Jaen, Spain
DIEGO FRANCO  Cardiovascular Development Group, Department of Experimental
Biology, University of Jaén, Jaen, Spain
ZHENXING FU  Department of Medicine, University of California, San Diego, San Diego,
CA, USA
ZHUOHUI GAN  Department of Bioengineering, University of California, San Diego,
La Jolla, CA, USA

ix
x Contributors

MICHAEL P. GANTIER  Department of Molecular and Translational Science, Monash

University, Clayton, VIC, Australia; Centre for Cancer Research, MIMR-PHI Institute
of Medical Research, Clayton, VIC, Australia
AGAPITO GIUSEPPE  Department of Surgical and Medical Sciences,
University of Catanzaro, Catanzaro, Italy
PIETRO HIRAM GUZZI  Department of Surgical and Medical Sciences,
University “Magna Graecia” of Catanzaro, Catanzaro, Italy
FRANCISCO HERNANDEZ-TORRES  Cardiovascular Development Group,
Department of Experimental Biology, University of Jaén, Jaen, Spain
JUGAL K. KALITA  University of Colorado, Colorado Springs, CO, USA
KAMER KAYA  Computer Science and Engineering, Sabancı University, Istanbul, Turkey
LILIANA LÓPEZ KLEINE  Departamento de Estadı́stica, Universidad Nacional de Colombia,
Bogotá, DC, Colombia
L. KOUMAKIS  Department of Production and Management Engineering, Technical
University of Crete, Chania, Greece; Foundation for Research and
Technology—Hellas (FORTH), Institute of Computer Science, Heraklion, Greece
ONUR KÜÇÜKTUNÇ  Computer Science and Engineering, The Ohio State University,
Columbus, OH, USA
WEN-CHANG LIN  Institute of Biomedical Sciences, Academia Sinica, Taipei, Taiwan,
People’s Republic of China
LILIANA LÓPEZ-KLEINE  Departamento de Estadı́stica, Universidad Nacional
de Colombia, Bogotá, DC, Colombia
ESTEFANIA LOZANO-VELASCO  Cardiovascular Development Group,
Department of Experimental Biology, University of Jaén, Jaen, Spain
MARIA TERESA DI MARTINO  Department of Experimental and Clinical Medicine,
T. Campanella Cancer Center, Magna Graecia University and Medical Oncology Unit,
Catanzaro, Italy
ANDREW D. MCCULLOCH  Department of Bioengineering, University of California,
San Diego, La Jolla, CA, USA
MARIANNA MILANO  Department of Surgical and Medical Sciences,
University of Catanzaro, Catanzaro, Italy
ROSA MONTAÑO  Departamento de Matemática y Ciencia de la Computación,
Universidad de Santiago de Chile, Santiago, Chile
V. MOUSTAKIS  Department of Production and Management Engineering,
Technical University of Crete, Chania, Greece
KRISTIAN OVASKA  Biomedicum Helsinki (B524a), University of Helsinki, Helsinki,
Finland
G. POTAMIAS  Foundation for Research and Technology—Hellas (FORTH),
Institute of Computer Science, Heraklion, Greece
FRANK L. POWELL  Department of Medicine, University of California, San Diego,
San Diego, CA, USA
SWARUP ROY  Department of Information Technology, North-Eastern Hill University,
Shillong, India
JENNIFER C. STOWE  Department of Bioengineering, University of California, San Diego,
La Jolla, CA, USA
PIEROSANDRO TAGLIAFERRI  Department of Experimental and Clinical Medicine,
T. Campanella Cancer Center, Magna Graecia University and Medical Oncology Unit,
Catanzaro, Italy
 Contributors xi

PIERFRANCESCO TASSONE  Department of Experimental and Clinical Medicine,

T. Campanella Cancer Center, Magna Graecia University and Medical Oncology Unit,
Catanzaro, Italy; Sbarro Institute for Cancer Research and Molecular Medicine,
Center for Biotechnology, College of Science and Technology, Temple University,
Philadelphia, PA, USA
FRANCISCO TORRES-AVILÉS  Departamento de Matemática y Ciencia de la Computación,
Universidad de Santiago de Chile, Santiago, Chile
M. TSIKNAKIS  Foundation for Research and Technology—Hellas (FORTH),
Institute of Computer Science, Heraklion, Greece; Department of Applied Informatics
and Multimedia, Technological Educational Institute, Heraklion, Greece
PIERANGELO VELTRI  Bioinformatic Bioinformatics Laboratory, Department of Surgical
and Medical Sciences, University Magna Græcia of Catanzaro, Catanzaro, Italy
JIANGXIN WANG  Laboratory of Synthetic Microbiology, School of Chemical Engineering
and Technology, Tianjin University, Tianjin, People’s Republic of China; Key Laboratory
of Systems Bioengineering, Ministry of Education of China, Tianjin, People’s Republic
of China; Collaborative Innovation Center of Chemical Science and Engineering,
Tianjin, People’s Republic of China
GANG WU  University of Maryland at Baltimore Country, Baltimore County, MD, USA
DI WU  Department of Statistics, Harvard University, Cambridge, MA, USA;
Centre for Cancer Research, MIMR-PHI Institute of Medical Research, Clayton, VIC,
Australia
M. ZERVAKIS  Department of Electronic and Computer Engineering, Technical University
of Crete, Chania, Greece
WEIWEN ZHANG  Laboratory of Synthetic Microbiology, School of Chemical Engineering
and Technology, Tianjin University, Tianjin, People’s Republic of China; Key Laboratory
of Systems Bioengineering, Ministry of Education of China, Tianjin, People’s Republic
of China; Collaborative Innovation Center of Chemical Science and Engineering,
Tianjin, People’s Republic of China
Methods in Molecular Biology (2016) 1375: 1–10
DOI 10.1007/7651_2015_239
© Springer Science+Business Media New York 2015
Published online: 14 May 2015

Normalization of Affymetrix miRNA Microarrays

for the Analysis of Cancer Samples
Di Wu and Michael P. Gantier

Abstract
microRNA (miRNA) microarray normalization is a critical step for the identification of truly differentially
expressed miRNAs. This is particularly important when dealing with cancer samples that have a global
miRNA decrease. In this chapter, we provide a simple step-by-step procedure that can be used to normalize
Affymetrix miRNA microarrays, relying on robust normal-exponential background correction with cyclic
loess normalization.

Keywords: microRNA, miRNA microarray, Normalization, Cancer samples, Affymetrix

1 Introduction

Variation in microRNA (miRNA) levels is a common feature of

cancer cells (1). It can result from mutations leading to increased
expression or chromosomal amplification of the miRNA gene—as
seen with the miR-17–92 cluster amplified in diffuse large B-cell
lymphoma patients (2)—or defective expression, processing, and
export of miRNA precursors (3–6).
Interestingly, early contradictions rapidly arose regarding the
overall profile of miRNA expression in cancer cells, with a number
of reports published that suggested a global decrease (7, 8), while
others observed an equal distribution of upregulated and down-
regulated miRNAs (9, 10). It is now well established that a signifi-
cant proportion of cancer cells exhibit alteration of the miRNA
biogenesis machinery (4–6, 11), resulting in a global miRNA
decrease and poorer survival outcomes (6, 12, 13).
This suggested a potential bias of miRNA microarray technol-
ogies that failed to identify global miRNA decreases (9, 10), and
prompted us to investigate the reliability of miRNA microarrays to
correctly identify samples with a global miRNA decrease. Profiling
of mouse embryonic fibroblasts following the induced genetic
deletion of Dicer1, the last processing enzyme in the miRNA
biogenesis pathway, allowed us to assess the suitability of Affymetrix
miRNA microarrays to detect global miRNA decrease (14).

1
2 Di Wu and Michael P. Gantier

Unexpectedly, we demonstrated that standard robust multichip

average (RMA) background correction and quantile normalization
of these miRNA microarrays, while aimed at decreasing the varia-
tions in log2 intensities between the replicate arrays, strongly biased
the identification of downregulated miRNAs (14). These observa-
tions underline the importance of array preprocessing in miRNA
microarray analyses. Critically, the previous lack of identification of
global miRNA decrease could have been, in fact, related to the
inappropriate use of normalization procedures, with the example
of median normalization assuming that few miRNAs are upregu-
lated or downregulated, thereby strongly biasing the possible
detection of a global decrease (9, 10).
In this chapter, we detail the step-by-step use of ‘R’ to apply
robust normal-exponential background correction with cyclic loess
normalization for the preprocessing of Affymetrix miRNA micro-
arrays, which was the best normalization procedure for detecting
global miRNA decreases in our mouse embryonic fibroblast model
and prostate cancer samples (14).

2 Materials

2.1 ‘R’ Software ‘R’ can be downloaded from http://cran.us.r-project.org. Once

and Bioconductor the most recent version for your operating system is installed on
your computer, start ‘R’ (see Note 1). To install the statistical
packages, required for the analyses described below, type in:
install.packages

To install bioconductor (while connected to the internet), type

in the following:
source("http://bioconductor.org/biocLite.R")
biocLite()

If prompted: ‘Update all/some/none? [a/s/n]:’, type

in ‘a’. These commands will download and install the statistical
packages required for the microarray analyses presented hereafter.

2.2 miRNA The command lines provided below are specifically designed for our
Affymetrix Microarray published dataset from Dicer-deficient cells, to be used as an exam-
(Version 1.0 or Later) ple of the overall normalization procedure. The nine .CEL files
(from GSM1118272_MG1.CEL to GSM1118280_MG9.CEL)
can be downloaded from Gene Expression Omnibus (GEO), acces-
sion number GSE45886. Briefly, miRNA levels were detected by
Affymetrix miRNA v1.0 microarray, at day 2, 3, and 4 after genetic
deletion of Dicer1. Each condition (t2, t3, and t4) was replicated in
biological triplicate (A, B, and C) (14). Our normalization proce-
dure relies on different weights being applied to different types of
probes present on the arrays. As such, the correct definition of the
 Affymetrix miRNA Microarray Normalization 3

non-miRNA small RNA probes is critical, and the microarray anno-

tation files should be downloaded from Affymetrix’s ‘Support’
section (use ‘miRNA 1.0 Annotations, Unsupported, CSV format’
for our case study). Importantly, our method has also been used
with more recent versions of Affymetrix miRNA arrays, which also
contain non-miRNA small RNA probes.

3 Methods

In this chapter, we present the microarray processing methods,

broken down into three major steps: background correction,
normalization, and summarization. Before proceeding to the first
step, however, the microarray files need to be loaded in ‘R’. This is
executed with the following:
library(limma)
library(affy)
library(MASS)

Importantly, the location of the .CEL files needs to be speci-

fied. In this example, the nine array files from GSE45886 have been
placed in the ‘/Documents’ directory.
setwd(’~/Documents/’)
affy2<-ReadAffy()
pm.raw<-pm(affy2,geneNames(affy2)) (see Note 2)

We can then proceed with the loading of the ‘design matrix’.

A design matrix defines how the microarrays are grouped in differ-
ent conditions/treatments. The design matrix relies on a .txt
‘target’ file, tabulated to identify the conditions of each array. In
our analysis of GSE45886, we use ‘targets-mirna.txt’ as the design
matrix. To create this file, we write the following in a blank text file:
Filename time dish
GSM1118272_MG1.CEL t2 A
GSM1118273_MG2.CEL t2 B
GSM1118274_MG3.CEL t2 C
GSM1118275_MG4.CEL t3 A
GSM1118276_MG5.CEL t3 B
GSM1118277_MG6.CEL t3 C
GSM1118278_MG7.CEL t4 A
GSM1118279_MG8.CEL t4 B
GSM1118280_MG9.CEL t4 C

This document is saved as a .txt file, named ‘targets-mirna.txt’

and placed in the same folder as the .CEL files, i.e., in the
~/Documents directory, before being loaded with the following
commands (see Note 3):
4 Di Wu and Michael P. Gantier

{
targets <- read.delim("targets-mirna.txt",stringsAs
Factors¼FALSE, sep¼" ")
}
des <- model.matrix(~0+as.factor(time),
data¼targets)

3.1 Robust Normexp For background correction, our procedure relies on normexp back-
Background ground correction using the ‘nec’ function in ‘R’. In addition, we
Correction use the ‘robust’ argument in ‘nec’ that determines background
mean and standard deviation, as we found it increased the sensitiv-
ity of the detection of differentially expressed miRNAs (14).
Nonetheless, robust can be disabled using ‘robust ¼ FALSE’ in
the command below.
Normexp background correction relies on the negative control
probes in the Affymetrix array—annotated as ‘BkGR’ in the man-
ufacturer’s annotation file. The following lines define which probes
are used as control probes, from the Affymetrix annotations.
bkgr.idx.pm<-grep("BkGr",rownames(pm.raw))
status<-rep("regular",nrow(pm.raw))
status[bkgr.idx.pm]<-"negative"
table(status)

This will print the amount of negative and regular probes in the
arrays (negative: 8221 and regular: 38006 when using GSE45886).
nec.pm.raw.r<-nec(pm(affy2),status¼status,negctrl¼
"negative",
regular¼"regular", offset¼16, robust¼TRUE)
summary(nec.pm.raw.r)

This will print the raw intensities for each microarray divided in:
Min./1st Qu./Median/Mean/3rd Qu./Max values.

3.2 Definition The first step is to obtain the probe annotations from the appropri-
of Non-miRNA Small ate annotation file from Affymetrix. The file should be placed in the
RNA Probes Used working directory—i.e., ‘/Documents’ in our case (see Note 4).
in Cyclic Loess ann<-read.csv("miRNA-1_0.annotations.20081203.
Normalization csv",skip¼11)
data.frame(table(ann$Sequence.Type))

This will print the features present on the arrays.

idx.probe<-indexProbes(affy2)
probe.name<-probeNames(affy2)
table(geneNames(affy2) %in% as.character(ann$Probe.Set.
ID))
identical(names(idx.probe),(geneNames(affy2)))
m<-match(names(idx.probe),as.character(ann$Probe.Set.
ID))
ann.m<-ann[m,]
 Affymetrix miRNA Microarray Normalization 5

ann.miRNA<- which(ann.m$Sequence.Type¼¼"miRNA")
mirna<-as.character(ann.m$Probe.Set.ID[ann.miRNA])
ann.affyctlseq<- which(ann.m$Sequence.Type¼¼"Affymetrix
Control Sequence")
affyctlseq<-as.character(ann.m$Probe.Set.ID[ann.
affyctlseq])
ann.spikein<- which(ann.m$Sequence.Type¼¼"Oligonucleo
tide spike-in controls")
spikein<-as.character(ann.m$Probe.Set.ID[ann.spikein])
ann.rrna<- which(ann.m$Sequence.Type¼¼"5.8 s rRNA")
rrna<-as.character(ann.m$Probe.Set.ID[ann.rrna])
ann.cdbox<- which(ann.m$Sequence.Type¼¼"CDBox")
cdbox<-as.character(ann.m$Probe.Set.ID[ann.cdbox])
ann.hacabox<- which(ann.m$Sequence.Type¼¼"HAcaBox")
hacabox<-as.character(ann.m$Probe.Set.ID[ann.hacabox])
ann.scarna<- which(ann.m$Sequence.Type¼¼"scaRna")
scarna<-as.character(ann.m$Probe.Set.ID[ann.scarna])
ann.snorna<- which(ann.m$Sequence.Type¼¼"snoRNA")
snorna<-as.character(ann.m$Probe.Set.ID[ann.snorna])
idx.pm.mirna<-which(match(probe.name,mirna)!¼"NA")
length(idx.pm.mirna)

The last command will print the amount of miRNA probes on

the array—this is 26,812 for miRNA.1_0.
identical(unique(probe.name[idx.pm.mirna]),mirna)
o.sml<-c(cdbox,hacabox,scarna,snorna)
idx.pm.sml<-which(match(probe.name,o.sml)!¼"NA")
length(idx.pm.sml)

This will print the amount of non-miRNA ‘other small RNA’

probes on the array—this is 10,090 for miRNA.1_0.
identical(sort(unique(probe.name[idx.pm.sml])),sort(o.
sml))
idx.pm.spk<-which(match(probe.name,spikein)!¼"NA")
identical(unique(probe.name[idx.pm.spk]),spikein)
idx.pm.rrna<-which(match(probe.name,rrna)!¼"NA")
identical(unique(probe.name[idx.pm.rrna]),rrna)
idx.pm.ctls<-which(match(probe.name,
affyctlseq)!¼"NA")
identical(unique(probe.name[idx.pm.ctls]),affyctlseq)
idx.pm.ctls.hyb<-idx.pm.ctls[-grep("BkGr",probe.name
[idx.pm.ctls])]
status.spot<-rep("NA",nrow(pm.raw))
status.spot[idx.pm.mirna]<-"miRNA"
status.spot[idx.pm.sml]<-"other.small.RNA"
status.spot[bkgr.idx.pm]<-"BkGr.ctl"
status.spot[idx.pm.ctls.hyb]<-"hyb.ctl"
status.spot[idx.pm.spk]<-"spike.in"
status.spot[idx.pm.rrna]<-"human.5.8s.rRNA"
table(status.spot)
6 Di Wu and Michael P. Gantier

This will print the different categories of probes now defined—

BkGr.ctl: 8221; human.5.8s.rRNA: 110; hyb.ctl: 774; miRNA:
26,812; other.small.RNA: 10,090; and spike.in: 220, for
miRNA_1.0.

3.3 Cyclic Loess The next step is cyclic loess normalization—which attributes
Normalization heavier weight to non-miRNA small RNA probes than miRNA
probes defined in the previous step to normalize the differences
between arrays. By using a much higher weight for non-miRNA
small RNA probes (100 vs. 0.01 for miRNAs), we found that we
greatly increased the accuracy of the normalization (14).
affy2.temp<-affy2
pm(affy2.temp)<-nec.pm.raw.r
w<-rep(1,nrow(pm(affy2.temp)))
w[status.spot¼¼"miRNA"]<- 0.001
w[status.spot¼¼"other.small.RNA"]<-100
norm3<- normalizeCyclicLoess(log2(pm(affy2.temp)),
weights¼w,
iteration¼5) (see Note 5)
pm(affy2.temp)<-2^(norm3)

3.4 RMA The last step of our procedure is RMA summarization—which

Summarization summarizes the previous normalization analyses in a data matrix
(‘exprs2’ in this case).
tmp2<-rma(affy2.temp,normalize¼FALSE,
background¼FALSE)
exprs2<-exprs(tmp2)
summary(exprs2)

This will print the quartile intensities for each normalized

microarray: Min./1st Qu./Median/Mean/3rd Qu./Max values.
Because human cancer samples are very heterogeneous, it is
advisable to introduce different estimated array weights in the
analysis of differentially expressed miRNAs. We have found that
the use of array weights gives a higher number of significantly
downregulated miRNAs in Dicer1-deficient samples than the pro-
cedure without array weights—consistent with a global impairment
of miRNA biogenesis (14). Therefore, we generally suggest the use
of array weights when analyzing microarrays from tumor samples.
Importantly, array weights are restricted to the miRNA probes of
the species of interest—mouse or ‘mmu’ in our Dicer1-deficient
samples. The ‘mmu’ should be changed to ‘hsa’ when looking at
human samples in the following command lines (see Note 6).
mmu.idx<-grep("mmu",rownames(exprs2))
w.des<-arrayWeightsSimple(exprs2[mmu.idx,],design¼des)
names(w.des)<-colnames(exprs2)
 Affymetrix miRNA Microarray Normalization 7

To compare the samples on the basis of a given variable, for

example the ‘time’ after Dicer1 deletion in our case study, in a linear
model, we define the ‘contrast’ in the variable in which we are
interested. Refer to the ‘limma User Guide’ for more details on
how to define the contrast (see Note 7).
c.matrix<-cbind(T3vs2¼c(-1,1,0),T4vs2¼c(-1,0,1),
T4vs3¼c(0,-1,1))

The linear model is subsequently fitted with the array weights

determined previously.
fit.w<-lmFit(exprs2,design¼des, weights¼w.des)
fit.w<-contrasts.fit(fit.w,c.matrix)
fit.w<-eBayes(fit.w)
summary(decideTests(fit.w[mmu.idx,],p.value¼0.1))

This will print the number of miRNAs that are downregulated

(1), unchanged (0), or upregulated (1) in the different conditions
of the experiment—in our case comparing T3vs2, T4vs2, and
T4vs3 as follows, with a p value of 0.1. In our example, the follow-
ing will be printed in ‘R’ (see Note 8):
T3vs2 T4vs2 T4vs3
-1 32 87 12
0 575 516 596
1 2 6 1

Finally, a table of differentially expressed miRNAs can be

retrieved with the following lines. Note that ‘top1’ corresponds
to differentially expressed probes (from mouse here as specified by
‘mmu’) between T3vs2—i.e., in the first column printed previously.
‘top2’ and ‘top3’ match the second and third columns, respectively.
The p value can also be changed—here set to p < 0.1.
top1<- topTable(fit.w[mmu.idx,],coef¼1,number¼Inf,p.
value¼0.1)
top2<- topTable(fit.w[mmu.idx,],coef¼2,number¼Inf,p.
value¼0.1)
top3<- topTable(fit.w[mmu.idx,],coef¼3,number¼Inf,p.
value¼0.1)
write.table(top1, file¼"topTab1.csv", row.names¼TRUE,
sep¼",")
write.table(top2, file¼"topTab2.csv", row.names¼TRUE,
sep¼",")
write.table(top3, file¼"topTab3.csv", row.names¼TRUE,
sep¼",")

Files with the indicated names will appear in the working

directory—‘/Documents’ in our case—containing the lists of
miRNAs differentially expressed, with normalized log2 fold change.
8 Di Wu and Michael P. Gantier

4 Notes

1. In this analysis we rely on ‘R’ version 3.1.0 (2014-04-10),

‘Spring Dance’. ‘R’ relies on command lines, which you need
to type after the ‘>’ symbol. Importantly, several lines of
commands can be copied and pasted at the same time in ‘R’,
and successively executed by pressing ‘enter/return’. When
doing so, care should be taken with quotes (‘’ and “”), which
can be modified by your operating system and alter the mean-
ing of the ‘R’ command—generally resulting in an error
message.
2. The last command might result in warning messages such as:
‘replacing previous import by ‘utils::head’ when loading ‘mir-
na10cdf” This indicates that the same names were included in
the different packages loaded. However, this can be ignored:
warnings in ‘R’ can usually be ignored without impacting on the
processing of the data.
3. The variable studied in our example is identified by the “time”
column from our targets-mirna.txt file, while the “dish” col-
umn refers to replicates. When creating another design matrix,
the previous command should be altered to reflect the variable
in the ‘as.factor(variable)’ expression.
4. Because the files for each version of miRNA arrays are slightly
different, the argument ‘skip’ has to be changed as follows:
skip ¼ 11 for ‘miRNA-1_0.annotations.20081203.csv’;
skip ¼ 13 for ‘miRNA-2_0.annotations.20101222.csv’; skip ¼
4 for ‘miRNA-3_0-st-v1.annotations.20140513.csv’ and
‘miRNA-4_0-st-v1.annotations.20140513.csv’.
5. This step will take about a minute to run, depending on your
processor, due to the five iterations.
6. The Affymetrix miRNA arrays contain many other species in
addition to human and mouse. You can check the nomencla-
ture for each species (for instance, ‘mmu’ for mouse, ‘has’ for
human, ‘gga’ for chicken, ‘eca’ for horse) at miRbase.org.
7. The following section will detail how to define the ‘design
matrix’ and ‘contrast’ of a variable when dealing with only
two groups of samples, which is particularly useful when com-
paring normal and tumor samples. For this purpose, we remove
the files GSM1118275_MG4.CEL, GSM1118276_MG5.
CEL, and GSM1118277_MG6.CEL from the working folder
(/Documents). In addition, we modify the targets-mirna.txt
file by deleting the lines corresponding to time 3 (t3). As such,
we will now detail how to compare samples with decreased
miRNA levels (t4) versus more normal samples (t2), mimicking
 Affymetrix miRNA Microarray Normalization 9

tumor versus normal samples. We make a design matrix that

contains the contrast data as follows:
a<-c("t2","t2","t2","t4","t4","t4")
designMatrix<-model.matrix(~0+as.factor(a))
colnames(designMatrix)
colnames(designMatrix)<-c("t2","t4")
contrast.matrix<- makeContrasts(t4-t2, levels¼
designMatrix)
contrast.matrix

This will print the contrasts (i.e., 1 for level t2 and 1 for level t4).
fit.w<-lmFit(exprs2,design¼designMatrix, weights¼
w.des)
fit.w<-contrasts.fit(fit.w, contrast.matrix)
fit.w<-eBayes(fit.w)
summary(decideTests(fit.w[mmu.idx,],p.value¼0.1))

This will print the following results for p < 0.1 (where 1
defines the number of probes downregulated at t4 versus t2;
0 defines the number of unchanged probes; +1 defines the
number of upregulated probes). Noteworthy, these differ
slightly from what is obtained with the analyses of the nine
microarrays due to statistical variations with fewer arrays.
t4 - t2
-1 68
0 538
1 3
Finally, the miRNAs that are significantly different at the two
time points can be retrieved with the following commands:
top1<- topTable(fit.w[mmu.idx,],coef¼1,number¼
Inf,p.value¼0.1)
write.table(top1, file¼"topTab1.csv", row.names¼
TRUE, sep¼",")

8. Please note that the values stated might change slightly with
the different releases of the statistical packages used.

Acknowledgments

The authors thank Frances Cribbin for her help with the redaction
of this review. The authors are supported by funding from the
Australian NHMRC (1022144 and 1062683 to MPG and
1036541 to DW) and the Victorian Government’s Operational
Infrastructure Support Program.
10 Di Wu and Michael P. Gantier

References
1. Melo SA, Esteller M (2011) Dysregulation of expression profiles classify human cancers.
microRNAs in cancer: playing with fire. FEBS Nature 435(7043):834–838
Lett 585(13):2087–2099 8. Gaur A, Jewell DA, Liang Y, Ridzon D, Moore
2. Ota A, Tagawa H, Karnan S, Tsuzuki S, Karpas JH, Chen C, Ambros VR, Israel MA (2007)
A, Kira S, Yoshida Y, Seto M (2004) Identifica- Characterization of microRNA expression
tion and characterization of a novel gene, levels and their biological correlates in human
C13orf25, as a target for 13q31-q32 amplifi- cancer cell lines. Cancer Res 67(6):2456–2468
cation in malignant lymphoma. Cancer Res 64 9. Volinia S, Calin GA, Liu C-G, Ambs S, Cim-
(9):3087–3095 mino A, Petrocca F, Visone R, Iorio M, Roldo
3. Calin GA, Dumitru CD, Shimizu M, Bichi R, C, Ferracin M, Prueitt RL, Yanaihara N, Lanza
Zupo S, Noch E, Aldler H, Rattan S, Keating G, Scarpa A, Vecchione A, Negrini M, Harris
M, Rai K, Rassenti L, Kipps T, Negrini M, CC, Croce CM (2006) A microRNA expres-
Bullrich F, Croce CM (2002) Frequent dele- sion signature of human solid tumors defines
tions and down-regulation of micro-RNA cancer gene targets. Proc Natl Acad Sci U S A
genes miR15 and miR16 at 13q14 in chronic 103(7):2257–2261
lymphocytic leukemia. Proc Natl Acad Sci U S 10. Yanaihara N, Caplen N, Bowman E, Seike M,
A 99(24):15524–15529 Kumamoto K, Yi M, Stephens RM, Okamoto
4. Melo SA, Moutinho C, Ropero S, Calin GA, A, Yokota J, Tanaka T, Calin GA, Liu C-G,
Rossi S, Spizzo R, Fernandez AF, Davalos V, Croce CM, Harris CC (2006) Unique
Villanueva A, Montoya G, Yamamoto H, microRNA molecular profiles in lung cancer
Schwartz S, Esteller M (2010) A genetic defect diagnosis and prognosis. Cancer Cell 9(3):
in exportin-5 traps precursor microRNAs in 189–198
the nucleus of cancer cells. Cancer Cell 18 11. Kumar MS, Pester RE, Chen CY, Lane K, Chin
(4):303–315 C, Lu J, Kirsch DG, Golub TR, Jacks T (2009)
5. Melo SA, Ropero S, Moutinho C, Aaltonen Dicer1 functions as a haploinsufficient tumor
LA, Yamamoto H, Calin GA, Rossi S, Fernan- suppressor. Genes Dev 23(23):2700–2704
dez AF, Carneiro F, Oliveira C, Ferreira B, Liu 12. Karube Y, Tanaka H, Osada H, Tomida S,
C-G, Villanueva A, Capella G, Schwartz S, Tatematsu Y, Yanagisawa K, Yatabe Y, Takami-
Shiekhattar R, Esteller M (2009) A TARBP2 zawa J, Miyoshi S, Mitsudomi T, Takahashi T
mutation in human cancer impairs microRNA (2005) Reduced expression of Dicer associated
processing and DICER1 function. Nat Genet with poor prognosis in lung cancer patients.
41(3):365–370 Cancer Sci 96(2):111–115
6. Merritt WM, Lin YG, Han LY, Kamat AA, 13. Grelier G, Voirin N, Ay A-S, Cox DG, Cha-
Spannuth WA, Schmandt R, Urbauer D, Pen- baud S, Treilleux I, Leon-Goddard S, Rimokh
nacchio LA, Cheng J-F, Nick AM, Deavers R, Mikaelian I, Venoux C, Puisieux A, Lasset C,
MT, Mourad-Zeidan A, Wang H, Mueller P, Moyret-Lalle C (2009) Prognostic value of
Lenburg ME, Gray JW, Mok S, Birrer MJ, Dicer expression in human breast cancers and
Lopez-Berestein G, Coleman RL, Bar-Eli M, association with the mesenchymal phenotype.
Sood AK (2008) Dicer, Drosha, and outcomes Br J Cancer 101(4):673–683
in patients with ovarian cancer. N Engl J Med 14. Wu D, Hu Y, Tong S, Williams BR, Smyth GK,
359(25):2641–2650 Gantier MP (2013) The use of miRNA
7. Lu J, Getz G, Miska EA, Alvarez-Saavedra E, microarrays for the analysis of cancer samples
Lamb J, Peck D, Sweet-Cordero A, Ebert BL, with global miRNA decrease. RNA 19(7):
Mak RH, Ferrando AA, Downing JR, Jacks T, 876–888
Horvitz HR, Golub TR (2005) MicroRNA
Methods in Molecular Biology (2016) 1375: 11–23
DOI 10.1007/7651_2015_238
© Springer Science+Business Media New York 2015
Published online: 12 June 2015

Methods and Techniques for miRNA Data Analysis

Francesca Cristiano and Pierangelo Veltri

Abstract
Genomic data analysis consists of techniques to analyze and extract information from genes. In particular,
genome sequencing technologies allow to characterize genomic profiles and identify biomarkers and
mutations that can be relevant for diagnosis and designing of clinical therapies. Studies often regard
identification of genes related to inherited disorders, but recently mutations and phenotypes are considered
both in diseases studies and drug designing as well as for biomarkers identification for early detection.
Gene mutations are studied by comparing fold changes in a redundancy version of numeric and string
representation of analyzed genes starting from macromolecules. This consists of studying often thousands
of repetitions of gene representation and signatures identified by biological available instruments that
starting from biological samples generate arrays of data representing nucleotides sequences representing
known genes in an often not well-known sequence.
High-performance platforms and optimized algorithms are required to manipulate gigabytes of raw data
that are generated by the so far mentioned biological instruments, such as NGS (standing for Next-
Generation Sequencing) as well as for microarray. Also, data analysis requires the use of several tools and
databases that store gene targets as well as gene ontologies and gene–disease association.
In this chapter we present an overview of available software platforms for genomic data analysis, as well as
available databases with their query engines.

Keywords: Next-generation sequencing, Bioinformatics, microRNA, Gene target, Databases,

Ontologies

1 Introduction

The analysis of biological data is increasing the interests of clinicians

and health operators, due to the possibility of gathering informa-
tion about patient treatments from genetic-based analysis. The
increasing reliability and efficiency of biological sample analysis
and information extraction from them has resulted in the availabil-
ity of clinically interesting information to health operators. For
instance, drug reaction as well as protein expression in blood sam-
ples or gene expression analysis to overcome the gene target pres-
ence has captured the interests of health operators that may move
from a study and research target use of genomic and proteomic
analysis to a patient-bed oriented application. In the first case,
research allows to study genes and their expressions in in vivo

11
12 Francesca Cristiano and Pierangelo Veltri

(as well as in vitro) biological sample, in an off-line way, i.e., in a not

well-defined time interval. In the second case, when a patient needs
to receive treatment, genomic (as well as proteomic) data analysis
has to produce results (and thus information) useful for defining
treatments, in a limited (and often short) time interval.
Today, the availability of efficient computational platforms
allows to guarantee the production of reliable and well-defined
information extracted from genomic analysis in a time interval
that is reasonable with respect to the patient treatment. This has
always led to more and more frequent interest in genomic technol-
ogies and analysis also in clinical studies and applications. Obviously
the main interests is related to study of macromolecules activities
and biological studies to identify biomarkers related to chronic and
severe diseases.

2 Microarray Data Analysis

Biological analysis of blood and tissue samples generates a huge

volume of data that requires high-performance analysis techniques
both in terms of hardware architecture and optimized software.
Therefore, it is commonly recognized that both characterizing
biological samples and identifying macromolecules in biological
samples are main tasks for biomarker identifications. Such techni-
ques require software tools and storage techniques to extract inter-
esting information from a huge amount of data. Also, on line
available databases have to be queried to retrieve available and/or
previously published results, related to the analyzed biological
samples. The main techniques that are used with the aim of analyz-
ing the expression profile of a tissue or organism are the RT-PCR,
microarray and next-generation sequencing (1). Microarray analy-
sis technique is used to gather information and to understand raw
data generated from experiments on DNA, RNA, and proteins. The
technique is based on use of microarray devices to study genes
starting from samples. Each microarray is a 2D solid array where
large amounts of biological samples can be positioned. By using
detection methods biological contents are associated to raw data
(2). Often microarrays are used in order to explore the differences
in expression between tissues or organisms, as well as between a
healthy control and treated, or to characterize a given disease and
discover new mechanisms of regulation (3). These large data
amount can be difficult to analyze, especially in case of lack of
gene annotation. Depending on the type of application and on
the biological sample, microarrays are formed by a support that
consists of thousands of spots, each containing the molecules of the
probe. In a microarray experiment for the analysis of gene expres-
sion, the starting sample is RNA, and the output must then neces-
sarily be normalized and analyzed statistically in order to obtain a
 Methods and Techniques for miRNA Data Analysis 13

list of miRNAs or more in general of genes and the associated

expression values. The analyzed data can then be stored in suitable
format which enables interoperability and exchange of data, the
MIAME (Minimum Information About a Microarray Experiment),
a standard that allows you to describe properly a microarray experi-
ment. Minimum information about a microarray experiment means
the accurate description of the following points:
l Experiment design.
l Array design.
l Samples.
l Hybridization, procedures and parameters.
l Measurement (as the images produced by scanner).
l Normalization.
Each of these sections has to be compiled using a vocabulary
already structured, and adding notes and comments in the free text (4).

3 NGS Data Analysis

NGS technique produces a huge amount of data (e.g., mRNA) that

requires bioinformatic analysis tools to extract useful information
from experiments (both in vivo and in vitro), as well as to predict
functionalities. NGS related research shows how computer scien-
tists have been studying possible solutions to support both infor-
mation extraction and result representation, to provide available
and useful information to clinicians and biologists, each with their
own interests. In particular, tools to simplify result reading to
biologists have been designed.
NGS technology allows to extract information from samples in
a faster way, producing a large amount of data. It is currently used
also to analyze RNA or small fragments of RNA, say microRNAs or
miRNAs. Large amounts of data need to be analyzed with different
tools and platforms. miRNAs are small fragments of RNA com-
posed of 21–23 nucleotides and are involved in many biological
processes. The interest of bioinformaticians for these molecules is
related to their potential function as biomarkers for many diseases
(5). miRNA-seq analysis requires the use of several tools, and there
exist many databases for storing prediction gene targets and
gene–disease associations. They are responsible of inhibiting the
mRNA (RNA messenger) functions, and thus for instance protein
production.
The NGS technologies are used to sequence in parallel DNA or
RNA samples allowing to obtain the number of counts of genes
found in the sample. NGS platforms are for example Roche-454 (6)
Illumina-Solexa (7), and SOLiD—Applied Biosystems (8). Each of
14 Francesca Cristiano and Pierangelo Veltri

them use the methods for sequencing the samples on the basis of
the length or type of sequence (paired end, single ended, etc.)
Nowadays the next-generation sequencing produces a large
amount of data and information difficult to manage and therefore
requires the use of efficient and high-performance tools in order to
conduct an analysis in a very short time (9). The output of the
sequencing is in FastQ format, and each file can reach an average
size of almost 1 GB, producing more than one FastQ file. Many ad
hoc pipelines are developed by software engineers to analyze the
produced data, but the process of installing, configurating, and
managing the software requires computer skill that users (often
doctors and biologists) usually do not have.

4 miRNA-seq Analysis in NGS

miRNA-seq data output are mainly used to quantify miRNAs

abundance levels and their expression values in the samples. Gener-
ally, raw data from sequencing platforms generate fastQ files. The
first step is to evaluate the goodness of the generated files. It is a
textual file that contains several read sequences and for each read
there are four lines that indicate sequence ID beginning with @ and
gives information about instrument, flow cell line, barcode, and
sequence type, i.e., paired end or single ended, second lines is the
read sequence, and then, there are a plus sign and quality of the read
known as Phred score.
@HWI-1KL111:71:C3UBGACXX:1:1101:10246:2477 1:N:0:CA
GATCACGTTCCCGTGGTGGAATTCTCGGGTGCCAAGG
AACTCCAGTCACCAGA + CCCFFFFFHHHHFHJJJJJJJJJJJ?
FGIJJIIIJJIIJJJJJJJJJJJ
MiRNA-seq analysis consists of some steps that can be per-
formed by available tools such as Galaxy, miRDeep, and
StrandNGS.
Generally these steps consist of:
l Viewing the quality plots using FastQC software (10). FastQC
checks the quality on raw sequences after the sequencing and
provide to correct them if there are some errors. It is possible to
generate an html report with graphs and tables related for
example to basic statistics, per base sequence quality or quality
scores, duplicate sequences and overrepresented sequence, and
adapter content. FastQC allows to obtain information in order
to improve the read sequences in the preprocessing.
l Preprocessing of raw data: before aligning the reads to the refer-
ence genome is necessary to improve the quality of the sequences
by using preprocessing. This step includes the removal of Adapter
sequence and the low quality reads (the tools usually have a list of
 Methods and Techniques for miRNA Data Analysis 15

common adapters). There exist many algorithms that perform

removing and trimming of some insignificant reads, for example
Cutadapt (11) or TrimGalore (12). Cutadapt removes adapter
sequences from raw data and reduces the sequences if they are too
long. In fact next-generation sequencing produces reads with a
rate of 50 up to 100 bps (base pairs) and smallRNA that are
shorter than this length. TrimGalore removes adapter sequence
and uses several Illumina standard adapters to adapt trimming
and then it is possible to use FastQC to recheck the quality of the
reads.
l Collapsing: Identical reads are collapsed into one read and their
values of frequency is considered for the next steps.
l Aligning and statistical/bioinformatics analysis.

5 Software Platform Analysis

NGS technique has been introduced and allows to generate data

obtained from DNA, RNA, and small-RNA samples, similarly to
microarray but allowing to generate multiple copies of the same
genes and to perform the analysis in fast time. Such new technique
is attracting lots of interests thanks to the fact that it is able to
generate many results from sample analysis. Nevertheless, there is
lots of works for analyzing processed data. The information extrac-
tion requires the support of bioinformaticians due to the difficulty
to automatize the analysis process. For instance, installing and
using an open source tool such as Galaxy (13) requires many
manual steps that cannot be performed by biologists that should
be supported by informatics experts. For NGS data analysis, soft-
ware such as Galaxy (13), Strand NGS (formerly Avadis NGS) (14),
GeneSpring (15), and miRDeep (16) can be used.

5.1 Galaxy The large amount of data that is produced with the next-generation
sequencing requires that data be stored and managed in an efficient
manner.
Galaxy (13, 17, 18) is an open and Web-based workbench that
enables users to perform statistical and bioinformatic analysis on NGS
data. Galaxy platform can be downloaded and installed locally, and
there are many tools that can be integrated as plugins.
Galaxy is a tool that is used mostly by researchers who have not
computer science skills. It provides a simple Web interface and
plugins that can be used in order to make an analysis. In particular,
the available modules to perform the analysis can be used in
sequence. However, it is possible to install a local version of Galaxy
and the various available plugins manually. MiRNA-seq for exam-
ple, can be analyzed following a simple workflow (19). It is neces-
sary to import the sequenced files in Galaxy and view the reads
16 Francesca Cristiano and Pierangelo Veltri

present in them, in order to detect the possible presence of

contaminants. The reads can be cleared through the various tools
available under NGS TOOLBOX and NGS:QC and Manipulation.
Using the Barcode Splitter (20) the barcode can be split from the
reads, where the barcode is an A/C/G/T/ sequence. Subse-
quently to assess the quality of the sequences, FastQC:Read QC
(10) might be used. The tool performs a check on raw reads, in
particular allows to import data in various formats such as SAM,
BAM, or FastQ, and provides detailed reports that allow the user to
view and correct manually results, providing also a series of useful
reports. Moreover in the NGS: QC and Manipulation module,
there are several tools that allows to: show other statistical reports
(as a result of importing fastQ files); clean sequences (such as
adapter removal), trim sequences; eliminate artifacts; filter
sequence on the quality of the reads; convert formats (i.e., from
FastQ to fasta or from BAM to SAM). The next step of analysis
consists in aligning sequences to the reference genome. This can be
made by importing or selecting the genome of interest among
those present. Bowtie is used for the alignment of small size
sequences also. Among the tools available in Galaxy, it is possible
to use miRDeep2 (21) for discovering miRNA sequences using
miRBase and helps identify novel miRNAs. miRDeep2 is a pipeline
that performs NGS data analysis and can be used to align sequences
and miRNA expression profiling.
For other type of sequences, i.e., RNA, the alignment can be
made using TopHat that performs the alignment of the sequences
to the reference genome (by using Bowtie) and the reads are
subsequently analyzed with the aim to detect splice junctions (22,
23). The TopHat output is a BAM file that must be appropriately
converted to other formats for the next steps, i.e., SAM. Last steps
are related to the count of the mapped reads using Cufflinks (24)
and for differential gene expression analysis; Galaxy offers tools
such as DESeq (25).

5.2 Strand NGS Strand NGS (14) is a commercial software that can be used to
(Formerly Avadis NGS) perform NGS analysis on DNA, RNA, or small RNA. This suite
allows to create two type of experiments including alignment and
statistical and bioinformatics analysis one. A smallRNA alignment
consists in importing the dataset (FastQ file in the tool) related the
sequencing experiment, define the appropriate reference genome,
i.e., mouse, human, and select from the entries, the library type and
the platform used during the sequencing. Before performing the
alignment, the program requires a preprocessing phase (pre-
alignment) to allow the increasing in the number of sequence that
has to be aligned with the considered genome. Even in Strand
NGS, you can view the report on the quality of the produced
sequences. If the reads present an adapter, a trimming set para-
meters is necessary to trim adapter and poor sequences. There is
 Methods and Techniques for miRNA Data Analysis 17

also the possibility to insert a number of bases to trim from

beginning to end of sequence. Usually, it is important also to create
a screening database with the aim of deleting contaminants. When
the alignment ends, you can see the results as alignment statistics
and report that contain information about total number of reads,
aligned and unaligned reads, read type, and read distribution, i.e.,
their position on chromosome and finally create an analysis experi-
ment. To identify miRNAs within the sequences previously
mapped, small RNA annotations must be defined and downloaded
to select the used genome (the same as the previous step). Then it is
necessary to filter reads among the small RNA regions, those of
interest (e.g., microRNAs). The navigation menu provides to the
quantification step that allows to count reads and to discover novel
miRNAs. After the quantification, the counts can be filtered by
their signal intensity values. To perform a differential expression
analysis, the samples that have biological or technical replicates can
be grouped. The interpretation allows users to group samples that
can be under the same experimental conditions. Subsequently, the
fold change analysis can be performed by selecting two ways to
perform the analysis, i.e., all conditions against a single condition.
For miRNA analysis, additional options are related to the target
gene search through the prediction database, by selecting as a input
miRNAs that have significant values of fold change. Last point of
this analysis can be the annotation of the genes with Gene Onto-
logy (26).
Creating an RNA-seq analysis experiment using Strand NGS is
not complicated; indeed, it is not necessary to know all the required
parameters, and it is possible to perform a standard analysis leaving
the default values. In the quantification step there are three choices
for the normalization algorithm (RPKM, DESeq, and TMM) and
the suite allows to show the count of raw data and the normaliza-
tion values used for calculating the fold change.

6 Gene Expression Data Analysis

The starting point of the analysis of gene expression data is repre-

sented by a numerical matrix. The matrix generally consists of a
number of rows representing genes and a number of columns
representing the different experimental conditions that can be
for example time intervals in the case of experiments related to
drug releases, or the comparisons between two samples, as treated
(sick subject) and control (healthy subjects). To biologically ana-
lyze the numerical values, the matrix content can be converted
into different formats, for example in a graphically and more
representative form, as the one defined as heat map. A heat map
is an image that is used to represent the analysis of fold change; the
fold change is a parameter that allows to define the genes within
18 Francesca Cristiano and Pierangelo Veltri

the expression matrix, differentially expressed, and thus allows to

identify upregulated and downregulated genes. The genes coex-
pressed are instead identified by cluster analysis. A cluster is a set of
objects with similar features. A cluster of genes, therefore, is
developed on the basis of the principle of distance metrics,
which allows to group genes that are neighbors, from the
biological point of view, among them. A refinement of the clus-
tering technique is the biclustering, which identifies the genes
belonging to a bicluster and exclude those that do not belong to
any bicluster, such as noise. In particular a bicluster is created by
selecting from the rows of the data matrix, genes that show a
similar behavior only within a subset of conditions, while cluster
analysis requires that genes belong to all conditions (27). This
analysis could lead, for example, to the identification of novel
biological samples or the discovery of new gene functions. Ana-
lyzing the data obtained as a result of an experiment can some-
times be a fairly complex task for bioinformatics. In reference to
the miRNA data, few Web available software are able to carry out a
comprehensive and efficient analysis. This is due in part to the
recent discovery of miRNA molecules and in part to the lack of
standards for the adjustment of the phases of analysis.
More specifically bioinformatics analysis consists of several
steps:
l Identification of miRNAs and mRNAs differentially expressed.
l Search of target genes by prediction database.
l Identification of miRNA–mRNA relations extracted from
experiments.
l Enrichment analysis of genes by using ontological database.
l Development of miRNA–mRNA networks representing (the
more) relevant relations.

7 Databases and Genome Query Languages

Bioinformatics provides the researcher with software tools and

biological databases to analyse a huge quantity of data in a very
short period of time, e.g., the recent sequencing techniques (NGS)
or nucleic acid sequence or protein search tools, possibly accom-
panied with information on available results. Indeed, data set
obtained by performing experimental analysis are stored and pub-
lished in huge data volumes in different databases (consider for
instance data obtained while sequencing the human genome). The
main bioinformatic database for biologists and researchers is
BLAST that allows to align locally genes and proteins against
those present in the NCBI system to identify similar sequences
(28). Similarly, ENTREZ can be considered a search engine that
 Methods and Techniques for miRNA Data Analysis 19

maintains biological and biomedical information (29). PubMed

allows to search articles and magazines of interest (30), while
EMBL (European Bioinformatics Institute) is a powerful source
of tools, tutorials, and different services offered to researchers,
created mainly with the aim of guiding the studies and contributing
to the advancement of research (31). There exist several databases
hosting the biological relation among miRNAs and the
corresponding set of mRNAs (i.e., their targets) that can be inhib-
ited or interested by miRNA function. For instance miRBase (32,
33) stores known miRNAs from human tissues as well as animals or
plants, as well as the correlation with mRNA targets. When biolo-
gists obtain miRNAs from tissues, they use such databases to
extract information on functionalities to (eventually) correlate
with molecular function and diseases. Similarly pharmacologists
are using miRNAs to design or test new drugs.

7.1 miRNA–mRNA The interest of studying miRNAs and their role with respect to
Associations chronic diseases has been recently shown (e.g., in refs. (34) and
(35) for chronic diseases) as well as in representing new target for
different therapies and drugs. miRNA functions are related to
(subset of) genes that can be regulated by them. There are many
tools available online that, given a set of miRNAs, are in charge of
searching gene targets as well as proteins involved and that are able
to predict the mRNAs target of miRNAs. There exist different
miRNA–gene target associations databases, for example: miRDIP
(36) is a database for miRNA and mRNA that integrates a large
number of prediction tools results; such results are obtained by
using different prediction tools such as DIANA microT (37),
MicroCosm Target (formerly miRBase) (32), microRNA.org
(38), PicTar (39), and TargetScan (40). mirDIP gets as input a
list of miRNAs and returns miRNA–mRNA interactions on the
basis of the accuracy level. It is possible to select both database
and prediction accuracy (i.e., high, medium, and low accuracy). It is
possible to select automatically the prediction database by tuning
accuracy or prediction parameters. The results can be stored in a file
and contain miRNAs with associated mRNAs and the database used
for the prediction with rispective accuracy measure. Another exam-
ple of miRNA target database is miRDB (41) that contains several
genes from different organisms. miRBD uses machine learning
algorithms to predict the gene targets of miRNAs; a query by
example interface can be used to compose a query starting from
single or multiple miRNAs and linking them to gene targets.
Finally, miRWalk (42) is a tool that allows to select predicted
genes and validated (from literature) genes from rat, human, and
mouse genome.
20 Francesca Cristiano and Pierangelo Veltri

8 Ontologies

Gene Ontology (GO) (26) is a project aiming to unify the gene

description for each organism by using databases. Structurally Gene
Ontology is a directed acyclic graph where each gene is described
using terms and annotations and ontologies.
The three ontologies defined in Gene Ontology are:
l Cell Component.
l Molecular Function.
l Biological Process.
Queries in GO can be performed by considering that each term
(or list of genes) can be associated to biological processes or path-
ways. It is possible starting from biological description, afind set of
genes involved into the process. Microarray or NGS experiments
generate large number of genes; GO can be used to reduce the
dataset and identify subset of genes that can be considered of
interest. Such a process can be done by annotating the genes and
biological processes. Similarly to GO, genes can be also associated
by using their relation with diseases. Genes can thus be used to
identify biomarkers related to pathologies. For this aim, it is neces-
sary to associate genes with pathologies by using available data-
bases, also increasing the number of information associated to
genes. A possible application allowing to associate genes to disease
is the Disease Ontology (43), where each disease (or a group of
diseases) can be associated to a graph based representation repre-
senting is A and inclusion relations as a parent/child one. Each
disease description is represented in a hierarchical form to allow a
simple disease navigation. Also, the graphical user interface allow to
visualize diseases and related information such as disease ontology
ID (DOID), pathology name, a short description, various syno-
nyms, and other information such as MeSH (44) and ICD 9 (45).
DisGenet (46), also available online and as Cytoscape plugin, allows
to search gene and disease associations extracted from literature,
predicted associations and databases. It also uses a numerical rank
value (among 0 and 1) as defined in ref. (47).

9 Graphical Results Representation

The miRNA–mRNA interaction targets are used to generate the

interaction network.
Relations can be represented as graphs mapping predictions of
mRNAs activated or inhibited by miRNAs. The data can be
imported to a tool such as Cytoscape (48) that is an open-source
software that allows to see complex molecular interaction networks
 Methods and Techniques for miRNA Data Analysis 21

and integrating the data with additional information. To present

results to physicians, next step is to associate additional information
about functions of mRNA targets that can be involved in the
biological process. Such information can be hosted by different
available biological ontologies. Each network can be analyzed by
looking for subnetwork of miRNA to mRNA connections where
miRNAs are involved in a number of connections above a threshold
or selecting those mRNAs that interests different processes
(biological process, cell component, etc.). Thank to the availability
of several plugins, Cytoscape can be considered as one of the main
tool for the analysis of biological data. For instance ReactomeFI
(49) is one of the most used Cytoscape plugins that identifies and
creates sub-networks from the main biological network (see ref.
(50) for a Reactome application). ReactomeFI analyzes the net-
work by filtering the data based on their molecular function, path-
way, and biological process. Clustering techniques as well as data
integration techniques can be used to manipulate networks (with
graph tools) and to extract meaningful information for biological
analysis. References (51) and (52) are examples of works where
mining techniques have been used to extract patterns from graphs
also applied in ref. (53) for miRNAs. Also information about
clinical results or mRNA activities extracted from ontologies
(such as Gene Ontology (26) and Mesh (44)) can be integrated
and used to analyze different networks. For instance in ref. (54)
integration protocols are used to merge information obtained from
different networks each representing pairs of clinical information
(e.g., healthy versus nonhealthy patients). Relations among miR-
NAs and regulated mRNAs can be clustered with respect to pathol-
ogies, e.g., for chronic diseases.

10 Conclusions

The analysis of biological data produced in a high-performance

laboratory analysis environment requires bioinformatics tools and
platforms. For instance, the analysis of microarray data and NGS
sometimes requires high-performance evaluation tools. Neverthe-
less, often the available tools require specific knowledge in bioinfor-
matics or computer science. Thus, more simple-to-use tools need to
be designed and developed to allow simple analysis and result repre-
sentation. There are many tools provided by the bioinformatics
community especially following the sequencing of the human
genome, as well as query tool to crawl and navigate through the
huge amount of biological data produced. Moreover, the increasing
number of available dataset has been associated to the possibility of
relating genes to diseases as potential biomarkers. The identification
of microRNA signature could lead for example to discover the causes
and cures of diseases. Thus, the need for high-performance and
22 Francesca Cristiano and Pierangelo Veltri

simple-to-use bioinformatics tools is currently attracting many

researchers, as well as software tools able to query available databases
and enrich available information by using predictions, annotations to
produce additional information on biological laboratory results.
In this chapter we report the most used and available software
tools and techniques for managing and analyzing gene data.

References
1. Zhang X, Zeng Y (2011) Performing custom GX/?cid¼AG-PT- 130&tabId¼AG-PR-
microRNA microarray experiments. J Vis Exp 1061
56:e3250. doi:10.3791/3250 16. Friedl€ander MR, Chen W et al (2008) Discov-
2. Schena M, Shalon D et al (1995) Quantitative ering microRNAs from deep sequencing data
monitoring of gene expression patterns with a using miRDeep. Nat Biotechnol 26
complementary DNA microarray. Science 270 (4):407–415. doi:10.1038/nbt1394
(5235) 17. Blankenberg D, Von Kuster G, et al (2010)
3. Yin JQ, Zhao RC et al (2008) Profiling micro- Current protocols in molecular biology.
RNA expression with microarrays. Trends Bio- Chapter 19:Unit 19.10.1-21
technol 26(2):70–76. doi:10.1016/j.tibtech. 18. Giardine B, Riemer C et al (2005) Galaxy: a
2007.11.007 platform for interactive large-scale genome
4. Brazma A, Hingamp P et al (2011) Minimum analysis. Genome Res 15(10):1451–1455
information about a microarray experiment 19. http://training.bioinformatics.ucdavis.edu/
(MIAME): toward standards for microarray docs/2012/09/BSC/ThuPM-miRNA.html
data. Nat Genet 29(4):365–371 20. http://hannonlab.cshl.edu/fastx_toolkit/
5. David P, Bartel (2009) MicroRNAs: target rec- commandline.html#fastx_barcode_splitter_
ognition and regulatory functions. Cell 136 usage
(2):215–233. doi:10.1016/j.cell.2009.01.002 21. Friedl€ander MR, Mackowiak SD et al (2012)
6. http://www.454.com miRDeep2 accurately identifies known and
7. http://technology.illumina.com/technology/ hundreds of novel microRNA genes in seven
next-generation-sequencing/solexatechnology. animal clades. Nucleic Acids Res 40(1):37–52.
html doi:10.1093/nar/gkr688
8. http://www.appliedbiosystems.com/absite/ 22. Trapnell C, Pachter L et al (2009) TopHat:
us/en/home/applications-technologies/solid- discovering splice junctions with RNA-Seq.
next-generation-sequencing.html Bioinformatics 25(9):1105–1111. doi:10.
9. Pop M, Salzberg SL (2008) Bioinformatics 1093/bioinformatics/btp120
challenges of new sequencing technology. 23. Kim D, Pertea G et al (2013) TopHat2:
Trends Genet 24(3):142–149. doi:10.1016/j. accurate alignment of transcriptomes in the
tig.2007.12.006 presence of insertions, deletions and gene
10. http://www.bioinformatics.babraham.ac.uk/ fusions. Genome Biol 14:R36. doi:10.1186/
projects/fastqc/ gb-2013-14-4-r36
11. http://journal.embnet.org/index.php/ 24. http://cole-trapnell-lab.github.io/cufflinks/
embnetjournal/article/view/200/479 25. Anders S, Huber W (2010) Differential expres-
12. http://www.bioinformatics.babraham.ac.uk/ sion analysis for sequence count data. Genome
projects/trim_galore/ Biol 11:R106. doi:10.1186/gb-2010-11-10-
13. Goecks J, Nekrutenko A et al (2010) Galaxy: a r106
comprehensive approach for supporting acces- 26. Gene ontology (2014) http://www.
sible, reproducible, and transparent computa- geneontology.org/
tional research in the life sciences. Genome Biol 27. Biclustering of gene expression data. Jesùs S.
11(8):R86 Aguilar-Ruiz
14. Strand Life Sciences Pvt. Ltd. Strand NGS- 28. BLAST. http://blast.ncbi.nlm.nih.gov/Blast.cgi
formerly Avadis NGS, 2012, Version 1.3.0. 29. ENTREZ. http://www.ncbi.nlm.nih.gov/gquery/
San Francisco, CA: Strand Genomics, Inc. 30. PubMed. http://www.ncbi.nlm.nih.gov/pubmed/
15. http://www.genomics.agilent.com/en/Micro 31. EMBL. http://www.embl.org
array-Data-Analysis-Software/GeneSpring-
 Methods and Techniques for miRNA Data Analysis 23

32. Kozomara A, Griffiths-Jones S (2013) miR- 43. Kibbe WA, Arze C et al (2014) Disease ontol-
Base: annotating high confidence microRNAs ogy 2015 update: an expanded and updated
using deep sequencing data. Nucleic Acids Res database of human diseases for linking biomed-
42:D68–D73. doi:10.1093/nar/gkt1181 ical knowledge through disease data. Nucleic
33. Kozomara A, Griffiths-Jones S (2011) miR- Acids Res 43:D1071–D1078, pii: gku1011
Base: integrating microRNA annotation and 44. Medical subject headings. http://www.nlm.
deep-sequencing data. Nucleic Acids Res 39 nih.gov/mesh/
(Database issue):D152–D157. doi:10.1093/ 45. ICD. http://www.who.int/classifications/icd
nar/gkq1027 46. Bauer-Mehren A, Bundschus M et al (2011)
34. Ellison GM, Vicinanza C et al (2013) Adult c- Gene-disease network analysis reveals func-
kit(pos) cardiac stem cells are necessary and tional modules in Mendelian, complex and
sufficient for functional cardiac regeneration environmental diseases. PLoS One 6(6):
and repair. Cell 154(4):827–842 e20284
35. Leidinger P, Backes C et al (2013) A blood 47. http://www.disgenet.org/web/DisGeNET/
based 12-mirna signature of Alzheimer disease v2.1/dbinfo
patients. Genome Biol 14:R78. doi:10.1186/ 48. Shannon P, Markiel A et al (2003) Cytoscape: a
gb-2013-14-7-r78 software environment for integrated models of
36. Shirdel EA, Xie W et al (2011) Navigating the biomolecular interaction networks. Genome
micronome. using multiple microRNA predic- Res 13(11):2498–2504
tion database to identify signalling pathway- 49. Reactome Fi Cytoscape Plugin. http://www.
associated microRNAs. PLoS One 6(2): reactome.org
e17429. doi:10.1371/journal.pone.0017429
50. Guanming W, Feng X et al (2010) A human
37. Paraskevopoulou MD et al (2013) Diana- functional protein interaction network and its
microt web server v5.0: service integration application to cancer data analysis. Genome
into mirna functional analysis workflows. Biol 11(53)
Nucleic Acids Res 41(Web Server issue):
W169–W173. doi:10.1093/nar/gkt393 51. Gade S, Porzelius C et al (2011) Graph based
fusion of mirna and mrna expression data
38. Betel D, Wilson M et al (2008) The micro- improves clinical outcome prediction in pros-
RNA.org resource: targets and expression. tate cancer. BMC Bioinformatics 12:488
Nucleic Acids Res 36(Database Issue):
D149–D153 52. Tian Z, Greene AS et al (2008) MicroRNA
target pairs in the rat kidney identified by
39. Pictar. http://pictar.mdc-berlin.de microRNA microarray, proteomic, and bioin-
40. TargetScan microRNA target prediction. formatic analysis. Genome Res 18:404–411
http://www.targetscan.org/ 53. Pietro Hiram Guzzi, Pierangelo Veltri et al
41. Wang X (2008) miRDB: a microRNA target (2012) Unraveling multiple miRNA-mRNA
prediction and functional annotation data- associations through a graph-based approach.
base with a wiki interface. RNA 14 In: ACM BCB
(6):1012–1017 54. Bo W, Mezlini Aziz M et al (2014) Similarity
42. Dweep H, Sticht C et al (2011) miRWalk: network fusion for aggregating data types on a
database—prediction of possible miRNA bind- genomic scale. Nat Methods 11:333–337.
ing sites by “walking” the genes of 3 genomes. doi:10.1038/nmeth.2810
J Biomed Inform 44:839–847
Methods in Molecular Biology (2016) 1375: 25–39
DOI 10.1007/7651_2015_236
© Springer Science+Business Media New York 2015
Published online: 12 April 2015

Bioinformatics and Microarray Data Analysis on the Cloud

Barbara Calabrese and Mario Cannataro

Abstract
High-throughput platforms such as microarray, mass spectrometry, and next-generation sequencing are
producing an increasing volume of omics data that needs large data storage and computing power. Cloud
computing offers massive scalable computing and storage, data sharing, on-demand anytime and anywhere
access to resources and applications, and thus, it may represent the key technology for facing those issues. In
fact, in the recent years it has been adopted for the deployment of different bioinformatics solutions and
services both in academia and in the industry. Although this, cloud computing presents several issues
regarding the security and privacy of data, that are particularly important when analyzing patients data, such
as in personalized medicine. This chapter reviews main academic and industrial cloud-based bioinformatics
solutions; with a special focus on microarray data analysis solutions and underlines main issues and problems
related to the use of such platforms for the storage and analysis of patients data.

Keywords: Cloud computing, Bioinformatics, Microarray data analysis

1 Introduction

High-throughput platforms for the investigation of the cell

machinery, such as mass spectrometry, microarray, and next-
generation sequencing, yielded to the so-called “omics” sciences.
In particular, genomics regards the study of the activity of genes,
proteomics the study of the activity of proteins, and interactomics
the study of protein interactions inside a cell. Pharmacogenomics is
an important branch of genomics that studies the impact of genetic
variation (e.g., Single Nucleotide Polymorphisms—SNPs) on drug
response in patients and is at the basis of the so-called “personalized
medicine,” where drugs are chosen or optimized to meet the
genetic profile of each patient.
The availability of such high-throughput technologies and the
application of genomics and pharmacogenomics studies of large
populations, are producing an increasing amount of experimental
and clinical data, as well as specialized databases spread over the
Internet. However, the storage, preprocessing, and analysis of
experimental data are becoming the main bottleneck of the analysis
pipeline.

25
26 Barbara Calabrese and Mario Cannataro

Managing omics data requires both space for data storing and
services for data preprocessing, analysis, and sharing. The resulting
scenario comprises a set of bioinformatics tools, often implemented
as Web services, for the management and analysis of data stored in
geographically distributed biological databases.
Cloud computing is a computing model that has spread very
rapidly in recent years for the supply of IT resources (hardware and
software) of different nature, through services accessible via the
network. The resources that a cloud system provides to users
include: CPU, memory, networks, operating systems, middleware,
and applications. The cloud resources are dynamically scalable,
virtualized, and accessible on the Internet (1). This model provides
new advantages related to massive and scalable computing
resources available on demand, virtualization technology, and pay-
ment for use as needed (2).
Thus, cloud computing may play an important role in many
phases of the bioinformatics analysis pipeline, from data manage-
ment and processing, to data integration and analysis, including
data exploration and visualization.
Despite the many benefits associated with cloud computing,
there are also several management, technology, security, and legal
issues to be addressed. In fact, cloud computing currently presents
some issues and open problems such as privacy and security, geo-
graphical localization of data, legal responsibilities in the case of
data leaks, that are particularly important when managing sensitive
data such as the patients data stored and processed in genomics and
pharmacogenomics studies, and more in general when clinical data
are transferred to the cloud.
The aim of this chapter is to describe and discuss the most
significant applications of cloud computing in the bioinformatics
with special focus on microarray data analysis. The chapter focuses
on specific requirements and issues of such applications on cloud
computing. The chapter is organized as follows: in Section 2 cloud
computing definition is discussed. Service and delivery models are
presented in order to define the cloud-related background. Succes-
sively, in Section 3 the chapter focuses on the application of cloud
computing in bioinformatics and microarray data analysis. Section 4
summarizes the main problems to be faced when moving bioinfor-
matics applications on the cloud and underlines open problems
related to the full adoption of cloud computing in the bioinformat-
ics data analysis pipeline.

2 Materials

Even though cloud computing is now becoming the key technol-

ogy for the storage and analysis of large data sets both in academia
and industry, it is not a totally new concept. In fact, it has some
 Bioinformatics and Microarray Data Analysis on the Cloud 27

relations with grid computing and other technologies such as utility

computing, clustering, virtualization systems, and distributed
systems. There are many definitions of cloud computing. The first
definition comes from the work of Mell and Grance (1) and is a
popular working definition of cloud computing from the National
Institute of Standards and Technology, US Department of Com-
merce. Their definition focuses on computing resources that can be
accessed from anywhere and may be provisioned online. It also
specifies five characteristics of cloud computing (i.e., on-demand
self-service, broad network access, resource pooling, rapid elasticity,
and measured service), three service models (i.e., Software as a
Service, Platform as a Service, and Infrastructure as a Service) and
four deployment methods (i.e., private cloud, community cloud,
public cloud, and hybrid cloud). Most of the other definitions do
not mention deployment methods. In contrast to other definitions,
this one does not explicitly mention virtualization as a key
technology.
Vaquero et al. (3) collected 22 excerpts from previous works
and fused these into a single definition by studying the common
properties of cloud computing. This definition emphasizes the
importance of Service Level Agreements (SLA) in order to increase
confidence in the cloud environment and defines virtualization as
the key enabler of cloud computing.

2.1 Service Models Cloud services can be classified into three main models:
l Infrastructure as a Service (IaaS): this service model is offered in
a computing infrastructure that includes servers (typically vir-
tualized) with specific computational capability and/or storage.
The user controls all the storage resources, operating systems,
and applications deployed to, while he/she has limited control
over the network settings. An example is Amazon’s Elastic
Compute Cloud (EC2), which allows the user to create virtual
machines and manage them, and Amazon Simple Storage Ser-
vice (S3), which allows storing and accessing data, through a
Web-service interface.
l Platform as a Service (PaaS): it allows the development, instal-
lation and execution on its infrastructure of user-developed
applications. Applications must be created using programming
languages, libraries, services, and tools supported by the pro-
vider that constitute the development platform provided as a
service. An example is Google Apps Engine, which allows
developing applications in Java and Python and provides for
both languages the SDK (Software Development Kit) and uses
a plugin for the Eclipse development environment.
l Software as a Service (SaaS): customers can use the applications
provided by the cloud provider infrastructure. The applications
are accessible through a specific interface. Customers do not
28 Barbara Calabrese and Mario Cannataro

manage the cloud infrastructure or network components,

servers, operating systems, or storage. In some cases, it is
possible to manage specific configurations of the application.

2.2 Delivery Models Cloud services can be made available to users in different ways. In
the following, a brief description of the delivery models is
presented:
l Public Cloud: vendors who provide the users/customers the
hardware and software resources of their data centers offer
public cloud services. Examples of public clouds are Amazon,
Google Apps, and Microsoft Azure.
l Private Cloud: private cloud is configured by a user or by an
organization for its exclusive use. Computers that are in the
domain of the organization supply services. To install a private
cloud, several commercial and free tools are available (e.g.,
OpenStack, Eucalyptus, Open Nebula, Terracotta, and
VMware Cloud).
l Community Cloud: it is an infrastructure on which are installed
cloud services shared by a community or by a set of individuals,
companies and organizations that share a common purpose and
that have the same needs. The cloud can be managed by the
community itself or by a third party (typically a cloud service
provider).
l Hybrid Cloud: the cloud infrastructure is made up of two or
more different clouds using different delivery models, which,
while remaining separate entities, are connected by proprietary
or standard technology that enables the portability of data and
applications.

3 Methods

High-throughput platforms for the investigation of the cell

machinery, such as mass spectrometry, microarray, and next-
generation sequencing, are producing an overwhelming amount
of the so-called “omics” data (4). In particular, genomics regards
the study of the activity of genes, proteomics the study of the
activity of proteins, and interactomics the study of protein interac-
tions inside a cell (5).
Pharmacogenomics is an important branch of genomics that
studies the impact of genetic variation (e.g., Single Nucleotide
Polymorphisms—SNPs) on drug response in patients and is at
the basis of the so-called “personalized medicine,” where drugs
are chosen or optimized to meet the genetic profile of each
patient.
 Bioinformatics and Microarray Data Analysis on the Cloud 29

Pharmacogenomics correlates gene expression or SNPs with

the toxicity or efficacy of a drug, with the aim to improve drug
therapy with respect to the patients’ genotype, to ensure maximum
efficacy with minimal adverse effects. Many works demonstrated a
correlation between the presence/absence of SNPs and the devel-
opment of diseases, as well as the effectiveness of drugs (6).
Thus the presence (or the absence) of specific SNPs may be
used as a clinical marker for the prediction of drug effectiveness,
foreseeing the response of individuals with different SNPs to drugs.
The availability of such high-throughput technologies and the
application of genomics and pharmacogenomics studies of large
populations, are producing an increasing amount of experimental
and clinical data, as well as specialized databases spread over the
Internet. However, the storage, preprocessing, and analysis of
experimental data are becoming the main bottleneck of the analysis
pipeline.
Managing omics data requires both space for data storing and
procedures for data preprocessing, analysis, and sharing. The result-
ing scenario comprises a set of bioinformatics tools, often imple-
mented as Web services, for the management and analysis of data
stored in geographically distributed biological databases.
The main challenges regard: (1) the efficient storage, retrieval,
and integration of experimental data; (2) their efficient and high-
throughput preprocessing and analysis; (3) the building of repro-
ducible “in silico” experiments; (4) the annotation of omics data
with preexisting knowledge stored into ontologies (e.g., Gene
Ontology) or specialized databases; (5) the integration of omics
and clinical data.
Cloud computing may play an important role in many phases of
the analysis pipeline, from data management and processing, to
data integration and analysis, including data exploration and
visualization.
Currently, high performance computing is used to face the
large processing power required when processing omics data,
while Web services and workflows are used to face the complexity
of the bioinformatics pipeline that comprises several steps. Cloud
computing may be the glue that put together those mainstreams
technologies already used in bioinformatics (parallelism, service
orientations, knowledge management), with the elasticity and
ubiquity made available by the cloud.
Cloud computing represents a cost-effective solution for the
problems of storing and processing data in the context of bioinfor-
matics. Classical computational infrastructure for data processing
has become ineffective and difficult to maintain (7, 8). Dudley and
his colleagues (9) demonstrated that cloud computing is a viable
and cheaper technology that enables large-scale integration and
analysis for studies in genomic medicine.
30 Barbara Calabrese and Mario Cannataro

On the other hand, cloud computing presents some issues and

open problems such as privacy and security, geographical localiza-
tion of data, legal responsibilities in the case of data leaks, that are
particularly important when managing sensitive data as the patients
data stored and processed in genomics and pharmacogenomics
studies and more in general when clinical data are transferred to
the cloud.
In the following sections, the main cloud-based applications
proposed in the fields of bioinformatics are illustrated and open
problems related to the full adoption of cloud computing in
bioinformatics are underlined.

3.1 Cloud-Based The traditional bioinformatics analysis involves downloading of

Bioinformatics public datasets (e.g., NCBI, Ensembl), installing software locally
Solutions and analysis in-house. By entering the data and software in the
cloud and providing them as a service, it is possible to get a level
of integration that improves the analysis and the storage of bioin-
formatics big-data. In particular, as a result of this unprecedented
growth of data, the provision of data as a service (Data as a Service,
DaaS) is of extreme importance. DaaS provides data storage in a
dynamic virtual space hosted by the cloud and allows to have
updated data that are accessible from a wide range of connected
devices on the Web. An example is represented by the DaaS of
Amazon Web Services (AWS, http://aws.amazon.com/public
datasets), which provides a centralized repository of public data
sets, including archives of GenBank, Ensembl, 1000 Genomes
Project, Unigene, Influenza Virus (10).
In the following subsections, examples of SaaS, PaaS, and IaaS
for several tasks in bioinformatics domain are presented.

3.2 Bioinformatics In recent years, there have been several efforts to develop cloud-
Tools Deployed as based tools to execute different bioinformatics tasks (11), e.g.,
SaaS mapping applications, sequences alignment, gene expression analy-
sis (12). Some examples of SaaS bioinformatics tools are reported in
the following.
In ref. (13), the authors propose an efficient Cloud-based
Epistasis cOmputing (eCEO) model for large-scale epistatic inter-
action in genome-wide association study (GWAS). Given a large
number of combinations of SNPs (Single-nucleotide polymor-
phism), eCEO model is able to distribute them to balance the
load across the processing nodes. Moreover, eCEO model can
efficiently process each combination of SNPs to determine the
significance of its association with the phenotype. The authors
have implemented and evaluated eCEO model on their own cluster
of more than 40 nodes. The experiment results demonstrate that
the eCEO model is computationally efficient, flexible, scalable, and
practical. In addition, the authors have also deployed the eCEO
model on the Amazon Elastic Compute Cloud.
 Bioinformatics and Microarray Data Analysis on the Cloud 31

STORMSeq (Scalable Tools for Open—Source Read Mapping)

(14), is a graphical interface cloud computing solution that
performs read mapping, read cleaning and variant calling and anno-
tation with personal genome data. At present, STORMSeq costs
approximately 2 dollars and 510 h to process a full exome sequence
and 30 dollars and 38 days to process a whole genome sequence.
The authors provide this open-access and open-source resource as a
user-friendly interface in Amazon EC2.
CloudBurst (15) and CloudAligner (16) are parallel read-
mapping algorithms optimized for mapping next-generation
sequence (NGS) data to the human genome and other reference
genomes, for use in a variety of biological analyses including SNP
discovery, genotyping, and personal genomics. They use the open-
source Hadoop implementation of MapReduce to parallelize exe-
cution using multiple compute nodes. Specifically, CloudAligner
has been designed for more long sequences. An other Hadoop-
based tool is Crossbow (17) that combines the speed of the short
read aligner Bowtie with the accuracy of the SNP caller SOAPsnp to
perform alignment and SNP detection for multiple whole-human
datasets per day.
VAT (Variant Annotation Tool) (18) has been developed to
functionally annotate variants from multiple personal genomes at
the transcript level as well as to obtain summary statistics across
genes and individuals. VAT also allows visualization of the effects of
different variants, integrates allele frequencies and genotype data
from the underlying individuals and facilitates comparative analysis
between different groups of individuals. VAT can either be run
through a command-line interface or as a Web application. Finally,
in order to enable on-demand access and to minimize unnecessary
transfers of large data files, VAT can be run as a virtual machine in a
cloud-computing environment.
FX (19) is an RNA-Seq analysis tool, which runs in parallel
on cloud computing infrastructure, for the estimation of gene
expression levels and genomic variant calling. FX allows analysis
of RNA-Seq data on cloud computing infrastructures, support-
ing access through a user-friendly Web interface. An other cloud-
computing pipeline for calculating differential gene expression in
large RNA-Seq datasets is Myrna (20). Myrna integrates short
read alignment with interval calculations, normalization, aggre-
gation, and statistical modeling in a single computational pipe-
line. After alignment, Myrna calculates coverage for exons,
genes, or coding regions and differential expression using either
parametric or nonparametric permutation tests. Myrna exploits
the availability of multiple computers and processors where pos-
sible and can be run on the cloud using Amazon Elastic MapRe-
duce, on any Hadoop cluster, or on a single computer (bypassing
Hadoop entirely).
32 Barbara Calabrese and Mario Cannataro

PeakRanger (21) is a software package for the analysis in

Chromatin Immunoprecipitation Sequencing (ChIP-seq) tech-
nique. This technique is related to NGS and allows investigating
the interactions between proteins and DNA. Specifically, PeakRan-
ger is a peak caller software package that can be run in a parallel
cloud computing environment to obtain extremely high perfor-
mance on very large data sets.
For spectrometry-based proteomics research, ProteoCloud
(22) is a freely available, full-featured cloud-based platform to
perform computationally intensive, exhaustive searches using five
different peptide identification algorithms. ProteoCloud is entirely
open-source and is built around an easy-to-use and cross-platform
software client with a rich graphical user interface. This client
allows full control of the number of cloud instances to initiate
and of the spectra to assign for identification. It also enables the
user to track progress, and to visualize and interpret the results in
detail.
An environment for the integrated analysis of microRNA and
mRNA expression data is provided by BioVLAB-MMIA (23).
Recently, a new version called BioVLAB-NGS, deployed on both
Amazon cloud and on a high performance, publically available
server called MAHA, has been developed (24). By utilizing next
generation sequencing (NGS) data and integrating various bioin-
formatics tools and databases, BioVLAB-MMIA-NGS offers several
advantages, such as a more accurate data sequencing for determin-
ing miRNA expression levels or the implementation of various
computational methods for characterizing miRNAs.
Cloud4SNP (25) is a novel Cloud-based bioinformatics tool for
the parallel preprocessing and statistical analysis of pharmacoge-
nomics SNP microarray data. Cloud4SNP is able to perform statis-
tical tests in parallel, by partitioning the input data set and using the
virtual servers made available by the Cloud. Moreover, different
statistical corrections such as Bonferroni, False Discovery Rate, or
none correction, can be applied in parallel on the Cloud, allowing
the user to choice among different statistical models, implementing
a sort of parameter sweep computation.

3.3 Bioinformatics Currently, the most used platform (PaaS) for bioinformatics appli-
Platforms Deployed cations is Galaxy Cloud, which is a Galaxy cloud-based platform for
as PaaS the analysis of data at a large scale. It allows anyone to run a private
Galaxy installation on the Cloud exactly replicating functionality of
the main site, but without the need to share computing resources
with other users. With Galaxy Cloud, unlike software service solu-
tions, the user can customize their deployment as well as retain
complete control over their instances and associated data; the anal-
ysis can also be moved to other cloud providers or local
resources, avoiding concerns about dependence on a single vendor.
Currently, a public Galaxy Cloud deployment, called CloudMan, is
 Bioinformatics and Microarray Data Analysis on the Cloud 33

provided on the popular Amazon Web Services (AWS) cloud; how-

ever, it is compatible with Eucalyptus and other clouds (26).
CloudMan (27) enables individual bioinformatics researchers to
easily deploy, customize, and share their entire cloud analysis envi-
ronment, including data, tools, and configurations.
In ref. (28), a modular and scalable framework called Eoulsan,
based on the Hadoop implementation of the MapReduce algo-
rithm dedicated to high-throughput sequencing data analysis, is
presented. Eoulsan allows users to easily set up a cloud computing
cluster and automate the analysis of several samples at once using
various software solutions available. Tests with Amazon Web Ser-
vices demonstrated that the computation cost is linear with the
number of instances booked as is the running time with the increas-
ing amounts of data. Eoulsan is implemented in Java, supported on
Linux systems and distributed under the LGPL License.

Cloud-based bioinformatics applications

Project’s Services
name/ref models Task URL
eCEO SaaS Sequencing www.comp.nus.edu.sg
(genome
resequencing)
STORMSEQ SaaS Sequencing http://www.stormseq.org
(genome
resequencing)
Crossbow SaaS Sequencing http://bowtie-bio.
(genome sourceforge.net/
resequencing) crossbow/index.shtml
CloudBurst SaaS Sequencing: http://sourceforge.net/
genome projects/cloudburst-
resquencing, bio/
short-read
aligner
CloudAligner SaaS Sequencing: http://sourceforge.net/
genome projects/cloudaligner/
resquencing,
short-read
aligner
VAT SaaS Sequencing: vat.gersteinlab.org
genome
resquencing,
variant
annotation
(continued)
34 Barbara Calabrese and Mario Cannataro

(continued)

Cloud-based bioinformatics applications

Project’s Services
name/ref models Task URL
FX SaaS Sequencing: fx.gmi.ac.kr
RNA-seq
Myrna SaaS Sequencing: http://bowtie-bio.
RNA-seq sourceforge.net/
myrna/index.shtml
PeakRanger SaaS Sequencing: http://ranger.
ChIP SEQ sourceforge.net
ProteoCloud SaaS Mass https://code.google.
spectrometry: com/p/proteocloud/
MS-based
proteomics
YunBE SaaS Transcriptomics: http://lrcv-crp-sante.s3-
gene set website-us-east-1.
analysis amazonaws.com
BioVLAB- SaaS Analysis of https://sites.google.
MMIA microRNA and com/site/biovlab/
mRNA
expression data
Cloud4SNP SaaS Microarray: SNP Not available
Analysis
CloudMan PaaS A public Galaxy wiki.galaxyproject.org
cloud
deployment for
bioinformatics
Eoulsan PaaS A framework for http://transcriptome.ens.
high- fr/eoulsan/
throughput
sequencing
data analysis
Bionimbus IaaS A cloud-based bionimbus.openscience
infrastructure openscience
for managing,
analyzing and
sharing
genomics
datasets.
CloVR IaaS A virtual machine http://clovr.org
for automated
and portable
microbial
(continued)
 Bioinformatics and Microarray Data Analysis on the Cloud 35

(continued)

Cloud-based bioinformatics applications

Project’s Services
name/ref models Task URL
sequence
analysis
CloudBioLinux IaaS Genome analysis cloudbiolinux.org
resources for
cloud
computing
platforms

3.4 Bioinformatics Bionimbus (29) is an open-source cloud-computing platform used

Tools Deployed as IaaS by a variety of projects to process genomics and phenotypic data. It
is based primarily upon OpenStack, which manages on-demand
virtual machines that provide the required computational
resources, and GlusterFS, which is a high-performance clustered
file system. Bionimbus also includes Tukey, which is a portal, and
associated middleware that provides a single entry point and a
single sign on for the various Bionimbus resources; and Yates,
which automates the installation, configuration, and maintenance
of the software infrastructure required.
Cloud Virtual Resource, CloVR, (30) is a new desktop applica-
tion for push-button automated sequence analysis that can utilize
cloud computing resources. CloVR is implemented as a single
portable virtual machine (VM) that provides several automated
analysis pipelines for microbial genomics, whole genome and meta-
genome sequence analysis. The CloVR VM runs on a personal
computer, utilizes local computer resources, and requires minimal
installation, addressing key challenges in deploying bioinformatics
workflows. In addition CloVR supports use of remote cloud com-
puting resources to improve performance for large-scale sequence
processing.
Cloud BioLinux (31) is a publicly accessible Virtual Machine
(VM) that enables scientists to quickly provision on-demand infra-
structures for high-performance bioinformatics computing using
cloud platforms. Users have instant access to a range of preconfi-
gured command line and graphical software applications, including
a full-featured desktop interface, documentation and over 135
bioinformatics packages for applications including sequence align-
ment, clustering, assembly, display, editing, and phylogeny. Besides
the Amazon EC2 cloud, the authors started instances of Cloud
BioLinux on a private Eucalyptus cloud and demonstrated access to
the bioinformatics tools interface through a remote connection to
EC2 instances from a local desktop computer.
36 Barbara Calabrese and Mario Cannataro

4 Notes

Genomics data extracted by patients’ samples, as in pharmacoge-

nomics studies, as well as other clinical data and exams (e.g., bio-
images), are sensitive data and present unprecedented requirements
of privacy and security.
On the other hand, cloud computing presents some issues and
open problems, such as privacy and security, geographical localiza-
tion of data, legal responsibilities in the case of data leaks, that are
particularly important when managing such sensitive data.
In general, genomics and clinical data managed through a
cloud are susceptible to unauthorized access and attacks. Specifi-
cally, the chapter (32) claims that storing huge volumes of patients’
sensitive medical data in third-party cloud storage is susceptible to
loss, leakage, or theft. The privacy risk of cloud environment
includes the failure of mechanisms for separating storage, memory,
routing, and even reputation between different tenants of the
shared infrastructure. The centralized storage and shared tenancy
of physical storage space means the cloud users are at higher risk of
disclosure of their sensitive data to unwanted parties.
Threats to the data privacy in the cloud include spoofing
identity, tampering with the data, repudiation, and information
disclosure. In spoofing identity attack, the attacker pretends to be
a valid user, whereas data tampering involves malicious alterations
and modification of the content. Repudiation threats are concerned
with the users who deny after performing an activity with the data.
Information disclosure is the exposure of information to the enti-
ties having no right to access information. The same threats prevail
for the health data stored and transmitted on the third-party cloud
servers.
Therefore, confidentiality and integrity of the stored health
data are the most important challenges elevated by the health-care
and biomedicine cloud-based systems.
A secure protection scheme will be necessary to protect the
sensitive information of the medical record. There is considerable
work on protecting data from privacy and security attacks. NIST
(33) has developed guidelines to help consumers to protect their
data in the Cloud. The work reported in ref. (34) evidences that
using cryptographic storage significantly enhances security of the
data. The chapter discusses the main mechanisms to be adopted in
order to guarantee and satisfy the previous cited issues. Specifically,
the authors present and discuss the utility of cryptographic and
non-cryptographic approaches. The cryptographic approaches to
mitigate the privacy risks utilize certain encryption schemes and
cryptographic primitives. Conversely, non-cryptographic
approaches mainly use policy based authorization infrastructure
that allows the data objects to have access control policies. Particu-
larly, in the public cloud environment operated by the commercial
 Bioinformatics and Microarray Data Analysis on the Cloud 37

service providers and shared by several other customers, data pri-

vacy and security are the most attended requirements.
Abbas and Khan (35) summarized the security and privacy
requirements for cloud-based applications in the following way:
l Integrity: it is needed to ensure that the health data captured by
a system or provided to any entity is true representation of the
intended information and has not been modified in any way.
l Confidentiality: the health data of patients is kept completely
undisclosed to the unauthorized entities.
l Authenticity: the entity requesting access is authentic. In the
healthcare systems, the information provided by the healthcare
providers and the identities of the entities using such informa-
tion must be verified.
l Accountability: an obligation to be responsible in light of the
agreed upon expectations. The patients or the entities nomi-
nated by the patients should monitor the use of their health
information whenever that is accessed at hospitals, pharmacies,
insurance companies etc.
l Audit: it is needed to ensure that all the healthcare data is secure
and all the data access activities in the e-Health cloud are being
monitored.
l Non-repudiation: repudiation threats are concerned with the
users who deny after performing an activity with the data. For
instance, in the healthcare scenario neither the patients nor the
doctors can deny after misappropriating the health data.
l Anonymity: it refers to the state where a particular subject cannot
be identified. For instance, identities of the patients can be made
anonymous when they store their health data on the cloud so that
the cloud servers could not learn about the identity.
l Unlinkability: it refers to the use of resources or items of
interest multiple times by a user without other users or subjects
being able to interlink the usage of these resources. More
specifically, the information obtained from different flows of
the health data should not be sufficient to establish linkability
by the unauthorized entities.
Finally, in the cloud, physical storages could be widely
distributed across multiple jurisdictions, each of which may have
different laws regarding data security, privacy, usage, and intellec-
tual property. For example, the US Health Insurance Portability
and Accountability Act (HIPAA) restricts companies from disclos-
ing personal health data to nonaffiliated third parties. Similarly, the
Canadian Personal Information Protection and Electronic Docu-
ments Act (PIPEDA) limits the powers of organizations to collect,
use, or disclose personal information in the course of commercial
activities. However, a provider may, without notice to a user, move
38 Barbara Calabrese and Mario Cannataro

the users’ information from jurisdiction to jurisdiction. Data in the

cloud may have more than one legal location at the same time, with
different legal consequences.

5 Conclusions

Applications and services in bioinformatics and microarray data

analysis pose quite demanding requirements. The fulfillment of
those requirements could results in the development of compre-
hensive bioinformatics data analysis pipeline easy to use, available
through the Internet, that may increase the knowledge in biology
and medicine. As shown and discussed in this chapter, cloud com-
puting may play a key role in many phases of the bioinformatics and
microarray data analysis pipeline. In particular, cloud computing
may be the glue that put together the parallelism, service orienta-
tion, and knowledge management technologies already used in
bioinformatics, with the elasticity, ubiquity, and pay-per-use char-
acteristics of the cloud.
Naturally, the adoption of this technology with its benefits will
determine a reduction of costs and the possibility of also providing new
services. However, it is important to emphasize that the use of cloud in
these fields is featured still by a number of open issues and problems,
such as privacy and security, geographical localization of data, legal
responsibilities in the case of data leaks, that are particularly important
when managing patients’ data stored and processed in the cloud.

References

1. Mell P, Grance T. The NIST definition of 7. Schadt EE, Linderman MD, Sorenson J et al
cloud computing. Recommendations of the (2011) Cloud and heterogeneous computing
National Institute of Standards and Technol- solutions exist today for the emerging big data
ogy, Special Publication, 800–145 http://csrc. problems in biology. Nat Rev Genet 12(3):224
nist.gov/publications/PubsSPs.html 8. Grossmann RL, White KP (2011) A vision for a
2. Armbrust M, Fox A, Griffith R et al (2010) A biomedical cloud. J Intern Med 271(2):
view of cloud computing. Commun ACM 53 122–130
(4):50–58 9. Dudley JT, Pouliot Y, Chen JR et al (2010)
3. Vaquero LM, Rodero-Merino L, Caceres J et al Translational bioinformatics in the cloud: an
(2009) A break in the clouds: towards a cloud affordable alternative. Genome Med 2:51
definition. Comput Comm Rev 39:50–55 10. Fusaro VA, Patil P, Gafni E et al (2011) Bio-
4. Calabrese B, Cannataro M, Cloud Computing medical cloud computing with Amazon web
in Healthcare and Biomedicine, Scalable Com- services. PLoS Comput Biol 7(8):e1002147.
puting: Practice and Experience 16(1):1–18. doi:10.1371/journal.pcbi.1002147
doi:10.12694/scpe.v16i1.1057 11. Dai L, Gao X, Guo Y et al (2012) Bioinformat-
5. Cannataro M, Guzzi PH, Veltri P (2010) Pro- ics clouds for big data manipulation. Biol
tein-to-protein interactions: technologies, data- Direct 7:43. doi:10.1186/1745-6150-7-43
bases, and algorithms. ACM Comput Surv 43 12. Zhang L, Gu S, Wang B et al (2012) Gene set
(1):1–36 analysis in the cloud. Bioinformatics 28
6. Phillips C (2009) SNP databases. In: Komar (2):294–295
AA (ed) Single nucleotide polymorphisms, vol 13. Wang Z, Wang Y, Tan KL et al (2011) eCEO:
578. Humana, Totowa, NJ, pp 43–71, ch. 3 an efficient Cloud Epistasis cOmputing model
 Bioinformatics and Microarray Data Analysis on the Cloud 39

in genome-wide association study. Bioinfor- Proceedings of the International Conference

matics 27(8):1045–1051 on Bioinformatics, Computational Biology
14. Karczewski KJ, Fernald GH, Martin AR et al and Biomedical Informatics (BCB’13)
(2014) STORMSeq: an open-source, 26. Afgan E, Baker D, Coraor N et al (2011) Har-
user-friendly pipeline for processing personal nessing cloud computing with Galaxy Cloud.
genomics data in the cloud. PLoS One 9(1): Nat Biotechnol 29(11):972–974
e84860. doi:10.1371/journal.pone.0084860 27. Afgan E, Chapman B, Taylor J (2012) Cloud-
15. Schatz MC (2009) CloudBurst: highly sensi- Man as a platform for tool, data and analysis
tive read mapping with MapReduce. Bioinfor- distribution. BMC Bioinformatics 13:315.
matics 25(11):1363–1369 doi:10.1186/1471-2105-13-315
16. Nguyen T, Shi W, Ruden D (2011) CloudA- 28. Jourdren L, Bernard M, Dillies MA et al (2012)
ligner: a fast and full-featured MapReduce Eoulsan: a cloud computing-based framework
based tool for sequence mapping. BMC Res facilitating high throughput sequencing ana-
Notes 4:171. doi:10.1186/1756-0500-4-171 lyses. Bioinformatics 11(28):1542–1543
17. Langmead B, Schatz MC, Lin J et al (2009) 29. Heath P, Greenway M, Powell R et al (2014)
Searching for SNPs with cloud computing. Bionimbus: a cloud for managing, analyzing
Genome Biol 10:R134. doi:10.1186/gb- and sharing large genomics datasets. Int J
2009-10-11-r134 Med Inform 21(6):969–975. doi:10.1136/
18. Habegger L, Balasubramanian S, Chen DZ amiajnl-2013-002155
et al (2012) VAT: a computational framework 30. Angiuoli SV, Matalka M, Gussman A et al
to functionally annotate variants in personal (2011) CloVR: a virtual machine for auto-
genomes within a cloud-computing environ- mated and portable sequence analysis from
ment. Bioinformatics 28(17):2267–2269 the desktop using cloud computing. BMC
19. Hong D (2012) FX: an RNA-Seq analysis tool Bioinformatics 12:356. doi:10.1186/1471-
on the cloud. Bioinformatics 28(5):721–723 2105-12-356
20. Langmead B, Hansen KD, Leek JT (2010) 31. Krampis K, Booth T, Chapman B et al (2012)
Cloud-scale RNA-sequencing differential Cloud BioLinux: pre-configured and on-
expression analysis with Myrna. Genome Biol demand bioinformatics computing for the
11:R83. doi:10.1186/gb-2010-11-8-r83 genomics community. Bioinformatics 13:42.
doi:10.1186/1471-2105-13-42
21. Feng X, Grossman R, Stein L (2011) PeakRan-
ger: a cloud-enabled peak caller for ChIP-seq 32. Johnson ME (2009) Data hemorrhages in the
data. BMC Bioinformatics 12:139. doi:10. health-care sector, Financial Cryptography and
1186/1471-2105-12-139 Data Security, Lecture Notes in Computer Sci-
ence Volume 5628, pp. 71–89. doi:10.1007/
22. Muth T, Peters J, Blackburn J et al (2013)
978-3-642-03549-4_5
ProteoCloud: a full-featured open source pro-
teomics cloud computing pipeline. J Proteo- 33. Guidelines on security and privacy in public
mics 88:104–108 cloud computing. National Institute of
Standards and Technology (NIST), U.S.
23. Lee H, Yang Y, Chae H et al (2012) BioVLAB-
Department of Commerce. Special Publication,
MMIA: a cloud environment for microRNA
800–144. http://csrc.nist.gov/publications/
and mRNA integrated analysis (MMIA) on
nistpubs/800- 144/SP800-144.pdf
Amazon EC2. IEEE Trans Nanobioscience 11
(3):266–272 34. Kamara S, Lauter K (2010) Cryptographic
Cloud Storage, Financial Cryptography and
24. Chae H, Rhee S, Nephew KP et al (2014)
Data Security, Lecture Notes in Computer Sci-
BioVLAB-MMIA-NGS: MicroRNA-mRNA
ence Volume 6054, pp. 136–149. doi:10.
integrated analysis using high throughput
1007/978-3-642-14992-4_13
sequencing data. Bioinformatics 31:265–267.
doi:10.1093/bioinformatics/btu614 35. Abbas A, Khan SU (2014) A review on the
state-of-the-art privacy preserving approaches
25. Agapito G, Cannataro M, Guzzi PH et al
in the e-health clouds. IEEE J Biomed Health
(2013) Cloud4SNP: distributed analysis of
Inform 18(4):1431–1441
SNP microarray data on the cloud. In:
Methods in Molecular Biology (2016) 1375: 41–54
DOI 10.1007/7651_2015_240
© Springer Science+Business Media New York 2015
Published online: 12 March 2015

Classification and Clustering on Microarray Data

for Gene Functional Prediction Using R
Liliana López Kleine, Rosa Montaño, and Francisco Torres-Avilés

Abstract
Gene expression data (microarrays and RNA-sequencing data) as well as other kinds of genomic data can be
extracted from publicly available genomic data. Here, we explain how to apply multivariate cluster and
classification methods on gene expression data. These methods have become very popular and are
implemented in freely available software in order to predict the participation of gene products in a specific
functional category of interest. Taking into account the availability of data and of these methods, every
biological study should apply them in order to obtain knowledge on the organism studied and functional
category of interest. A special emphasis is made on the nonlinear kernel classification methods.

Keywords: Microarrays, Functional prediction, Multivariate data analysis, Clustering, Classification

1 Introduction

The methods presented here use clustering and classification methods

in order to determine groups of genes with functional relationships.
Although both types of methods are used to construct groups (of
genes in this case), clustering methods construct them without using
previous knowledge (unsupervised classification) and classification
methods using previous knowledge (supervised learning).

2 Materials

The procedures described here for gene functional prediction based

on microarrays can be applied to all kinds of gene expression data
organized in an n
p table containing the amount of RNA
messenger as shown in Table 1. Moreover, this kind of table can
be obtained from RNA sequencing data after reads are mapped on
the genes of the organism of interest. Using RNA sequencing
technique the amount of RNA transcripts represents also the RNA
quantity. Both raw data tables need to be normalized and trans-
formed in order to make them comparable before further proces-
sing and data analysis (1, 2). All procedures are presented in R (3).
41
42 Liliana López Kleine et al.

Table 1
Typical microarray data table

Gene ID Microarray condition (1) Microarray condition (p)

Gene (1) RNA quantity
Gene (n)

Fig. 1 Prediction of new genes belonging to the virulence category through classification

3 Methods

Below we describe briefly all multivariate clustering and classification

methods we have found useful for functional prediction from micro-
array data. We use the example of prediction of virulence factors
throughout the chapter as shown in Fig. 1. Nevertheless, these
methods are useful also for other functional categories of interest
and other organisms, for which genomic data is available. We have
concluded in previous studies (4–7) that more than one method
should be used and coincident predictions should be taken into
account in order to emit biological hypothesis and plan further in
silico or wet-lab validation experiments.

3.1 Functional Data This kind of method is straightforward for classifying genes into
(Known Gene two categories (e.g., virulence factors and not virulence factors,
Categories) and immunity related genes and not). These categories need to be
Training Sets constructed prior to applying the here proposed methods. They
can be constructed based on literature or extracted from genomic
databases. They should be represented as a vector indicating for
each of the genes to which category it belongs.
For supervised classification like support vector machine classifi-
cation (SVM) and linear discriminant analysis (LDA), we used a
training set. The genes belonging to the training set are chosen at
random from the two known categories and should represent approx-
imately a third part of all genes. The ratio between both classes in the
overall data set should be maintained in the training set.
 Microarray Classification for Functional Prediction 43

3.2 Preprocessing Several preprocessing methods for gene expression data exist. Any
Microarray Data of them can be used in order to normalize gene expression data and
to make experiments comparable. We recommend using the
method proposed by Huber et al. (8).
Below the code with an example data contained in Huber’s vsn
package. This package is an R package and makes a part of the
Bioconductor packages ((9), www.bioconductor.org) developed
especially for gene expression data.
source("http://bioconductor.org/biocLite.R") #installa-
tion of this package from Bioconductor
biocLite("vsn")
library(vsn) # this library contains Huber’s (2003) method
implemented
citation("vsn") # here is how you should cite the library if
you use it
data("lymphoma") #this is an available microarray data set
in R we are going to use to illustrate all methods
class(lymphoma) # this command returns the type of object;
this is a special object for microarray data
dim(exprs(lymphoma)) #returns the dimension of the data
table containing the gene expression data
boxplot(exprs(lymphoma)) #constructs a boxplot of gene
expression values for each of the 16 samples
par(mfrow¼c(1,2)) #prepares graphic window for two plots
hist(exprs(lymphoma)[,1],main¼"green") # plots a histo-
gram of the first sample
hist(exprs(lymphoma)[,2],main¼"red") # plots a histogram
of the second sample
lym2¼justvsn(lymphoma) #applies normalization method and
creates a new table
meanSdPlot(lym2, ranks¼TRUE) #shows the result of normal-
ization plotting mean against variance; #higher mean values
should not implicate higher standard deviation (sd).
#A horizontal red line is expected.
boxplot(exprs(lym2)) #boxplots after normalization
par(mfrow¼c(1,2))
hist(exprs(lym2)[,1]) #histograms after normalization
hist(exprs(lym2)[,2])

3.3 Clustering Clustering methods allow grouping observations, with the aim of
Methods reducing the variability inside each cluster and identifying homoge-
neous subgroups. Therefore, the observations inside each group
will present similar characteristics, essentially numerical. This meth-
odology has been widely used in many applications and is classified
as a multivariate technique in statistics.
A key element of this analysis is the similarity metric to be used in
order to quantify how similar individuals (here genes) are based on
the data at hand. The most common is the Euclidean metric.
44 Liliana López Kleine et al.

Nevertheless, other popular distance metrics are worth mentioning:

Manhattan, Canberra, and Binary. Details about these measures can
be found in the books edited by Rencher and Christensen (10) and
Izenman (11).
For the illustration of these methods, we use the Lymphoma
data set presented in the previous subsection.

3.3.1 Hierarchical Agglomerative Hierarchical clustering methods are the most

Methods common ones for unsupervised classification. In this method each
of the n observations (here genes) starts as one different cluster,
and on each iteration, several pairs of clusters are merged until, as a
final stage, a big cluster is formed.
The most recommended criterion, is the “Ward” criterion,
which is highly optimal when groups present a spherical behavior.
It is based on variance reduction in each new cluster created during
iteration, conducting to an optimal number of clusters minimizing
variability inside clusters (10).
Other choice to perform an unsupervised clustering is the
Divisive Analysis. This algorithm initially starts with all observations
in one single cluster and divides at each step the clusters until each
cluster contains just one single observation (12). It is called DIANA
in most of the references and is one of a few representatives of the
divisive hierarchical approach for clustering analysis. The main
elements for its implementation are the dissimilarity matrix and
the clustering fusion criteria.
A rule to define the number of clusters using this technique is
proposed by Mojena (13) and can be represented by a parallel line
that cuts the distance axis of the dendrogram. A correct number of
clusters are those with distances less than a constant. This constant
is computed from the mean plus k times the standard deviation of
the distances used to form all groups. Milligan and Cooper (14)
suggested k ¼ 1.25 as the most satisfactory criterion.
Given below is the code in order to apply the methods
described here to the example data of the vsn package (8).
Even when this is not the methodology to obtain the most
homogeneous clusters, it is possible to use it to detect initial
number of clusters and their mean vectors (cluster centers or cen-
troids) and use them as initial information useful in other better
procedures.
lymph_express<-as.data.frame(exprs(lymphoma))
dim(lymph_express)
Dmatrix<-dist(lymph_express)
hc_lymph<-hclust(Dmatrix, "ward")
group_member1<- cutree(hc_lymph, 10)
center1<- NULL
for(k in 1:10)}
 Microarray Classification for Functional Prediction 45

center1<- rbind(center1, colMeans(lymph_express

[group_ member1 ¼¼ k,, drop¼FALSE]))
}
table(group_member1)
hc_cutree<- hclust(dist(center1), method¼"ward", mem-
bers¼table(group_member1))
plot(hc_cutree)
group_member2<- cutree(hc_lymph, 3)
center2<- NULL
for(k in 1:3){
center2<- rbind(center2, colMeans(lymph_express
[group_ member2 ¼¼ k,, drop¼FALSE]))
}
table(group_member1)
hc_cutree2<- hclust(dist(center2), method¼"ward", mem-
bers¼table(group_member2))
plot(hc_cutree2)
STDlymphoma<- justvsn(lymphoma)
STDlymph_express<-as.data.frame(exprs(STDlymphoma))
dim(STDlymph_express)
DmatrixSTD<-dist(STDlymph_express)
hc_lymphSTD<-hclust(DmatrixSTD, "ward")
###################################################
mediadist<-mean(hc_lymphSTD$height)
dstddist<-sd(hc_lymphSTD$height)
mojenaC<- mediadist+1.75*dstddist
plot(hc_lymphSTD, main¼"", ylab¼"")#,ylim¼c(0,50))
abline(h¼mojenaC,col¼"green")
group_member_STD<- cutree(hc_lymphSTD, 8)
center_STD<- NULL
for(k in 1:8){
center_STD<- rbind(center_STD, colMeans(STDlymph_
express[group_member_STD ¼¼ k,, drop¼FALSE]))
}
table(group_member_STD)
hc_cutree_STD<- hclust(dist(center_STD), method¼"ward",
members¼table(group_member_STD))
plot(hc_cutree_STD)

3.3.2 K-Means Algorithm The K-means algorithm is the most common partition method
used in clustering analysis. This is a dynamic iterative method
which minimizes the within-class sum of squares for a given num-
ber of clusters (15). The algorithm starts with a vector of initial
centroids and each observation is placed in the cluster to which it is
the closest. The algorithm can be classified as dynamic because the
centroids are updated on each iteration. In this case, iteration is
defined as each complete stage once the groups are formed. The
process is repeated until the cluster centers stabilize.
46 Liliana López Kleine et al.

Regarding the initial choice of the groups or centroids, there is

at least a pair of alternatives the user can select. In the first one, the
initial centroids can be generated randomly (10). Instead, they can
also be obtained from the hierarchical cluster analysis (using Ward’s
criterion), and therefore, a k-means combined with hierarchical
clustering hybrid algorithm is obtained (16). This last method can
be applied using the code below.
KMc<-kmeans(STDlymph_express, center_STD, iter.max¼100)
STDmemberKM<-matrix(KMc$cluster,nrow
(STDlymph_express),1)
table(group_member_STD)
table(STDmemberKM)
table(group_member_STD, STDmemberKM)

3.3.3 Kohonen Self The genesis of this method was proposed by T. Kohonen (17, 18).
Organizing Maps The so-called Kohonen self-organizing maps, or simply SOM, are
highly appreciated for its ability to map and visualize high-
dimensional data in two dimensions. This algorithm is classified as
an unsupervised learning technique based on the neural networks
theory (18).
This neural network has a competitive unsupervised learning,
that is, no additional information is available for the classification of
the data, where all neurons compete in order to carry out a specific
task. Therefore, under an input pattern, only one of the output
neurons (or a group of neighbors) is activated. Therefore, activated
neurons compete until a winning neuron is assigned.
Initially, self-organized network use all information as input
data (communalities, regularities, correlations or categories) to
incorporate them into its internal structure connections. So the
neurons must self-organize based on the provided data.
In this method an input information vector is connected to an
intermediate layer, where each neuron or node is compared to the
input through weights computed from predefined functions.
Finally, an exit occurs when the result obtained is compared with
the final nodes and the winner node (activated neuron) is that one
that produced the smaller output. Code is presented below.
library(kohonen)
Koh_lymph<-som(scale(STDlymph_express), grid¼somgrid
(4, 2, "rectangular"))
Koh_member<-Koh_lymph$unit.classif
table(Koh_member)
par(mfrow¼c(4,2))
plot(Koh_lymph, type¼"codes")
plot(Koh_lymph, type¼"changes")
plot(Koh_lymph, type¼"counts")
plot(Koh_lymph, type¼"dist.neighbours")
####### Prediction
 Microarray Classification for Functional Prediction 47

percentage_training<- floor(.8*nrow(STDlymph_express))
Lymph_train<- sample(nrow(STDlymph_express), percentage_
training)
Lymph_training<- as.matrix(STDlymph_express[Lymph_train,])
Lymph_test<- as.matrix(STDlymph_express[-Lymph_train,])
som.STDlymphoma<- som(Lymph_training, grid¼somgrid(4, 2,
"hexagonal"))
som.Lymph.prediction<- predict(som.STDlymphoma, newdata¼
Lymph_test,
trainX¼Lymph_training,
trainY¼factor(som.STDlymphoma$unit.classif
[Lymph_train]))
table(som.STDlymphoma$unit.classif)
table(som.Lymph.prediction$unit.classif)

3.4 Supervised These methods use a part of the data in order to train a classifier,
Classification Methods which is a function allowing to place new individuals in one of the
previously known groups. For these methods, a training set is
needed as presented in Section 3.1.

3.4.1 Linear Discriminant The linear discriminant analysis (LDA) is a supervised method in
Analysis which a linear classifier is adjusted to classify known objects into
previously known groups, in this case for two groups. The method-
ology looks for linear combinations of variables which best explain
and separate the data, and explicitly attempts to model the differ-
ence between the given data classes (19). There are some assump-
tions related to the correct application of this method: normal
distribution and equality of variances between groups. Neverthe-
less, the latter coincides with the Fisher’s rule, a nonparametric
approach, when the response variables contain two categories.
Therefore, normality can be omitted, but the constraint of equally
variance should be evaluated.
In order to illustrate the method, we use the Breast cancer NKI
dataset from Bioconductor package (http://www.bioconductor.
org/packages/release/data/experiment/html/breastCancerNKI.
html). An auxiliary file is incorporated with the gene categories (R,
NR), obtained from the work developed by van’t Veer et al. (20).
source("http://bioconductor.org/biocLite.R")
biocLite("breastCancerNKI")
library(breastCancerNKI)
data(nki)
show(nki)
fData(nki)
pData(nki)
Rep<-read.table("nkiR.txt",header¼T)
reporter<-Rep[,4]
48 Liliana López Kleine et al.

mxpr<- as.data.frame(exprs(nki))
dim(mxpr)
cl_mxpr<- data.frame(mxpr,R¼as.data.frame(reporter))
dim(cl_mxpr)
names(cl_mxpr)
cl_mxpr$reporter
library(MASS)
train<- sample(1:24481, round(24481*9/10))
table(cl_mxpr$reporter[train])
ytest<- lda(reporter~., cl_mxpr, prior¼c(1,1)/2, subset
¼train)
ypred<- predict(ytest, cl_mxpr[-train,])
ctable<- table(cl_mxpr[-train,]$reporter, ypred$class)
gclass<- sum(diag(ctable))/sum(ctable)

3.4.2 Linear Support Support Vector Machines (SVM) are kernel methods. The particu-
Vector Machines larity of kernel methods is that algorithms are performed after the
data is transformed and therefore projected into a different high
dimensional space called feature space. So the expression data will
be projected into space H through the mapping Φ : X ! H where
X is the original space H is called feature space and all data points
will have their analogue in that space. For the case of kernel meth-
ods, this mapping does not need to be known and is achieved
through the kernel function K : X
X ! ℝ, ðx; yÞ ! K ðx; yÞ
(21). This is called the “kernel trick” and can be stated as follows:
Let X be the space of the function and consider a bivariate function
K defined as X
X. Let H be the associated feature space. 0
Then a
transformation Φ : X ! H exists, so that K ðx; yÞ=ΦðxÞ Φð yÞ.
An important result for the application of SVM is that any
X f defined on X can be expressed as follows:
function
f ðÞ= αi K ðx i ; Þ this is called the Reproducibility Kernel Hilbert
i
Space (RKHS).
Several types of kernel functions exist, being the linear kernel,
the simplest. The linear kernel is directly the inner product between
two data vectors x and y on the same p individuals:
0 Pp
K ðx; yÞ = x y = i¼1 x i y i . Other common kernels are the polyno-
mial kernel K ðx; yÞ = ðscale 〈x, y〉 þ cte Þg and the Gaussian kernel
K ðx; yÞ = exp (σjjxyjj2).
A linear classifier will allow separating individuals (here genes)
as shown in Fig. 2. If data is not linearly separable, a nonlinear
classifier needs to be constructed (Fig. 3).  
Consider a the sample ðx1 ; y 1 Þ, ðx2 ; y 2 Þ,   , xn ; y n ; where
xi 2 ℝ p and y i 2 f1, þ 1g, (the response variable). These sample
is named learning sample in machine learning theory.
The idea behind these kinds of models is that they make it
possible to separate the two classes 1 and +1 by a hyperplane.
 Microarray Classification for Functional Prediction 49

Fig. 2 Classifier: the support vectors are the objects that are placed on the dotted
lines representing the margin

Fig. 3 Schematic representation of the nonlinear SVM classifier and its margin

This hyperplane is constructed using the so-called support vectors as

is shown in Fig. 2. Clarke et al. (22) proved that the distance
between hyperplane and support vectors is Ma ¼ jjWjj1
. Ma is called
the margin described by W vector (it is the distance between the
hyperplane and the nearest individual in both classes). A large
50 Liliana López Kleine et al.

margin implies flexibility and a small margin precision. Therefore its

width should be chosen trading off these two properties.
When the highest value of M is a constraint to a correct classifi-
cation, the optimization problem can be expressed as follows:
1  0 
minW, b jjWjj2 constraint to y i W xi þ b  1 f or all i ¼ 1, 2, . . . , n
2
ð1Þ
Clarke et al. (22) show the last optimization problem can achieve to
the dual formulation, because it allows a less complex solution by
quadratic programming:

1X n X n
0
Xn
minα αi α j y i y j xi x j  αi
2 i¼1 j ¼1 i¼1

X
n
Constrain to αi y i ¼ 0 and 0  αi  C, i ¼ 1, 2, . . . , n ð2Þ
i¼1

Where the αi 2 ℝ for i=1, 2,   , n are Lagrange multipliers

associated with Eq. 1, finally the support vectors exist, based on
the following equations:
 0 
If α ¼ 0 where y i W xi þ b > 1
And
 0 
If α > 0 where y i W xi þ b ¼ 1
The vectors xi that satisfy αi > 0 are support vectors (22).
The example code we use below is adapted from JP Vert’s
practical session on SVM (http://cbio.ensmp.fr/~jvert/svn/
tutorials/practical/svmbasic/svmbasic_notes.pdf).
n <- 150 # number of data points
p <- 2 # dimension
sigma <- 1 # variance of the distribution
meanpos <- 0 # centre of the distribution of positive
examples
meanneg <- 3 # centre of the distribution of negative
examples
npos <- round(n/2) # number of positive examples
nneg <- n-npos # number of negative examples
# Generate the positive and negative examples
xpos <- matrix(rnorm(npos*p,mean¼meanpos,sd¼sigma),
npos,p)
xneg <- matrix(rnorm(nneg*p,mean¼meanneg,sd¼sigma),
npos,p)
x <- rbind(xpos,xneg)
# Generate the labels
y <- matrix(c(rep(1,npos),rep(-1,nneg)))
 Microarray Classification for Functional Prediction 51

# Visualize the data

plot(x,col¼ifelse(y>0,1,2))
legend("topleft",c(’Positive’,’Negative’),col¼seq(2),
pch¼1,text.col¼seq(2))
## Prepare a training and a test set (30 % of the data) ##
ntrain <- round(n*0.7) # number of training examples
tindex <- sample(n,ntrain) # indices of training samples
xtrain <- x[tindex,]
xtest <- x[-tindex,]
ytrain <- y[tindex]
ytest <- y[-tindex]
istrain¼rep(0,n)
istrain[tindex]¼1
plot(x,col¼ifelse(y>0,1,2),pch¼ifelse
(istrain¼¼1,1,2))
legend("topleft",c(’Positive Train’,’Positive Test’,’-
Negative Train’,’Negative Test’),
col¼c(1,1,2,2),pch¼c(1,2,1,2),text.col¼c(1,1,2,2))
# load the kernlab package
library(kernlab)
# train the SVM with C¼100
svp <- ksvm(xtrain,ytrain,type¼"C-svc",kernel¼’vanilladot’,
C¼100,scaled¼c())
# General summary
svp
# Use the built-in function to plot the classifier
plot(svp,data¼xtrain)
# Predict labels on test
ypred ¼ predict(svp,xtest)
table(ytest,ypred)
# Compute accuracy
sum(ypred¼¼ytest)/length(ytest)

3.4.3 Nonlinear Support Here we consider again the function to optimize (Eq. 1), but with
Vector Machine the mapped points in H done by the Φ function:
1  0 
minW, b jjWjj2 constrain to y i W Φðxi Þ þ b  1 f or all
2
i ¼ 1, 2, . . . , n ð3Þ
A new parameter C to solve the optimization problem is needed.
It controls the individuals that the model cannot classify in the
correct class:

X
n
minW, b jjWjj2 þ C ξi
i¼1
 0 
Constrain to y i W Φðxi Þ þ b  1  ξi and ξi  0 f or all

i ¼ 1, 2, . . . , n ð4Þ
52 Liliana López Kleine et al.

Similarly as in Eq. 2 the dual formulation is

1X n X n   X n
minα αi α j y i y j K xi ; x j  αi
2 i¼1 j ¼1 i¼1

X
n
Constrain to αi y i ¼ 0 and 0  αi  C to i ¼ 1, 2, . . . , n ð5Þ
i¼1

Where αi are Lagrange multipliers and xi are support vectors. The

inequality 0<α< has to be satisfied. The result of the optimization
is the nonlinear classifier model, where the support vectors are
those objects that are placed on the dotted lines representing the
limit of the margin (Fig. 3).
# Train a nonlinear SVM
svp <- ksvm(x,y,type¼"C-svc",kernel¼’rbf’,kpar¼list
(sigma¼1),C¼1)
plot(svp,data¼x)
# Train a nonlinear SVM with automatic selection of sigma by
heuristic
svp <- ksvm(x,y,type¼"C-svc",kernel¼’rbf’,C¼1)
plot(svp,data¼x)
#Prediction of a class of tumor using the publicly available
dataset of gene expression.
# Load the ALL dataset
library(ALL)
data(ALL)
# Inspect them
?ALL
show(ALL)
print(summary(pData(ALL)))
#prediction of the type of the disease (B-cell or T-cell).
x <- t(exprs(ALL))
y <- substr(ALL$BT,1,1)
y <- ALL$BT
print(y)
n<-dim(exprs(ALL))[2]
ntrain <- round(n*0.7) # number of training examples
tindex <- sample(n,ntrain) # indices of training samples
xtrain <- x[tindex,]
xtest <- x[-tindex,]
ytrain <- y[tindex]
ytest <- y[-tindex]
istrain¼rep(0,n)
istrain[tindex]¼1
svp <- ksvm(xtrain,ytrain,type¼"C-svc",kernel¼’vanilladot’,
C¼100,scaled¼c())
# General summary
svp
# Predict labels on test
ypred ¼ predict(svp,xtest)
 Microarray Classification for Functional Prediction 53

table(ytest,ypred)
# Compute accuracy
sum(ypred¼¼ytest)/length(ytest)

3.5 Extracting Most reliable prediction methods (based on the classification

Functional Predictions errors) are chosen to assemble a list of genes whose product could
have the function of interest (i.e., virulence, immunity, . . .), based
on the hypothesis that genes with similar gene expression profiles
participate in the same molecular function. Moreover, genes that
are predicted by several methods, are more likely to be accurately
predicted.
Assuming the list of genes that wants to be combined are list1
and list2, a simple code to extract common elements is shown
below.
list1<-c(1,2,3,4,5,6,7,8,9,10)
list2<-c(4,5,6,7,8,9,10,11)
combilist<-list1[list1%in%list2]

4 Notes

First difficulties can arise when genomic data is extracted from

publicly available databases, because gene expression data has dif-
ferent formats and because gene names between data tables do not
always match. For example, if you are using microarray data from
NCBI (http://www.ncbi.nlm.nih.gov/) and information on par-
ticipation in metabolic functions from KEGG (http://www.
genome.jp/kegg/), gene names could differ. This needs to be
corrected previously.
For the unsupervised methods of classification, it is convenient
to obtain initial objective groups using the hierarchical methods,
implemented in most of the specialized software packages. These
centers can be used as prior knowledge to apply the K-means
algorithm, which presents a better performance in contrast with
its hierarchical alternative. On the other hand, Kohonen’s method
is a better choice when nonlinear behaviors are suspected.
Supervised methods are useful when the categories to predict
are known for some genes. Problems arise here when the categories
are known only for a very small amount of genes. In that case it is
suggested to filter genes based on any objective criterion (such as
variance or mean expression) and predict only for a subgroup of
genes. Linear discriminant analysis is an useful method to explore
potential linear relationships among data and the dependent
categorical variable that represents the groups. If the covariance
matrices among groups are not equal, the method could present a
poor performance and a bad classification.
To study the performance of the methods proposed here,
especially for the supervised methodologies, it is necessary to
54 Liliana López Kleine et al.

estimate rates of good classification (or prediction) and, sensitivity

and/or specificity rates, to study the performance of the empirical
proposed rule.
Finally, we recommend a previous training on R before these
methods are applied in order to cope with common errors
successfully.

References
1. Dudoit S, Yang YH, Callow MJ, Speed TP computational biology and bioinformatics.
(2002) Statistical methods for identifying dif- Genome Biol 5(10):R80
ferentially expressed genes in replicated cDNA 10. Rencher AC, Christensen WF (2012) Methods
microarray experiments. Stat Sinica 12 of multivariate analysis, 3rd edn. Wiley, Hobo-
(1):111–140 ken, NJ
2. Moguerza JM, Muñoz A (2006) Support vec- 11. Izenman AJ (2008) Modern multivariate sta-
tor machines with applications. Statist Sci 21 tistical techniques: regression, classification,
(3):299–426 and manifold learning. Springer, New York
3. R Core Team (2014) R: a language and envi- 12. Kaufman L, Rousseeuw PJ (1990) Finding
ronment for statistical computing. R Founda- groups in data: an introduction to cluster anal-
tion for Statistical Computing, Vienna, http:// ysis. Wiley, New York
www.R-project.org/ 13. Mojena R (1977) Hierarchical grouping meth-
4. López-Kleine L1, Molano N, Ospina L. Int J ods and stopping rules: an evaluation. Comput
Bioinform Res Appl. 2013;9(3):285–300. doi: J 20(4):359–363. doi:10.1093/comjnl/20.4.
10.1504/IJBRA.2013.053607. Using multi- 359
variate methods to infer knowledge from geno- 14. Glenn W, Milligan GW, Cooper MC (1985) An
mic data examination of procedures for determining the
5. López-Kleine L, Torres-Avilés F, Tejedor FH, number of clusters in a data set. Psychometrika
Gordillo LA (2012) Virulence factor predic- 50(2):159–179
tion in Streptococcus pyogenes using classifi- 15. Hartigan JA, Wong MA (1979) A k-means
cation and clustering based on microarray clustering algorithm. Appl Statist 28:100–108
data. Appl Microbiol Biotechnol
93:2091–2098. doi:10.1007/s00253-012- 16. Leiva-Valdebenito S, Torres-Avilés F (2010) A
3917-3 review of the most common partition algo-
rithms in cluster analysis: a comparative study.
6. López-Kleine L, Romeo J, Torres-Avilés F Rev Colomb Estad 33(2):321–339
(2013) Gene functional prediction using clus-
tering methods for the analysis of tomato 17. Kohonen T (1982) Self-organizing formation
microarray data. In: Mohamad MS et al (eds) of topologically correct feature maps. Biol
7th International conference on PACBB, Cybern 43:59–69
AISC, vol 222, pp 1–6 18. Kohonen T (2001) Self-organizing maps, 3rd
7. Romeo JS, Torres-Avilés F, López-Kleine L edn. Springer, New York
(2013) Detection of influent virulence and 19. Friedman JH (1989) Regularized discriminant
resistance genes in microarray data through analysis. JASA 84:165–175
quasi likelihood modeling. Mol Genet Geno- 20. van’t Veer LJ, Dai H, van de Vijver MJ, He YD,
mics 288(1–2):49–61. doi:10.1007/s00438- Hart AAM, Mao M, Peterse HL, van der Kooy
012-0730-8 K, Marton MJ, Witteveen AT, Schreiber GJ,
8. Huber W, von Heydebreck A, Sueltmann H, Kerkhoven RM, Roberts C, Linsley PS, Ber-
Poustka A, Vingron M (2003) Parameter esti- nards R, Friend SH (2002) Gene expression
mation for the calibration and variance stabili- profiling predicts clinical outcome of breast
zation of microarray data. Stat Appl Genet Mol cancer. Nature 415:530–536
2(1):Article 3 21. Schölkopf B, Smola A (2002) Learning with
9. Gentleman RC, Carey VJ, Bates DM, Bolstad Kernels: support vector machines, regulariza-
B, Dettling M, Dudoit S, Ellis B, Gautier L, Ge tion, optimization, and beyond. The MIT
Y, Hornik K, Gentry J, Hothorn T, Huber W, Press, Cambridge
Iacus S, Irizarry R, Leisch F, Li C, Maechler M, 22. Clarke B, Fokoué E, Zhang H (2009) Princi-
Rossini AJ, Sawitzki G, Smith C, Smyth G, ples and theory for data mining and machine
Tierney L, Yang JYH, Zhang J (2004) Biocon- learning. Springer, New York
ductor: open software development for
Methods in Molecular Biology (2016) 1375: 55–74
DOI 10.1007/7651_2015_246
© Springer Science+Business Media New York 2015
Published online: 02 December 2015

Querying Co-regulated Genes on Diverse Gene

Expression Datasets Via Biclustering
Mehmet Deveci, Onur K€
uç€
uktunç, Kemal Eren, Doruk Bozdağ,
€
Kamer Kaya, and Umit V. Çataly€
urek

Abstract
Rapid development and increasing popularity of gene expression microarrays have resulted in a number of
studies on the discovery of co-regulated genes. One important way of discovering such co-regulations is the
query-based search since gene co-expressions may indicate a shared role in a biological process. Although
there exist promising query-driven search methods adapting clustering, they fail to capture many genes that
function in the same biological pathway because microarray datasets are fraught with spurious samples or
samples of diverse origin, or the pathways might be regulated under only a subset of samples. On the other
hand, a class of clustering algorithms known as biclustering algorithms which simultaneously cluster both
the items and their features are useful while analyzing gene expression data, or any data in which items are
related in only a subset of their samples. This means that genes need not be related in all samples to be
clustered together. Because many genes only interact under specific circumstances, biclustering may recover
the relationships that traditional clustering algorithms can easily miss. In this chapter, we briefly summarize
the literature using biclustering for querying co-regulated genes. Then we present a novel biclustering
approach and evaluate its performance by a thorough experimental analysis.

Keywords: Biclustering, Microarray, Gene expression, Clustering

“What we call chaos is just patterns we haven’t recognized. What we

call random is just patterns we can’t decipher.”
— Chuck Palahniuk, Survivor

1 Introduction

The microarray technology enables large-scale genomic research by

allowing the measurement of the expression levels of thousands of
genes in parallel. Expression levels of genes in various samples are
collected and stored in a gene expression matrix. Mining these gene
expression matrices can provide insights into gene functions and
aids in the development and treatment of complex diseases. The
discovery of related genes is a challenging task and has been the
focus of many research studies [1–4] that search for more sophisti-
cated analysis methods. Most of the time, however, researchers
55
56 Mehmet Deveci et al.

focus on a specific gene or a gene set rather than exploring the

whole dataset. Query-based search algorithms [5–10] are proven to
be very useful when the objective is to rank the genes according to
how strongly they are correlated with the queried gene(s). For
example, several genes in S. cerevisiae database are categorized and
annotated by Hibbs et al. [6]. Similarly, top-ranked genes co-
regulated with breast cancer associated tumor suppressors,
BRCA1 and BRCA2, are found to be regulating the mitotic spindle
and cytokinesis by Bozdağ et al. [9]. In analyzing this torrent of
new data, unsupervised learning methods such as clustering are
important as the first step. In particular, a class of clustering algo-
rithms known as biclustering is useful for analyzing gene expression
data, or any data whose items are related in only a subset of their
samples. Biclustering methods cluster both the items and their
features simultaneously. In gene expression context, this means
that genes need not be related in all samples to be clustered
together. Because many genes only interact under specific circum-
stances, biclustering may recover relationships that traditional clus-
tering algorithms can miss.
In this chapter, first, we briefly survey the literature on biclus-
tering and proposed algorithms. Then we introduce a novel biclus-
tering algorithm, Correlated Patterns Biclustering (CPB), which
attempts to find genes that are related on a subset of their features
with a query gene. As mentioned above, identifying the genes co-
regulated with a gene of important function is crucial to understand
biochemical and genetic pathways in which the gene participates.
To quantify gene relationships, CPB uses the Pearson correlation
coefficient (PCC), an effective and widely used metric in this type of
analysis to quantify co-regulation between pairs of genes [2, 4].
CPB’s novel approach avoids costly pairwise correlation calcula-
tions in a manner that also increases its accuracy. It also allows
assigning genes to multiple biclusters, because many genes partici-
pate in multiple biological pathways. We further introduce a unique
method for combining results from multiple datasets, which is
important for uncovering uncommon genetic relationships. Initial
testing on artificial data shows that CPB outperforms other biclus-
tering methods in finding multiple types of biclusters. CPB’s per-
formance for querying the microarray data is similarly promising: it
was able to find many genes that have high correlation with
BRCA1, BRCA2, and p53. Of those genes, half are already
known to be involved in cancer processes, and the others are
promising new candidates for further investigation. The source
code of the framework, documentation, and sample datasets is
available at http://bmi.osu.edu/hpc/software/cpb/.
The methods in this chapter extend the framework proposed by
Bozdağ et al. [9] to increase the efficiency of the algorithms as well
as the consistency and relevancy of the results. The novelty of the
proposed algorithm can be summarized as follows:
 Querying Co-regulated Genes on Diverse Gene Expression Datasets Via Biclustering 57

l A grid-based method is used for generating initial biclusters,

which covers the whole dataset.
l Results are investigated and the statistically insignificant biclus-
ters are filtered out with a non-parametric scheme.
l The biclustering method is tested on various models, noise
levels, and overlap ratios; compared with other techniques.
l Correlation scores of the genes are computed and combined
more efficiently.
The key advantages of the proposed query-based search frame-
work are:
l It finds co-regulated genes with a given reference gene on a
number of diverse microarray datasets having the same genes.
This is the case for data obtained from a single microarray.
l PCC-based biclustering technique is able to discover constant-
row, shift, scale, and shift-scale models with positive and nega-
tive correlations.
l CPB is extremely efficient compared to other PCC-based meth-
ods because of a novel correlation calculation.
l Filtering step increases the relevance of the results while elim-
inating insignificant and overlapping biclusters.
The rest of the chapter is organized as follows: In Section 2,
biclustering algorithms from the literature are briefly surveyed.
Section 3 describes the CPB algorithm. The results of the proposed
algorithm and framework’s experimental evaluation are given in
Section 4. Section 5 concludes the chapter.

2 Biclustering of the Microarray Data

Biclustering refers to a class of methods that perform simultaneous

clustering of both rows and columns of a data matrix. It was first
introduced to gene expression data analysis by Cheng and Church
[11]. This initial algorithm was followed by numerous biclustering
algorithms to identify additive, multiplicative [12, 13], or more
complex relationships [14–22] between the rows and columns of a
data matrix that correspond to genes and samples, respectively.
A straightforward two-phase approach to identify the biclusters
is applying standard clustering algorithms to the genes and samples
separately in the first step, and combine the results in the second
one [23]. However, the research on biclustering is focused to a
more integrated approach in which the genes and samples are
analyzed simultaneously. Several randomized or deterministic algo-
rithms based on both novel and existing techniques from various
domains, such as independent component analysis, singular value
decomposition, simulated annealing, and local search, have been
58 Mehmet Deveci et al.

proposed, i.e., [24–33], and evaluated on the gene expression data

for many diseases including the complex ones such as cancer. Some
algorithms in this set use greedy techniques, i.e., [11, 34], and
some employ evolutionary techniques [35–40]. In addition,
graphs, modeling pairwise gene–gene interactions, have also been
employed to design novel biclustering methods. For example, a
local, correlated structure in the graph obtained by the gene expres-
sion data is shown to be promising to be used as a bicluster [41].
In the literature, bicluster models that a biclustering algorithm
seeks for can be divided into two categories. Global biclusters are
defined by comparing a metric within the bicluster to the outside of
the bicluster. Up-regulated biclusters with higher expression values
compared to background, and down-regulated biclusters with
lower expression values than the background are examples of global
biclusters. Many algorithms have been proposed to capture the
global biclusters such as SAMBA [42], ISA [43], Spectral [44],
BiMax [45], QUBIC [46], COALESCE [47], BBK [48]. On the
other hand, local biclusters can be defined by the relationships
within the bicluster columns and rows such as constant, additive,
and multiplicative biclusters. Additive models are useful for captur-
ing shifting patterns (see Fig. 1b), whereas multiplicative models are
useful for capturing scaling patterns (see Fig. 1c) in the data. How-
ever, neither can simultaneously identify the shifting and scaling
patterns. In this chapter, we will seek biclusters fitting the shift-scale
model (see Fig. 1d) which covers both additive and multiplicative
patterns as special cases.
How to evaluate the quality of the biclusters is also an impor-
tant problem: for example, the classical mean squared resi-
due (MSR) has been shown to be successful at finding constant,
and additive biclusters, while it is not suitable for multiplicative
biclusters. In addition, it is claimed to be biased toward flat biclus-
ters with low row variance, and hence, different scoring schemas
have been proposed Bryan and Cunningham [49]. Cheng and
Cherch propose a deterministic greedy algorithm that seeks to
find the biclusters with low variance, as defined by the MSR [11].

a b c d

Fig. 1 Sample biclusters with various models: (a) constant-row, (b) shift, (c) scale, and (d) shift-scale. In
pattern expressions, aij represents expression level of gene i in sample j, π j a base value, αi scaling, and βi
shifting patterns. The parameters are selected as αi ¼ ½1, 2, 3,  1T , βi ¼ ½2, 3, 4, 1T , π j ¼ ½1, 2, 1, 4.
Shift-scale is the most general model, as it has shift and scale models as special cases and can represent both
positive and negative correlation
 Querying Co-regulated Genes on Diverse Gene Expression Datasets Via Biclustering 59

Similarly, the xMOTIFs algorithm has been proposed to capture

conserved gene expression motifs that are the biclusters with con-
served rows in discretized dataset [50]. A more complex relation-
ship among the genes has been later studied in order-preserving
submatrix problem (OPSM) [1, 51]. The authors propose a deter-
ministic greedy algorithm that seeks biclusters for which the col-
umns can be sorted in increasing order for all rows in the bicluster.
Although additive and multiplicative biclusters can be captured by
OPSM algorithm, it fails to capture constant biclusters. Surveys on
biclustering of gene expression data, the proposed algorithms, and
their evaluation via bicluster validation from a biological point of
view can be found in [3, 45, 52–56].

2.1 PCC-Based PCC is a measure that evaluates positive and negative linear rela-
Biclustering tionships between vectors. It is commonly used in clustering gene
expression data [2, 4] due to its power in capturing both shifting
and scaling patterns. For a PCC-based biclustering on gene expres-
sion dataset, the correlation of two genes is calculated on some
specified columns since those genes may or may not be correlated
on every experiment. Therefore, our PCC-based similarity measure
between rows r and s on selected columns Y is calculated with:
 
X 
 
 ðr  r Þðs  s Þ 
i∈Y i i

pccðr, s, Y Þ ¼ rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X X , ð1Þ
ðr i  r Þ2 ðs i  s Þ2
i∈Y i∈Y

where the equation runs on select columns, and the absolute value
of the expression gives a result in [0, 1] interval.
PCC-based biclustering was recently proposed in [9, 57]. In [57],
the authors present the bi-correlation clustering algorithm (BCCA),
which tries to find biclusters using Pearson correlation. They also
discuss the complexity of computing pairwise PCCs, and the ineffi-
ciency of the method. Bozdağ et al. [9] discuss potential complexity
issues of an exhaustive search using PCC, and propose that, instead of
computing all pairwise PCC values, a center-like vector (tendency
vector) is sufficient and more efficient at finding correlated rows.

3 Correlated Pattern Biclusters

Given a query gene and a set of microarray datasets, we compute a

ranked list of co-regulated genes in three steps. Here we give the
details of these steps: In the first step, the CPB algorithm recovers
a set of biclusters (Section 3.1). In the next step, we filter out
statistically insignificant biclusters (Section 3.2). Finally, the corre-
lation scores gathered from different datasets (Section 3.3). The
overview of the framework is given in Fig. 2.
60 Mehmet Deveci et al.

Fig. 2 Overview of the proposed framework

3.1 The CPB Let R and C denote the set of rows and columns of a data matrix A,
Algorithm respectively. Each element arc ∈ A represents the relation between
row r and column c. A bicluster B ¼ (X, Y ) is a subset of rows
X ¼ { x1, . . ., xn} and a subset of columns Y ¼ { y1, . . ., ym}, where
n  N, and m  M.
Definition 1 (Correlated Pattern Biclusters Algorithm). Given a data
matrix A, reference row rr, PCC threshold ρ, and minimum number of
columns γ, CPB finds a bicluster B ¼ (X, Y) such that rr ∈ X, m  γ,
8x i , x j ∈X pccðx i , x j , Y Þ  ρ.

CPB starts with an initial bicluster B ¼ (X, Y ) and improves it

by iteratively moving rows and columns in and out of the bicluster
using a search technique similar to local search methods. Algo-
rithm 1 outlines the proposed biclustering algorithm. Important
steps, i.e., generation of the initial biclusters, computing tendency
vector T and normalization parameters, updating rows and columns
are described in detail in the following subsections.

Algorithm 1 Correlated Pattern Biclusters

1: function CPB(A, B, rr , w, γ, ρ )
2: B = (X, Y ) is an initial bicluster s.t. rr ∈ X
3: ρc ← 2/3ρ ; ρΔ = 1/12ρ ; γc = m; γΔ = m−γ 4
4: repeat
5: step ← 0
6: repeat
7: step ← step + 1; Bsave ← B
8: Compute T, αi , βi
9: if step mod 2 = 1 then
10: Update X such that
11: ∀xi ∈ X, pcc(xi , T, Y ) > ρc
12: else
13: Find row r with smallest pcc(r, T, Y ) > ρc
14: Update Y such that
15: ∀yk ∈ Y, ERROR(yk ) > ERROR(r)
16: end if
17: until step > 20 or B = Bsave
18: ρc ← ρc + ρΔ ; γc ← γc − γΔ
19: until ρc > ρ
20: return B = (X, Y )
21: end function
 Querying Co-regulated Genes on Diverse Gene Expression Datasets Via Biclustering 61

3.1.1 Generating Initial Selecting the rows and columns of the initial bicluster is important
Biclusters since the algorithm converges to a more stable one by adding and
removing rows and columns to this bicluster. In [9], initial biclus-
ters were chosen randomly, and the algorithm runs efficiently when
discovering small number of biclusters embedded in synthetic data-
sets. However, we observe that when there are multiple biclusters
this approach does not provide a consistent mechanism to return
multiple biclusters with good coverage of the whole dataset.
In CPB, we generate initial biclusters with a grid-based
approach. We first shuffle the row and column numbers of the
dataset, and then partition the dataset into a coarse-grain grid of
10  2 initial biclusters. The query gene rr is inserted into each
bicluster, if necessary. At the end, all genes and conditions in the
dataset are assigned to at least one initial bicluster. Repeating the
process gives us enough initial biclusters to find co-regulated genes
and corresponding conditions. In addition, different runs obtain
more than 75 % of the top-ranked co-regulated genes with the
grid-based initialization, even though the generation of the initial
biclusters is randomized.

3.1.2 Computing In order to avoid making pairwise comparisons of all rows, we

Normalization Parameters compute a tendency vector that represents an average of the rows
and Tendency Vector of the bicluster. We compute a normalized data value
a x i y k  αx i
ãx i y k ¼
βx i

for each xi ∈ X and yk ∈ Y, where αx i and βx i are shifting and

scaling parameters associated with row xi, respectively. Then, each
element tk of tendency vector T is computed as the arithmetic mean
of ãx i y k on all rows xi ∈ X.
To ensure that the reference row rr has a larger impact on
decision mechanisms of the algorithm, we assign a larger weight,
ω, to the reference row when computing the vector T. Total con-
tribution from rows except rr is multiplied by (1 ω) and contribu-
tion from rr is multiplied by ω, where ω is an input parameter. Large
values for ω allow discovering patterns that resemble rr more
closely, whereas small values reduce sensitivity, hence offer a higher
tolerance to noise. Therefore, if a reference row and ω specified, the
elements are calculated with
X
ω  ãx i y r r þ ð1  ωÞ  ãx i y k
k∈X fr r g
tk ¼ : ð2Þ
jX j

We compute T, αx i and βx i using an iterative process. Initially we

set αx i ¼ 0 and βx i ¼ 1, and compute T. Then, we apply least
squares fitting on pairs fðt 1 , a x i y 1 Þ, . . . , ðt m , a x i y m Þg to obtain the
best shifting and scaling parameters that maximize alignment of
62 Mehmet Deveci et al.

each row xi with the tendency vector T. We assign intercept and

slope obtained in least squares fitting to αx i and βx i , respectively. T is
updated using these parameters, and the process iterates until
convergence.

3.1.3 Updating the Rows For a row r to be included in X, we require pcc(r, xi, Y ) > ρ for all
of a Bicluster xi ∈ X. To avoid testing this condition against all xi ∈ X, we
utilize the tendency vector T, and only test whether pcc(r, T, Y) is
greater than another threshold ρ0 instead. ρ0 is selected such that
pcc(r, T, Y) > ρ0 must ensure pcc(r, xi, Y ) > ρ for all xi ∈ X.
However, PCC lacks transitivity property [58] and has a complex
formula that strongly depends on the values and the length of the
vectors. Although it is analytically difficult to compute a lower
bound for ρ0, it was empirically shown that there exists a lower
bound proportional to ρ [9].
In Algorithm 1, we start with a relaxed threshold and slowly
tighten it at Line 18. While tightening ρ0 , we relax the constraint on
minimum number of columns. This allows sweeping the search
space between two extreme combinations of these parameters.
The algorithm uses five tightening steps and initial values of
0 0
ρc ¼ 2=3ρ and γ c ¼ jY j (Line 3).

3.1.4 Updating the Using PCC to measure the coherence between the columns is too
Columns of a Bicluster restrictive. For example, although the rows in Fig. 1d are perfectly
correlated, Pearson correlation between columns is less than 1.
Therefore, we use root mean square error to assess the coherence
of the columns. It is computed as:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1X n
ERRORðy k Þ ¼ ðãx y  t k Þ2 , ð3Þ
n i¼1 i k

where yk ∈ Y and n ¼ jY j. For a column c 2 = Y, we compute

ERROR(c) in a similar way, by using a value tc analogous to tk
that quantifies tendency of rows xi ∈ X in column c.
In CPB, only the columns having ERROR below a threshold ε
are included in the bicluster. In order to have comparable ERROR
threshold for the column selection with respect to row addition, we
select ε in relation to ρ0 . To establish this relation, first we note that
ERROR can also be computed for rows, and it is a comparable
metric for rows and columns. For a row xi ∈ X, ERROR(xi)
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X m
is computed as 1
m ðãx i y k  t k Þ2 . Then, it is observed that
k¼1
ERROR(r) generally implies a high pcc(r, T, Y) [9]. Therefore,
by setting ε to the ERROR of row r that has the smallest
pcc(r, T, Y) above threshold ρ0 c (Line 13), we prevent the algorithm
from returning imbalanced biclusters (i.e., very small or very high
number of columns).
 Querying Co-regulated Genes on Diverse Gene Expression Datasets Via Biclustering 63

3.2 Filtering Any dataset contains small biclusters with a high Pearson correla-
Biclusters Found by tion value by random chance. Although we specify a lower bound
Random Chance for PCC ρ0 and minimum number of columns γ, especially when γ is
small, in addition to larger biclusters, CPB recovers such
small biclusters. To eliminate randomly found biclusters in a
non-parametric fashion, we developed following method. Suppose
 ¼ fB 1 , B 2 , . . . , B z g be the set of biclusters found by different
runs of CPB on a data matrix A. We first generate A0 by shuffling
the elements of A. Then, we find the bicluster Bmax with the highest
number of rows in A0 , and use its dimension n0 as a threshold to
filter biclusters in . Algorithm 2 summarizes the filtering process.
Note that the parameter n0 is unique for each dataset, but this
method empirically finds a lower bound for n0 . The more biclusters
generated from the shuffled dataset, the better the estimate of n0 .

Algorithm 2 Filter Random Biclusters.

1: function FilterRandomBiclusters(A, B, γ, ρ )
2: A ← Shuffle(A); B ← {}
3: for each row ri in A do
4: B ← B ∪ CPB(A , ri , 0.5, γ, ρ )
5: end for
6: n ← argmaxBi ∈B ni
7: for each bicluster Bi in B do
8: (ni , mi ) ← size[Bi ]
9: if ni ≤ n then
10: B ← B \ {Bi }
11: end if
12: end for
13: return B
14: end function

In addition to filtering out the statistically insignificant biclus-

ters, we also remove those that have substantial overlaps. For any
bicluster pair that has an overlap of 75 % or more, we remove the
smaller bicluster.

3.3 Combining CPB often produces different resulting biclusters due to the ran-
Correlation dom selection of initial biclusters. Information from these biclus-
Information ters, each including the reference row rr, is merged to score each
row’s relationship with rr.
In [9], bicluster uniqueness (BU) measure was proposed to
calculate the correlation score of the genes. Although BU is able
to capture the information redundancy caused by overlapping
biclusters, we present a similar but more efficient scoring function
to be used instead.
Let  ¼ fB 1 , B 2 , . . . , B z g be the set of biclusters found by
different runs of CPB on a data matrix A, and with reference row
rr. Suppose IR(r) and IC(c) denote the maximal subset of  that
contain the given row and column, respectively.
64 Mehmet Deveci et al.

Definition 2 (Correlation Score (CS)). A score is assigned to a row r

based on the number of experiments in which r is co-regulated with rr by:
X
CSðrÞ ¼ j I RðrÞ \ I CðcÞ j : ð4Þ
c∈C

To increase significance and consistency of our findings, we

apply our method on different datasets separately and combine
correlation scores. To achieve this in a meaningful way, we require
datasets to have the same row labels. In gene expression data
analysis, this requirement can be met by merging results only
from datasets obtained using the same microarray chip.
Definition 3 (Gene Correlation Score (GCS)). Let
 ¼ fA1 , A2 , . . . , Ap g be the set of (microarray) datasets with the same
row labels (genes). Given a reference row rr and datasets , gene
correlation score G C S of a row r is calculated with
X CSðrÞ
GCSðr, r r Þ ¼ : ð5Þ
A ∈
CSðr r Þ
v

4 Experimental Results

We test CPB on the probes of three tumor suppressor genes (i.e.,

BRCA1, BRCA2, and p53) as queries to reveal co-regulated genes
involved in the complex process of tumor formation. Experiments
on 40 large datasets, each with 22,283 probe sets, show that the
results are remarkably enriched for genes that have a role in cancer
progression, tumor growth and metastasis regulation, and DNA
degradation and repair. We also compare our results with another
query-based framework to see how successfully each method finds
known and unexplored genes co-regulated with tumor suppressors.
While the ratios of known cancer-related genes are similar, the
proposed framework finds more unexplored genes that are likely
to be missed by earlier clustering-based methods. CPB’s perfor-
mance for querying the microarray data is promising: it was able to
capture many genes that are highly correlated with BRCA1,
BRCA2, and p53. We observed that of those genes, half are already
known to be involved in cancer processes, and the others are
promising new candidates for further investigation.
We first define some evaluation metrics and test CPB on syn-
thetic datasets generated with biclusters with (1) different models,
(2) increasing noise levels, and (3) increasing overlaps between
embedded biclusters. We selected four other biclustering
algorithms; δ-biclusters [11], OPSM [1], BBC [17], and
BCCA [57], for comparison, due to their success at capturing
shift-scale biclusters.
We test CPB on a number of human microarray datasets using
four probes of breast cancer associated BRCA1 and BRCA2 genes,
 Querying Co-regulated Genes on Diverse Gene Expression Datasets Via Biclustering 65

and two probes of p53 tumor suppressor gene as queries. The

correlation scores of the genes are combined, and the top-ranked
genes are further studied. We also compare our results with MEM
framework [8] in terms of the algorithms’ effectiveness of retrieving
undiscovered cancer-related genes.

4.1 Experiments on We first define recovery and relevance metrics to evaluate the results
Synthetic Datasets of biclustering algorithms. For each experiment, a synthetic dataset
is generated with 1000 rows and 200 samples. Then two 60  60
biclusters with the given model are embedded into the dataset. The
average score of 100 replication of the same experiment is reported.

4.1.1 Evaluation Metrics Similar to recall and precision metrics, recovery and relevance
scores are proposed to evaluate the biclustering results. These
measures can be defined to compare a single found bicluster against
an expected one, as well as a set of found biclusters against a set of
expected ones.
Let e and f be expected and found biclusters, respectively. The
recovery score of a found bicluster against an expected one is
calculated by dividing the intersection area by the area of the
expected bicluster:
je\f j
recðe, f Þ ¼ , ð6Þ
jej
where the recovery score reaches to 1 if and only if e  f .
Similarly, relevance score is calculated by dividing the intersec-
tion area by the area of the found bicluster:
je\f j
relðe, f Þ ¼ , ð7Þ
jf j
where the relevance score reaches to 1 if and only if f  e. Examples
of how these scores are computed are given in Fig. 3.
Using these rec and rel measures, we define recovery and rele-
vance scores to compare two sets. Let E and F be a set of expected

Fig. 3 Example expected/found biclusters with their recovery and relevance scores
66 Mehmet Deveci et al.

and found biclusters, respectively. The set-based recovery score is

calculated by taking the mean of the maximum recovery score for
each expected bicluster. An equivalent approach is used for
relevance.
1 X
RECðE, F Þ ¼ max recðe, f Þ ð8Þ
j E j e∈E f ∈F

1 X
RELðE, F Þ ¼ max relðe, f Þ ð9Þ
j F j f ∈F e∈E

4.1.2 Effects of the Biclustering methods often focus on detecting specific types of
Bicluster Model biclusters, as mentioned in Background section. In this experiment,
we compare the success rate of CPB with other algorithms on
detecting biclusters generated with various models. Constant-row,
shift, scale, and shift-scale models were chosen for this experiment.
Examples of these models are given in Fig. 1.
The resulting recovery and relevance scores (see Fig. 4a–d)
show that CPB is the only algorithm that can fully recover biclusters
generated with all four models with a high relevance score. BBC
was able to find shifted and constant-row biclusters with a slightly
lower relevance score. BCCA was expected to display similar results
to CPB since they both use Pearson correlation; however, it was
only able to fully recover shift biclusters. OPSM could not identify
shift, scale, or shift-scale biclusters although they are all valid order-
preserving submatrices. Our experiments show that when a base
row is scaled with a value between 1 and 1, the expression rank-
ings of the columns of a bicluster row lie in a narrow range along
the row; therefore, OPSM fails to discover it. Despite this limita-
tion, OPSM was able to identify one of the shifted biclusters, since
it reports a single bicluster for each size of column. The δ-biclusters
algorithm performed poorly on all the datasets, among which it can
partially recover only constant-row biclusters. The other models are
not captured by the metric, which was previously discussed in [53].
For the noise and overlap experiments, CPB and other meth-
ods were compared on shift biclusters since the shift model is
successfully recovered by most of the algorithms (see Fig. 4b).

4.1.3 Effects of the Noise Microarrays results are perturbed by many sources of noise. In
order to measure the sensitivity of CPB to noise, an error value ε
was added to each element of the synthetic datasets. The experi-
ments were run with various noise levels: each error value was
drawn from a normal distribution with zero mean and variance
equal to the chosen noise level.
Figure 4e shows the recovery and relevance scores of the algo-
rithms on datasets with varying noise levels. OPSM is dramatically
affected by noise since it may violate the order-preserving structure.
 Querying Co-regulated Genes on Diverse Gene Expression Datasets Via Biclustering 67

a b c d

Fig. 4 Experiments on synthetic datasets: (a–d) with different bicluster models, (e) under noise, and (g) with
overlapping biclusters
68 Mehmet Deveci et al.

Since BCCA checks for pairwise correlation score for each row, this
method is more likely to be affected by the increasing noise.
Although BBC seems to be insensitive to noise in recovery plot,
its relevance score drops slightly with noise addition (see Fig. 4e).
CPB is the second best algorithm that is resistant to noise even
though it has a linear metric. Moreover, the noise resistance of CPB
can be improved by adjusting a better PCC threshold. We fixed
ρ ¼ 0. 9 in order to be consistent with the rest of the experiments.
We also experimented with relative noise, in which the noise is
added to each element with respect to its expression value, i.e.,
element x becomes x + x ε. We observed results similar to the
previous experiment.

4.1.4 Effects A gene may take roles in several functions in a cell, each of which
of the Overlap may be occurring simultaneously in a given sample; therefore, there
might be overlaps between biclusters. In this experiment we test
how CPB and other algorithms perform with increasing overlaps of
biclusters. The datasets are generated with two overlapping biclus-
ters. The overlapping regions of these biclusters are increased by 10
rows and 10 columns at each step. The expression values in the
these regions are not assumed to be additive; instead, shift values
for rows and base vector are chosen in a way to allow both of the
biclusters to have the same expression value at overlapping regions.
Figure 4f shows the results of the overlap test. We observe that
CPB and BCCA are both insensitive to increasing overlap, while
BCCA fails to recover a very small portion of the biclusters. BBC is
affected more than BCCA in terms of recovery; also, its relevance
score drops with increasing overlap. Although OPSM recovers only
one of the biclusters, it increases its recovery score by including
more of the overlapping region with increasing overlap.

4.2 Identifying Genes In this experiment, we employ CPB to identify the most correlated
Co-regulated with genes with BRCA1, BRCA2, and p53, which are highly penetrant
BRCA1, BRCA2, p53 cancer specific tumor suppressors. CPB was run on 40 different
datasets obtained from the GPL96 series (GDS{1064, 1284, 1615,
2113, 2362, 2649, 2954, 3116, 3312, 3716, 1067, 1329, 1815,
2190, 2373, 2736, 3057, 3128, 3471, 534, 1209, 1375, 1956,
2255, 2519, 2767, 3096, 3233, 3514, 596, 1220, 1479, 1975,
2297, 2643, 2771, 3097, 3257, 3517, 987}), all of which have the
same set of probes. The results of each dataset are then combined
with Gene Correlation Score function.
Table 1 gives the top-ranked genes for the probes of BRCA1,
BRCA2, and p53. We observe that more than 50 % of the
genes found by our framework are already investigated in cancer
research, suggesting that CPB is indeed finding genes involved
with cancer.
 Querying Co-regulated Genes on Diverse Gene Expression Datasets Via Biclustering 69

Table 1
Associated top-ranked genes for 6 probes of BRCA1, BRCA2, and p53

BRCA1 BRCA2 p53

204531_s_at 211851_x_at 208368_s_at 214727_at 201746_at 211300_s_at

1 C1orf105 H49077 CHRNA4 C1orf105 C1orf105 C1orf105
2 GPR98 C1orf105 C1orf105 ACRV1 H49077 PCNXL2
3 H49077 ARID4B MTMR8 GFRA4 GFRA4 GFRA4
4 CHRNA4 MTMR8 ACRV1 MTMR8 ARID4B U88898
5 SLC17A1 MBD2 GFRA4 AGTR2 MTMR8 H49077
6 GPX5 AK022006 H49077 PRO2958 AK022006 CHRNA4
7 PCNXL2 SLC17A1 PCNXL2 H49077 UBQLN3 ACRV1
8 MTMR8 CSRP3 PRO2958 ACRV1 ACRV1 MTMR8
9 ARID4B PRO2958 U88898 U88898 GNRHR PPP3CC
10 GFRA4 GFRA4 SLC17A1 SLC17A1 CHRNA4 NKX3-1
11 MBD2 NOS1 NKX3-1 GNRHR GPR98 GNRHR
12 IL17A PPP3CC ACRV1 PPP3CC RNF185 ACRV1
13 AK022006 CHRNB3 PPP3CC MBD2 SLC17A1 MBD2
14 NOS1 MAPK11 AGTR2 SPINLW1 KLK10 GNPTAB
15 ALPI ACRV1 FAM55D GNPTAB U05589 CHRNB3
16 KLK10 GNPTAB SNX1 AK000787 GPX5 ACRV1
17 PRO2958 IL17A TREX2 AK023690 OR7C2 AK000787
18 CSRP3 NEK1 GPR98 OR5I1 P2RY4 GPR98
19 PPP1R3A U05589 LEP BTNL8 PPP3CC SPINLW1
20 ACRV1 IFNA5 M78162 EPAG PCNXL2 AL162044
21 GNPTAB GPX5 CSRP3 C22orf33 AK023690 ARID4B
22 AL117549 GNA11 KRT38 RPS6KA6 TREX2 ACRV1
23 NEK1 KLK10 TBR1 GPR98 PRO2958 AW139195
24 AK023690 GJB3 EPAG MYL1 RBMY2FP GPX5
25 MYL1 SLC7A11 MBD2 PCNXL2 IL17A AK022006
Genes that have cancer-related studies found in PubMed are highlighted

Among those, for example, MBD2 is shown to have a role in

cancer progression and can be therapeutically targeted in aggressive
breast cancers [59]. KLK10 provides important prognostic infor-
mation in early breast cancer patients [60]. CHRNA4 polymorph-
isms are found to activate factors that participate in DNA
70 Mehmet Deveci et al.

degradation and repair, specifically the level of p53 participating in

DNA repair [61].
Some genes, highly correlated with p53 in the list, are also
investigated in other cancer types. For instance, 4 messenger
RNA biomarkers, including ACRV1, may differentiate pancreatic
cancer patients from noncancer subjects [62]. GFRA4 is predomi-
nantly expressed in normal and malignant thyroid medullary
cells [63]. Of those, which are also correlated with BRCA1 and
BRCA2 genes, can be further studied to see if they are up- or down-
regulated in cancer patients.
The genes that have not already appeared in the literature may
play an as-yet unknown role in cancer-related processes. These
genes are open for further research.

4.3 Comparison with There are a limited number of studies on query-based discovery of
Other Query-Based co-regulated genes in the literature. Gene recommender [5] ana-
Frameworks lyzed Rb protein complex to find new co-regulated genes in worms
(specifically C. elegans) using a technique similar to biclustering.
SPELL [6], a PCC-based clustering framework was tested on S.
cerevisiae datasets, where several genes were categorized and anno-
tated. A probabilistic biclustering framework, QDB [7], was tested
on some synthetic and yeast microarray datasets. Adler et al. [8]
propose a query engine (MEM) to search for correlated genes
across many datasets. Zhao et al. proposed ProBic [10], a probabi-
listic biclustering algorithm, and tested on E. coli to detect high
quality biclusters in the presence of noise.
Among those query-driven search methods, we could compare
our framework on cancer-related genes with only MEM [8],
because the other studies are either specialized in non-human
organisms, or resource is not accessible. Using MEM framework,
we retrieved the genes correlated with the selected probes of
BRCA1, BRCA2, and p53 genes with Pearson correlation. Top-
ranked genes are then investigated to find whether the gene is
claimed to be cancer-related in a research study in medical
literature.
In Table 2 we compare our results with MEM’s based on how
successful each method is on finding known and unexplored genes.
Since all samples (columns) of a microarray dataset were included in
similarity calculations before biclustering, the top-ranked genes
discovered by MEM are expected to be investigated before. While
the ratios of known cancer-related genes are similar, we argue that
our framework finds more unexplored genes that are likely to be
missed by earlier clustering-based methods.
Although PCC is the default similarity measure, we also run
MEM with absolute Pearson correlation, which is expected to
capture negative correlations as in our pcc function (see Eq. (1)).
However, the results on six probes between two runs of MEM
framework are 87 % overlapping. Absolute PCC could only
 Querying Co-regulated Genes on Diverse Gene Expression Datasets Via Biclustering 71

Table 2
Ratio of known and unexplored genes found within top-25 results of MEM (with Pearson and absolute
Pearson correlation) and our framework

BRCA1 BRCA2 p53

204531 211851 208368 214727 201746 211300 Average

_s_at (%) _x_at (%) _s_at (%) _at (%) _at (%) _s_at (%) (%)
MEM Known 56 64 64 64 68 68 64.0
Pearson New 32 16 28 32 24 24 26.0
Duplicate 12 20 8 4 8 8 10.0
MEM Known 64 64 72 68 76 72 69.3
jPearsonj New 32 16 24 32 20 20 24.0
Duplicate 4 20 4 0 4 8 6.6
CPB Known 60 64 52 32 48 48 50.6
New 40 36 44 64 52 40 46.0
Duplicate 0 0 4 4 0 12 3.3
Only the unique gene names are considered. Probes of the same gene after the first one are counted as redundant
(duplicate) information

introduce 20 new genes, where 60 % of them are already found to

be cancer-related. We conclude that altering the similarity measure
(in this case, taking the absolute value to capture negative correla-
tions) is not as effective for finding undiscovered correlated genes as
applying biclustering.

5 Conclusion

In this chapter, we briefly survey the biclustering algorithms in the

literature and introduce a method for querying co-regulated genes
using a novel biclustering method, the CPB. Initial testing on
artificial data confirms that CPB is capable of finding such biclusters
and that it outperforms other biclustering methods in finding
multiple types of biclusters. CPB’s performance for querying the
microarray data is promising: it finds many genes that have high
correlations with BRCA1, BRCA2, and p53. Of those genes, half
are already known to be involved in cancer processes, and the others
are promising new candidates for further investigation.
There are many possible extensions to CPB that may yet be
explored. For instance, PCC is only one of the well-known metrics
for evaluating similarity. CPB approach may be extended to use
other metrics and benefit from their unique properties. CPB’s
iterative optimization process may likewise be improved by
72 Mehmet Deveci et al.

choosing initial biclusters differently or using a mathematical opti-

mization method to avoid the local maximas.

6 Acknowledgments

This work was supported in parts by National Institutes of Health/

National Cancer Institute (grant number R01CA141090); by the
Department of Energy (grant number DE-FC02-06ER2775); and
by the National Science Foundation (grants numbers CNS-
0643969, OCI-0904809, OCI-0904802).

References

1. Ben-Dor A, Chor B, Karp R, Yakhini Z (2002) 8. Adler P, Kolde R, Kull M, Tkachenko A, Peter-
Discovering local structure in gene expression son H, Reimand J, Vilo J (2009) Mining for
data: The order-preserving submatrix problem. coexpression across hundreds of datasets using
In: Proceedings of the International Confer- novel rank aggregation and visualization meth-
ence on Computational Biology, pp 49–57 ods. Genome Biol 10:R139
2. Jiang D, Pei J, Zhang A (2003) DHC: a 9. Bozdağ D, Parvin JD, Çataly€ € (2009) A
urek UV
density-based hierarchical clustering method biclustering method to discover co-regulated
for time series gene expression data. In: Pro- genes using diverse gene expression datasets.
ceedings IEEE Symposium on BioInformatics In: Proceedings of 1st International Confer-
and Bioengineering, pp 393–400 ence on Bioinformatics and Computational
3. Madeira SC, Oliveira AL (2004) Biclustering Biology, pp 151–163
algorithms for biological data analysis: a survey. 10. Zhao H, Cloots L, Van den Bulcke T, Wu Y, De
IEEE/ACM Trans Comput Biol Bioinform 1 Smet R, Storms V, Meysman P, Engelen K,
(1):24–45 Marchal K (2011) Query-based biclustering
4. Pujana MA, Han J-DJ, LM Starita, Stevens of gene expression data using probabilistic rela-
KN, Tewari M, Ahn JS, Rennert G, Moreno tional models. BMC Bioinf 12(Suppl 1):S37
V, Kirchhoff T, Gold B, Assmann V, ElShamy 11. Cheng Y, Church GM (2000) Biclustering of
WM, Rual J-F, Levine D, Rozek LS, Gelman expression data. In: Proceedings of Interna-
RS, Gunsalus KC, Greenberg RA, Sobhian B, tional Conference on Intelligent Systems for
Bertin N, Venkatesan K, Ayivi-Guedehoussou Molecular Biology, pp 93–103
N, Sole X, Hernandez P, Lazaro C, Nathanson 12. Segal E, Taskar B, Gasch A, Friedman N, Koller D
KL, Weber BL, Cusick ME, Hill DE, Offit K, (2001) Rich probabilistic models for gene expres-
Livingston DM, Gruber SB, Parvin JD, Vidal sion. Bioinformatics 17(suppl_1):S243–S252
M (2007) Network modeling links breast can- 13. Wang H, Wang W, Yang J, Yu PS (2002) Clus-
cer susceptibility and centrosome dysfunction. tering by pattern similarity in large data sets. In:
Nat Genet 39(11):1338–1349 Proceedings of ACM SIGMOD
5. Owen AB, Stuart J, Mach K, Villeneuve AM, 14. Lazzeroni L, Owen A (2000) Plaid models for
Kim S (2003) A gene recommender algorithm gene expression data. Tech. Rep., Stanford
to identify coexpressed genes in C. elegans. University
Genome Res 13(8):1828–1837
15. Mitra S, Banka H (2006) Multi-objective evo-
6. Hibbs MA, Hess DC, Myers CL, Huttenhower lutionary biclustering of gene expression data.
C, Li K, Troyanskaya OG (2007) Exploring the Pattern Recognit 39(12):2464–2477
functional landscape of gene expression:
directed search of large microarray compendia. 16. Mejı́a-Roa E, Carmona-Saez P, Nogales R,
Bioinformatics 23:2692–2699 Vicente C, Vázquez M, Yang XY, Garcı́a C,
Tirado F, Pascual-Montano A (2008)
7. Dhollander T, Sheng Q, Lemmens K, De bioNMF: a web-based tool for nonnegative
Moor B, Marchal K, Moreau Y (2007) matrix factorization in biology. Nucleic Acids
Query-driven module discovery in microarray Res 36(suppl 2):W523–W528
data. Bioinformatics 23:2573–2580
 Querying Co-regulated Genes on Diverse Gene Expression Datasets Via Biclustering 73

17. Gu J, Liu JS (2008) Bayesian biclustering of 29. Liu J, Wang J, Wang W (2004) Gene ontology
gene expression data. BMC Genomics 9(Suppl friendly biclustering of expression profiles. In:
1):S4 Proceedings of IEEE Computational Systems
18. Hochreiter S, Bodenhofer U, Heusel M, Mayr Bioinformatics Conference, pp 436–447.
A, Mitterecker A, Kasim A, Khamiakova T, Van IEEE Computer Society
Sanden S, Lin D, Talloen W et al (2010) Fabia: 30. Madeira S, Oliveira A (2005) A linear time
factor analysis for bicluster acquisition. Bioin- biclustering algorithm for time series gene
formatics 26(12):1520–1527 expression data. In: Casadio R, Myers G (eds)
19. Painsky A, Rosset S (2012) Exclusive row Algorithms in bioinformatics. Lecture Notes in
biclustering for gene expression using a combi- Computer Science, vol 3692, pp 39–52,
natorial auction approach. In: Proceedings of Springer, Berlin/Heidelberg
the 2012 I.E. 12th International Conference 31. Pontes B, Giraldéz R, Aguilar-Ruiz JS (2013)
on Data Mining, pp 1056–1061. IEEE Com- Configurable pattern-based evolutionary
puter Society biclustering of gene expression data. Algo-
20. Joung J-G, Kim S-J, Shin S-Y, Zhang B-T rithms Mol Biol 8:4
(2012) A probabilistic coevolutionary biclus- 32. Yang W-H, Dai D-Q, Yan H (2011) Finding
tering algorithm for discovering coherent pat- correlated biclusters from gene expression
terns in gene expression dataset. BMC Bioinf data. IEEE Trans Knowl Data Eng
13(Suppl 17):S12 23:568–584
21. Flores JL, Inza I, Larrañaga P, Calvo B (2013) 33. Yoon S, Nardini C, Benini L, De Micheli G
A new measure for gene expression biclustering (2005) Discovering coherent biclusters from
based on non-parametric correlation. Comput gene expression data using zero-suppressed
Methods Prog Biomed 112(3):367–397 binary decision diagrams. IEEE/ACM Trans
22. Sun P, Speicher NK, Röttger R, Guo J, Baum- Comput Biol Bioinf 2:339–354
bach J (2014) Bi-force: large-scale bicluster 34. Angiulli F, Cesario E, Pizzuti C (2008) Ran-
editing and its application to gene expression dom walk biclustering for microarray data. Inf
data biclustering. Nucleic Acids Res. doi:10. Sci 178(6):1479–1497
1093/nar/gku201 35. Bryan K (2005) Biclustering of expression data
23. Chakraborty A (2005) Biclustering of gene using simulated annealing. In: Proceedings of
expression data by simulated annealing. In: the 18th IEEE Symposium on Computer-
Proceedings of Eighth International Confer- Based Medical Systems, CBMS’05, (Washing-
ence on High-Performance Computing in ton, DC, USA), pp 383–388. IEEE Computer
Asia-Pacific Region, 2005, pp 627–632 Society
24. Liew AW-C, Law N-F, Yan H (2011) Recent 36. Bryan K, Cunningham P, Bolshakova N (2006)
patents on biclustering algorithms for gene Application of simulated annealing to the
expression data analysis. Recent Pat DNA biclustering of gene expression data. Trans Inf
Gene Seq 5(2):117–125 Tech Biomed 10:519–525
25. Hussain SF (2011) Bi-clustering gene expres- 37. Bleuler S, Prelic A, Zitzler E (2004) An EA
sion data using co-similarity. In: Proceedings of framework for biclustering of gene expression
the 7th International Conference on Advanced data. In: Congress on Evolutionary Computa-
Data Mining and Applications - Volume Part I, tion, 2004 (CEC2004), vol 1, pp 166–173
ADMA’11, pp 190–200. Springer, Berlin/ 38. Divina F, Aguilar-Ruiz J (2006) Biclustering of
Heidelberg expression data with evolutionary computation.
26. An J, Liew AW-C, Nelson CC (2012) Seed- IEEE Trans Knowl Data Eng 18:590–602
based biclustering of gene expression data. 39. Nepomuceno JA, Troncoso A, Aguilar-Ruiz JS
PLoS ONE 7:e42431, 08 (2010) Correlation-based scatter search for
27. Kiraly A, Abonyi J, Laiho A, Gyenesei A (2012) discovering biclusters from gene expression
Biclustering of high-throughput gene expres- data. In: Proceedings of the 8th European
sion data with bicluster miner. In: IEEE 12th Conference on Evolutionary Computation,
International Conference on Data Mining Machine Learning and Data Mining in Bioin-
Workshops (ICDMW), 2012, pp 131–138 formatics, EvoBIO’10, pp 122–133. Springer,
28. Liu J, Wang J, Wang W (2004) Biclustering in Berlin/Heidelberg
gene expression data by tendency. In: Proceed- 40. Nepomuceno JA, Troncoso A, Aguilar-Ruiz JS
ings of IEEE Computational Systems Bioinfor- (2011) A comparative analysis of biclustering
matics Conference, pp 182–193. IEEE algorithms for gene expression data. BioData
Computer Society Mining 4:3
74 Mehmet Deveci et al.

41. Erten C, Sözdinler M (2009) Biclustering Algorithms. In: ACM International Confer-
expression data based on expanding localized ence on Bioinformatics and Computational
substructures. In: Rajasekaran S (ed) Bioinfor- Biology
matics and computational biology. Lecture 54. Chia BKH, Karuturi RKM (2010) Differential
Notes in Computer Science, vol 5462, co-expression framework to quantify goodness
pp 224–235. Springer, Berlin/Heidelberg of biclusters and compare biclustering algo-
42. Tanay A, Sharan R, Shamir R (2002) Discover- rithms. Algorithms Mol Biol 5(1):8
ing statistically significant biclusters in gene 55. Eren K, Deveci M, K€ uç€
uktunç O, Çataly€urek
expression data. Bioinformatics 18(Supple- € (2012) A comparative analysis of bicluster-
UV
ment 1):136–144 ing algorithms for gene expression data. Brief
43. Bergmann S, Ihmels J, Barkai N (2003) Itera- Bioinform
tive signature algorithm for the analysis of 56. Oghabian A, Kilpinen S, Hautaniemi S, Czei-
large-scale gene expression data. Phys Rev E zler E (2014) Biclustering methods: Biological
Stat Nonlinear Soft Matter Phys 67:031902 relevance and application in gene expression
44. Kluger Y, Basri R, Chang JT, Gerstein M analysis. PloS one 9(3):e90801
(2003) Spectral biclustering of microarray 57. Bhattacharya A, De RK (2009) Bi-correlation
data: coclustering genes and conditions. clustering algorithm for determining a set of
Genome Res 13(4):703–716 co-regulated genes. Bioinformatics 25
45. Prelić A, Bleuler S, Zimmermann P, Wille A, (21):2795–2801
B€uhlmann P, Gruissem W, Hennig L, Thiele L, 58. Casella G, Wells MT (1993) Is Pitman close-
Zitzler E (2006) A systematic comparison ness a reasonable criterion: comment. J Am
and evaluation of biclustering methods for Stat Assoc 88(421):70–71
gene expression data. Bioinformatics 59. Mian O, Wang S, Zhu S, Gnanapragasam M,
22:1122–1129 Graham L, Bear H, Ginder G (2011) Methyl-
46. Li G, Ma Q, Tang H, Paterson AH, Xu Y binding domain protein 2-dependent prolifer-
(2009) QUBIC: a qualitative biclustering algo- ation and survival of breast cancer cells. Mol
rithm for analyses of gene expression data. Cancer Res 9(8):1152–62
Nucleic Acids Res 37(15):e101 60. Kioulafa M, Kaklamanis L, Stathopoulos E,
47. Huttenhower C, Mutungu KT, Indik N, Yang Mavroudis D, Georgoulias V, Lianidou ES
W, Schroeder M, Forman JJ, Troyanskaya OG, (2009) Kallikrein 10 (KLK10) methylation as
Coller HA (2009) Detailing regulatory net- a novel prognostic biomarker in early breast
works through large scale data integration. cancer. Ann Oncol 20:1020–1025
Bioinformatics 25:3267–3274 61. Dorszewska J, Florczak J, Rozycka A,
48. Voggenreiter O, Bleuler S, Gruissem W (2012) Jaroszewska-Kolecka J, Trzeciak WH,
Exact biclustering algorithm for the analysis of Kozubski W (2005) Polymorphisms of the
large gene expression data sets. BMC Bioinf 13 CHRNA4 gene encoding the alpha4 subunit
(Suppl 18):A10 of nicotinic acetylcholine receptor as related to
49. Bryan K, Cunningham P (2006) Bottom-up the oxidative DNA damage and the level of
biclustering of expression data. In: IEEE Sym- apoptotic proteins in lymphocytes of the
posium on Computational Intelligence and patients with Alzheimer’s disease. DNA Cell
Bioinformatics and Computational Biology, Biol 24:786–794
2006 (CIBCB ’06), pp 1–8 62. Zhang L, Farrell JJ, Zhou H, Elashoff D, Akin
50. Murali T, Kasif S (2003) Extracting conserved D, Park N-H, Chia D, Wong DT (2010) Sali-
gene expression motifs from gene expression vary transcriptomic biomarkers for detection of
data. Pac Symp Biocomput 8:77–88 resectable pancreatic cancer. Gastroenterology
51. Liu J, Wang W (2003) Op-cluster: clustering 138(3):949–957, e1–7
by tendency in high dimensional space. In: 63. Lindahl M, Poteryaev D, Yu L, Arumae U,
Proceedings of IEEE International Conference Timmusk T, Bongarzone I, Aiello A, Pierotti
on Data Mining, p 187 MA, Airaksinen MS, Saarma M (2001) Human
52. Freitas AV, Ayadi W, Elloumi M, Oliveira J, glial cell line-derived neurotrophic factor
Oliveira J, Hao J-K (2013) Survey on bicluster- receptor alpha 4 is the receptor for persephin
ing of gene expression data, pp 591–608. and is predominantly expressed in normal and
Wiley, New York malignant thyroid medullary cells. J Biol Chem
53. Bozdağ D, Kumar A, Çataly€ € (2010)
urek UV 276:9344–9351
Comparative Analysis of Biclustering
Methods in Molecular Biology (2016) 1375: 75–89
DOI 10.1007/7651_2015_237
© Springer Science+Business Media New York 2015
Published online: 11 April 2015

MetaMirClust: Discovery and Exploration of Evolutionarily

Conserved miRNA Clusters
Wen-Ching Chan and Wen-chang Lin

Abstract
Recent emerging studies suggest that a substantial fraction of microRNA (miRNA) genes is likely to form
clusters in terms of evolutionary conservation and biological implications, posing a significant challenge for
the research community and shifting the bottleneck of scientific discovery from miRNA singletons to
miRNA clusters. In addition, the advance in molecular sequencing technique such as next-generation
sequencing (NGS) has facilitated researchers to comprehensively characterize miRNAs with low abundance
on genome-wide scale in multiple species. Taken together, a large scale, cross-species survey of grouped
miRNAs based on genomic location would be valuable for investigating their biological functions and
regulations in an evolutionary perspective. In the present chapter, we describe the application of effective
and efficient bioinformatics tools on the identification of clustered miRNAs and illustrate how to use the
recently developed Web-based database, MetaMirClust (http://fgfr.ibms.sinic.aedu.tw/MetaMirClust) to
discover evolutionarily conserved pattern of miRNA clusters across metazoans.

Keywords: MetaMirClust, microRNA cluster, Data mining, Synteny

1 Introduction

MicroRNAs (miRNAs) are, of 21–23 nucleotides (nt) long in their

mature forms, a recently identified class of endogenous small non-
coding RNA molecules, which play important roles in gene regula-
tion via the RNA interference pathway (1–4). In 1993, when the
first miRNA lin-4 was identified in Caenorhabditis elegans, the
negative regulation pair between lin-4 and its target lin-14 was
thought as an individual case (5). As a result, miRNAs have not
gained the attention of researchers until a second similar system of
let-7 was observed (6), and then its homologous transcripts were
extensively investigated in animal and plant genomes. In these two
decades, a considerable body of evidence suggests that miRNAs
play important gene-regulatory roles related to organism develop-
ment, cell differentiation, and tumor progression and oncogenesis
(7–11). Currently, newly discovered miRNA genes either by exper-
imental or computational approaches have steadily increased as
evident by the amount of records in the miRBase registry (12)
and other resources (13, 14). In recent years, many studies have
75
76 Wen-Ching Chan and Wen-chang Lin

attempted to gain insights into the biogenesis, expression,

targeting, and evolution of individual miRNA gene in different
species. Some well-studied examples in human are, for instance,
mir-34b and mir-129 which serve as tumor-suppressor miRNAs
connected to DNA methylation-associated silencing in gastric can-
cer (10); mir-196a is overexpressed in primary gastric cancer tissues
compared to adjacent normal ones (9); three individual loci of mir-
9 are simultaneously hypermethylated in gastric cancer and are
likely to serve as tumor suppressive miRNAs (8). Correspondingly,
a substantial amount of literature has demonstrated miRNAs as
crucial negative regulators in diverse physiological and develop-
mental processes at the posttranscriptional level (15).
Up to date, a handful of miRNA clusters have been reported in
animal genomes. To the best of our knowledge, Altuvia et al. was
the first group that identified conserved regions of miRNA clusters
systematically (16). Then, Yu et al. adopted the same method to
enlarge the extent of conserved miRNA cluster (17), and thus
checked the expression profile of identified human miRNA clusters.
Accumulating studies have illustrated that clustered miRNA genes
located on polycistronic transcripts might be expressed at similar
levels and coordinately involve in an intricate regulatory network.
These miRNA clusters are usually derived from polycistrons within
the length from few hundred nucleotides to almost million base
pairs (18–21). For instance, mir-17 cluster and its paralogous clus-
ters are one of the well-studied cases. In 2004, Tanzer et al. have
tried to reconstruct the phylogenetic evolution of mir-17 cluster
family mainly in nine metazoan genomes and have revealed at least
three paralogous clusters related to the mir-17 cluster family, which
are mir-17-92, mir-106-92, and mir-106-25, and governed by tan-
dem duplications (22).
A growing range of studies has further demonstrated that the
aberrant expression of miRNAs in cluster families plays an impor-
tant role in cancer oncogenesis and metastasis (23–25). In addition
to the known function of mir-92a as negative regulator of angio-
genesis, an overexpression pattern of the mir-17-92 cluster
(13q31.3) comprising seven miRNAs has been discovered in 19
lung cancer cell lines (26). In renal cell carcinoma (RCC), the
restoration of the downregulated mir-143/145 cluster (5q32) in
RCC cells revealed significant inhibition of cancer cell proliferation
and invasion via a putative target gene, hexokinase-2 (HK2) (27).
In bladder cancer (BC), five downregulated clusters: mir-1/133a
(18q11.2 and 20q13.33), mir-206/133b (p12.2), let-7c/mir-99a
(21q21.1), mir-143/145 (5q32), and mir-195/497 (17p13.1),
were identified from 950 candidates by the genome-wide miRNA
expression signature analysis, and the following transfection assay
of mir-195/497 into BC cell lines has confirmed their function as
tumor suppressors in BC (25). It is believed that miRNAs in
clusters might represent putative bifunctional regulators, of which
 miRNA Cluster Discovery in Metazoan Genomes 77

miRNAs in high expression level can act as oncogenes by repressing

tumor suppressors, and when in low level they can turn over to
behave as tumor suppressors through a negative regulation of
oncogenes (28). Although the entire regulatory mechanisms of
clustered miRNA genes remain largely uncharacterized, it is likely
that these miRNA clusters may function more efficiently in a com-
plicated miRNA-mediated network than individual miRNA alone
(29). Therefore, identification of evolutionary conserved miRNA
clusters is an important first step for the research society toward
elucidating miRNA-cluster-mediated pathways in cancer research
and might provide new insights into the potential miRNA-based
therapeutics for cancer.
Many resources were developed to investigate miRNA genes.
However, only a handful of resources dedicate to an efficient and
extensive investigation of miRNA clusters (20, 21). Generally,
miRNA clusters were arbitrarily defined by a fixed distance (e.g.,
10 Kb) (12), and only few studies systematically investigating the
conservation patterns of clustered miRNA genes across metazoan
species (20). Here, we illustrate the synergistic potential of Meta-
MirClust and miRBase for exploring miRNA clusters conserved
across species in evolution.
The remainder of the chapter is organized as follows. First, the
Materials section highlights the technical prerequisites for the iden-
tification of miRNA clusters used in MetaMirClust; second, we give
an overview of available databases that enlarge the scope of miRNA
genes; third, we introduce how to identify miRNA clusters (Mir-
Clust) in different maximum inter-miRNA distances (MIDs) as well
as a simple case study of using and browsing MirClust; fourth, we
outline the use of MetaMirClust for exploring metazoan conserved
miRNA clusters and their hierarchically evolutionary structure;
fifth, we describe an advanced case study that uses bioinformatics
tools and additional annotation files to uncover the synteny regions
flanking miRNA clusters between human and mouse. Finally, in the
Notes section, we briefly comment on practical issues and highlight
potential pitfalls of the methods that are outlined in this chapter.

2 Materials

MetaMirClust is a Web-based database and can be browsed via a

user-friendly interface implemented according to the protocols of
HyperText Markup Language (HTML) and Cascading Style Sheets
(CSS). For general users who focus on browsing data in MetaMir-
Clust, the community user can easily access it using a computer
with an Internet network. For instance, it can be a desktop com-
puter running Microsoft Windows, an Apple computer running
Mac OS, or a LINUX platform. A few commonly used browsers
include (1) Mozilla Firefox (http://www.firefox.com/), (2)
78 Wen-Ching Chan and Wen-chang Lin

Microsoft Internet Explorer (http://www.microsoft.com/ie/), (3)

Apple Safari (http://www.apple.com/de/safari/), and (4) Google
Chrome (https://www.google.com/intl/en/chrome/).
For advanced users who want to re-perform the whole analysis
procedure and/or follow-up analyses (i.e., the identification of
synteny regions between human and mouse), beyond the essential
Web browser, it is recommended to install an advanced text editor,
e.g., Sublime Text 2/3 (http://www.sublimetext.com/) or Pro-
grammer’s Notepad (http://www.pnotepad.org/), which can
effectively and efficiently facilitate scripting jobs and which manip-
ulates large files (e.g., table-delimited BED files with gene models
from UCSC Table Browser) and/or data format conversions.
When dealing with BED format files, BEDTools 2 (https://
github.com/arq5x/bedtools2) is one of fundamental tools,
which efficiently manages the operations like merging, intersecting,
and/or subtracting between two BED files. In addition, to build a
SQL-like environment to contain data downloaded from public
resources like miRBase or to store intermediate results generated
through the pipeline, MySQL is one of the best choices for a fast,
multi-threads/users and robust database management system. In
MetaMirClust, we introduced a data mining approach, i.e., FP-
growth (30, 31), to efficiently discover highly conserved sets of
miRNA genes upon miRNA clusters (MirClust). The implementa-
tion version of FP-growth algorithm by Borgelt is available to
download (http://www.borgelt.net/fpgrowth.html). Similarly,
the final results after the mining procedure are restored into
MySQL database for querying and browsing via the Web-based
interface. Finally, for visualization, it is also useful to install the
perl models like GD (http://search.cpan.org/~lds/GD/) as well
as the R statistics software (http://www.r-project.org/) to present
results in image files for visual inspection.

3 Methods

3.1 Homology Search A comprehensive understanding of miRNA clusters will require an

of miRNA Genes extensive survey of the coverage of miRNA genes in genomes.
Previously, miRNA genes were identified through cloning and
sequencing of small-RNA libraries. However, miRNA genes could
be overlooked due to low expression levels. In this decade, the ever-
growing data adopted next-generation sequencing (NGS) tech-
nique to identify miRNA genes has been incorporated into public
databases like miRBase. Since those studies were mainly focusing
on a small set of species, it is still necessary to conduct an extensive
homology search based on known miRNA genes collected in miR-
Base to enlarge the scope of miRNA genes across mammals. The
current version of MetaMirClust has been performed based on
known miRNA genes reported in miRBase (Release 16: Sept
 miRNA Cluster Discovery in Metazoan Genomes 79

2010) and predicted homologous miRNA genes in ZooMir

(http://insr.ibms.sinica.edu.tw/zoomir/) (14). The data of the
ZooMir version used in the current MetaMirClust are dumped
from MySQL and can be downloaded (http://insr.ibms.sinica.
edu.tw/ZooMir/ZooMir.Candidates_3.tar.bz2). Using the char-
acteristics of sequence- and structure-conservation of miRNA
genes, additional 14,989 homologous precursor miRNA candi-
dates in 56 genomes have been identified according to 11,839
animal miRNA entries reported in miRBase 16.0. In addition, we
classified miRNA genes by reassigning miRNA classes based on the
sequence similarity with same prefix of their entry names without
considering species abbreviations used in miRBase.

3.2 Identification Recent studies have revealed that the clustering propensity of
of miRNA Clusters miRNA genes is higher than previously evaluated and they usually
(MirClust) occur on polycistronic transcripts (17, 32–36). To investigate clus-
tered miRNA genes derived from the same polycistronic transcript,
researchers usually adopt adjacent miRNA genes located on the
same strand to form miRNA clusters. Two or more consecutive
miRNA genes on the same strand of individual chromosome are
considered to form a cluster according to their adjacent distance. In
miRBase, 10 Kb is used to report clustered miRNAs when users
browse an individual miRNA gene. Take hsa-mir-25
(chr7:99,691,183-99,691,266:-) as example, miRBase will display
hsa-mir-93 (chr7:99,691,391-99,691,470:-) and hsa-mir-106b
(chr7:99,691,616-99,691,697:-) as adjacent miRNA genes within
10 Kb as shown in Fig. 1. As a result, using different adjacent
distance might result in a different data set of miRNA clusters.
Meanwhile, the clustered miRNAs reported in miRBase are lack
of evolutionary conservation across species. Four different maxi-
mum inter-miRNA distances (MIDs); 1 Kb, 3 Kb, 10 Kb, and
50 Kb, were commonly used to identify clustered miRNA genes
(MirClust). To illustrate the procedure of identification of miRNA
clusters (MirClust), we prepared two BED file composed of human
(hg19) precursor/mature miRNA genes (reported in miRBase v.16
or ZooMir) (http://fgfr.ibms.sinica.edu.tw/MetaMirClust/data/
pre.mir.bed; http://fgfr.ibms.sinica.edu.tw/MetaMirClust/data/
mat.mir.bed) as a sample data set for readers to identify miRNA
clusters (MirClust) in human. In addition, the BED file of individ-
ual mature miRNA genes was prepared for the retrieval of miRNA

Fig. 1 The hsa-mir-25-106b cluster reported in miRBase. By the default MID of 10 Kb used in miRBase, this
snapshot figure shows two adjacent miRNA genes, hsa-mir-106b and hsa-mir-93, when querying hsa-mir-25
80 Wen-Ching Chan and Wen-chang Lin

clusters with their corresponding mature miRNAs. The individual

processes were listed as follows.
1. Sort precursor miRNA genes:
sort -k1,1 -k2,2n -k6,6 pre.mir.bed > pre.mir.sort.bed
2. Group miRNA genes to form miRNA clusters based on user-
defined MID
bedtools merge -s -d 10000 -c 4,6,4 -o collapse,distinct,count -
i pre.mir.sort.bed > mir.clust.bed
3. Remove singleton miRNA clusters
awk ’BEGIN{OFS¼FS¼"\t"}{if ($6 > 1) {print $0}}’ mir.
clust.be > mir.clust.filter.bed
4. (Optional) Retrieve mature miRNA genes for each miRNA
cluster
bedtools intersect -wo -a mir.clust.filter.bed -b mat.mir.bed >
mir.clust.mat.bed
The above command would create two intermediate files (i.e.,
pre.mir.sort.bed and mir.clust.bed) plus one output file for the final
result of miRNA clusters (i.e., mir.clust.filter.bed). First, to prepare
the sorted BED file as the input file of BEDTools in the following
process, the human miRNA genes in the BED file were sorted
according to their genomic location plus strand information. Sub-
sequently, using the merge command in the BEDTools package,
adjacent miRNA genes were grouped according to the user-defined
MID (here, 10 Kb). The grouped miRNAs passing the third step by
filtering singleton miRNA clusters will create miRNA clusters in
human in this sample example. Correspondingly, the whole proce-
dure can be achieved by piping into one command line: bedtools
merge -s -d 10000 -c 4,6,4 -o collapse,distinct,count -i <(sort -k1,1 -
k2,2n -k6,6 pre.mir.bed) | awk ’BEGIN{OFS¼FS¼"\t"}{if ($6 >
1) {print $0}}’ > mir.clust.filter.bed.
By comparing miRNA clusters discovered in a short MID to
those in a longer one, three scenarios are discovered: (1) forming a
new miRNA cluster by merging singleton miRNA genes, (2) enlarg-
ing a small miRNA cluster by recruiting singleton miRNA genes,
and (3) producing a large miRNA cluster by merging at least two
small miRNA clusters. According to our previous observation (20),
when considering a long MID, these newly involved clustered
miRNA genes are apt to generate new miRNA clusters instead of
enlarging miRNA clusters in a short MID. It is suggestive of that
miRNA genes are prone to form clusters, and those miRNA clusters
are separately located far away from each other. Table 1 shows the
distributions of numbers of miRNA clusters (MirClust) identified
using four different MID in nine representative species, including
Caenorhabditis elegans (worm, ce6), Drosophila melanogaster
(fly, dm3), Danio rerio (zebrafish, danRer6), Gallus gallus (chicken,
 miRNA Cluster Discovery in Metazoan Genomes 81

Table 1
Distributions of numbers of identified miRNA clusters in nine representative species

MID

Species UCSC accession 1 Kb 3 Kb 10 Kb 50 Kb

Caenorhabditis elegans ce6 13 18 26 38
Drosophila melanogaster dm3 19 18 21 33
Danio rerio danRer6 38 55 61 73
Gallus gallus galGal3 22 41 54 72
Canis familiaris canFam2 50 49 57 74
Bos taurus bosTau4 56 54 61 81
Mus musculus mm9 78 65 69 84
Rattus norvegicus rn4 57 60 63 70
Homo sapiens hg19 66 74 79 100

Fig. 2 Two miRNA clusters identified on chromosome 13 in human. Based on miRNA genes identified in
miRBase and ZooMir and the use of MID of 10 Kb, two miRNA clusters can be revealed on chromosome 13.
One is mir-15a/16 (13q14.2) in the length of 229 nt on the plus strand and the other is mir-17-92 (13q31.3) in
the length of 787 nt on the minus strand

galGal3), Canis familiaris (dog, canFam2), Bos taurus

(cow, bosTau4), Mus musculus (mouse, mm9), Rattus norvegicus
(rat, rn4), and Homo sapiens (human, hg19). According to the
sample example, Fig. 2 lists two miRNA clusters (MirClust) identi-
fied on chromosome 13 in human according to the MID of 10 Kb,
which are mir-16-1/15a (13q14.2) and mir-17-92 (13q31.3). The
community users can retrieve the detailed information of individual
mature miRNA genes for each miRNA clusters through the forth,
optional command listed above. Correspondingly, through our
Web-based interface (http://fgfr.ibms.sinica.edu.tw/MetaMir
Clust/MirClustStat.php), the community users can browse related
information of mir-17-92 (13q31.3) as shown in Fig. 3. The links to
external browsers like UCSC Genome Browser (https://genome.
ucsc.edu/) are provided to obtain more information about miRNA
clusters (e.g., conservation levels and transcriptions in RefSeq or
GenBank). Figure 4 shows several default tracks in the genomic
location flanking mir-17-92 (13q31.3) in UCSC Genome Browser.
Fig. 3 The mature miRNA genes located in mir-17-92. The mir-17-92 (13q31.3) cluster located on chromo-
some 13 consists of six precursor miRNA genes and will encode 12 mature miRNA genes

Fig. 4 MetaMirClust data of mir-17-92 shown in UCSC Genome Browser. Viewing the cluster information of
human mir-17-92 (13q31.3) cluster using public genome browser like UCSC Genome Browser, additional
pieces of evidence such as transcriptional regions, histone modifications, and conservation level and so on
can facilitate users to gain more insights of miRNA clusters of interest
 miRNA Cluster Discovery in Metazoan Genomes 83

3.3 Discovery Most previous works only focused on studying the evolutionary
of Metazoan and functional implications of limited specific miRNA clusters
miRNA Clusters among a few species. No systematic and efficient approach has
(MetaMirClust) been performed before MetaMirClust to analyze the conservation
by FP-Growth pattern of miRNA clusters on global-wide scale. To interrogate the
Algorithm conservation level of the clusters of miRNA genes in large numbers
of metazoan genomes, we adopted a data mining approach to
discover the conserved co-occurrence modules of miRNA genes
upon miRNA clusters identified under the same MID. Filtering
singleton miRNA clusters identified in MirClust as mentioned in
the previous procedure, we conducted the analysis by utilizing the
FP-growth algorithm implemented by Borgelt (http://www.
borgelt.net/fpgrowth.html) to detect the conserved co-occurrence
sets of miRNA genes in terms of miRNA clusters defined within the
same MID. These frequent co-occurrence sets present highly con-
served combinations of miRNA genes through miRNA clusters in
metazoan species, which are defined as metazoan miRNA clusters.
Based on nine representative species same as listed in Table 1, we
prepared an aggregate file (http://fgfr.ibms.sinica.edu.tw/
MetaMirClust/data/nine.mir.clust.csv) consisting of all miRNA
clusters using the previous procedure to identify MirClust. The
following command can be used to discover co-occurred miRNA
genes across selected species.
1. Discover co-occurred miRNA genes across species
fpgrowth -s-7 -q0 nine.mir.clust.csv nine.meta.mir.clust.csv
According to the output result (i.e., nine.meta.mir.clust.csv),
there are 84 evolutionarily conserved miRNA clusters (MetaMir-
Clust) identified in at least seven out of nine representative species.
Among those evolutionarily conserved miRNA clusters, mir-17-92
(13q31.3) is the largest group containing five miRNA classes with
six precursor miRNA genes. Figure 5 shows the conservation pat-
tern of mir-17-92 (13q31.3) in MetaMirClust. The length of the
mir-17-92 (13q31.3) cluster varies from 717 (Loxodonta africana)
to 1,028 (Gasterosteus aculeatus) nucleotides (nt) in 20 metazoan
genomes, which confirmed the estimation of the mir-17 cluster
length as 1 kb reported previously.
In MetaMirClust, to investigate the recruitment process
between evolutionarily conserved miRNA clusters, we also recon-
structed the hierarchical structure using the sets of co-occurred
miRNA genes. The community users can directly select one of
evolutionarily conserved miRNA clusters of interest from the Meta-
MirClust list (http://fgfr.ibms.sinica.edu.tw/MetaMirClust/Met
aMirClustStat.php) or select one of miRNA classes from the
search page in MetaMirClust (http://fgfr.ibms.sinica.edu.tw/
MetaMirClust/MetaMirClustSearch.php) to obtain the hierarchi-
cal information involving the selected miRNA cluster and the
84 Wen-Ching Chan and Wen-chang Lin

Fig. 5 The conservation pattern of mir-17-92 across metazoan. The mir-17-92 (13q31.3) cluster has been
revealed to conserve across 20 species in terms of evolution in our data set and occurs 24 instances among
these species

occurrence in each species under different MIDs. Take mir-25 as

example, the search result under the MID of 10 Kb is shown as
Table 2 with all evolutionarily conserved miRNA clusters contain-
ing the target mir-17 miRNA. For visualization, the drawing of
conservation pattern upon genomes across species has been
provided in MetaMirClust as shown in Fig. 6.

4 Notes

4.1 Data Preparation In miRNA research, miRBase is the most critical repository, in which
from Diverse Sources computational and experimental miRNA genes have been collected,
and a searchable database. Recently, due to the advance in molecular
sequencing technique like next-generation sequencing (NGS),
miRBase have obtained ever-growing miRNA genes identified
from the screening experiments (37). Currently, the miRBase data-
base provides two major formats of archive files: raw-text and SQL-
like files. The former includes dat and fa files in EMBL and fasta
formats, respectively. They are easily for the community users to
check the RNA sequences of precursor and mature miRNA genes.
On the other hand, the SQL-like files dumped directly from miR-
Base contain more information, which is normalized and store into
individual tables in terms of database management. For advanced
users, the latter files will be more efficient to retrieve related data
Table 2
Hierarchical structure of different recruitment of mir-25-106
miRNA Cluster Discovery in Metazoan Genomes
85
86 Wen-Ching Chan and Wen-chang Lin

Fig. 6 The evolutionarily conserved patterns of mir-25-106. This figure shows the conservation pattern of the
mir-25-106 (7q22.1) across 15 species according to the proportion of genomic distance
 miRNA Cluster Discovery in Metazoan Genomes 87

from joining tables by using the SQL language. For our predicted
miRNA genes across metazoans, the dumped data from ZooMir
(http://insr.ibms.sinica.edu.tw/ZooMir/ZooMir.Candidates_3.
tar.bz2) can be easily incorporated into the latest version of
miRBase.

4.2 Understanding In recent years with the large-scale and genome-wide data generated
the Basics of Data by ever-developing molecular biology technique, the huge amount
Mining and Machine of data have become the major challenge for biologists to manipu-
Learning late and analyze them using conventional approaches. Increasing
evidence suggests that data mining and machine learning
approaches can facilitate researchers to efficiently and effectively
conquer the massive number of data like in biological research.
For instance, in MetaMirClust we introduced a data mining
approach to efficiently discover highly conserved sets of miRNA
genes upon miRNA clusters. By treating miRNA genes as items,
FP-growth algorithm can be utilized to mining the frequent item
sets without using candidate generations, of which it can dramati-
cally improve performance in terms of memory space and running
time. The algorithm first compresses the input data into a tree-based
structure, FP-tree, in which all frequent item sets can be retrieved
after easily tracing the entire tree. By iteratively tracing the sub FP-
tree based on conditional frequent item sets, the algorithm can
efficiently reduce the search costs by avoiding the problem intro-
duced in other approaches to look for short fundamental patterns
recursively. Subsequently, the identified frequent item sets using the
FP-growth algorithm are equivalent to the frequently co-occurred
miRNA genes in terms of clusters. Based on those conservation sets
of miRNA genes, we can further reconstruct the hierarchical struc-
ture of conservation patterns across metazoans to facilitate the
community users to gain more insights into the recruitment process
of miRNA genes in clusters in evolution perspective.

4.3 Investigation of To test whether miRNA clusters are co-conserved with their flanking
Conservation Between protein-coding genes, we have conducted a downstream analysis, in
miRNA Clusters and which the linkage of known protein-coding genes in the vicinity of
Flanking Protein- evolutionarily conserved miRNA clusters between human and
Coding Genes mouse were interrogated. We focused only on the nearest adjacent
known genes located in the upstream/downstream regions of con-
served miRNA clusters upon the same strand between those two
species. The genomic information of the protein-coding genes in
human (hg19) and mouse (mm9) were downloaded from the UCSC
Genome Browser (https://genome.ucsc.edu/). In addition, the lift-
Over program (http://genome.ucsc.edu/cgi-bin/hgLiftOver)
downloaded from UCSC Genome Browser was utilized to find the
best mapping of genomic locations between human and mouse if a
miRNA cluster occurs in multiple locations. The homologous
annotations between known protein-coding genes were identified
88 Wen-Ching Chan and Wen-chang Lin

according to the HomoloGene release 64 from NCBI (http://www.

ncbi.nlm.nih.gov/homologene). As a result, our result demon-
strated that 24 out of 37 genomic regions were co-conserved accord-
ing to the evolutionarily conserved miRNA clusters and their
corresponding adjacent protein-coding genes. Nine out of thirty-
seven genomic regions were partially con-served with either
upstream or downstream protein-coding genes. Intriguingly, all six
conserved miRNA clusters located in the intronic regions were
entirely conserved with their host protein-coding genes. This may
suggest that the conservation pattern could be largely extended from
miRNA clusters to their adjacent protein-coding genes.

References

1. Ambros V (2004) The functions of animal 14. Li SC et al (2010) Identification of homolo-

microRNAs. Nature 431(7006):350–355 gous microRNAs in 56 animal genomes. Geno-
2. Bartel DP (2004) MicroRNAs: genomics, bio- mics 96(1):1–9
genesis, mechanism, and function. Cell 116 15. Wu HH, Lin WC, Tsai KW (2014) Advances in
(2):281–297 molecular biomarkers for gastric cancer: miR-
3. He L, Hannon GJ (2004) MicroRNAs: small NAs as emerging novel cancer markers. Expert
RNAs with a big role in gene regulation. Nat Rev Mol Med 16:e1
Rev Genet 5(7):522–531 16. Altuvia Y et al (2005) Clustering and conserva-
4. Lee Y et al (2002) MicroRNA maturation: tion patterns of human microRNAs. Nucleic
stepwise processing and subcellular localiza- Acids Res 33(8):2697–2706
tion. EMBO J 21(17):4663–4670 17. Yu J et al (2006) Human microRNA clusters:
5. Lee RC, Feinbaum RL, Ambros V (1993) The genomic organization and expression profile in
C. elegans heterochronic gene lin-4 encodes leukemia cell lines. Biochem Biophys Res Com-
small RNAs with antisense complementarity mun 349(1):59–68
to lin-14. Cell 75(5):843–854 18. Sewer A et al (2005) Identification of clustered
6. Reinhart BJ et al (2000) The 21-nucleotide let-7 microRNAs using an ab initio prediction
RNA regulates developmental timing in Caenor- method. BMC Bioinformatics 6:267
habditis elegans. Nature 403(6772):901–906 19. Hertel J et al (2006) The expansion of the
7. Tsai KW et al (2010) Epigenetic regulation of metazoan microRNA repertoire. BMC Geno-
miR-196b expression in gastric cancer. Genes mics 7:25
Chromosomes Cancer 49(11):969–980 20. Chan WC et al (2012) MetaMirClust: discov-
8. Tsai KW et al (2011) Aberrant hypermethyla- ery of miRNA cluster patterns using a data-
tion of miR-9 genes in gastric cancer. Epige- mining approach. Genomics 100(3):141–148
netics 6(10):1189–1197 21. Mathelier A, Carbone A (2013) Large scale
9. Tsai KW et al (2012) Aberrant expression of chromosomal mapping of human microRNA
miR-196a in gastric cancers and correlation structural clusters. Nucleic Acids Res 41
with recurrence. Genes Chromosomes Cancer (8):4392–4408
51(4):394–401 22. Tanzer A, Stadler PF (2004) Molecular evolu-
10. Tsai KW et al (2011) Epigenetic regulation of tion of a microRNA cluster. J Mol Biol 339
miR-34b and miR-129 expression in gastric (2):327–335
cancer. Int J Cancer 129(11):2600–2610 23. Calin GA, Croce CM (2006) MicroRNA sig-
11. Berezikov E (2011) Evolution of microRNA natures in human cancers. Nat Rev Cancer 6
diversity and regulation in animals. Nat Rev (11):857–866
Genet 12(12):846–860 24. Laddha SV et al (2013) Genome-wide
12. Griffiths-Jones S (2004) The microRNA regis- analysis reveals downregulation of miR-379/
try. Nucleic Acids Res 32(Database issue): miR-656 cluster in human cancers. Biol Direct
D109–D111 8:10
13. Li SC et al (2010) Discovery and characteriza- 25. Itesako T et al (2014) The microRNA
tion of medaka miRNA genes by next genera- expression signature of bladder cancer by deep
tion sequencing platform. BMC Genomics 11 sequencing: the functional significance of the
(Suppl 4):S8 miR-195/497 cluster. PLoS One 9(2):e84311
 miRNA Cluster Discovery in Metazoan Genomes 89

26. Hayashita Y et al (2005) A polycistronic 32. Megraw M et al (2007) miRGen: a database for
microRNA cluster, miR-17-92, is overexpressed the study of animal microRNA genomic orga-
in human lung cancers and enhances cell prolif- nization and function. Nucleic Acids Res 35
eration. Cancer Res 65(21):9628–9632 (Database issue):D149–D155
27. Yoshino H et al (2013) Tumor-suppressive 33. Lai EC et al (2003) Computational identifica-
microRNA-143/145 cluster targets tion of Drosophila microRNA genes. Genome
hexokinase-2 in renal cell carcinoma. Cancer Biol 4(7):R42
Sci 104(12):1567–1574 34. Lagos-Quintana M et al (2003) New micro-
28. Esquela-Kerscher A, Slack FJ (2006) Onco- RNAs from mouse and human. RNA 9(2):
mirs: microRNAs with a role in cancer. Nat 175–179
Rev Cancer 6(4):259–269 35. Berezikov E et al (2005) Phylogenetic shadow-
29. Zhang Y, Zhang R, Su B (2009) Diversity and ing and computational identification of human
evolution of MicroRNA gene clusters. Sci microRNA genes. Cell 120(1):21–24
China C Life Sci 52(3):261–266 36. Alexiou P et al (2010) miRGen 2.0: a database
30. Han JW et al (2004) Mining frequent patterns of microRNA genomic information and regu-
without candidate generation: a frequent- lation. Nucleic Acids Res 38(Database issue):
pattern tree approach. Data Min Knowl Discov D137–D141
8(1):53–87 37. Kozomara A, Griffiths-Jones S (2014) miR-
31. Chen L, Liu W (2013) Frequent patterns Base: annotating high confidence microRNAs
mining in multiple biological sequences. Com- using deep sequencing data. Nucleic Acids Res
put Biol Med 43(10):1444–1452 42(Database issue):D68–73
Methods in Molecular Biology (2016) 1375: 91–103
DOI 10.1007/7651_2015_280
© Springer Science+Business Media New York 2015
Published online: 10 September 2015

Analysis of Gene Expression Patterns Using Biclustering

Swarup Roy, Dhruba K. Bhattacharyya, and Jugal K. Kalita

Abstract
Mining microarray data to unearth interesting expression profile patterns for discovery of in silico biological
knowledge is an emerging area of research in computational biology. A group of functionally related genes
may have similar expression patterns under a set of conditions or at some time points. Biclustering is an
important data mining tool that has been successfully used to analyze gene expression data for biologically
significant cluster discovery. The purpose of this chapter is to introduce interesting patterns that may be
observed in expression data and discuss the role of biclustering techniques in detecting interesting
functional gene groups with similar expression patterns.

Keywords: Data mining, Expression patterns, Bi-clustering, Microarray

1 Introduction

With the rapid growth of DNA microarray technology, it is now

possible to analyze expression patterns of many genes in a system-
atic and comprehensive manner at the genomic level [1]. The study
of expression patterns of genes in different experimental conditions
may enable one to understand the dynamic behavior of genes and
pathways involved in biological processes. A gene expression level is
a numerical value that measures how a particular gene is over-
expressed or under-expressed in comparison with its activity in
normal conditions. Analysis of expression patterns can be helpful
in discovering groups of genes that participate in similar biological
processes or functions. Various biotechnology laboratories and
pharmaceutical companies involved in in silico drug design can
identify molecular targets that may interact with the drugs. Micro-
array analysis can assist drug companies in choosing the most
appropriate candidates for participation in clinical trials of new
drugs [2]. Wide availability of diagnostic DNA microarrays has
positively impacted cancer research compared to other recent tech-
nologies since they are relatively easy to make and use.
One major goal of analyzing expression data is to discover
functionally similar genes. Co-regulation is a common phenome-
non in gene expression. Expression patterns with similar tendency
or behavior are normally termed as positively regulated and
91
92 Swarup Roy et al.

inverted behavior as negatively regulated [3]. Finding positively

and negatively co-regulated gene clusters from gene expression
data is a real need. A group of co-regulated genes may form gene
clusters that can encode proteins, which interact amongst them-
selves and take part in common biological processes. Genes with
similar (or inverted) expression profiles are very likely to be regula-
tors of one another or be regulated by some other common parent
gene [4, 5]. It has been observed that small sets of genes are co-
regulated and co-expressed under certain conditions, their behavior
being almost inactive for other conditions. Discovering groups of
genes with similar or inverted expression profiles under a set of
conditions leads to the concept of biclustering expression data. We
discuss here various expression patterns identified in microarray
data and how, based on these patterns, biological knowledge can
be extracted in the form of biclusters.

1.1 Patterns in Gene With the help of microarray experiments one can simultaneously
Expression Data monitor the expression levels of genes at a genome scale. Data
generated from microarray experiments, measuring relative expres-
sion levels of genes in a sample and in a controlled population can
be represented in the form of a matrix or vector [6], often called
gene expression matrix. Formally, it can be defined as follows.
Definition 1 (Gene Expression Data). Let G ¼ {G1, G2, . . ., Gm} be a
set of m genes and R ¼ {T1, T2, . . ., Tn} be the set of n conditions or
time points at which the genes’ expression levels are recorded in a
microarray dataset. The gene expression dataset X can be represented
as an m  n matrix, Xmn where each entry xi, j in the matrix corre-
sponds to the logarithm of the relative abundance of mRNA
corresponding to a gene.
To gain better understanding of genes and their behavior inside
the cell, various patterns can be derived by analyzing the change in
expression levels of the genes. The notion of patterns in microarray
data is introduced in [7] as below.
Definition 2 (Expression Pattern). Given a gene Gi, its expression
values under a single condition or a series of varying conditions lie
within a certain range. Gi is a vector of real numbers within the range
[a, b], denoted as Gi@[a, b], and is called an item. The values in Gi are
limited inclusively between a and b.
A set containing one single item is called a pattern. A set of several
items, which come from different genes is also called a pattern. So, a
pattern looks like:

fG i1 @½a i1 , b i1 , . . . , G ik @½aik , b ik g

where it 6¼ is, 1  t, s  k, if k > 1.

 Gene Expression Patterns 93

Table 1
Sample gene expression data from Homo sapiens

ORF C1 C2 C3 C4
GALNT5 3.474 3.837 4.644 5.059
APOE 2 1.943 1.786 1.737
IDH3B 1.449 1.299 0.993 0.832

Homo sapiens
2

1
GALNT5
0
Expression level

APOE
−1 IDH3B

−2

−3

−4

−5

0.5 1 1.5 2 2.5 3 3.5 4 4.5

Conditions

Fig. 1 Profile plot of Homo sapiens expression data

Example data (Table 1) from Homo sapiens microarray dataset,

GDS825, taken from NCBI1 and their respective profile plots are
shown in Fig. 1.
From a biological point of view, patterns play an important role
in discovering functions of genes, disease targets, or gene interac-
tions [8]. A number of different patterns have been identified in
biologically significant gene groups.

1.1.1 Shifting and In shifting patterns [7], the gene profiles show similar trends, but
Scaling Patterns distance-wise, they may not be close to each other (see Fig. 2).
In terms of expression values, the gene patterns are separated
by more or less constant vertical distances among them. Formally,
shifting patterns can be defined as follows
Definition 3 (Shifting Pattern). Given two gene expression profiles
Gi ¼ {Ei1, Ei2, . . ., Eik} and Gj ¼ {Ej1, Ej2, . . ., Ejk} with k expression
values, a profile is called a shifting pattern with respect to another

1
www.ncbi.nlm.nih.gov.
94 Swarup Roy et al.

Shifting Patterns Scaling Patterns

90
60 G1 G1
80
G2 G2
50 G3 70 G3
Expression level

Expression level
60
40
50
30 40

20 30
20
10
10
0 0
0 2 4 6 0 2 4 6
Conditions Conditions

Fig. 2 Expression profile plot shows Shifting and Scaling patterns

profile, if expression value Eip can be related to Ejp with a constant

additive factor π ij for the p ¼ 1. . . k. For genes Gi and Gj, the fact can
be represented as follows.




E ip  E jp
 π ij : ð1Þ

Similarly, scaling patterns in gene expression roughly have mul-

tiplicative distance among the patterns. Scaling pattern can be
defined as follows.
Definition 4 (Scaling Pattern). Given two gene expression profiles
Gi ¼ {Ei1, Ei2, . . ., Eik} and Gj ¼ {Ej1, Ej2, . . ., Ejk} with k expression
values, a profile is called a scaling pattern with respect to another
profile, if expression value Eip can be related to Ejp with constant
multiplicative factor ζ ij for the p ¼ 1. . . k. For genes Gi and Gj, the
fact can be represented as follows.

E ip =E jp  ζij or E jp =E ip  ζ ij : ð2Þ

As shown in Fig. 2, values of G2 are roughly three times larger

than those of G3, and values of G1 are roughly three times larger
than those of G2. In nature, it may so happen that due to different
environmental stimuli or conditions, the pattern G3 responds to
these conditions similarly, although G1 is more responsive or more
sensitive to the stimuli than the other two.

1.1.2 Coherent Patterns A group of genes showing similar pattern tendency across different
conditions is called coherent. Such a group shows predominantly
one kind of co-expression in the expression profiles of all member
genes. Co-expressed genes are likely to be involved in the same
cellular processes. In practice, co-expressed genes may belong to
the same or similar functional categories indicating co-regulated
 Gene Expression Patterns 95

families [5]. Coherent gene expression patterns may characterize

important cellular processes and may provide a foundation for
understanding regulation mechanisms in the cells [9]. The patterns
shown in Fig. 2 are examples of coherent patterns.

1.1.3 Co-regulated Often, coherent patterns are divided into two categories, namely,
Patterns positively regulated patterns and negatively regulated or inverted
patterns. Sometimes, a group of genes that are positively or nega-
tively regulated are also called co-regulated genes. In Fig. 1, human
genes, GLANT5 and IDH3B, show similar patterns or positively
regulated patterns. On the other hand, IDH3B and GLANT5 show
inverted or negative patterns with APOE. Biologically all three
genes are very significant. As suggested Gene Ontology, the three
genes are involved in regulation of plasma lipoprotein particle levels
and triglyceride-rich lipoprotein particle remodeling. Pronounced
inverted or negative patterns can be observed in Fig. 3, taken
from NCBI Rat dataset GDS3702. Gene Ontology suggests that
both are responsible for regulation of interferon-beta production. A
group of genes may share a combination of both positive and
negative co-regulation under a few conditions or at some time
points.
Thus, gene expression data analysis involves pattern finding.
Data mining is the study of techniques that extract patterns from
large amounts of data. As a result, data mining provides the pri-
mary tools for gene expression data analysis. Biclustering is an
important data mining tool for analyzing biologically significant
gene groups. Below we present a brief discussion of biclustering
techniques.

GDS3702 (Rat)
500
450 Mrps26
Pfn2
400
350
Expression level

300
250
200
150
100
50

2 4 6 8 10 12
Conditions

Fig. 3 Expression profile of RAT genes showing negative-regulation

96 Swarup Roy et al.

1.2 Biclustering of Clustering is a popular data analysis tool in genomic studies, partic-
Co-regulated Genes ularly in the context of gene-expression microarrays [10–12]. Each
microarray provides expression measurements for thousands of
genes and clustering is a useful exploratory technique to analyze
gene expression data since it groups similar genes together and
allows biologists to identify groups of potentially meaningful
genes, which have related functions or are co-regulated, which in
turn helps find the relationships among them in the form of gene
regulatory networks [5]. It has frequently been observed that sub-
sets of genes are co-regulated and co-expressed under a subset of
environmental conditions or time points [13]. Biclustering algo-
rithms tackle the problem of finding a set of sub-matrices where
each sub-matrix or bicluster meets a certain homogeneity criterion.
Given a gene expression dataset DNM, where G ¼ {G1, G2,
. . . GN} is a set of N genes and R ¼ {T1, T2, . . ., TM} is the set of M
conditions or time points, biclusters can be defined as follows.
Definition 5 (Biclusters). Biclusters are a set of sub-matrices of the
matrix D ¼ (N, M) with dimensions I1  J1, . . . Ik  Jk such that
Ii  N, Ji  M 8i{1, . . ., k}, where each sub-matrix (bicluster) meets a
given homogeneity criterion.
Madeira and Oliveira [14] identify four different categories of
biclusters based on homogeneity criterion, namely:
1. Constant biclusters,
2. Biclusters with constant values on either columns or rows,
3. Biclusters with coherent values, and
4. Biclusters with coherent evolutions.
A comprehensive survey of different biclustering techniques for
gene expression data clustering can be found in [15, 16]. In gene
expression analysis, patterns play a more important role than
expression values [17]. As a result, the value based homogeneity
criterion mentioned above may not be suitable for grouping bio-
logically significant genes.

2 Materials

Technological improvements in high-throughput DNA microarray

technology is instrumental in the tremendous growth of publicly
available gene expression data. This growing amount of expression
data requires concurrent development of adequate bioinformatics
tools for comprehensive analysis of the data for extracting
biological knowledge. A number of online and offline tools are
available for biclustering of gene expression data. We mention
here a few of the leading, freely available biclustering packages
(Table 2).
 Gene Expression Patterns 97

Table 2
Freely available Biclustering software packages

Package Availability Web site Platform Method(s) Reference

Expander 6.3 Download http://acgt.cs.tau.ac.il/expander/ Java Samba [18]
Bic_AT Plus Download http://people.ee.ethz.ch/s̃op/bicat/ Java BiMax, CC, [19]
ISA, xMotif,
OPSM
BiGGEsTS Download http://kdbio.inesc-id.pt/software/ Java CCC, e-CCC, [20]
biggests/
CC-TSB
BiVisu Download http://www.eie.polyu.edu.hk/ Matlab BiVisu [21]
ñflaw/Biclustering/
QServer Online http://csbl.bmb.uga.edu/ Web QUBIC [22]
publications/materials/ffzhou/
QServer/
PAGE Download http://www.niehs.nih.gov/research/ Java q-Clustering [23]
resources/software/biostatistics/
page/
CoBi Download https://sites.google.com/site/ Java CoBi [24]
swarupnehu/publications/
resources

2.1 Data Sources A plethora of real expression data produced by different biotech-
nology labs are freely available online. In this chapter, we use some
datasets from Table 3 for experimentation and demonstration.

2.2 Evaluating From the point of view of biological data analysis, a cluster is
Quality of Biclusters biologically significant if it can produce functionally enriched
groups of genes. A majority of the literature on biclustering evalu-
ates and reports results based on functional enrichment of the
clusters against Gene Ontology (GO). To determine the statistical
significance of the association of a particular GO term with a group
of genes in a cluster, various online tools from the GO Project2 are
available. In Table 4, we report some freely available tools.
These tools use the hypergeometric distribution to calculate
the p-value or q-value, which evaluates whether the clusters have
significant enrichment in one or more function groups. The p-value
is computed as follows:

2
http://www.geneontology.org.
98 Swarup Roy et al.

Table 3
Short description of data sources

No. of No. of
Organism Dataset genes samples Source
YeastDB 2884 17 http://arep.med.harvard.edu/biclustering/yeast.matrix
Yeast Sporulation 474 7 http://cmgm.stanford.edu/pbrown/sporulation
Yeast_KY 237 17 http://faculty.washington.edu/kayee/cluster/
YeastCho 384 17 http://faculty.washington.edu/kayee/cluster
(cell cycle)
Rat Rat_CNS 112 9 http://faculty.washington.edu/kayee/cluster
Human GDS3712 325 12 NCBI
Fibroblast 517 13 http://www.sciencemag.org/feature/data/984559.hsl/
Serum
Mouse GDS958 308 12 NCBI
Rice Thaliana 138 8 http://homes.esat.kuleuven.be/s̃istawww/bioi/thijs/
Work/Clustering.html

Table 4
GO-based cluster evaluation tools

Tool Platform Type Url

FuncAssociate 2 Web Online http://llama.mshri.on.ca
Fatigo Web Online http://fatigo.bioinfo.cnio.es
GOTermFinder Web Online http://go.princeton.edu,
http://db.yeastgenome.org/cgi-bin/GO/goTermFinder
OntoExpress Web Online http://vortex.cs.wayne.edu
GeneMANIA Web Online www.genemania.org
DAVID 6.7 Web Online http://david.abcc.ncifcrf.gov
AGO Matlab Offline www.k-space.org/alakwa/AGO/AGO.zip

  
f gf
X
k
i ni
p ¼1   : ð3Þ
i¼0
g
n

The p-value gives the probability of seeing at least k genes out

of the total n genes in a cluster annotated with a particular GO
 Gene Expression Patterns 99

termf, given the total number of genes in the whole genome g and
the number of genes in the whole genome that are annotated with
that GO term f. It is important to note that p-value measures
whether a cluster is enriched with genes from a particular category
to a greater extent than what would be expected by chance. If the
majority of genes in a cluster appears in one category, the p-value of
the category is small. That is, the closer the p-value to zero, the
more the probability that the particular GO term is associated with
the group of genes. The Q-value is the minimal False Discovery
Rate (FDR) at which this gene appears significant. Q-values are
estimated using the Benjamini Hochberg procedure [25].

3 Methods

This approach to clustering was originally introduced by Hartigan

[26] and later applied by Cheng and Church [27] to expression
data to capture the coherence of a subset of genes under a subset of
conditions. Several techniques have been proposed to find quality
biclusters from expression data. In Cheng and Church’s approach,
the degree of coherence is measured using the concept of mean
squared residue (MSR) and the algorithm greedily inserts/removes
rows and columns to arrive at a certain number of biclusters,
achieving some predefined residue score. The lower the score, the
stronger the coherence exhibited by the biclusters, and better is the
quality of the biclusters. Following Cheng and Church, a number
of biclustering techniques have been proposed [27–37] to deter-
mine quality biclusters.
A greedy iterative search [27, 28] based approach finds a local
optimal solution with an expectation to finally obtain a globally
good solution. A divide and conquer [26] approach divides the
whole problem into sub-problems and solves them recursively.
Finally, it combines all the solutions to solve the original problem.
In exhaustive biclustering [35], the best biclusters are identified
using exhaustive enumeration of all possible biclusters extant in the
data, in exponential time. A detailed categorization of heuristic
approaches is available in [29]. A number of techniques based on
metaheuristics, such as evolutionary and multi-objective evolution-
ary framework, have also been explored [30] to generate and itera-
tively refine an optimal set of biclusters. All of them use MSR as the
merit function.
An MSR based technique is effective in finding optimized max-
imal biclusters. From a biological point of view, the interest resides
in finding biclusters with subsets of genes showing similar behaviors,
not similar values. Interesting and relevant patterns from a
biological point of view, such as shifting and scaling patterns, may
not be detected using this measure as it considers only expression
values, not the patterns or tendencies of gene expression profile. It is
important to discover this type of patterns because frequently the
100 Swarup Roy et al.

genes can present similar behavior although their expression levels

vary in range or magnitude. Aguilar-Ruiz [31] proves that MSR is
not a good measure to discover patterns in data when the variance
among gene values is high, that is, when the genes present scaling
and shifting patterns. To detect biologically relevant biclusters with
scaling and shifting patterns, a scatter search based approach has
been proposed [32]. This method uses a fitness function based on
linear correlation among genes and an improvement method to
select just the positively correlated genes.
Often, it has been observed that genes share local rather than
global similarity in their gene expression profiles and only under a
few conditions or time points [13]. Thus, correlation based tech-
nique may not be effective when computing pair-wise similarity
among gene expression profiles. Other than that, various pattern-
based approaches have also been proposed [33, 34, 38, 39] for
discovery of biclusters where expression levels of genes rise and fall
at a subset of conditions or time points.
Recently, it has been observed that [3] co-regulated genes also
share negative patterns or inverted behaviors, which existing
pattern-based approaches are unable to detect. CoBi [24] (Core-
gulated Biclustering) captures biclusters among both positively and
negatively regulated genes as co-regulated genes. It considers both
up- and down-regulation trends and similarity in degrees of fluctu-
ation under consecutive conditions for expression profiles of two
genes as a measure of similarity between the genes. It uses a new
BiClust tree for generating biclusters in polynomial time that needs
a single pass over the dataset.

3.1 Performance We compare the performance of a few biclustering methods taking

Comparison into account functional enrichment of the biclusters. We consider
four biclustering techniques: Bimax [40], Cheng and Church
(CC) [27], OPSM [41], and CoBi [24]. For the purpose of com-
parison, we set the parameter values of other algorithms as recom-
mended in the original papers. The functional enrichment of each
bicluster is measured using Q-values associated with GO categories.
For each bicluster, we calculate average of the percentage of the
number of genes from the bicluster with a given function against all
genes in the genome with the function. Figure 4 shows the average
of functional enrichments of each bicluster obtained by different
biclustering algorithms from four different datasets [24].
From the graphs it is clearly evident that CoBi outperforms all
three algorithms in obtaining functionally enriched biclusters.
However, for the YeastCho dataset, the Cheng and Church (CC)
approach performs better than other algorithms.
 Gene Expression Patterns 101

Yeast Sporulation Yeast Cho

Avg.Percentage of Enriched Biclusters

Avg.Percentage of Enriched Biclusters

25 25

20 20

15 15

10 10

5 5

0 0
CoBi BiMax CC OPSM CoBi BiMax CC OPSM
Bicluster Algorithms Bicluster Algorithms

Yeast DB YeastKY
Avg.Percentage of Enriched Biclusters

Avg.Percentage of Enriched Biclusters

12 30

10 25

8 20

6 15

4 10

2 5

0 0
CoBi BiMax CC OPSM CoBi BiMax CC OPSM
Bicluster Algorithms Bicluster Algorithms

Fig. 4 Comparison on functionally enriched biclusters obtained by different biclustering techniques

4 Notes

Biclustering is a promising and important data mining tool for

analyzing gene expression data. A number of techniques are avail-
able for biclustering. Most are greedy in nature and often compu-
tationally expensive. Moreover, they ignore positive- and negative-
regulation patterns when performing biclustering. As mentioned in
[42], a bicluster is considered a quality bicluster when participating
genes exhibit consistent trends and similar degrees of fluctuation
under consecutive conditions. We consider both up- and down-
regulation trends and similarity in degrees of fluctuations under
consecutive conditions for expression profiles of two genes as a
measure of similarity between the genes. Compared to other meth-
ods discussed above, the design of CoBi has been motivated by a
desire to handle the outstanding issues mentioned above and as a
result, it exhibits promising results.
102 Swarup Roy et al.

References
1. Kurella M, Hsiao L, Yoshida T, Randall J, Chow 15. Kriegel HP, Kröger P, Zimek A (2009) Clus-
G, Sarang S, Jensen R, Gullans S (2001) Dna tering high-dimensional data: a survey on sub-
microarray analysis of complex biologic pro- space clustering, pattern-based clustering and
cesses. J Am Soc Nephrol 12:1072–1078 correlation clustering. ACM Trans Knowl Dis-
2. Kraljevic S, Stambrook PJ, Pavelic K (2004) cov Data (TKDD) 3:1
Accelerating drug discovery. EMBO Rep 16. Mahanta P, Ahmed H, Bhattacharyya D, Kalita
5:837–842 JK (2011) Triclustering in gene expression data
3. Yu H, Luscombe N, Qian J, Gerstein M (2003) analysis: a selected survey. In: 2011 2nd
Genomic analysis of gene expression relation- national conference on emerging trends and
ships in transcriptional regulatory networks. applications in computer science (NCETACS),
Trends Genet 19:422–427 IEEE pp 1–6
4. Gasch A, Eisen M et al (2002) Exploring the 17. Roy S, Bhattacharyya DK, Kalita JK (2014)
conditional coregulation of yeast gene expres- Reconstruction of gene co-expression network
sion through fuzzy k-means clustering. from microarray data using local expression
Genome Biol 3:1–22 patterns. BMC Bioinf 15:S10
5. Tavazoie S, Hughes J, Campbell M, Cho R, 18. Shamir R, Maron-Katz A, Tanay A, Linhart C,
Church G et al (1999) Systematic determina- Steinfeld I, Sharan R, Shiloh Y, Elkon R (2005)
tion of genetic network architecture. Nat Expander—an integrative program suite for
Genet 22:281–285 microarray data analysis. BMC Bioinf 6:232
6. Grant R (2004) Computational genomics: the- 19. Barkow S, Bleuler S, Prelić A, Zimmermann P,
ory and application. Horizon Bioscience, Zitzler E (2006) Bicat: a biclustering analysis
Cambridge toolbox. Bioinformatics 22:1282–1283
7. Li J, Wong L (2001) Emerging patterns and 20. Gonçalves JP, Madeira SC, Oliveira AL (2009)
gene expression data. Genome Inform Ser Biggests: integrated environment for bicluster-
12:3–13 ing analysis of time series gene expression data.
8. Alberts B, Johnson A et al (2002) Studying BMC Res Notes 2:124
gene expression and function. In: Molecular 21. Cheng KO, Law NF, Siu WC, Lau T (2007)
biology of the cell, 4th edn Bivisu: software tool for bicluster detection and
9. Spellman P, Sherlock G, Zhang M, Iyer V, visualization. Bioinformatics 23:2342–2344
Anders K, Eisen M, Brown P, Botstein D, 22. Zhou F, Ma Q, Li G, Xu Y (2012) Qserver: a
Futcher B (1998) Comprehensive identifica- biclustering server for prediction and assess-
tion of cell cycle-regulated genes of the yeast ment of co-expressed gene clusters. PloS one
saccharomyces cerevisiae by microarray hybri- 7:e32660
dization. Mol Biol Cell 9:3273–3297 23. Leung E, Bushel PR (2006) Page: phase-
10. Ben-Dor A, Shamir R, Yakhini Z (1999) Clus- shifted analysis of gene expression. Bioinfor-
tering gene expression patterns. J Comput Biol matics 22:367–368
6:281–297 24. Roy S, Bhattacharyya DK, Kalita JK (2013)
11. Chipman H, Hastie TJ, Tibshirani R (2003) Cobi: pattern based co-regulated biclustering
Clustering microarray data. In: Statistical anal- of gene expression data. Pattern Recogn Lett
ysis of gene expression microarray data, vol 34:1669–1678
1. Chapman & Hall/CRC, Boca Raton, 25. Benjamini Y, Hochberg Y (1995) Controlling
pp 159–200 the false discovery rate: a practical and powerful
12. Ahmed HA, Mahanta P, Bhattacharyya D, approach to multiple testing. J R Stat Soc Ser B
Kalita JK (2011) Gerc: tree based clustering (Methodological) 57:289–300
for gene expression data. In: 2011 I.E. 11th 26. Hartigan JA (1972) Direct clustering of a data
international conference on bioinformatics and matrix. J Am Stat Assoc 67:123–129
bioengineering (BIBE), IEEE, pp 299–302 27. Cheng Y, Church G (2000) Biclustering of
13. Mitra S, Banka H (2006) Multi-objective evo- expression data. In: Proceedings of 8th inter-
lutionary biclustering of gene expression data. national conference on intelligent systems
Pattern Recogn 39:2464–2477 for molecular biology, ICISMB’00, vol 8,
14. Madeira SC, Oliveira AL (2004) Biclustering pp 93–103
algorithms for biological data analysis: a survey. 28. Yang J, Wang H, Wang W, Yu P (2003)
IEEE/ACM Trans Comput Biol Bioinform Enhanced biclustering on expression data. In:
1:24–45 Proceedings of the 3rd IEEE symposium on
 Gene Expression Patterns 103

bioinformatics and bioengineering, 2003, pp co-regulated genes from gene expression

321–327 data. In: Proceedings of the 16th interna-
29. Madeira S, Oliveira A (2004) Biclustering algo- tional conference on KES12, FAIA, vol 243,
rithms for biological data analysis: a survey. pp 490–499
IEEE/ACM Trans Comput Biol Bioinf 1:24–45 37. Eren K, Deveci M, K€ uç€
uktunç O, Çataly€ urek
30. Banka H, Mitra S (2006) Evolutionary biclus- € (2013) A comparative analysis of bicluster-
UV
tering of gene expressions. Ubiquity 7:1–12 ing algorithms for gene expression data. Brief
31. Aguilar-Ruiz J (2005) Shifting and scaling pat- Bioinform 14:279–292
terns from gene expression data. Bioinformat- 38. Wang H, Wang W, Yang J, Yu P (2002) Clus-
ics 21:3840–3845 tering by pattern similarity in large data sets. In:
32. Nepomuceno J, Troncoso A, Aguilar-Ruiz J Proceedings of the international conference on
et al (2011) Biclustering of gene expression management of data. ACM SIGMOD’02,
data by correlation-based scatter search. Bio- ACM, pp 394–405
Data Min 4:3 39. Zhao Y, Yu J, Wang G, Chen L, Wang B, Yu G
33. Pei J, Zhang X Cho M, Wang H, Yu P (2003) (2008) Maximal subspace coregulated gene
Maple: a fast algorithm for maximal pattern- clustering. IEEE Trans Knowl Data Eng 20:
based clustering. In: Proceedings of the 3rd 83–98
IEEE international conference on data mining, 40. Prelić A, Bleuler S et al (2006) A systematic
2003 (ICDM’03), IEEE, pp 259–266 comparison and evaluation of biclustering
34. Wang H, Chu F, Fan W, Yu P, Pei J (2004) A methods for gene expression data. Bioinfor-
fast algorithm for subspace clustering by pat- matics 22:1122–1129
tern similarity. In: Proceedings of the 16th 41. Ben-Dor A, Shamir R, Yakhini Z (1999) Clus-
international conference on scientific and tering gene expression patterns. J Comput Biol
statistical database management, 2004, IEEE, 6:281–297
pp 51–60 42. Ji L, Mock K, Tan K (2006) Quick hierarchical
35. Tanay A, Sharan R, Shamir R (2002) Discover- biclustering on microarray gene expression
ing statistically significant biclusters in gene data. In: Proceedings of the 6th IEEE sympo-
expression data. Bioinformatics 18:S136–S144 sium on bioinformatics and bioengineering,
36. Roy S, Bhattacharyya DK, Kalita JK (2012) 2006 (BIBE’06), IEEE, pp 110–120
Deterministic approach for biclustering of
Methods in Molecular Biology (2016) 1375: 105–116
DOI 10.1007/7651_2015_241
© Springer Science+Business Media New York 2015
Published online: 14 May 2015

Using Semantic Similarities and csbl.go

for Analyzing Microarray Data
Kristian Ovaska

Abstract
Cellular phenotypes result from the combined effect of multiple genes, and high-throughput techniques
such as DNA microarrays and deep sequencing allow monitoring this genomic complexity. The large scale
of the resulting data, however, creates challenges for interpreting results, as primary analysis often yields
hundreds of genes. Gene Ontology (GO), a controlled vocabulary for gene products, enables semantic
analysis of such gene sets. GO can be used to define semantic similarity between genes, which enables
semantic clustering to reduce the complexity of a result set. Here, we describe how to compute semantic
similarities and perform GO-based gene clustering using csbl.go, an R package for GO semantic similarity.
We demonstrate the approach with expression profiles from breast cancer.

Keywords: Gene ontology, Semantic similarity, Measure, Hierarchical clustering, Expression

microarray, Data analysis

1 Introduction

Many normal and pathological cellular phenotypes result from the

combined effect of several genes and proteins. For instance, cancer
is caused by aberrations on several large protein pathways (1), and
there are more than 100 known cancer-driving genes (2). A goal of
life sciences is to identify and elucidate the mechanics of such gene
modules. High-throughput measurement techniques, such as
DNA microarrays and deep DNA sequencing, allow monitoring
entire or substantial portions of whole genomes in parallel, yielding
large-scale molecular data. For instance, DNA microarrays can be
used to measure the expression of all known genes of a species. The
high dimensionality and interconnectedness of such data, however,
creates challenges for analyzing and interpreting the results. After
primary statistical analysis, hundreds of result genes may need to be
evaluated. Manual inspection of such gene sets is time-consuming,
subjective, and error-prone. Automated bioinformatics methods
can help in interpretation by reducing the dimensionality and com-
plexity of gene sets.
One approach to reducing the complexity of a gene set is to
group similar genes together, which decreases the dimensionality
105
106 Kristian Ovaska

and allows inspecting shared features of similar genes (i.e., clusters).

This type of analysis requires the definition of a similarity measure,
which assigns high numerical values for similar genes and low values
for dissimilar genes. What, exactly, “similarity of genes” means is
subjective, and there are different approaches for defining such a
measure.
Several advanced similarity metrics are defined using Gene
Ontology (GO), a controlled vocabulary for describing gene prod-
uct features (3). GO allows expressing the knowledge on genes
using a formalism that is both human- and machine-readable. GO
is composed of three ontologies, which capture different aspects of
cell biology: biological processes (BP), cellular components (CC),
and molecular functions (MF). Each ontology consists of GO terms
(concepts), which are the “words” of the vocabulary, and their
hierarchical relationships. For example, the proto-oncogene epider-
mal growth factor receptor (EGFR) is described in GO by terms such
as signal transduction (BP), plasma membrane (CC), and protein
kinase activity (MF) (4). Annotating a gene with a specific term,
such as plasma membrane, also implies annotation with the less
specific parent term(s), such as membrane.

1.1 Semantic The formal structure of GO allows defining semantic similarity (SS)
Similarity Using Gene between genes using the idea that similar genes share similar GO
Ontology annotations. More than two dozen such measures have been
defined, and it is not always clear which one is the best for a given
purpose (5). However, usually the choice of a default measure is
sufficient. Here, we first illustrate the use of one simple and effective
SS measure, Resnik similarity (6, 7). Then, we briefly survey the
main features of other measures.
As an example, we compute the similarity of two hypothetical
genes, A and B, using the Resnik measure. Gene A is annotated
with the GO terms plasma membrane (CC) and signal transduction
(BP), and gene B with mitochondrial membrane (CC) and regula-
tion of signaling (BP) (Fig. 1). Resnik similarity is a “pairwise”
similarity measure, which means that it defines similarity between
individual GO terms, but not, as such, for genes. Gene similarity is
obtained in a second step from term similarities. Hence, we first
compute the pairwise similarities between GO terms of A and B.
The ontologies BP, CC, and MF are handled individually, so the
similarity of terms from different ontologies is zero.
To compute the term similarity between plasma membrane and
mitochondrial membrane, Resnik similarity first evaluates the speci-
ficity of all GO terms, as intuitively specific terms are more informa-
tive and contribute more to similarity than generic terms. In Resnik
similarity, specificity is based on empirical usage of the terms in gene
annotations: a less frequently used term is more specific than a
commonly used one. This is formalized as term information content
(IC) as follows. We compute the number of genes annotated with a
 Using Semantic Similarities and csbl.go for Analyzing Microarray Data 107

cellular component 100 biological process 80

membrane 20 signaling 12

A A B
plasma membrane 5 organelle membrane 11 signal transduction 6 regulation of signaling 7

B
mitochondrial membrane 3

Fig. 1 Illustration of a GO network structure that is used to compute semantic similarity between genes A and
B using the Resnik measure. Small subsets of the ontologies cellular component and biological process
are shown here. Links between GO terms denote “is a” or “part of” relationships so that child terms are more
specific than parent terms. The annotations for genes A and B are shown next to the relevant GO terms. The
figures next to each GO term are usage frequencies in a hypothetical annotation corpus, used for computing
information content. For example, 20 genes are annotated with membrane or one of its child terms. Usage
frequencies are decreasing when following paths from the root node to leaf nodes because annotation with a
child term implies annotation with ancestor terms

certain GO term (n), and compare it to the number of all genes in

the annotation corpus (N). Then, the probability that a random
gene is annotated with the term is p ¼ n/N. Now, the IC of the
term is defined using an information theoretical approach: IC ¼
log2 p. In our example (Fig. 1), 20 genes, out of a corpus of 100
genes with annotation in CC, are annotated with membrane. The IC
of membrane is then log2 20/100 ¼ 2.32, and the IC of the root
term cellular component is log2 100/100 ¼ 0.
The next step in Resnik similarity is to use the network struc-
ture of GO to find the common ancestors between plasma mem-
brane and mitochondrial membrane to capture the “semantic
overlap.” In our example, the common ancestors are membrane
and cellular component. Out of these ancestors, Resnik similarity
selects the one with the highest IC to quantify the similarity. In our
case, the ICs are 2.32 and 0, respectively. Thus, membrane is the
most informative common ancestor (MICA). The Resnik similarity
is then the IC of that term, i.e., 2.32. The similarity between signal
transduction and regulation of signaling is computed in a
similar fashion. The common ancestors of these terms are signaling
(IC ¼ log2 12/80 ¼ 2.74) and biological process (IC ¼ 0), and
the Resnik similarity is 2.74.
Pairwise term similarities establish the matrix shown in Table 1.
There are several methods to obtain a gene similarity from this
matrix. A simple and, in many cases, effective method is to
define gene similarity as the maximum of the matrix. Hence, the
“Resnik/max” similarity between genes A and B is max(2.32, 0,
0, 2.74) ¼ 2.74.
108 Kristian Ovaska

Table 1
Pairwise GO term Resnik similarities for genes A (columns) and B (rows)

Plasma Signal
membrane (A) transduction (A)
Mitochondrial membrane (B) 2.32 0
Regulation of signaling (B) 0 2.74

1.2 Semantic Resnik similarity is a good first choice for an SS measure, but for
Similarity Measure specific applications other measures may need to be considered.
Dimensions Understanding the differences between various measures is
facilitated by categorization of measures along feature dimensions
(5, 8, 9). Most differences between SS measures can be understood
along three such dimensions. First, a major differentiator between
SS measures is whether they directly define an SS measure for
complete GO term sets (groupwise), or if they define an SS
measure for individual GO terms (pairwise), which must then be
transformed into gene similarity. Resnik similarity belongs to the
pairwise family.
Second, measures differ in how they evaluate GO term speci-
ficity. Resnik similarity uses the empirical information content
obtained from an annotation corpus. An alternative is to use the
GO graph structure so that a term with high depth (long distance
to ontology root node) is considered to be more specific.
Some measures do not explicitly consider term specificity at all.
IC-based methods are considered the most accurate option (5).
Third, some measures, such as Resnik, consider only one
common ancestor to evaluate the similarity of two terms; for
IC-based methods, this is usually the MICA. Other measures
consider all common ancestors, which may provide additional
value over MICA (5).

2 Materials

In Materials, we demonstrate how to use a semantic similarity

measure package for R, csbl.go (10), to cluster genes obtained
from an expression microarray experiment.

2.1 Breast Cancer The clustering demonstration is done using example data from
Example Data Set invasive breast cancer from The Cancer Genome Atlas (TCGA)
(11). The publicly available data consist of Agilent G4502A expres-
sion microarrays, including both tumors (TCGA codes 01 and 06)
and adjacent healthy tissue (TCGA code 11) for control. Prepro-
cessed level 3 data, containing cy5/cy3 log-ratios, were downloaded
 Using Semantic Similarities and csbl.go for Analyzing Microarray Data 109

from TCGA on 2014-06-23, resulting in 529 tumor and 61 control

samples. Using the log-ratios for tumor samples, we obtained
Bonferroni-corrected p-values using a two-sided one-sample t-test.
Genes with a p-value below 0.0001 and linear fold change above
8 (or below 1/8) were considered as differentially expressed. These
strict criteria were used to obtain a relatively small gene set suitable
for illustrating GO-based analysis. The analysis resulted in 168
differentially expressed genes (DEGs).

2.2 Custom Data The csbl.go package requires (1), minimally, a gene set, and (2),
Sets optionally, an expression or other quantitative matrix for the gene
set. The genes are represented as a set of database identifiers. Any
genome database can be used, as long as identifiers in the gene set,
expression matrix, and GO annotation source match.

2.3 Computational csbl.go requires a Windows or a Linux machine (32 or 64 bit), with
Environment an installation of the R programming environment (http://www.
r-project.org/). R version 2.12+ or 3.0+ is recommended.

3 Methods

3.1 Installing csbl.go Running an up-to-date version of csbl.go is recommended, as the

GO structure and annotations are frequently updated.
1. Install the packages required by csbl.go. In Windows, open the
graphical R environment; in Linux, open the interactive R shell
(using administrator privileges). From the Bioconductor (12)
library, csbl.go requires the Biobase, annotate, and GO.db
packages. From CRAN, cluster and RUnit are required.
These are all installed with the following:
source("http://bioconductor.org/biocLite.R")
biocLite(c("Biobase", "annotate", "GO.db"))
install.packages(c("cluster", "RUnit"))
2. Download csbl.go from http://csbi.ltdk.helsinki.
fi/csbl.go/. Make sure you download the correct archive
for your platform (Linux, Windows).
3. In Windows, using the graphical R environment, install the
package from the downloaded local ZIP file (Packages !
Install package(s) from local zips files). In Linux, open the
system shell (e.g., Bash), change to the download directory,
and install using R CMD INSTALL (file).tar.gz (with
administrator privileges).
4. It is recommended to ensure successful installation by running
built-in tests by entering the interactive R prompt and typing:
library(csbl.go)
run.tests.csbl.go()
110 Kristian Ovaska

3.2 Obtaining GO Using csbl.go requires that the genes under analysis are annotated
Annotations for with GO terms. This information is obtained from genome or
Differentially proteome databases, or directly from the Gene Ontology database.
Expressed Genes In our breast cancer example, we use Biomart from R to annotate
the 168 DEGs using the Uniprot database.
First, install the biomaRt Bioconductor package using:
source("http://bioconductor.org/biocLite.R")
biocLite("biomaRt")

DEGs, together with their expression values, are in a tab-

delimited file deg.txt, which looks like the following example
(numeric values are for demonstration purposes):
Hybridization REF TCGA-XX-YYYY-01A-ZZZ TCGA-XX-
YYYY-01A-ZZZ . . .
Gene1 2.55 6.59
Gene2 0.98 4.33
...
The important column here is the first one, which contains
HUGO gene identifiers. Other columns contain expression values
for TCGA samples.
The genes are annotated with the following R script, which
writes annotated.txt that contains entries for those genes that
have one or more GO annotations:
library(biomaRt)
IN <- "deg.txt"
OUT <- "annotated.txt"
ORGANISM <- "Homo sapiens"

uniprot <- useMart("unimart", "uniprot")

table.in <- read.table(IN, header¼TRUE, sep¼"\t")
table.out <- data.frame(GeneName¼table.in[,1], GO¼NA)

for (row.index in 1:nrow(table.in)) {

gene.name <- table.in[row.index, 1]
annotation <- getBM("go_id",
c("gene_name", "proteome_name"),
list(gene.name, ORGANISM), uniprot)
table.out[row.index, "GO"] <- paste(annotation
[,1], collapse¼" ")
}
table.out <- table.out[nchar(table.out$GO) > 0,]
write.table(table.out, OUT,
row.names¼FALSE, col.names¼FALSE, quote¼FALSE)
 Using Semantic Similarities and csbl.go for Analyzing Microarray Data 111

The output is in a format supported by csbl.go, in which the

first column contains a gene name and the rest of the line contains
one or more GO identifiers:
TF GO:0008199 GO:0006826 GO:0006879 GO:0005576
SNCA GO:0014059 GO:0005737
PVALB GO:0005509 GO:0005634 GO:0043234 GO:0051480
GO:0030424 GO:0005737
...

3.3 Clustering Genes Combining expression values and GO annotations of DEGs, we

Using GO Similarity compute semantic similarities between genes and use these for
and Expression hierarchical clustering. Breast cancer samples are, in turn, clustered
using expression values. Clustering is done with the following R
script:
library(csbl.go)
EXPRESSION <- "deg.txt"
ANNOTATION <- "annotated.txt"
CUT <- 0.4

set.prob.table(organism¼TAXONOMY.HUMAN,
type¼"similarity")
expr <- read.table(EXPRESSION, header¼TRUE, row.
names¼1,
sep¼"\t", stringsAsFactors¼FALSE)
expr <- expr[, 1:20] # For demonstration
result <- go.heatmap(expr, ANNOTATION, metric¼
"Resnik",
go.cut¼CUT, margins¼c(15, 5))
print(result$members[[6]])
print(result$desc[[6]])

In csbl.go, we first specify the species from which the gene set is
obtained with set.prob.table. Then, we read the expression
matrix into memory and take a subset of the first 20 samples to
obtain a more readable clustering visualization for demonstration
purposes. GO-based clustering is performed with go.heatmap,
which takes the in-memory expression matrix and the GO annota-
tion file as mandatory arguments. The expression matrix must have
row (gene) and column (sample) names. Clustering yields (1) a heat
map visualization shown in Fig. 2, and (2) a data structure describ-
ing the clusters obtained. Details on go.heatmap can be obtained
from its help page, shown with ?go.heatmap in R.

3.4 Interpreting GO From the right margin of Fig. 2, we see that the GO-based cluster-
Clustering Results ing resulted in six gene clusters, G1 to G6. Each cluster contains
genes that are semantically similar (i.e., share similar GO annota-
tions). From the heat map, we can observe expression values within
the gene clusters. On the X axis, the three leftmost samples are
112 Kristian Ovaska

low high

G6
G5
G4
G3

G2

GO groups
G1

cut
TCGA.BH.A0BV.11A.31R.A089.07
TCGA.BH.A18Q.11A.34R.A12D.07
TCGA.E2.A15M.11A.22R.A12D.07
TCGA.A8.A07C.01A.11R.A034.07
TCGA.AO.A0JL.01A.11R.A056.07
TCGA.E2.A15A.06A.11R.A12D.07
TCGA.B6.A0RM.01A.11R.A084.07
TCGA.B6.A0RI.01A.11R.A056.07
TCGA.A1.A0SD.01A.11R.A115.07
TCGA.B6.A0IE.01A.11R.A034.07
TCGA.BH.A18R.01A.11R.A12D.07
TCGA.A8.A08C.01A.11R.A00Z.07
TCGA.A8.A09N.01A.11R.A00Z.07
TCGA.E2.A15C.01A.31R.A12D.07
TCGA.AO.A0JA.01A.11R.A056.07
TCGA.A8.A09D.01A.11R.A00Z.07
TCGA.E2.A108.01A.13R.A10J.07
TCGA.B6.A0WS.01A.11R.A115.07
TCGA.A2.A0T1.01A.21R.A084.07
TCGA.BH.A0DG.01A.21R.A12P.07

Samples

Fig. 2 Results of GO-based clustering for TCGA breast cancer samples. On the X axis, a subset of 20 samples
is shown; they are clustered using expression values, which are visualized in the heat map. The Y axis
represents genes, which are clustered using Resnik similarity. The selected cut point in the dendrogram on the
left results in six gene clusters, denoted G1 to G6

healthy adjacent tissue, as seen from their TCGA sample codes

(TCGA.XX.YYYY.11.); these samples were used in the visualization
for demonstration. Their expression profiles are clearly different
from the tumor samples, seen both visually and in the dendrogram
on the top.
The data structure result returned by go.heatmap contains
details about the gene clusters. result$members is a list of gene
identifiers for each cluster, and result$desc is a list of R data
 Using Semantic Similarities and csbl.go for Analyzing Microarray Data 113

frames that describe common GO terms shared by members of the

cluster. In the following example, we see the details for cluster G6:
# result$members[[6]]
[1] "PCOLCE2" "MSR1"
# result$desc[[6]]
goid priori information desc
4 GO:0004872 0.04280044 4.546231 receptor activity
2 GO:0016020 0.30805812 1.698726 membrane

This cluster contains two genes, PCOLCE2 and MSR1, that

both code membrane proteins with receptor activity. The priori
column is the probability of the GO term in the annotation corpus,
and information is the corresponding information content.
Clusters sharing high-IC GO terms are generally the most
interesting.

3.5 Computing Gene In addition to GO-based clustering, csbl.go contains lower level
Similarity Matrix functionality for computing similarities. The following R script
produces a similarity matrix sim between all the genes, using
Resnik/max similarity.
library(csbl.go)
ANNOTATION <- "annotated.txt"
set.prob.table(organism¼TAXONOMY.HUMAN,
type¼"similarity")
ent <- entities.from.text(ANNOTATION)
sim <- entity.sim.many.allont(ent, "Resnik", "max")

4 Notes

4.1 Creating Custom The csbl.go package is bundled with GO term probability tables for
GO Probability Tables Homo sapiens, Saccharomyces cerevisiae, Caenorhabditis elegans,
Drosophila melanogaster, Mus musculus, Rattus norvegicus, Arabi-
dopsis thaliana, and Xenopus tropicalis. The GO annotation corpora
for these are obtained from the database provided by the Gene
Ontology Consortium (GOC). Building a custom GO probability
table may be necessary if: (a) you work with a species other than the
above, or (b) you use a different database for obtaining GO anno-
tations for your gene set than GOC. Whereas condition (a) is
obvious, (b) is more subtle. GO analysis gives best results when
the background probabilities (ICs) are computed from the same
database that is used for annotation. If different databases are used,
results may be inaccurate because the expected (a priori) GO term
usage is different from actual usage. The degree of inaccuracy
114 Kristian Ovaska

depends on the degree of differences in the two databases, so it is

difficult to give general rules on (b). To create a custom GO table:
1. Obtain GO annotations for all genes or genes products in your
organism from the database of your choice. The annotations
are stored in a text format similar to the one in
Section Obtaining GO annotations for differentially expressed
genes.
2. A script (prob_tables_from_text.r) to create probability
tables from text-based GO annotations is in the inst/tools
directory of the csbl.go source package, or tools in the Win-
dows binary package. This needs to be invoked from the shell.
First change to the directory containing the script.
3. Assuming here that the GO annotations are in annotation.
txt, create the custom GO table with:

R --slave --args annotation.txt test 9606

"My annotation" < prob_tables_from_text.r

Note that there should be no newline after 9606. Here, test is

a name for your table, 9606 is the NCBI taxonomy ID (in this
case, human), and the text in quotes is a description of the
table. The output is written to test-similarity.csv.
4. In R, load the custom GO table with:
set.prob.table(filename¼"test-similarity.
csv").

4.2 Loading GO For loading a bundled GO probability table into memory, csbl.go
Tables for Bundled requires an NCBI taxonomy identifier as a parameter to set.
Species prob.table. The following constants represent the identifiers of
bundled species: TAXONOMY.HUMAN, TAXONOMY.YEAST, TAXON-
OMY.C_ELEGANS, TAXONOMY.DROSOPHILA, TAXONOMY.MOUSE,
TAXONOMY.RAT, TAXONOMY.ARABIDOPSIS and TAXONOMY.
XENOPUS.

4.3 Optimizing the Hierarchical clustering uses a “cut” parameter with range from 0 to
Dendrogram Cut 1 to determine how to obtain gene clusters from the clustering
Parameter dendrogram. This parameter is named go.cut in the go.heatmap
for Clustering function and corresponds to the h parameter in cut.dendrogram.
A lower value favors smaller clusters. Often, it is necessary to try
different values of go.cut to obtain a suitable set of clusters.
The cut threshold can be seen in the left margin of the heat map,
aiding manual optimization.

4.4 Alternative In addition to Resnik similarity, csbl.go supports several other

Semantic Similarity semantic similarity measures (10). The measure is set with the
Measures metric and multiple arguments of go.heatmap; a full list is
available with ?go.heatmap. As of June 2014, the supported
 Using Semantic Similarities and csbl.go for Analyzing Microarray Data 115

methods are Resnik (6), ResnikGraSM, Lin (13), LinGraSM, Jiang-

Conrath (14), JiangConrathGraSM, Relevance (15), Kappa (16),
Cosine (17), WeightedJaccard (18), and CzekanowskiDice (19).
The GraSM variants are an improvement over the basic measures
(20). For pairwise measures (Resnik, Lin, JiangConrath, and Rele-
vance), the method for obtaining a gene similarity from GO term
similarity also needs to be specified. Options include max (default),
avg, and rcmax (15).

4.5 Alternative In addition to csbl.go, alternative software packages for semantic

Software Tools for similarity computation include GOSemSim (21), GOSim (22), and
Computing Similarity SML (23).

4.6 Running Without GO-based clustering can be executed also when expression or other
Expression Data quantitative data is not available, using only the annotated gene set.
This is done by supplying NULL as the first argument to go.heat-
map, i.e., go.heatmap(NULL, ANNOTATION.FILE).

4.7 Integration with The core functionality of csbl.go has been integrated with the
Anduril Workflow Anduril workflow framework (24) available from http://anduril.
Framework org. The relevant Anduril components are GOClustering
(GO-based clustering) and GOProbabilityTable (creating custom
GO tables). GO annotations can be fetched using, for example,
BiomartAnnotator (various databases) or KorvasieniAnnotator
(Ensembl).

Acknowledgements

I thank Tiia Pelkonen for proofreading.

References
1. Hanahan D, Weinberg RA (2011) Hallmarks assessment with biological features and issues.
of cancer: the next generation. Cell Brief Bioinform 13:569–585
144:646–674 6. Resnik P (1995) Using information content to
2. Vogelstein B, Papadopoulos N, Velculescu VE evaluate semantic similarity in a taxonomy. Pro-
et al (2013) Cancer genome landscapes. Sci- ceedings of the 14th international joint confer-
ence 339:1546–1558 ence on artificial intelligence, vol 1, pp
3. Ashburner M, Ball C, Blake J et al (2000) Gene 448–453
ontology: tool for the unification of biology. 7. Lord P, Stevens R, Brass A et al (2003) Inves-
Nat Genet 25:25–29 tigating semantic similarity measures across the
4. Rebhan M, Chalifa-Caspi V, Prilusky J et al gene ontology: the relationship between
(1998) GeneCards: a novel functional geno- sequence and annotation. Bioinformatics
mics compendium with automated data mining 19:1275–1283
and query reformulation support. Bioinfor- 8. Mazandu GK, Mulder NJ (2013) Information
matics 14:656–664 content-based gene ontology semantic similar-
5. Guzzi PH, Mina M, Guerra C et al (2012) ity approaches: toward a unified framework
Semantic similarity analysis of protein data: theory. BioMed Res In 2013:292063
116 Kristian Ovaska

9. Harispe S, Sánchez D, Ranwez S et al (2014) A associative relations in the gene ontology. Pac
framework for unifying ontology-based seman- Symp Biocomput 2005:91–102
tic similarity measures: a study in the biomedi- 18. Pesquita C, Faria D, Bastos H et al (2008)
cal domain. J Biomed Inform 48:38–53 Metrics for GO based protein semantic similar-
10. Ovaska K, Laakso M, Hautaniemi S (2008) ity: a systematic evaluation. BMC Bioinformat-
Fast gene ontology based clustering for micro- ics 9:S4
array experiments. BioData Mining 1:11 19. Brun C, Chevenet F, Martin D et al (2004)
11. The Cancer Genome Atlas Network (2012) Functional classification of proteins for the
Comprehensive molecular portraits of human prediction of cellular function from a
breast tumours. Nature 490:61–70 protein-protein interaction network. Genome
12. Gentleman RC, Carey VJ, Bates DM et al Biol 5:6
(2004) Bioconductor: open software develop- 20. Couto FM, Silva MJ, Coutinho PM (2007)
ment for computational biology and bioinfor- Measuring semantic similarity between
matics. Genome Biol 5:R80 gene ontology terms. Data Knowl Eng
13. Lin D (1998) An information-theoretic defini- 61:137–152
tion of similarity. Proceedings of the 15th inter- 21. Yu G, Li F, Qin Y et al (2010) GOSemSim: an
national conference on machine learning, R package for measuring semantic similarity
pp 296–304 among GO terms and gene products. Bioinfor-
14. Jiang J, Conrath D (1997) Semantic similarity matics 26:976–978
based on corpus statistics and lexical taxonomy. 22. Frohlich H, Speer N, Poustka A et al (2007)
Proceedings of international conference on GOSim – an R-package for computation of
research in computational linguistics, pp 19–33 information theoretic GO similarities between
15. Schlicker A, Domingues F, Rahnenf€ uhrer J et al terms and gene products. BMC Bioinformatics
(2006) A new measure for functional similarity 8:166
of gene products based on gene ontology. 23. Harispe S, Ranwez S, Janaqi S et al (2014) The
BMC Bioinformatics 7:302 semantic measures library and toolkit: fast
16. Huang D, Sherman B, Tan Q et al (2007) The computation of semantic similarity and related-
DAVID gene functional classification tool: a ness using biomedical ontologies. Bioinformat-
novel biological module-centric algorithm to ics 30:740–742
functionally analyze large gene lists. Genome 24. Ovaska K, Laakso M, Haapa-Paananen S et al
Biol 8:R183 (2010) Large-scale data integration framework
17. Bodenreider O, Aubry M, Burgun A (2005) provides a comprehensive view on glioblastoma
Non-lexical approaches to identifying multiforme. Genome Med 2:65
Methods in Molecular Biology (2016) 1375: 117–121
DOI 10.1007/7651_2015_249
© Springer Science+Business Media New York 2015
Published online: 14 May 2015

Ontology-Based Analysis of Microarray Data

Agapito Giuseppe and Marianna Milano

Abstract
The importance of semantic-based methods and algorithms for the analysis and management of biological
data is growing for two main reasons. From a biological side, knowledge contained in ontologies is more
and more accurate and complete, from a computational side, recent algorithms are using in a valuable way
such knowledge. Here we focus on semantic-based management and analysis of protein interaction net-
works referring to all the approaches of analysis of protein–protein interaction data that uses knowledge
encoded into biological ontologies.
Semantic approaches for studying high-throughput data have been largely used in the past to mine
genomic and expression data. Recently, the emergence of network approaches for investigating molecular
machineries has stimulated in a parallel way the introduction of semantic-based techniques for analysis and
management of network data. The application of these computational approaches to the study of micro-
array data can broad the application scenario of them and simultaneously can help the understanding of
disease development and progress.

Keywords: Data mining, Expression patterns, Bi-clustering, Microarray

1 Introduction

The accumulation of data about proteins, genes, and small

molecules on a large scale caused the possibility to look at molecular
machineries on a system level scale. After the rise of the systems
biology, more recently the network biology (1), i.e., the discipline
that bring together molecular biology and network theory, has
gained a big interest.
In this scenario, data about genes constitute the fundamental
building blocks (2) used to grow models and theories.
Let us consider for instance interactions among proteins,
named protein–protein interactions (PPI). Proteins play their role
usually by interacting with them or other macromolecules. An
interaction usually involves a contact with surfaces of two or more
proteins.
Due to the introduction of high-throughput techniques, many
experimental datasets have been produced causing the introduction
of computer science methods to manage, store, and analyze PPI
data (3). The whole set of protein interactions of a single species are

117
118 Agapito Giuseppe and Marianna Milano

also referred to as Protein to Protein Interaction Network (PIN).

PINs have been easily modeled by using undirected graphs (4)
where nodes are associated with proteins and edges represent inter-
actions among proteins.
PPI data have been collected in many public databases.
Usually, PPI databases contain raw data, e.g., the identifiers of
the interacting proteins, and some annotation related to the reli-
ability of the stored data.
The accumulation of raw experimental data about genes and
proteins have been accompanied by the accumulation of functional
information, i.e., knowledge about function. The assembly, organi-
zation, and analysis of this data have given a considerable impulse to
research (5). Usually, biological knowledge is encoded by using
annotation terms, i.e., terms describing for instance function or
localization of genes and proteins. Such annotations are often
organized into ontologies, which offer a formal framework to
organize in a formal way biological knowledge.
For instance, Gene Ontology (GO) (6) provides a set of anno-
tations (namely GO Terms) of biological aspects, structured into
three main taxonomies: Molecular function (MF), Biological Pro-
cess (BP), and Cellular Component (CC). Annotations are often
stored in publicly available databases, for instance, a main resource
for GO annotations is the Gene Ontology Annotation (GOA)
database (7). The availability of well-formalized functional data
enabled the development of algorithms and methods to analyze
proteins and genes from a semantic perspective.
Historically, first approaches were referred to as functional
enrichment algorithms. They have been developed to determine
the statistical significance of the presence (or the absence) of a GO
term in a set of gene products or proteins (8). Despite the existence
of more than 60 freely available tools, the functional analysis
of large list is still a challenge. Classical algorithms referred to
Gene Enrichment Algorithms (GEA) or Gene Set Enrichment
Algorithms (GSEA), do not cope with the topological information
contained in protein or gene interaction network. More recently,
network enrichment analysis (NEA) approaches that extends the
classical approaches to network links between genes in the experi-
mental set and those in the functional categories (9).
More recently, a set of algorithms, referred to as Semantic
Similarity Measures (SSMs), have been developed to compare in a
quantitative way set of terms belonging to the same ontology. SSMs
take in input two or more ontology terms and produce as output a
value representing their similarity.
This enabled the possibility to use such formal instruments for
the comparison and analysis of proteins and genes (10, 11). Many
works have focused on: (1) the definition of ad-hoc semantic mea-
sures tailored to the biological scenario (12); (2) the introduction
of algorithms for the functional analysis of interactomics data (13);
 Ontology-Based Analysis of Microarray Data 119

and (3) finally the building of semantic similarity networks (SSN),

i.e., edge-weighted graph whose nodes are genes or proteins, and
edges represent semantic similarities among them (14).

2 Semantic Similarity Measures

While sequence or structure-based similarity of genes and proteins

has been largely investigated in the past, the similarity based on
functions presents a more complex scenario. In fact, while primary
and tertiary structures can be compared in terms of number of
shared amino acids or in terms of spatial conformation. The com-
parison of the functions needs the introduction of a comparison
metrics between terms that are often expressed in natural language.
The adoption of ontologies for managing annotations provides
a means to compare entities on aspects that would otherwise not be
comparable. For instance, if two gene products are annotated
within the same schema, we can compare them by comparing the
terms with which they are annotated (15, 16).
The annotations of biological concepts are currently organized
in simple taxonomies or more complex ontologies, such as Gene
Ontology or Open Biomedical Ontologies (OBO), (17). The use
of ontologies enables the comparison of annotations in terms of
analysis of the ontology schema. Thus, the problem to define the
semantic similarity of two terms can be solved in terms of analysis of
the underlying ontology. While the semantic similarity among two
biomedical or biological concepts is not a trivial problem, the
semantic similarity among terms that come from a common
schema, e.g., a taxonomy has been largely investigated and can be
solved in an efficient way. In the same way, if two biological con-
cepts, e.g., proteins are annotated with terms organized by using
an ontology, the problem of the determination of their semantic
similarity can be solved in terms of semantic similarity of the anno-
tating terms.
Several approaches are available to quantify the semantic simi-
larity between terms or annotated entities in an ontology repre-
sented as a directed acyclic graph (DAG) such as GO.
We here presents a brief categorization on the basis of accord-
ing to the strategy used for the calculation: (1) Term Information
Content (IC), (2) Term Depth, (3) based on a common ancestor,
(4) based on all common ancestors, (5) Path Length, and (6)
Vector Space Models (VSM). Measures based on Term Depth and
IC evaluate terms similarity on the basis of the specificity of the
terms. Measures based on a common ancestor first select a common
ancestor of two terms according to its properties, and then evalu-
ates the semantic similarity on the basis of the distance among the
terms and their common ancestor and the properties of the com-
mon ancestor. Techniques based on Path Length correlate
120 Agapito Giuseppe and Marianna Milano

measures of the length of the path connecting the two terms.

VSM-based measures initially represent the set of the annotations
of proteins as vectors. Then, the similarity is evaluated by consider-
ing the distance among vectors that are defined using topological
considerations.
Proteins and genes are annotated with a set of GO terms, so to
assess the functional similarity between gene products it is necessary
to compare sets of terms rather than single terms. All the proposed
approaches are based on the comparison of terms and on the
combination of the results, i.e., the pairwise similarity of annota-
tions calculated using an existing measure. The simplest way to
measure the semantic similarity between two gene products is to
calculate the pairwise semantic similarity among the terms that
annotate the gene products and successively to combine such pair-
wise similarity by using some formulas such as the average, the
maximum, or the sum. Other approaches are based on the repre-
sentation of two gene products as the induced subgraph of annota-
tion or as a point in a vector space induced by annotations (18, 19).
Semantic similarity measures are affected by three main pro-
blems (20):
Annotation length. The number of annotations per protein (i.e., the
GO Terms associated with each protein) is highly variable within
the same GO taxonomy and over different species. Consequently,
the resulting similarity score is affected by this variability. Conse-
quently, comparing proteins with few annotations is more likely to
return low similarity scores, even if the proteins are related.
Evidence codes. The task of associating with proteins the GO Terms
that describe their functions and properties, called annotation, is
performed with different methods. Without entering into details,
they range from experimentally verified to electronIcally infErred
Annotations (IEA). SSMs usually do not weight annotations on the
basis of their ECs, and one has to choose between including poten-
tially unreliable annotations to increase the number of annotations
at the expenses of the quality or ignoring them but drastically
reducing the number of annotations considered.
Shallow annotations. Several proteins are annotated with generic
GO terms. These annotations do not identify the specific role or
function of the protein, but only suggest the area in which the
proteins operate.
 Ontology-Based Analysis of Microarray Data 121

References
1. Barabasi AL, Oltvai ZN (2004) Network 12. Pesquita C et al (2009) Semantic similarity in
biology: understanding the cell’s functional biomedical ontologies. PLoS Comput Biol.
organization. Nat Rev Genet 5(2):101–113 doi:10.1371/journal.pcbi.1000443
2. Cannataro M, Guzzi PH, Veltri P (2010) Pro- 13. Dai X et al (2014) A comprehensive semantic
tein-to-protein interactions: technologies, similarity measurement for predicting the func-
databases, and algorithms. ACM Comput tion of gene products. J Bionanosci 8
Surv. doi:10.1145/1824795.1824796 (4):287–292
3. Ciriello G et al (2012) AlignNemo: a local 14. Agapito G, Guzzi PH, Cannataro M (2013)
network alignment method to integrate Visualization of protein interaction networks:
homology and topology. PLoS One. doi:10. problems and solutions. BMC Bioinformatics.
1371/journal.pone.0038107 doi:10.1186/1471-2105-14-S1-S1
4. West DB (2000) Introduction to graph theory, 15. Guzzi PH, Cannataro M (2012) Cyto-sevis:
2nd edn. Prentice Hall, New York semantic similarity-based visualisation of pro-
5. Blake JA, Bult CJ (2006) Beyond the data tein interaction networks. EMB-Net J doi:
deluge: data integration and bio-ontologies. J http://dx.doi.org/10.14806/ej.18.A.397
Biomed Informat 39(3):314–320 16. Cannataro M et al (2007) Using ontologies for
6. Harris MA et al (2004) The gene ontology preprocessing and mining spectra data on the
(go) database and informatics resource. grid. Future Generat Comput Syst 23
Nucleic Acids Res 32:258–261 (1):55–60
7. Barrell D et al (2009) The GOA database in 17. Smith B et al (2007) The OBO foundry: coor-
2009-an integrated gene ontology annotation dinated evolution of ontologies to support bio-
resource. Nucleic acids Research. doi:10. medical data integration. Nat Biotechnol 25
1093/nar/gkn803 (11):1251–1255
8. Huang DW, Sherman BT, Lempicki RA (2009) 18. Popescu M, Keller JM, Mitchell JA
Bioinformatics enrichment tools: paths toward (2006) Fuzzy measures on the gene ontology
the comprehensive functional analysis of large for gene product similarity. IEEE/ACM
gene lists. Nucleic Acids Res 37(1):1–13 Trans Comput Biol Bioinformatics 3
9. Alexeyenko A et al (2012) Network enrichment (3):263–274
analysis: extension of gene-set enrichment anal- 19. Yu G, Li F et al (2010) GOSemSim: an R
ysis to gene networks. BMC Bioinformatics. package for measuring semantic similarity
doi:10.1186/1471-2105-13-226 among GO terms and gene products. Bioinfor-
10. Guzzi PH et al (2012) Semantic similarity analysis matics 26(7):976–978
of protein data: assessment with biological fea- 20. Guzzi PH, Mina M (2012) Towards the assess-
tures and issues. Brief Bioinform 13(5):569–585 ment of semantic similarity analysis of protein
11. Smoot ME et al (2011) Cytoscape 2.8: new data: main approaches and issues. ACM SIG-
features for data integration and network visu- Bioinformatics Rec 2(3):17–18
alization. Bioinformatics 27(3):431–432
Methods in Molecular Biology (2016) 1375: 123–136
DOI 10.1007/7651_2015_242
© Springer Science+Business Media New York 2015
Published online: 12 March 2015

Integrated Analysis of Transcriptomic and Proteomic

Datasets Reveals Information on Protein Expressivity
and Factors Affecting Translational Efficiency
Jiangxin Wang, Gang Wu, Lei Chen, and Weiwen Zhang

Abstract
Integrated analysis of large-scale transcriptomic and proteomic data can provide important insights into the
metabolic mechanisms underlying complex biological systems. In this chapter, we present methods to
address two aspects of issues related to integrated transcriptomic and proteomic analysis. First, due to the
fact that proteomic datasets are often incomplete, and integrated analysis of partial proteomic data may
introduce significant bias. To address these issues, we describe a zero-inflated Poisson (ZIP)-based model to
uncover the complicated relationships between protein abundances and mRNA expression levels, and then
apply them to predict protein abundance for the proteins not experimentally detected. The ZIP model takes
into consideration the undetected proteins by assuming that there is a probability mass at zero representing
expressed proteins that were undetected owing to technical limitations. The model validity is demonstrated
using biological information of operons, regulons, and pathways. Second, weak correlation between
transcriptomic and proteomic datasets is often due to biological factors affecting translational processes.
To quantify the effects of these factors, we describe a multiple regression-based statistical framework
to quantitatively examine the effects of various translational efficiency-related sequence features on
mRNA–protein correlation. Using the datasets from sulfate-reducing bacteria Desulfovibrio vulgaris, the
analysis shows that translation-related sequence features can contribute up to 15.2–26.2 % of the total
variation of the correlation between transcriptomic and proteomic datasets, and also reveals the relative
importance of various features in translation process.

Keywords: Transcriptome, Proteome, Correlation, Zero-inflated Poisson regression, Prediction,

Undetected proteins, Translation, Sequence features

1 Introduction

Due to revolutionary improvements in high-throughput DNA

sequencing technologies, several thousand microbial genomes
from almost all known major phylogenetic lineages have been fully
sequenced, and many more are nearing completion (1–3). Various
computational-based annotation and comparative genomic analyses
of DNA sequences have provided biologists with information
regarding gene function, genome structures, biological pathways,
metabolic and regulatory networks, and evolution of microbial
genomes, which has greatly enhanced our understanding of

123
124 Jiangxin Wang et al.

microbial metabolism (4–8). However, to fully elucidate microbial

metabolism and its responses to environmental factors, it is
necessary to include functional characterization and accurate quan-
tification of all levels of gene products, mRNA, proteins and
metabolites, as well as their interaction. In the past decade, signifi-
cant efforts in improving analytical technologies pertaining to
measuring mRNA, proteins, and metabolites have been made.
These efforts have led to the generation of several new “omics”
research fields: transcriptomics, proteomics, metabolomics,
interactomics, and so on (9–14). To date, although a great deal of
information regarding cellular metabolism has been acquired
through application of individual “omics” approaches (15), it is
also becoming clear that any single “omics” approach may not be
sufficient to characterize the complexity of biological systems (16).
Moreover, in cells many levels of regulation occur after genes have
been transcribed, such as posttranscriptional, translational, and
posttranslational regulation, and all forms of biochemical control
such as allosteric or feedback regulation. Taking this view into
account, it is hard to believe that functional genomics can stop at
the mRNA level or any single level of information. In fact,
integrated multi-“omics” approaches have been applied recently
and the studies have enabled researchers to unravel global
regulatory mechanisms and complex metabolic networks in various
eukaryotic organisms (17–20).
One of the key tasks for integrated transcriptomic and proteomic
analysis is to identify relationships between protein abundances and
concentrations of their cognate mRNA. Although one would
hypothesize that the correlation between mRNA expression levels
and protein abundance will be strong based on the central dogma of
molecular genetics, support from experimental data is not immedi-
ately apparent. Most recent studies have either failed to find a corre-
lation between protein and mRNA abundances (16) or have observed
only a weak correlation (21–23). It is now accepted that modest
correlation between mRNA expression and protein abundance in
large-scale datasets is explained in part by experimental challenges,
such as technological limitations, and in part by fundamental
biological factors in the transcription and translation processes
(24–26).
While transcriptomic analysis produces data on transcript levels
for most genes in a given genome, proteomic datasets are often
incomplete due to the imperfect identification of coding sequences
within a genome and the limited sensitivity of current peptide
detection technologies (27). Current technologies allow detection
of only one-half to two-third of all coded proteins (22, 28, 29).
In prior comparisons of transcriptomic and proteomic data,
undetected proteins were often assigned a concentration value of
zero, and excluded from the correlation analysis. This unrealistic
simplification could adversely affect interpretation of relationships
Integrated Analysis of Transcriptomic and Proteomic Datasets Reveals. . . 125

between transcriptomic and proteomic data. For instance, current

technologies for proteomic analysis tend to be biased towards
detection of relatively abundant proteins. Correlation patterns
between transcriptomic and proteomic data for these highly
expressed genes are unlikely valid for the entire genome since
correlation patterns may be different for lowly expressed genes.
Hence, improved methods of coping with missing protein abun-
dance values are necessary for integrative analysis of transcriptomic
and proteomic datasets. To address issues with the missing proteo-
mics data, one recent tactic was to integrate Gene Ontology (GO)
information into the data imputation; the approach could enhance
the imputation even when the missing fraction is large (30).
Using transcriptomic and proteomic datasets collected from
Desulfovibrio vulgaris, we recently proposed a novel Zero-inflated
Poisson (ZIP) regression model in which we assumed that
100
p % (0 < p < 1) of the genes with a proteomic abundance
level of zero could be unexpressed genes or expressed genes that
were undetected due to technical limitations (31).
Efficiency of protein biosynthesis depends on many factors, (1)
initial anchoring of ribosomes onto the mRNA depends on com-
plementary binding of the Shine–Dalgarno (SD) sequence ~10
bases upstream of the start codon and a sequence close to the
30 end of the 16S rRNA in the 30S ribosomal subunit; (2) nonran-
dom use of synonymous codons in the coding region of highly
expressed Escherichia coli genes indicates that sequences further
downstream of the start codon could be of importance for transla-
tion efficiency (32); (3) translation efficiency also depends on the
availability of various amino acids. Among 20 amino acids, costs of
synthesis vary from 12 to 74 high-energy phosphate bonds per
molecule (33). The evidence of natural selection of amino acid
usage to enhance metabolic efficiency has been found in the
proteomes of E. coli and Bacillus subtilis (33); (4) translation
termination depends upon the attachment of a release factor (RF)
in the place of a tRNA in the ribosomal complex. In addition,
studies showed that nucleotide distribution around the stop
codons, especially the base following the stop codon, is significantly
biased and is related to translation termination efficiency (34).
In this chapter, we describe statistical protocols to address the
two aspects of the issues related to integrated transcriptomic and
proteomics datasets: (1) a Zero-inflated Poisson (ZIP) regression
model to infer protein abundance for proteins undetected by
proteomic analysis, probably due to technical limitation; and (2) a
multiple regression model to quantify effects of biological factors
(i.e., sequence features) related to translational efficiency on corre-
lation between transcriptomic and proteomic datasets. Using the
sample transcriptomic and proteomic datasets collected from
Desulfovibrio vulgaris grown under various conditions, the (ZIP)
126 Jiangxin Wang et al.

model can predict protein abundance for all undetected proteins in

D. vulgaris, and the multiple regression analysis of all translation-
related sequence features show that they together contribute up to
15.2–26.2 % of the total variation of mRNA–protein correlation.

2 Materials

2.1 Datasets The transcriptomic and proteomic datasets collected from

Desulfovibrio vulgaris are used. The datasets consist of the whole-
genome mRNA expression and LC–MS/MS proteome abundance
data from D. vulgaris in two different growth stages—log and
stationary—and under two distinct types of media: lactate- or
formate-based. To minimize variations between microarray and
proteomic measurements, identical cell samples from each growth
condition were split and used to isolate both the RNA and proteins
for analyses. A detailed description of the datasets is provided in our
previous publication (26, 29, 35, 36). The raw intensity values
from both datasets are normalized with a quantile normalization
using an R package (caret) available through the R project (http://
www.r-project.org/).
Quality of the datasets is assessed by calculating Pearson
correlation coefficients among multiple replicates. The analysis
shows that correlation coefficients of the microarray experiments
are from 0.97 to 0.99 among replicate samples (31, 37), and
correlation coefficients of LC–MS/MS measurements normalized
by amino acid composition are 0.86–0.92 among replicates. The
correlation between mRNA expression and normalized protein
abundance is modest: 0.54–0.63 (p-value, 0.001) by Pearson
correlation coefficient for all conditions.

2.2 Genome The cellular functional categories of all genes in the target genome
Information are downloaded from the Comprehensive Microbial Resource of
TIGR (http://cmr.tigr.org) (35) and NCBI (http://www.ncbi.
nlm.nih.gov/). On the basis of the original annotation, the
genes/proteins are classified into 19 cellular functional categories.
On the basis of the original annotation, the genes and proteins
are classified into different cellular functional categories. These
categories are included in the model as possible predictors of
protein abundance. Gene annotation attributes such as sequence
length, protein length, molecular weight, and GC content and
triple codon counts of all genes in the target genome are down-
loaded from the TIGR or NCBI resource. Continuous numerical
values are gathered for the molecular weight of each gene. The GC
content reflects the proportion of nucleotides G or C in the target
genome. The triple codon information includes counts for all 64
triple codon combinations in the genetic code.
 Integrated Analysis of Transcriptomic and Proteomic Datasets Reveals. . . 127

The complete genome of target species and its ORF calls and
annotation of D. vulgaris are downloaded from NCBI Genbank
and the TIGR resource. Genes transcribed in the same direction
having intergenic regions <15 bp are defined as one operon. Gene
lists of all metabolic pathways defined for target genomes of interest
were downloaded from the KEGG database (http://www.genome.
jp/kegg/kegg2.html).

3 Methods

3.1 Zero-Inflated The Poisson regression model, one of the so-called generalized
Poisson Regression linear models (38), was used to model the relationship between
Model proteomic abundance and mRNA expression.
3.1.1 Model Construction 1. In the Poisson regression model, for protein abundances (Y ),
and Validation we assume that the mean (λ) of the Poisson distribution depends
on log-scaled mRNA abundance (X), and therefore λ ¼ exp
(α + β
X), which ensures that the expected value is nonneg-
ative. This Poisson regression model provides a valid framework
to integrate two types of expression data; however, it provides
no explanation for the fact that ~83 % genes have zero proteo-
mic abundance.
2. We then ascribe the high percentage of proteins with zero
abundance to technical limitations in the proteomic analyses,
such as detection sensitivity. Therefore, a nonstandard mixture
model, the ZIP regression model (39), is proposed to analyze
the data. In this model, we assume that 100
p % of the genes
with proteomic abundance level of 0 may be unexpressed genes
or expressed genes that were undetected owing to the technical
limitations. Thus, the proteomic abundance, y, is distributed as
follows: y ¼ 0, probability mass at zero, with probability p;
where y follows a Poisson regression distribution with probabil-
ity (1  p). Therefore the observed protein abundance (y)
follows a mixture model:
 1δ
 λy
f ð y Þ ¼ p þ ð1  pÞ
expðλÞ δ ð1  pÞexpðλÞ
y!
ð1Þ
log itð pÞ ¼ log½ p=ð1  pÞ ¼ α0 þ β0 x ;

(a) where the indicator δ ¼ 1 if y ¼ 0; otherwise δ ¼ 0. We

also assume that p is dependent on the mRNA level (x)
through a logit model:
(b) where α0 and β0 are the intercept and slope in the logit
model.
128 Jiangxin Wang et al.

3. Parameters in the ZIP regression model are estimated with

maximum likelihood methods through SAS Proc NLMIXED
(SAS code is available upon request). This model allows
prediction of protein abundances for all genes, even under
the current technical limitations.
4. To describe the variation within a dataset, such as “molar
abundance” of proteins within one operon, the coefficient of
variation (CV) for each set of proteins is computed. The CV is
defined as the ratio of the standard deviation and the mean of
the “molar abundance” for a set of proteins (40) where the
calculation of CV score is independent of the sample size.
5. Based on the ZIP model, we are able to predict the abundance
of proteins that were undetectable owing to current technical
limitations. The prediction of the protein abundance for a gene
is expressed by

0
p þ expðα þ β
x Þ
ð1 pÞ,

where x is the mRNA abundance of that gene on a log scale and

p = expðα0 þ β0
x Þ= 1 þ expðα0 þ β0
x Þ, which is the
probability that the product of the gene was not expressed or
not detected simply owing to technical limitations.
6. For those expressed genes, we can further develop the ZIP
model into a Potential inference from ZIP regression (P-ZIP)
model to predict potential proteomic abundance. In this case,
100 % of the prediction should rely on the true distribution of
the protein abundance from the Poisson distribution instead of
assigning the probability p to 0 mass; i.e., the prediction is
expðα þ β
x Þ.

3.1.2 Validation 1. Cross validation: Cross validation is a technique for model

assessment that includes randomization. Input data is parti-
tioned into K equal parts where K – 1 sets are used to
train the model and the other set is used to calculate prediction
errors (41). This is repeated K times, yielding K prediction
errors values. An average and standard deviation (SD) can be
extracted to select the most representative model for future
prediction. Once the best model has been selected based on
cross validation, it is evaluated based on its coefficient of deter-
mination (R2) that represents the variation explained by the
model. The coefficient of determination (R2) is a statistical
measure representing the percentage of variance explained by
the model. R2 values range from 0 to 1. The closer the R2 to 1
the better the model is explaining the variance of the data.
Furthermore, as an alternative means to assess the goodness
of the model, we study the predictions of small sets of genes
grouped based on pathway, operon, and regulon information.
 Integrated Analysis of Transcriptomic and Proteomic Datasets Reveals. . . 129

In order to describe the variation within a dataset, such as

“molar abundance” of proteins within one operon, we com-
pute the coefficient of variation (CV) for each set of proteins.
2. Validation by biological knowledge: The information used
included gene organization information such as operon, and
gene function information such as regulon and pathway. We
test the mode prediction by assuming that relationships
between genes in operons, regulons, and pathways are tighter
than those between random gene sets. The “molar abundance”
of all proteins using protein abundance divided by molecular
weight is calculated, and it is hypothesized that the “molar
abundance” of pathway member proteins, i.e., ribosomal pro-
teins, should be roughly at the same level. To evaluate the
similarity of the “molar abundance” among the ribosomal
protein set, the CV values are calculated and compared with
that calculated for the whole genome excluding the genes from
the set. The validation is conducted by calculating the CV
within conditions for every operon, regulon, and pathway of
target species. These groups of genes are thought to have less
dispersion than a random set of genes by virtue of their intrinsic
biological relationship. To compare CV values we also perform
a permutation test in the following way. A CV is computed
from the protein prediction values for a set of randomly
selected genes. This step is repeated a thousand times through
resampling of genes without replacement. Repeating this
calculation a thousand times provided a CV-distribution to
calculate mean, SD, and percentile scores for groups with
random genes per condition.

3.2 Effects of Two different methods are used:

Sequences on
1. Free energy-based method:
Correlation of
Transcriptomic and (a) An analysis similar to what was described by Osada et al.
Proteomic Datasets
(42) is performed first, where the base-pairing potentials
between the 30 tail of 16S rRNA and 50 -UTR of all genes are
3.2.1 Identification and calculated and averaged by positions to view the overall
Analysis of Shine–Dalgarno trend in the whole genome.
Sequences (b) Sequences cctgcggctggatcacctccttt (NC_002937 and
NC_005863) and cctgcggttggatcacctcctta (U00096)
from the 30 end of 16S rRNA are extracted and used to
calculate the free energy values for D. vulgaris and E. coli,
respectively.
(c) The C programs used to perform the calculation are
provided by Dr. Y. Osada of the Institute for Advanced
Biosciences of Keio University.
130 Jiangxin Wang et al.

(d) To determine the effects of SD sequence during protein

translation, the 25-base and 50-base nucleotide sequences
immediately upstream of the start codon of each gene are
extracted, and the free energy for base pairing of 16S rRNA
with SD sequence for each gene is calculated. Each
extracted sequence is aligned with the 30 tail of 16S rRNA
to compute the minimal free energy (MFE) with the
Java Applet at http://www.mag.keio.ac.jp/%7Ersaito/
Research/BasePAP/BasePAP.html (implemented by Dr.
Y. Osada) (42).
(e) Since it is possible that various lengths of 16S rRNA tails
used in the calculation might affect the accuracy of the
MFE values, different sequence lengths (such as 13, 20,
and 23 bp) are used.
2. Probabilistic method:
(a) This method use a “seed” sequence to train a probabilistic
model of SD sequences, which was then used to find the SD
sequences in regions upstream of start codons of all genes.
A good seed sequence is the 30 end of the 16S rRNA (43).
(b) Two window sizes, 25 and 50, are used to search for SD
sequences using the RBSFinder program (43).

3.2.2 Identification of The identity of each start codon and stop codon is treated as a
Start Codon, Stop Codon, categorical variable during multiple regression analysis. The start
and Their Contexts codon context is defined as the upstream 30 bases and downstream
9 codons of the start codon. Therefore, each sequence of start
codon context is 60 bases long, including the start codon. To
evaluate the potential of each start codon context to form a stable
mRNA secondary structure, the minimum free energy of this
region is computed with the Vienna package RNAfold (44, 45).
The stop codon and the base immediately downstream of the stop
codon are regarded as the stop codon context. Each combination is
treated as a categorical variable in multiple regression analysis
described below.

3.2.3 Analyses of the The major trends in codon usage and amino acid usage are revealed
Overall Codon Usage and with a correspondence analysis. The relative synonymous codon
Amino Acid Usage usage (RSCU) is used in the correspondence analysis to remove
the effects of amino acid usage. For amino acid usage, the raw
codon counts are added up for each amino acid and used as input
in the correspondence analysis. The CodonW software (http://
codonw.sourceforge.net) is used for the correspondence analysis,
generating four major axes accounting for most of the variations in
codon usage or amino acid usage of D. vulgaris genes or proteins,
respectively (46).
 Integrated Analysis of Transcriptomic and Proteomic Datasets Reveals. . . 131

3.2.4 Correlation and 1. Correlation coefficients, such as Pearson’s correlation coefficient

Regression Analysis and Spearman’s rank correlation coefficient, are computed
(47). To obtain a reliable correlation between mRNA and
protein abundance, only proteins with variations among mea-
surement replicates less than threefold were included.
2. Single regression analyses are performed to measure correlation
pattern between mRNA and protein abundance as described
before (31, 48). Fold change in original scale is equivalent to
the arithmetic difference in the log scale,
also called range
of
samples (49). For instance,max y 1 ; y 2 ; y 3
=min y 1 ; y 2 ; y 3 =
3
is equivalent to log max y 1 ; y 2 ; y 3 -logmin y 1 ; y 2 ; y 3 
= logð3Þ. Previously, we reported the proteomic-mRNA corre-
lation through RmRNA2 from a simple regression:
y i ¼ α þ mRNAi
β ð2Þ
where mRNAi is the log of mRNA expression level for the ith
gene (31).
3. Multiple regression analyses are performed to measure correla-
tion pattern between mRNA and protein abundance and quan-
titative effects of sequence features according to the following
equation:
k
y i ¼ α þ mRNA i
β þ ∑ β j x i j ; ð3Þ
j ¼1

where xij refers to jth covariate (measuring sequence

feature such as codon usage, k is the number of covariates of
a particular sequence feature) of the ith gene, βj represents
the slope for the jth covariate (31). Particularly, the
R2mRNA, sequences R2mRNA
1R2mRNA
as the adjusted R2 is used for the mRNA–
protein correlation. For each covariate, the standard F-test can
be used to examine whether they are significant (p-value of the
F-test reported) (48).
4. Since the variation from each of these effects may not be
additive, their joint effects on the mRNA–protein correlation
are analyzed in a single multiple regression analysis by using the
following equation:
m
y i ¼ α þ mRNA i
β þ ∑ β j x i j ð4Þ
j ¼1

where all sequence features are included as covariates (m is the

total number of all covariates). P-values associated with each
covariate are also measured.
5. To evaluate the multiple regression model itself, bootstrap tests
are run by keeping sequence features unchanged for all genes,
while randomly permutating their proteomic abundance
among the genes so that the proteomic abundance of a given
132 Jiangxin Wang et al.

gene is randomly assigned to a different gene. The bootstrap

tests are run by randomly selecting 1,000 permutations for
each test. For each permutation, a multiple regression is fitted
and R2 is reported. The bootstrap p-value is reported as the
probability that the simulated R2 is larger than the R2 asso-
ciated with the real data. A smaller p-value suggests the R2
obtained for the real model is statistically more significant.
6. The two null models for the bootstrap tests are:
(a) The contribution by mRNA levels and all sequence features
is not larger than mRNA level alone.
(b) The contribution by mRNA levels and all sequence features
is no larger than that by mRNA levels and initiation-related
sequence features (excluding elongation and termination
related sequence features), respectively.

4 Conclusion

High-throughput experimentation measuring mRNA and protein

expression provides rich sources of information for better under-
standing of the metabolic mechanisms underlying complex
biological systems. With only partial proteomic data, the power of
integrative transcriptomic and proteomic analysis could be limited
and the analyses could be biased. There exists, therefore, an urgent
need to develop methodologies to accurately estimate missing
proteomic data to provide deeper insight into metabolic mechan-
isms underlying complex biological systems. Estimating missing
proteomic data is not a trivial task (31, 50). The data-driven ZIP
regression-based was developed model for integrative analysis of
these two different types of large-scale genomic data. This approach
is a significant improvement over previous methods since it allows
undetected proteins (those with an assigned protein abundance
value of zero) to be assigned a predicted abundance based on the
mRNA levels. This allows us to include the abundance of proteins
that were undetected owing to experimental or technical limita-
tions in our investigations. Moreover, the validity of this model is
evaluated using bioinformatics approaches. For example, in a com-
parison of the predicted protein abundance patterns of genes
belonging to the same operons (representing groups of proteins
that are expected to have similar molar abundance values), the
results demonstrate that the CV of estimated protein abundance
values within operons is indeed smaller than that for random
groups of proteins.
Although it is widely accepted that gene regulation in prokar-
yotes occurs also at the level of translation (51), few systematic
 Integrated Analysis of Transcriptomic and Proteomic Datasets Reveals. . . 133

quantitative analysis has been performed on the effects of various

translation-related sequence features on mRNA–protein correlation,
and it remains unclear how strong the translation regulation is.
Using the multiple regression method described above, a quantitative
measurement of effects of various sequence features on translation
efficiency can be performed using transcriptomic and proteomic
datasets. Using transcriptomic and proteomic datasets from D. vul-
garis grown under three conditions, multiple regression analyses of all
sequence features showed that the sequence features together
contribute up to 15.2–26.2 % of the total variation of mRNA–protein
correlation, suggesting that regulation at the translational level is
indeed involved in determining mRNA–protein correlation in D.
vulgaris. In addition, the analysis of all sequence features in the single
unified statistical framework also allows quantitative comparison of
the contribution of various features, which may lead to new biological
insight on microbial metabolism. For example, it has been suggested
that translation initiation is a rate-limiting step when compared with
the elongation and termination stages of protein biosynthesis
(52, 53). However, the study using D. vulgaris transcriptomic and
proteomic datasets found that features related to translation initiation
(start codon and start codon context) play only a minor role in
determining mRNA–protein correlation; to the contrary, the
sequence features involved in translation elongation, such as codon
usage and amino acid usage, may be more important in determining
mRNA–protein correlation in D. vulgaris. Although further valida-
tion is still needed, this result is consistent with an early study that
shows codon bias has the greatest influence on protein expression
levels (54).
Finally, caution should be taken when biological interpretations
of the predicted protein expression values and effects of sequence
features on translation process are conducted, as all the predictions
and calculation are constrained by the quality of the experimental
transcriptomic and proteomics data which are used as model input,
and further validation by either experimental or alternative compu-
tational methodologies is still needed.

5 Notes

ZIP model: In a previous study using multiple regressions, Nie et al.

(37) found that mRNA abundance alone can explain only 20–28 %
of the total variation of protein abundance, suggesting mRNA–
protein correlation cannot be determined solely on the basis of
mRNA abundance.
ZIP Advantages: First, for the proteins that have been experimen-
tally detected, the predicted protein abundance level values are
corrected with their mRNA levels being taken into consideration.
134 Jiangxin Wang et al.

Second, for the proteins undetected by experimental methods

(in the case of this study, >80 % of the proteins in the D. vulgaris
genome), the model predicts their protein abundance values.
Sequence features on correlation: Free energy-based method aligns
the 30 end of 16S rRNA and 50 -UTR of an mRNA and then uses a
dynamic programming algorithm to find the minimum free energy
of a window of specific size (42).
Sequence features on correlation: Although D. vulgaris is a GC-rich
species, the sequence of the 16S rRNA 30 tail is highly similar to
that of most other prokaryotes and it contains the typical anti-SD
sequence ctcct, which complements with the default seed sequence
aggag used in the RBSFinder program.
Sequence features on correlation: It is noteworthy that while our
previous studies have determined the effects of experimental
challenges and various physical properties of mRNA or proteins,
such as analytic variation, stability of mRNAs and proteins, and the
cellular functional category of genes/proteins (31), the method
described here focuses on the impact of the sequences to the
proteomic and mRNA correlation pattern. An attempt has also
been made to integrate all biological features and experimental
challenges into one multiple regression model, and we have found
that >71 % of mRNA–protein correlation variation can be account-
able (unpublished data), suggesting that using the methods
described here we have identified most of the factors that can
potentially affect mRNA–protein correlation.
Sequence features on correlation: F-test results show that the con-
tributions by translation-level sequence features are significantly
independent of the contributions by errors and protein stability
with a P-value <0.0001.
Sequence features on correlation: The results from the D. vulgaris
datasets show that the p-values for the first bootstrap null hypothesis
are 0 for the contributions computed in all growth conditions, and
the p-value for second bootstrap null hypothesis is less than 0.0001
for all three conditions. The results demonstrate that correlation of
mRNA expression and protein abundance is affected at a fairly
significant level by multiple sequence features related to transla-
tional efficiency in D. vulgaris.

References
1. Medini D, Serruto D, Parkhill J, Relman DA, 3. Uchiyama I, Mihara M, Nishide H, Chiba H
Donati C, Moxon R, Falkow S, Rappuoli R (2013) MBGD update 2013: the microbial
(2008) Microbiology in the post-genomic era. genome database for exploring the diversity of
Nat Rev Microbiol 6:419–430 microbial world. Nucleic Acids Res 41(Data-
2. Kyrpides NC (2009) Fifteen years of microbial base issue):D631–D635
genomics: meeting the challenges and fulfilling 4. Schoolnik GK (2001) The accelerating conver-
the dream. Nat Biotechnol 27:627–632 gence of genomics and microbiology. Genome
Biol 2: REPORTS4009
 Integrated Analysis of Transcriptomic and Proteomic Datasets Reveals. . . 135

5. Ward N, Fraser CM (2005) How genomics has protein composition, tissue diversity, and gene
affected the concept of microbiology. Curr regulation in mouse mitochondria. Cell
Opin Microbiol 8:564–571 115:629–640
6. Sharan R, Ideker T (2006) Modeling cellular 20. Alter O, Golub GH (2004) Integrative analysis
machinery through biological network com- of genome-scale data by using pseudoinverse
parison. Nat Biotechnol 24:427–433 projection predicts novel correlation between
7. Cardenas E, Tiedje JM (2008) New tools for DNA replication and RNA transcription. Proc
discovering and characterizing microbial diver- Natl Acad Sci U S A 101:16577–16582
sity. Curr Opin Biotechnol 19:544–549 21. Greenbaum D, Jansen R, Gerstein M (2002)
8. Rocha EP (2008) The organization of the bac- Analysis of mRNA expression and protein
terial genome. Annu Rev Genet 42:211–223 abundance data: an approach for the compari-
9. Fiehn O (2001) Combining genomics, meta- son of the enrichment of features in the cellular
bolome analysis, and biochemical modelling to population of proteins and transcripts. Bioin-
understand metabolic networks. Comp Funct formatics 18:585–596
Genomics 2:155–168 22. Ideker T, Thorsson V, Ranish JA, Christmas R,
10. Singh OV, Nagaraj NS (2006) Transcrip- Buhler J, Eng JK, Bumgarner R, Goodlett DR,
tomics, proteomics and interactomics: unique Aebersold R, Hood L (2001) Integrated geno-
approaches to track the insights of bioremedia- mic and proteomic analyses of a systematically
tion. Brief Funct Genomic Proteomic perturbed metabolic network. Science
4:355–362 292:929–934
11. Lin J, Qian J (2007) Systems biology approach 23. Washburn MP, Koller A, Oshiro G, Ulaszek G,
to integrative comparative genomics. Expert Plouffe D, Deciu C, Winzeler E, Yates JR III
Rev Proteomics 4:107–119 (2003) Protein pathway and complex cluster-
ing of correlated mRNA and protein expression
12. Kandpal R, Saviola B, Felton J (2009) The era analyses in Saccharomyces cerevisiae. Proc Natl
of omics unlimited. Biotechniques Acad Sci U S A 100:3107–3112
46:351–355
24. Greenbaum D, Colangelo C, Williams K, Ger-
13. Ishii N, Tomita M (2009) Multi-omics data- stein M (2003) Comparing protein abundance
driven systems biology of E. coli. In: Lee SY and mRNA expression levels on a genomic
(ed) Systems biology and biotechnology of scale. Genome Biol 4:117.1–117.8
Escherichia coli. Springer, Dordrecht, The
Netherlands, pp 41–57 25. Beyer A, Hollunder J, Nasheuer HP, Wilhelm
T (2004) Posttranscriptional expression regu-
14. Tang YJ, Martin HG, Myers S, Rodriguez S, lation in the yeast Saccharomyces cerevisiae on a
Baidoo EE, Keasling JD (2009) Advances in genomic scale. Mol Cell Proteomics
analysis of microbial metabolic fluxes via 13C 3:1083–1092
isotopic labeling. Mass Spectrom Rev
28:362–375 26. Nie L, Wu G, Zhang W (2006) Correlation of
mRNA expression and protein abundance
15. Park SJ, Lee SY, Cho J, Kim TY, Lee JW, Park affected by multiple sequence features related
JH, Han MJ (2005) Global physiological to translational efficiency in Desulfovibrio vul-
understanding and metabolic engineering of garis: a quantitative analysis. Genetics
microorganisms based on omics studies. Appl 174:2229–2243
Microbiol Biotechnol 68:567–579
27. Wilkins MR, Pasquali C, Appel RD, Ou K,
16. Gygi SP, Rochon Y, Franza BR, Aebersold R Golaz O, Sanchez J, Yan JX, Gooley AA,
(1999) Correlation between protein and Hughes G et al (1996) From proteins to pro-
mRNA abundance in yeast. Mol Cell Biol teomes: large scale protein identification by
19:1720–1730 two-dimensional electrophoresis and amino
17. Hegde PS, White IR, Debouck C (2003) Inter- acid analysis. Biotechnology (NY) 14:61–65
play of transcriptomics and proteomics. Curr 28. Scherl A, Francois P, Charbonnier Y, Deshusses
Opin Biotechnol 14:647–651 JM, Koessler T, Huyghe A, Bento M, Stahl-
18. Mootha VK, Lepage P, Miller K, Bunkenborg Zeng J, Fischer A et al (2006) Exploring
J, Reich M, Hjerrild M, Del-monte T, Ville- glycopeptide-resistance in Staphylococcus
neuve A, Sladek R et al (2003) Identification of aureus: a combined proteomics and transcrip-
a gene causing human cytochrome c oxidase tomics approach for the identification of
deficiency by integrative genomics. Proc Natl resistance-related markers. BMC Genomics
Acad Sci U S A 100:605–610 7:296
19. Mootha VK, Bunkenborg J, Olsen JV, Hjerrild 29. Zhang W, Gritsenko M, Moore RJ, Culley DE,
M, Wisniewski JR, Stahl E, Bolouri MS, Ray Nie L, Petritis K, Strittmat-ter EF, Camp DG,
HN, Sihag S et al (2003) Integrated analysis of
136 Jiangxin Wang et al.

Smith RD, Brockman FJ (2006) A proteomic inference, prediction. Springer, New York,
view of Desulfovibrio vulgaris metabolism as NY, USA
determined by liquid chromatography coupled 42. Osada Y, Saito R, Tomita M (1999) Analysis of
with tandem mass spectrometry. Proteomics base-pairing potentials between 16S rRNA and
6:4286–4299 50 UTR for translation initiation in various
30. Tuikkala J, Elo L, Nevalainen OS, Aittokallio T prokaryotes. Bioinformatics 15:578–581
(2006) Improving missing value estimation in 43. Suzek BE, Ermolaeva MD, Schreiber M, Salz-
microarray data with gene ontology. Bioinfor- berg SL (2001) A probabilistic method for
matics 22:566–572 identifying start codons in bacterial genomes.
31. Nie L, Wu G, Brockman FJ, Zhang W (2006) Bioinformatics 17:1123–1130
Integrated analysis of transcriptomic and pro- 44. Hofacker IL (2003) Vienna RNA secondary
teomic data of Desulfovibrio vulgaris: zero- structure server. Nucleic Acids Res
inflated Poisson regression models to predict 31:3429–3431
abundance of undetected proteins. Bioinfor- 45. Hofacker IL, Stadler PF (2006) Memory
matics 22:1641–1647 efficient folding algorithms for circular
32. Collins RF, Roberts M, Phoenix DA (1995) RNA secondary structures. Bioinformatics
Codon bias in Escherichia coli may modulate 22:1172–1176
translation initiation. Biochem Soc Trans 46. Wu G, Nie L, Zhang W (2006) Relation
23:76 between mRNA expression and sequence
33. Akashi H, Gojobori T (2002) Metabolic effi- information in Desulfovibrio vulgaris: combi-
ciency and amino acid composition in the pro- natorial contributions of upstream regulatory
teomes of Escherichia coli and Bacillus subtilis. motifs and coding sequence features to varia-
Proc Natl Acad Sci U S A 99:3695–3700 tions in mRNA abundance. Biochem Biophys
34. Tate WP, Poole ES, Dalphin ME, Major LL, Res Commun 344:114–121
Crawford DJ et al (1996) The translational 47. Devore J, Farnum N (2005) Applied statistics
stop signal: codon with a context, or extended for engineers and scientists. Thompson
factor recognition element? Biochimie Learning, Belmont, CA
78:945–952 48. Ott RY, Longnecker M (2001) An introduc-
35. Heidelberg JF, Seshadri R, Haveman SA, tion to statistical methods and data analysis.
Hemme CL et al (2004) The genome sequence Thompson Learning, Pacific Grove, CA
of the anaerobic, sulfate-reducing bacterium 49. Montgomery DC (2001) Introduction to sta-
Desulfovibrio vulgaris Hildenborough. Nat tistical quality control (Wiley series in statistics
Biotechnol 22:554–559 and probability). Wiley, New York
36. Zhang W, Culley DE, Scholten JC, Hogan M, 50. Nie L, Wu G, Culley DE, Scholten JC, Zhang
Vitiritti L, Brockman FJ (2006) Global tran- W (2007) Integrative analysis of transcriptomic
scriptomic analysis of Desulfovibrio vulgaris on and proteomic data: challenges, solutions and
different electron donors. Antonie Van Leeu- applications. Crit Rev Biotechnol 27:63–75
wenhoek 89:221–237
51. Lange R, Hengge-Aronis R (1994) The cellu-
37. Nie L, Wu G, Zhang W (2006) Correlation lar concentration of the S subunit of RNA
between mRNA and protein abundance in polymerase in Escherichia coli is controlled at
Desulfovibrio vulgaris: a multiple regression to the levels of transcription, translation, and pro-
identify sources of variations. Biochem Biophys tein stability. Genes Dev 8:1600–1612
Res Commun 339:603–610
52. Rocha EP, Danchin A, Viari A (1999) Transla-
38. McCullagh P, Nelder JA (1989) Generalized tion in Bacillus subtilis: roles and trends of
linear models. Chapman and Hall, Boca initiation and termination, insights from a
Raton, FL genome analysis. Nucleic Acids Res
39. Lambert D (1992) Zero-inflated Poisson 27:3567–3576
regression, with an application to defects in 53. Romby P, Springer M (2003) Bacterial transla-
manufacturing. Technometrics 34:1–14 tional control at atomic resolution. Trends
40. Johnson RA (2005) Miller and Freund’s prob- Genet 19:155–161
ability and statistics for engineers. Pearson 54. Lithwick G, Margalit H (2003) Hierarchy of
prentice Hall sequence-dependent features associated with
41. Hastie T, Tibshirani R, Friedman J (2001) The prokaryotic translation. Genome Res
elements of statistical learning-data mining, 13:2665–2673
Methods in Molecular Biology (2016) 1375: 137–153
DOI 10.1007/7651_2015_252
© Springer Science+Business Media New York 2015
Published online: 02 July 2015

Integrating Microarray Data and GRNs

L. Koumakis, G. Potamias, M. Tsiknakis, M. Zervakis, and V. Moustakis

Abstract
With the completion of the Human Genome Project and the emergence of high-throughput technologies,
a vast amount of molecular and biological data are being produced. Two of the most important
and significant data sources come from microarray gene-expression experiments and respective databanks
(e,g., Gene Expression Omnibus—GEO (http://www.ncbi.nlm.nih.gov/geo)), and from molecular path-
ways and Gene Regulatory Networks (GRNs) stored and curated in public (e.g., Kyoto Encyclopedia of
Genes and Genomes—KEGG (http://www.genome.jp/kegg/pathway.html), Reactome (http://www.
reactome.org/ReactomeGWT/entrypoint.html)) as well as in commercial repositories (e.g., Ingenuity
IPA (http://www.ingenuity.com/products/ipa)). The association of these two sources aims to give new
insight in disease understanding and reveal new molecular targets in the treatment of specific phenotypes.
Three major research lines and respective efforts that try to utilize and combine data from both of these
sources could be identified, namely: (1) de novo reconstruction of GRNs, (2) identification of Gene-
signatures, and (3) identification of differentially expressed GRN functional paths (i.e., sub-GRN paths that
distinguish between different phenotypes). In this chapter, we give an overview of the existing methods that
support the different types of gene-expression and GRN integration with a focus on methodologies that
aim to identify phenotype-discriminant GRNs or subnetworks, and we also present our methodology.

Keywords: Microarray, Gene expression, Gene regulatory networks, Pathways, Functional pathways,
Bioinformatics, Systems biology

1 Introduction

In recent years, high-throughput data capture technology, as with

microarray platforms, have vastly improved life scientists’ ability
to detect and quantify gene, protein, and metabolite expression.
The most common type, two-color microarrays, can measure the
expression of tens of thousands of genes with a single chip (1).
Applications include measuring gene expression in different devel-
opmental stages, identifying biomarkers for particular phenotypes
or diseases, and monitoring treatment response.
In the systems biology framework, scientists follow a “holistic”
approach in order to explore and study the behaviour of biological
components. System biology provides a global view of the dynamic
interactions in a biological system. On the molecular level, the
purpose of the underlying systems biology computational
approaches is to ascertain the interactions and dynamic behavior
137
138 L. Koumakis et al.

of molecules within a cell (2). The molecular mechanisms

determine how cells interact and how they develop and maintain
higher levels of organization and function. Systems biology tries to
formulate these mechanisms in mathematical models.
Currently bioinformatics community focuses on more
enhanced methods for gene selection on microarrays mainly by
adding and amalgamating knowledge from other sources, such as
GRNs. Integrating GRN information into the class comparison,
discovery, and prediction process is an important issue in bioinfor-
matics, mainly because the provided information possesses a true
biological content. By changing the focus from individual genes to
a set of genes or pathways, the gene set analysis (GSA) approach
enables the understanding of cellular processes as an intricate net-
work of functionally related components. A performance evalua-
tion of GSA methodologies (3) concluded that the inclusion of
additional biological features such as topology or covariates would
be more useful than simple gene selection approaches. In addition,
utilizing more domain knowledge is likely to reveal more insights in
the analysis.
Similar to bioinformatics, systems biology community took
advantage of the human genome and the microarray technology
to reconstruct and validate gene regulatory networks in an auto-
matic way. GRN reconstruction or reverse engineering aims toward
the inference GRN models from data (in most of the cases from
gene expression data). In the literature, a large number of compu-
tational methods are reported with the target of inferring gene
regulatory networks from expression data (4).
A relatively new line of research in the field is the identification
of the most discriminant GRNs, or parts of GRNs that differentiate
between specific phenotypes by coupling GRNs and microarray
data. Assessment of the discriminant power of (sub)networks is
based on the identification of those genes whose expression values
are consistent, i.e., could be justified, by their corresponding inter-
action pattern in the target GRN.
The study of the function, structure, and evolution of GRNs in
combination with microarray gene-expression profiles is essential
for contemporary biology research. Due to limitations in DNA
microarray technology—due to the different platforms utilised, to
the different experimental protocols, and mainly to small sample
sizes, higher differential expressions of a gene do not necessarily
reflect a greater likelihood of the gene being related to a disease and
therefore, focusing only on the candidate genes with the highest
differential expressions might not be the optimal procedure (5, 6).
Based on our knowledge, we propose a taxonomy of the meth-
odologies that combine gene-expression data and GRNs in order to
identify and assess discriminant pathway and subpathways (Fig. 1)
and a taxonomy of methodologies which identify and assess dis-
criminant pathway and sub-pathways (Fig. 2).
 Integrating Microarray Data and GRNs 139

Fig. 1 Integration of microarray data with gene regulatory networks

Fig. 2 Taxonomy of discriminant pathways and sub-pathways

A general observation concerns the different levels of knowl-

edge extraction from the GRNs employed by the different methods.
The first category naming pathway selection focuses on the
identification of differentially expressed pathways using microarray
data. Within this approach information about the topology, the
existing subpaths, as well as the reactions/relationships between
genes in a pathway are ignored. The second category subpathway
140 L. Koumakis et al.

selection using topology goes one step further and tries to identify the
discriminant pathways or subpathways. Within this approach iden-
tification and selection of the most discriminant paths ignore the
present gene relations/regulations. The last and most informative
category is the subpathway selection using regulatory mechanisms.
This approach takes advantage of the GRN topology as well as the
type of GRN gene relations (e.g., activation or inhibition).
Initial efforts used GRN information as groups (plain list) of
associated genes in order to identify the most discriminant and
phenotype-differentiating genes. Molecular pathways effectively
reduced the resulting sets of genes, extracted from a gene set analysis
approach, and in some cases improved prediction performance. But
GRNs encompass much more knowledge form just a plain list of
genes. Recently, more and more methods take advantage of the
GRNs topology and the underlying gene interaction patterns.
Pathway selection methodologies show similarities with gene
signatures in terms of the level of information used over the years.
Although GRNs hold important information about the structure
and correlation among genes that should not be neglected, most of
the currently available methods in pathway selection do not fully
exploit it. In the literature, one can find three categories of
methodologies that focus on the identification and selection
of discriminant pathways and subpathways, based on the different
levels of knowledge extraction from target GRNs. Initially the focus
was on the identification of differentially expressed pathways (as a
whole) using microarray data. Then the efforts concentrated on the
knowledge of the GRN topology using decomposition mechanisms
to reveal discriminant subpathways based on the graph theory
concepts and network visualization toolkits. Recently more
advanced methodologies are developed, which takes in consider-
ation not only the topology of the GRNs but also the regulation
type (activation/inhibition) of the interaction link that connects
two or more genes.
One can easily identify three main categories of methodologies
according to the level of the utilised GRN information. The cate-
gories are pathway selection using GRNs as list of genes, subpathway
selection using the topology of GRNs, and subpathway selection
methodologies using the underlying GRN gene regulatory interac-
tions. The last category—being in its infancy—exhibits the fewer
methodologies so far, but it takes the most out of GRNs and gene-
expression data compared to the other two, and is a promising
alternative for the identification of the regulatory mechanisms that
underlie and putatively govern various phenotypes.
The subpathway selection using the underlying GRN gene
regulatory interactions approach solves the major problem of the
set enrichment strategies that refers to the conflicting constrains
between GRNs and gene-expression data. A typical example of the
conflicting constrains is reflected in the situation when two
 Integrating Microarray Data and GRNs 141

significantly up-regulated genes increase the enrichment of the set

in microarray expression data, even if the first gene inhibits the
other in a GRN.

2 Method

We introduce a new methodology for the identification of

differentially expressed functional paths or subpaths within a gene
regulatory network (GRN) using microarray data analysis. The
analysis takes advantage of interactions among genes (e.g., activa-
tion, inhibition) as nodes of a graph network, which are derived
from expression data.
We propose a novel perception of GRNs and gene expression
data (Fig. 3). Initially we locate all functional paths encoded in
GRNs and we try to assess which of them are compatible with the
gene-expression values of samples that belong to different clinical
categories (diseases and phenotypes). The differential power of the
selected paths is computed and their biological relevance is assessed.
The whole approach is applied on a set of microarray studies with
the target of revealing putative regulatory mechanisms that govern
the treatment responses of specific phenotypes.
GRN and gene-expression data matching aims to differentiate
GRN paths and identify the most prominent functional sub-paths
for the given samples. In other words, the quest is for the subpaths
that exhibit high-matching scores for one of phenotypic class and
low-matching scores for the other. This is a paradigm shift from the
mining of differential genes to the mining of GRN functional
subpaths. The algorithm for differential subpath identification is
inherently simple.
Figure 4 provides an indicative example of the gene expression
limitation, where samples S1, S2, S3 belong to the “+” class and
samples S4, S5 belong to the “
” class. At the first sight, we can see
that no gene or no group of genes can discriminate 100 % our two
classes (“+” and “
”).
Figure 5 highlights the paradigm shift from the mining of
differential genes to the mining of GRN functional subpaths.
Given the same example as previously, we check our samples against
known sub-paths of GRNs.
The first path (IL-1R ! TRADD) satisfies samples 1,2,3,5.
Second path (IL-1R ! TRADD —| FLIP) satisfies samples S1,
S2, S3. Third path satisfies all samples and the fourth path doesn’t
satisfy any sample. The green arrow indicates that the second path
yields the maximum differential power, and it contains a potential
function differentiation since it contains only with samples that
belong to the “+” class (“!”: activation; “—|”: inhibition).
We rely on a novel approach for GRN processing that takes into
account all possible functional interactions in the network. Gene-
142 L. Koumakis et al.

Fig. 3 Flow of operations

 Integrating Microarray Data and GRNs 143

Fig. 4 Gene expression data example

Fig. 5 Matching functional sub-paths and gene-expression profiles

expression samples profiles and their phenotype assignments are

extracted form microarray data, and all targeted GRNs are evalu-
ated for the identification of the most informative ones.
The method unfolds into three modular steps.
1. Data preprocessing: On the one hand, gene expression values
are discretized into two states with values 1 and 0 for up-
regulated and down-regulated genes. On the other hand,
each target GRN is decomposed into its constituent subpaths.
2. Data annotation: Each subpath is interpreted on the basis of its
functional active-state, represented by a binary ordered-vector
with active states, resulting into its active-state ordered vector
<1,1,0> for the corresponding genes.
144 L. Koumakis et al.

Fig. 6 The gene discretization process

3. Analysis (data mining): The binary ordered-vector of each

subpath is aligned and matched against all (discretized) binary
gene-expression sample profiles. The subpaths are taking the
place of sample descriptor features and utilized for the con-
struction of subpath based phenotype prediction models.

2.1 Data We utilize discretization of the gene-expression continuous values

Preprocessing into the core of the gene-selection process. Discretization of a given
gene’s expression values means that each value is assigned to an
interval of numbers that represents the expression-level of the gene
in the given samples. A variable set of such intervals may be utilized
and assigned to naturally interpretable values e.g., low, high. Given
the situation that, in most of the cases, we are confronted with the
problem of selecting genes that discriminates between two classes
(i.e., disease-states) and we believe that it is convenient to follow a
two-interval discretization of gene-expression patterns. Below we
give a general statement of the discretization problem when two
classes are present, followed by an algorithmic process that heur-
istically solves it. Therefore, expression value represented with
0 indicates a nonexpressed or underexpressed gene, whereas value
of 1 indicates overexpressed gene. These values are being derived
using the following process (as also shown in Fig. 6):
 Integrating Microarray Data and GRNs 145

Fig. 7 Microarray discretization, an indicative example

1. The expression levels of gene A over the total number of

samples are sorted in descending order.
2. The midpoints between each two consecutive values are
calculated.
3. For each midpoint, the samples are clustered into two subgroups,
H and L.
4. For each midpoint, an information gain formula is applied,
which computes the entropy (7) of the system in respect to its
division into subgroups. IG(μκ) is the Information Gain of the
system for midpoint μκ. E(L) is the total entropy of the system
taking into account their prior assignment into classes (e.g.,
case–control), whereas E(L/μκ) ¼ E(Hκ,Lκ) is the entropy of
the system taking into account its division into subgroups
around midpoint μκ.
5. Finally, the midpoint that results in the highest information
gain is selected as the one which best discriminates against the
two subgroups, and all the samples in the H group are consid-
ered to be overexpressed getting a value of 1, whereas the ones
in the L group are the nonexpressed/underexpressed, getting a
value of 0.
This discretization process is applied to each gene separately,
and the final dataset is a matrix of discretized values. A similar
approach has been used before in other expression profiling studies
(8, 9). Figure 7 shows an indicative example of a “dummy” micro-
array with five genes (rows) and six samples (columns) categorized
into two classes, normal and diseased. To the left of the figure we
can see the absolute or normalized values of our “dummy” micro-
array and to the right we have the discretized matrix when we
applied the proposed methodology.
On the other hand, the origin of concurrent knowledge about
GRNs does not come from any concrete theoretic framework.
However, although incomplete, this knowledge covers almost
every biology function such as metabolism, genetic/environmental
146 L. Koumakis et al.

information processing, cellular processes, human diseases, and

drug development, while it is constantly under refinement and
enrichment. We chose to incorporate KEGG data for our analysis.
Since its first introduction in 1995, KEGG DB for pathways has
been widely used as a reference knowledge base for understanding
biological pathways and functions of cellular processes. The knowl-
edge from KEGG has proven of great value by numerous works in a
wide range of fields (10).
Although it has been shown that KEGG has some errors (11),
they are not so prominent and can be counterbalanced by the
simplicity, the variety and the standard ontology that KEGG pro-
vides. Through KEGG public database, pathways can be down-
loaded in KGML1 format. KGML (stands for KEGG Markup
Language) is an exchange format of KEGG graph objects including
GRNs. The GRN is described through standard graph annotation.
Nodes can be either genes, groups of genes, compounds, or other
networks. Edges can be one of the gene relations known from the
biology theory (activation, inhibition, expression, indirect, phos-
phorylation, diphosphorylation, ubiquination, association, and dis-
sociation). Each gene relation has a different semantic that depicts
the precise biology phenomenon that happens during the regula-
tion of the specific network.
Our approach relies on a novel processing for GRN that takes
into account all possible functional interactions of the network. The
different interactions correspond to the different functional sub-
paths that can be followed during the regulation of a target gene.
GRNs are downloaded from the KEGG repository. With an
XML parser (based on the specifications of KEGG’s KGML repre-
sentation of GRNs), we obtain all the internal network semantics.
In a subsequent step, all possible functional network subpaths are
extracted as exemplified in Fig. 8.

2.2 Data Annotation We exploit microarray experiments and respective gene-expression

data for which we expect (suspect) the targeted GRNs play an
important role. These paths uncover and present potential under-
lying gene regulatory mechanisms that govern the gene-expression
profile of the samples under investigation. Such a discovery may
guide the fine classification of samples as well as the reclassification
of diseases, based on the most prominent molecular evidence. The
samples of a binary transformed (discretized) gene-expression
matrix are matched against targeted molecular pathways and
respective GRN functional paths (retrieved form the pathway
decomposition).
A translation between the genes identifiers used in the gene
expression data to the corresponding KEGG identifiers is needed.

1
http://www.kegg.jp/kegg/xml/
 Integrating Microarray Data and GRNs 147

Fig. 8 Functional-path decomposition: Left: A target part of an artificial GRN; Right: The ten decomposed
functional sub-paths

Both the GRNs and the gene expression data have to use the same
ids. GRNs use gene ids while gene expression platforms use probes.
A probe is a specific segment of single-strand DNA that is comple-
mentary to a desired gene. For example, if the gene of interest
contains the sequence AATGGCACA, then the probe will contain
the complementary sequence TTACCGTGT. When added to the
appropriate solution, the probe will match and then bind to the
gene of interest.
Due to the large number of databases and associated IDs, the
conversion of gene identifiers is one of the initial and central steps in
many workflows related to genomic data analysis. In the literature
and the web, we can find several freely available ID conversion
tools. Although each tool has distinct features and strengths, as
reviewed by Khatri et al. (12), they all adopt a common core
strategy to systematically map a large number of interesting genes
in a list to the associated biological annotation.
The mapping from a thesaurus to another rises the many to one
issue which in our case many probes from the gene expression
dataset are assigned to the same KEGG gene ID. We check the
multiple probes for the gene and place a logic OR for the assess-
ment of the gene’s value. This is actually the selection of the value
of the probe with the highest intensity out of all the probes that
map to the same gene.
Then we need to identify the subpaths that exhibit high-
matching scores for one of phenotypic class and low-matching
148 L. Koumakis et al.

scores for the others. Each GRN subpath is interpreted according

to Kauffman’s principles and semantics (13):
1. The network is a directed graph with genes (inputs and outputs)
being the graph nodes and the edges between them represent-
ing the causal links between them, i.e., the regulatory reactions.
2. Each node can be in one of the two states, “ON,” the gene is
expressed or up-regulated (i.e., the respective substance being
present) or, “OFF,” the gene is not-Expressed or down-
regulated.
3. Time is viewed as proceeding in discrete steps—at each step the
new state of a node is a Boolean function of the prior states of
the nodes with arrows pointing towards it.
In order to cope with and reveal functional regulatory mechan-
isms we impose the following requirement over the formed sub-
paths: for a subpath to be considered as functional it should be
“active” during the GRN regulation process—in other words we
assume that all genes in a subpath are functional. For example,
consider the reaction A ! B, if A is “ON” then the activation/
expression (“!”) regulatory reaction is active, resulting into the
activation/expression of gene B (“ON”)—the same holds for an
inhibition (—|) reaction. In the case that gene A is “OFF” then the
reaction is considered as inactive with the state of the regulated
gene B to remain undetermined. Under this assumption, a path-
module is just a subpath (atomic or more complex) for which all its
reactions are considered as active. So, the state of all genes engaged
in a path-module that forms an ordered regulation pattern, e.g., the
pattern of the complex regulatory mechanism A ! D —| C is
<“ON,” “ON,” “OFF”>.
The samples of a binary transformed (discretized) gene-expression
matrix are matched against functional path-modules of target
GRNs. We follow an information-theoretic gene-expression discreti-
zation process.

2.3 Data Analysis As an example, assume the gene-expression binary profiles of six
artificial samples for genes A, B, D and C—with “1” to denote
“ON” and “0” to denote “OFF”—three of them are assigned to
phenotype-1 (S1, S2, and S3) and the other three to phenotype-2
(S4, S5, and S6)—refer to Fig. 9.
Furthermore, assume the artificial GRN shown in the left part
of Fig. 9, and its subpath A ! B ! D —| C (in bold). We follow a
logic-gates process that aims to match the path-module instance of
the subpath with the respective samples’ binary instances. The
process results into the formation of an ordered pattern that indi-
cate the samples for which the target sub-path is consistent with
(“1”s) or not (“0”s), i.e., the respective path-module
A¼“ON” ! B¼“ON” ! D¼“ON” —| C¼“OFF” is active.
 Integrating Microarray Data and GRNs 149

Fig. 9 Matching gene-expression sample profiles with GRN functional path-modules: a logic-gates approach

Note that for the finally inferred pattern of Fig. 9,

<1,0,1,0,0,0>, value “1” occurs in positions one and three,
which means that the examined path-module is active for samples
one and three; in all other samples it is inactive (“0”). As samples
one and three belong to phenotype-1, the target path-module
matches 2 out of 3 phenotype-1 samples, and zero phenotype-2
samples. In general, assume that there are S1 and S2 samples
that belong to phenotype-1 and phenotype-2, respectively, and
that path-module Pi matches Si;1 and Si;2 samples form
phenotype-1 and phenotype-2, respectively. Formula 1, computes
the differential power of a path-module with respect to the two
phenotypic classes;
Formula 1
S i ;1 =S 1
S i ;2 =S 2

The formula posses a polarity characteristic according the class

phenotype: positive for class S1 and negative for class S2; e.g., for
the above example, the differential power of path-module
A¼“ON” ! B¼“ON” ! D¼“ON” —| C¼“OFF” is (2/3)
– 0 ¼ 0.67, and as it positive it is interpreted and considered as a
regulation mechanism that governs phenotype-1.
After the decomposition of each of these pathways into its
functional components, each subpath has been matched against
the respective samples’ gene-expression profiles of the respective
microarray studies. The result is an array of sub-paths with binary
150 L. Koumakis et al.

values for every sample in the form of a discretized microarray.

Then using the machine learning library WEKA (14) we can extract
the most discriminant subpaths using ranking algorithms. A feasi-
bility study of the methodology approach is presented in the fol-
lowing section.

2.4 Experiments Most of breast cancer (BRCA) cases are estrogen responsive, imply-
ing the activation of a series of growth-promoting pathways, for
example, the estrogen receptor (ER) related ErbB signaling GRN.
In an effort to reveal the underlying regulatory mechanisms that
govern BRCA patients’ treatment responses we applied our meth-
odology on a public gene-expression study from the GEO, the
GSE73902 dataset targeting the ER phenotypic status of the
respective patients, i.e., ER+ (ER positive) vs. ER
(ER negative).
We targeted 14 pathways all of which are engaged within the
“Pathways in Cancer” integrated pathway of KEGG (hsa05200)
namely: ECM-receptor interaction (hsa04512), Cytocin-cytocin
receptor interaction (hsa04060), Adherens junction (hsa04520),
Wnt signaling (has04310), Focal adhesion (hsa04510), Jak-STAT
signaling (hsa04630), ErbB signaling (hsa04012), MAPK signaling
(hsa04010), mTOR signaling (hsa04150), VEGF signaling
(hsa04370), Apoptosis (hsa04210), p53 signaling (hsa04115),
Cell cycle (hsa04110), and TGF-β signaling (hsa04350).
The visualization of the results for the ErbB signaling
(hsa04012) can be found in Fig. 10 where with the help of the
Cytoscape3 graph library. The graph preserves the KEGG layout
topology. It is enriched with the expressed regulatory mechanisms
(relations) between genes that differentiate between the two
phenotypes and the color coding is as follows:
l Red indicates relations active at class 1 which in our example
is the ERpos.
l Blue indicates relations active at class 2 (ERneg).
l Magenta indicates overlapping relations in the two classes.
l Orange for subpaths that are always active.
The figure highlights only the “interesting” subpaths which in
our case are the most discriminant subpaths for the specific two
phenotypes.
Inspecting the reduced network, it is clear that there is a
pathway starting from NRG (1 and 2) and ends at inhibiting the
CDKN1B for ERpos phenotype; and a pathway starting from
TGFA or AREG or HBEFG that ends-up at inhibiting EIF4EBP1
for ERneg phenotype.

2
http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc¼gse7390
3
http://www.cytoscape.org/
 Integrating Microarray Data and GRNs 151

Fig. 10 Results of GSE7390 over 14 cancer related pathways

According to recent literature, the aforementioned results are

quite relevant to the estrogen-receptor status. Based on a search of
the related biomedical literature we focus our exploration on the
mechanisms underlying the resistance to pure estrogen antagonists
(e.g., fulvestrant). Recent studies show the significant role of both
ErbB3 and ErbB4 as alternative targets for the treatment of BRCA
patients. As Sutherland notes in ref. (15): “the initial growth inhib-
itory effects of fulvestrant appear compromised by cellular plasticity
that allows rapid compensatory growth stimulation via ErbB-3/4.
Further evaluation of pan-ErbB receptor inhibitors in endocrine-
resistant disease appears warranted.” In addition, Hutcheson et al.
in ref. (16) investigated whether induction of ErbB3 and/or ErbB4
may provide an alternative resistance mechanism to antihormonal
action. Their conclusion is that fulvestrant treatment is sensitive to
the actions of the ErbB3/4 ligand HRGb1 (NRG1) with enhanced
ErbB3/4-driven signaling activity, and significant increases in cell
proliferation.

3 Discussion and Conclusions

Current trend in GRNs and gene expression data is the subpathway

selection using regulatory mechanisms, which seems that it is at its
first steps and could possibly gain a momentum. Our assumption
152 L. Koumakis et al.

for that momentum amplifies with the similarities we can find

between the discriminant gene regulatory (sub)networks and
microarray gene selection methodologies.
Apart the proposed procedure, only four (4) other tools take
advantage of the underlying GRN gene regulation mechanisms,
naming GGEA (17), SPIA (18), TEAK (19), and PATHOME (20).
The main difference of the proposed methodology from these four
systems is the handling of the gene regulatory mechanisms. To our
knowledge all the other methodologies count with a +1 the activa-
tions and
1 the inhibitions. Each subpath gets a final score which is
also used as a ranking mechanism. Contrary, our approach strictly
checks and takes into account only subpaths that are functional
(according to the gene relations and the expression values). Our
approach is binary and leads to distinction between functional and
nonfunctional subpaths per sample instead of a representation of the
sub-path per class (the sum).
Our methodology relies on a novel approach for GRN proces-
sing that takes into account all possible functional interactions of
the network. The phenotype information is extracted from micro-
arrays and all the selected GRNs are evaluated for the identification
of the most informative GRNs at the specific phenotype. The
efficient ranking of subpaths provides the most differentiating and
prominent GRN functional subpaths for the respective target phe-
notypes. The formula posses a polarity characteristic according the
class phenotype, i.e., positive for class S1 and negative for class S2.
These subpaths present evidential molecular mechanisms that gov-
ern the disease itself, its type, its state or other targeted disease
phenotypes (e.g., positive or negative response to specific drug
treatment). The methodology was applied on a gene-expression
study with the target of identifying putative mechanisms that
underlie and govern the treatment response of breast cancer
patients according to their ER-status profiles. Results were quite
indicative and strongly supported by the relevant biomedical
literature.
It is known that integrating heterogeneous data sources is more
effective than working within the boundaries of a single scientific
technology/field. Bioinformatics and systems biology has proven
that taking advantage of the knowledge from each other can aid the
relevant scientific communities in their research endeavours or even
reveal and create new research domains. In most of the cases there
are levels of integration as well as levels of knowledge to be utilised.
Extracting out the most of the knowledge will always give us more
natural and meaningful, as well as more accurate results.
 Integrating Microarray Data and GRNs 153

Acknowledgment

This work was supported by the European Community’s Seventh

Framework Programme (FP7/2007-2013) under grant agreement
N 270089 and by the European Union (European Social Fund—
ESF) and by the European Union (European Social Fund—ESF)
and Greek national funds through the Operational Program “Edu-
cation and Lifelong Learning” of the National Strategic Reference
Framework (NSRF)—Research Funding Program: Heracleitus II
Investing in knowledge society through the European Social Fund.

References

1. Brown PO, Botstein D (1999) Exploring the 11. Ott MA, Gert V (2006) Correcting ligands,
new world of the genome with DNA microar- metabolites, and pathways. BMC Bioinformat-
rays. Nat Genet 21:33–37 ics 7(1):517
2. Huang Y, Zhao Z, Xu H, Shyr Y, Zhang B 12. Khatri P, Draghici S (2005) Ontological analy-
(2012) Advances in systems biology: computa- sis of gene expression data: current tools, lim-
tional algorithms and applications. BMC Syst itations, and open problems. Bioinformatics
Biol 6(3) 21:3587–3595
3. Hung J-H, Yang T-H, Zhenjun H, Weng Z, 13. Kauffman SA (1993) The origins of order: self-
DeLisi C (2012) Gene set enrichment analysis: organization and selection in evolution.
performance evaluation and usage guidelines. Oxford University Press, New York
Brief Bioinform 13(3):281–291 14. Hall M, Frank E, Holmes G, Pfahringer B,
4. Heckera M, Lambecka S, Toepferb S, van Som- Reutemann P, Ian H (2009) The WEKA data
erenc E, Guthke R (2009) Gene regulatory mining software: an update. SIGKDD Explora-
network inference: data integration in dynamic tions 11(1)
models—a review. Biosystems 96(1):86–103 15. Sutherland RL (2011) Endocrine resistance in
5. Ein-Dor L, Kela I, Getz G, Givol D, Domany E breast cancer: new roles for ErbB3 and ErbB4.
(2005) Outcome signature genes in breast can- Breast Cancer Res 13(3):106
cer: is there a unique set? Bioinformatics 21 16. Hutcheson IR et al (2007) Heregulin beta1
(2):171–178 drives gefitinib-resistant growth and invasion
6. Iwamoto T, Pusztai L (2010) Predicting prog- in tamoxifen-resistant MCF-7 breast cancer
nosis of breast cancer with gene signatures: are cells. Breast Cancer Res 9(4):50
we lost in a sea of data? Genome Med 2(11):81 17. Geistlinger L, Csaba G, K€ uffner R, Mulde N,
7. Shannon CEA (1948) Mathematical theory of Zimmer R (2011) From sets to graphs towards
communication. Bell Sys Tech J 27 a realistic enrichment analysis of transcriptomic
(3):379–423 systems. Bioinformatics 27(13):366–373
8. Potamias G, Koumakis L, Moustakis V (2004) 18. Tarca AL, Draghici S, Khatri P, Hassan SS,
Gene selection via discretized gene-expression Mittal P, Kim JS, Kim CJ, Kusanovic JP,
profiles and greedy feature-elimination. Meth Romero R (2009) A novel signaling pathway
Appl Artif Intelligence 3025:256–266 impact analysis. Bioinformatics 25(1):75–82
9. Li L, Weinberg CR, Darden TA, Pedersen LG 19. Judeh T, Johnson C, Kumar A, Zhu D (2013)
(2001) Gene selection for sample classification TEAK: Topology Enrichment Analysis frame-
based on gene expression data: study of sensi- worK for detecting activated biological sub-
tivity to choice of parameters of the GA/KNN pathways. Nucleic Acids Res 41(1):1425–1437
method. Bioinformatics 17(12):1131–1142 20. Nam S, Chang HR, Kim KT et al (2014)
10. Kanehisa M, Araki M, Goto S, Hattori M, Hir- PATHOME: an algorithm for accurately
akawa M, Itoh M, Yamanishi Y (2008) KEGG detecting differentially expressed subpathways.
for linking genomes to life and the environ- Oncogene 33(41):4941–4951
ment. Nucleic Acids Res 36:480–484
Methods in Molecular Biology (2016) 1375: 155–167
DOI 10.1007/7651_2015_284
© Springer Science+Business Media New York 2015
Published online: 28 October 2015

Biological Network Inference from Microarray Data,

Current Solutions, and Assessments
Swarup Roy and Pietro Hiram Guzzi

Abstract
Currently in bioinformatics and systems biology there is a growing interest for the analysis of associations
among biological molecules at a network level. A main research in this area is represented by the inference of
biological networks from experimental data. Biological network inference aims to reconstruct network of
interactions (or associations) among biological molecules (e.g., genes or proteins) starting from experi-
mental observations. The current scenario is characterized by a growing number of algorithms for the
inference, while few attention has been posed on the determination of fair assessments and comparisons.
Current assessments are usually based on the comparison of the algorithms using reference networks or
gold standard datasets. Here we survey some selected inference algorithms and we compare current
assessments. We also present a systematic listing of freely available inference and assessment tools for easy
reference. Finally we outline some possible future directions of research, such as the use of a prior
knowledge into the assessment process.

Keywords: Biological network inference, Assessment, Gene regulatory network, Gold standard, Gene
Ontology, Graph theory

1 Introduction

Bioinformatics and systems biology are recently moving towards the

modeling of the association among biological molecules, e.g., genes
or proteins, on a system level. The common assumption of this
research area is that biological molecules have a dense set of associa-
tions among them that need to be modeled and represented to
improve the knowledge in molecular biology. The best formalism
to represent such complex world comes from graph theory [1, 2].
There exist different representations on the basis of the consid-
ered molecules and kind of associations. Here we focus on gene
regulatory networks. In such kind of networks nodes represent
genes while edges represent association in terms of regulation
(i.e., positive or negative regulation) [3]. These networks are
often referred to as gene regulatory networks (GRNs).
The initial step of analysis is the inference of networks from
expression data derived from microarray experiments. Formally, a
biological network inference algorithm aims to reconstruct
155
156 Swarup Roy and Pietro Hiram Guzzi

network of interactions (or associations) among biological genes

starting from experimental observations. The reconstruction of
GRNs is an important area since it enables for instance the compar-
ison of GRNs of different states (e.g., health or disease) or the
investigation of the progression of diseases [4] as well as the visual
inspection of network properties [5]. To cope with this problem a
number of different algorithms has been introduced in the past
(see for instance [6]).
Once that networks have been determined, there is the need to
evaluate the ability of the algorithms in terms of reliability of the
network, i.e., how many false positive (or incoherent associations
among genes) and how many false negatives (or missing associa-
tions) are expected using a particular algorithm. The Dialogue on
Reverse Engineering Assessment and Methods (DREAM) [7] proj-
ect has introduced a rigorous methodology for an objective assess-
ment of biological network reconstruction algorithms. DREAM
assessment relies on three main steps: (1) periodically organizers
provide a set of experimental data, (2) researchers try their own
algorithms on these datasets and submit resulting networks, and
(3) organizers score networks on the basis of gold standard data-
sets. DREAM challenge use mainly yeast and worm data and the
scoring is obtained only with topological consideration (i.e., the
number of true/false-positive edges).
Limitations of this approach have been yet discussed on some
recent articles (see for instance Siegenthaler et al.).
Here we present some of the network inference algorithms and
we discuss main classes in which they may be categorized. Then we
discuss currently network assessment methods.

2 Basic of Gene Network Structure

Graph theory offers the best formalism to represent information

modeled as a network [8]. A graph G¼(V,E) is a structured com-
posed by a set of nodes V and a set of edges E, i.e., pair of nodes.
Graphs are usually distinguished by the kind of edges. A major
distinction is made among directed and undirected networks.
In directed networks edges are directed, i.e., are sorted pairs of
nodes, while in undirected network edges have no direction [9].
Gene regulatory networks can be represented by using graphs
[8] where nodes are associated to genes, and edges represent asso-
ciations among proteins. The simplest representation uses an undi-
rected graph, while more refined models use directed and weighted
edges to integrate the information about the kind of biochemical
association and its direction. The undirected graph represents sig-
nificant association or co-expression over a series of gene expression
measurements. They often referred as gene association network or
gene co-expression network [10]. The directed edges in GRNs
Biological Network Inference from Microarray Data, Current Solutions, and Assessments 157

YCR084C

YFL026W
YCL067C

YIL015W
YHR084W

YNL145W

YDR461W YMR043W

YPR113W

YKR097W

YJL157C YAL040C

Fig. 1 Shows a graph representing a portion of a real GRN

correspond to causal influences between gene activities (nodes).

These could include regulation of transcription by transcription
factors, but also less intuitive causal effects between genes involving
signal transduction or metabolism. For instance, Fig. 1 represents
the association of some selected genes in yeast. As evidences, graph
takes into account only the presence (or absence) of an association
among nodes. Differently, Fig. 2 is an example of the same network
corredated with the information about the direction of nodes. As
evident, it is possible to individuate which node influences and
which node is influenced by. More sophisticated models may
include the kind of influence (e.g., positive or negative).
A gene may directly influence the activity of other target gene or
gene product. Influence may be indirect by coding a transcription
factor (TF) that in turn regulates another gene. A possible causal
relationship in GRN is shown in Fig. 3. Apparently, four different
types of causal relationship may be possible in a living cell. Based on
the above figure we can derive following causal relationship [11].
1. A gene can enhance the activity of more than one gene (rela-
tionship between A, B, C, and D).
2. A gene’s activity may be influenced by more than one gene
(relationship between B, D, and F). Often F is referred as
Collider [12].
158 Swarup Roy and Pietro Hiram Guzzi

YCR084C

YFL026W

YCL067C

YIL015W
YHR084W

YNL145W

YDR461W YMR043W

YPR113W

YKR097W

YJL157C YAL040C

Fig. 2 Shows a graph representing a directed graph modeling of a GRN

Fig. 3 Shows possible causal dependency in GRN graph

 Biological Network Inference from Microarray Data, Current Solutions, and Assessments 159

3. Gene can also influence the activity of itself (node B).

4. A gene may inhibit activity of other gene (D inhibits E).
Inhibition or negative regulation may also follow above three
relationships, i.e., many to one, one-to-many, and self.
Once that a GRN is modeled by using graphs, the study of
biological properties can be done using graph-based algorithms
[1], and associating graph properties to the biological properties
of the modeled GRN [13].

3 Network Inference Algorithms

A number of techniques have been proposed for network inference.

Existing techniques for finding gene networks can be broadly cate-
gorized as (i) computational approaches, and (ii) literature-based
approaches. The computational approach mainly uses statistical,
machine learning, or soft-computing techniques [14, 15] as discov-
ery tools. On the other hand, a literature-based approach gathers
relevant published information on genes and their interrelation-
ships and constructs networks based on such documented informa-
tion. The literature-based approach is capable of building networks
with high biological relevance but is computationally expensive.
A biomedical literature search-based technique is used in [16, 17]
to construct gene relation networks by mapping literature knowl-
edge into gene expression data.
Network models such as Bayesian [18] and Boolean networks
[19] are used to infer interrelationships among genes. Graphical
Gaussian models (GGM) [20] and categorical Bayesian networks
(CBN) [21] are used to infer networks of differentially expressed
genes using Bayesian network. Kwon et al. [22], extract gene
regulatory relationships for cell cycle-regulated genes with activa-
tion or inhibition between gene pairs. Recently, Jump3 [23], a
hybrid Boolean network and decision tree-based method, has
been proposed to predict regulatory network from time-series
data. Regulatory relationships have also been deduced from corre-
lation of co-expressions, between DNA-binding transcription reg-
ulator and its target gene, by using a probabilistic expression model
[24]. Although standard statistical techniques for extracting rela-
tionships can come up with multiple models to fit the data, they
often require additional data to resolve ambiguities. Soft comput-
ing tools like fuzzy sets, neuro-computing, evolutionary comput-
ing, and their hybridization are alternatives for handling real-life
ambiguities. Mitra et al. [25] propose a bi-clustering technique to
extract simple gene interaction networks. They use continuous
column multi-objective evolutionary bi-clustering to extract rank
correlated gene pairs. Such pairs are used to construct the gene
network for generating relationship between a transcription factor
160 Swarup Roy and Pietro Hiram Guzzi

and its target’s expression level. Similarly, Jung and Cho [4] also
propose an evolutionary computation-based approach for con-
struction of gene (interaction) networks from gene expression
time-series data. It assumes an artificial gene network and compares
it with the reconstructed network from the gene expression time-
series data generated by the artificial network. Next, it employs real
gene expression time-series data to construct a gene network by
applying the proposed approach. Mutual information [26, 27] or
correlation coefficient [28, 29] based approaches have been pro-
posed for extracting gene-gene interaction networks. It has been
observed that a pair of genes with high mutual information are
nonrandomly associated with each other biologically or with
biological significance. Butte et al. [26] compute comprehensive
pair-wise mutual information for all genes in an expression dataset.
By picking a threshold mutual information and using only associa-
tions at or above the threshold, they construct relevance networks.
A number of additional mutual information-based approaches have
also been proposed. Some of the well-known algorithms in this
category are CLR [30], ARACNE [31], and MRNET [32].
GENIE3 [33] is a Random Forest-based method that model the
inference as a regression problem. Recently, GeCON [34] a pattern
based co-expression network inference method has been proposed
capable of detecting undirected co-regulated network with regula-
tion information (+ or ). A summery of various inference methods
discussed above is reported in Table 1.

4 Network Inference Assessment

Formally, the assessment of network inference algorithm refers to

the possibility to score (or rank) inferred networks in terms of
distance among inferred and real networks. A similar problem has
been posed in some other bioinformatics fields (such as protein
structure prediction). However in such fields, the existence of real
structures, i.e., proteins whose structure was determined and con-
firmed by in vitro experiments, made the assessment more simple.
The CASP challenge published for each edition the primary
sequence of an unknown protein. Then researcher try to determine
the spatial structure. Finally candidate structures are compared with
respect to the real structure revealed by in vitro experiments. More-
over there is not a gold-standard experiments for establishing the
true network structure (i.e., the ground truth) for real networks.
Consequently, all the assessment are designed on top of in silico
standard (i.e., simulated network) since existing networks cannot
be used since their identity may not be hidden [35–37].

4.1 In Silico As introduced before the assessment of GRN algorithms has been
Assessment Methods largely discussed within the Dialogue on Reverse Engineering
Assessment and Methods (DREAM) project [38]. The project is
 Biological Network Inference from Microarray Data, Current Solutions, and Assessments 161

Table 1
Synopsis of GRN inference methods

Inferred
network Assessment based
Algorithm Approach type Package Platform on (type of data)
GGM Bayesian network Undirected GeneNeta R Synthetic (statically simulated)
and real (breast cancer)
CBN Bayesian network Directed CatNetb R Real (breast, lung, gastric, and
renal cancer)
GENIE3 Random Forest Directed GENIE3c R, MatLab Synthetic (DREAM4) and
real (E. coli)
CLR Mutual information Undirected MINETd R Real (E. coli)
e
ARACNE Mutual information Undirected MINET R Synthetic (random network)
and real (human B cells)
MRNET Mutual information Undirected MINETf R Synthetic (sRogers and
SynTReN)
GeCON Expression pattern Undirected GeCONg Java Synthetic (DREAM4), real
similarity (yeast, human, rat, mouse,
rice)
JUMP3 Decision tree and Directed Jump3h Matlab DREAM4, IRMA
Boolean network
a
http://strimmerlab.org/software/genenet/
b
http://cran.r-project.org/web/packages/catnet
c
http://www.montefiore.ulg.ac.be/huynh-thu/software.html
d
http://minet.meyerp.com
e
http://minet.meyerp.com
f
http://minet.meyerp.com
g
https://sites.google.com/site/swarupnehu/publications/resources
h
http://homepages.inf.ed.ac.uk/vhuynht/misc/jump3.zip

organized on an annual basis and researchers from all the world may
participate on this context. The structure of the competition is
quite simple. For each year, organizer provides many test dataset
to the community of participant. Then researchers may test their
own algorithms on these datasets and they may determine a candi-
date network for each dataset. Once completed, researchers send
candidate network to the organizers for the analysis of results.
Network analysis is based on the comparison of candidate
network with a priori-determined network (one for each dataset)
that represents reference or gold standard network. The measure of
distance from the gold standard network is calculated by evaluating
the confusion matrix, i.e., a matrix containing the numbers of true
positives/negatives and false positives/negatives. All the submis-
sion are reviewed manually by the organizers and submission are
then scored by deriving receiver operating characteristics (ROC) or
precision-recall.
162 Swarup Roy and Pietro Hiram Guzzi

The DREAM challenge is currently the de-facto standard for

the assessment of biological network inference, but there are some
drawbacks that have been analyzed in some recent works.
As noted in [39], a missing assumption in DREAM challenge is
that the network inference problem is an undetermined problem
since experimental data do not containing all the needed informa-
tion to reconstruct network completely. In other word, a single
microarray experiment do not examine all the biological molecules.
For instance, some microarray platforms do not cover all the tran-
scriptome, or some transcripts are missing for experimental biases
such as errors during the experiments, noise in the data, and the
number and types of network gene perturbation experiments. All
these factors make the network inference an underdetermined
problem. Consequently, as investigated in [39], some gene inter-
actions cannot be inferred and more than one network can agree
with the data.
To cope with this problem, Siegenthaler et al. proposed a novel
assessment procedure that incorporates the inferability of gene
regulatory interactions by redefining the confusion matrix in
terms of inferability of the network, i.e., the possibility of the
network to be determined from data. The inferability of GRNs
was analyzed based on the causal information that could be
extracted from experiments. Authors used data from the DREAM
4 In Silico Network Inference Challenge to score their assessment
and to show the concept of inferability from experiments.
A new performance score was introduced based on a redefini-
tion of the confusion matrix, by taking into consideration non-
inferable gene regulations. Results confirmed the same best
performing teams as in the DREAM 4 inference challenge, but
they changed in a significant way the overall team rankings.

4.2 Assessment Limited availability of true gold standard network (of all living
Against Biological organisms) for validation of a candidate inference method imposes
Truth an additional challenge to the system biologist to select a suitable
inference method for their experimentation.
Literature-based known interactions are explored to validate an
inference method from biological significance point of view. Exper-
imentally validated regulators collected from publicly available
databases like RegulonDB [40] are used to assess the performance
of an inference method [33]. Pathways provide a way of linking the
functionality of groups of genes to specific biological processes.
Well-established methodologies such as Gene Set Enrichment
Analysis (GSEA) [41] help in differentiating pathways as functional
units from experimental populations. Manually curated pathways
based on expert knowledge and existing literature obtained from
the Kyoto Encyclopedia of Genes and Genomes (KEGG, http://
www.genome.jp/kegg/pathway.html) are another alternative mea-
sure used for validation [21].
 Biological Network Inference from Microarray Data, Current Solutions, and Assessments 163

To evaluate the biological significance of a inference method,

researchers explored an alternative measure based on Gene Ontol-
ogy (GO) against functional, biological enrichment of a group of
genes derived from inferred network modules [34]. More a method
is able to generate biologically significant modules (in terms of
p value) more they are relevant.
Next we highlight on some of the publicly available software
tools for in silico inference and validation of an inferred network.

5 Software for Network Reconstruction and Assessments

A number of software tools, both web and desktop versions are

available to facilitate the in silico reconstruction and visualization of
gene regulatory networks from microarray expression data. Some
of the tools also provide an integrated platform for assessment of
inferred network against gold standard synthetic network. Readers
may refer [42] for a comprehensive review on various available
software tools.
GeneNetWeaver (GNW) [43] is a Java-based reverse engineer-
ing tool for generating synthetic benchmark expression datasets
from gold standard DREAM challenge network. E. coli and Yeast
transcriptional regulatory networks are integrated as test case for
benchmark. Comparative assessment of inference algorithms
against DREAM challenge data can also be performed with the
help GNW. Cytoscape [44] is a powerful tool most suitable for
large-scale network analysis. Cytoscape supports directed, undi-
rected, and weighted graphs and comes along with powerful visual
styles thereby allowing users to change the properties of nodes or
edges. Plenty of elegant layout algorithms including cyclic and
spring-embedded layouts are available for visualization. GENeVis
[45] is a Java desktop application that allows visualization of gene
regulatory networks. STARNET2 [46] facilitates discovery of puta-
tive gene regulatory networks in a variety of species (Homo sapiens,
Rattus norvegicus, Mus musculus, Gallus gallus, Danio rerio, Dro-
sophila, C. elegans, S. cerevisiae, Arabidopsis, and Oryza Sativa) by
graphing networks of genes that are closely co-expressed across a
large heterogeneous set of preselected microarray experiments.
NetBioV [47] (Network Biology Visualization) is an R package
that allows visualization of large network data in biology and medi-
cine. Fast MEDUSA [48] is a parallel program to infer gene regu-
latory networks from gene expression and promoter sequences
which is implemented in C++. BANJO [49] a software application
and framework for structure learning of static and dynamic Bayes-
ian networks. LegumeGRN [50] is a specialized tool for legume
species. The tool is pre-loaded with gene expression data for Med-
icago, Lotus, and Soybean. A summery of few of the tools are listed
in Table 2.
164 Swarup Roy and Pietro Hiram Guzzi

Table 2
Few free GRN inference and visualization tools

Tool Platform Availability Visualization Plug-ins Url

GNW Java Off-line Yes No http://gnw.sourceforge.net/
GENeVis Java Off-line Yes Yes http://www.win.tue.nl/
mwestenb/genevis/
Cytoscape Java Off-line/ Yes Yes http://www.cytoscape.org/
Online
GeneNetwork C++ Online Yes No http://www.genenetwork.
org/webqtl/main.py
STARNET2 web application Online Yes No http://vanburenlab.tamhsc.
edu/starnet2.html
LegumeGRN J2EE Online Yes No http://legumegrn.noble.org
Osprey Java Online Yes No http://biodata.mshri.on.ca/
osprey/servlet/Index
NetBioV R Off-line Yes No http://bivi.co/visualisation/
netbiov
FastMEDUSA C++ Off-line Yes No https://wiki.nci.nih.gov/
display/NOBbioinf/
FastMEDUSA

6 Conclusion

The analysis of associations among biological molecules at a system

level is a large research area in bioinformatics and systems biology.
The workflow of research in this area starts with molecular biology
experiments that investigate the behavior of molecules. Then a set
of analysis algorithms have to be used to infer associations among
molecules. This research has a main role in determination of
mechanisms of regulations of the expression of genes. In such a
context algorithms aim to reconstruct network of interactions (or
associations) in genes starting from experimental observations
obtaining gene regulatory network. The current scenario is char-
acterized by a growing number of algorithms for the inference,
while few attention has been posed on the determination of fair
assessments and comparisons. Current assessments are usually
based on the comparison of the algorithms using reference net-
works or gold standard datasets. Here we surveyed some selected
inference algorithms and we compare current assessments. To the
best of our knowledge there are some open problems and chal-
lenges. First of all, since the structure of real regulatory networks is
far to be complete, the network inference remains still an
 Biological Network Inference from Microarray Data, Current Solutions, and Assessments 165

undetermined problem. Consequently the investigation of alterna-

tive methods of assessment under the absence of a gold standard
methods is a key area. Moreover the research may be helped by the
introduction of prior knowledge (e.g., biological ontologies) on
the assessment and during the inference may help to filter out
possible low quality networks [53]. The scenario we envision is
characterized by the use of ontologies both in the assessment
phase and during the inference of network. Considering the assess-
ment phase it should be pointed out that low attention has been
posed on the evaluation of the inferred network with respect to
diseases or phenotypes. Briefly speaking, all the current assessment
methods are not able to quantify how much inferred networks are
related to biology since they are related only to topology. The
problem has been investigated in other fields such as biological
network alignment [51, 52] Moreover, ontologies may also be
used during the generation of candidate networks in a similar way
as proposed for inferable network by Siegenthaler et al. In that work
the inferability notion is used to filter out low scoring network,
while in the scenario we envision ontologies may be used to filter
out low-scoring network from a biological point of view.

References

1. Cannataro M, Guzzi PH, Veltri P (2010) Pro- toward parameter identification for whole-cell
tein-to-protein interactions: technologies, models. PLoS Comput Biol 11(5):e1004096
databases, and algorithms. ACM Comput 8. Godsil C, Royle GF (2013) Algebraic graph
Surv (CSUR) 43(1):1 theory, vol 207. Springer Science & Business
2. Cannataro M, Guzzi PH, Sarica A (2013) Data Media, New York
mining and life sciences applications on the 9. Cannataro M, Guzzi PH, Veltri P (2010)
grid. WIREs Data Mining Knowl Discov 3 Impreco: distributed prediction of protein
(3):216–238 complexes. Futur Gener Comput Syst 26
3. Levine M, Davidson EH (2005) Gene regu- (3):434–440
latory networks for development. Proc Natl 10. Fuente ADI (2010) What are gene regulatory
Acad Sci U S A 102(14):4936–4942 networks? Handbook of research on computa-
4. Jung SH, Cho K-H (2004) Identification of tional methodologies in gene regulatory net-
gene interaction networks based on evolution- works. IGI Global, Hershey, PA, pp 1–27
ary computation, AIS. Springer, New York, pp 11. Roy S, Das D, Choudhury D, Gohain GG,
428–439 Sharma R, Bhattacharyya DK (2013) Causality
5. Agapito G, Guzzi PH, Cannataro M (2013) inference techniques for in-silico gene regu-
Visualization of protein interaction networks: latory network, Mining intelligence and knowl-
problems and solutions. BMC Bioinformatics edge exploration. Springer, New York,
14(Suppl 1):S1 pp 432–443
6. Marbach D, Prill RJ, Schaffter T, Mattiussi C, 12. Olsen C, Meyer PE, Bontempi G (2009) Infer-
Floreano D, Stolovitzky G (2010) Revealing ring causal relationships using information the-
strengths and weaknesses of methods for gene oretic measures. In Proceedings of the 5th
network inference. Proc Natl Acad Sci 107 Benelux Bioinformatics Conference (BBC09)
(14):6286–6291 13. Mina M, Guzzi PH (2014) Improving the
7. Karr JR, Williams AH, Zucker JD, Raue A, robustness of local network alignment: design
Steiert B, Timmer J, Kreutz C, Wilkinson S, and extensive assessment of a Markov
Allgood BA, Bot BM et al (2015) Summary of clustering-based approach. IEEE/ACM Trans
the DREAM8 parameter estimation challenge: Comput Biol Bioinformatics 11(3):561–572
166 Swarup Roy and Pietro Hiram Guzzi

14. Mitra S, Das R, Hayashi Y (2011) Genetic net- genome-wide expression patterns. Proc Natl
works and soft computing. IEEE/ACM Trans Acad Sci 95(25):14863–14868
Comput Biol Bioinformatics 8(1):94–107 28. Tong AHY, Lesage G, Bader GD, Ding H, Xu
15. Nagrecha S, Lingras PJ, Chawla NV (2013) H, Xin X, Young J, Berriz GF, Brost RL, Chang
Comparison of gene co-expression networks M et al (2004) Global mapping of the yeast
and Bayesian networks, Intelligent Information genetic interaction network. Science 303
and Database Systems. Springer, New York, pp (5659):808–813
507–516 29. Kuo WP, Mendez E, Chen C, Whipple ME,
16. Karopka T, Scheel T, Bansemer S, Glass Ä Farell G, Agoff N, Park PJ (2003) Functional
(2004) Automatic construction of gene rela- relationships between gene pairs in oral squa-
tion networks using text mining and gene mous cell carcinoma, AMIA annual symposium
expression data. Med Inform Internet Med 29 proceedings. American Medical Informatics
(2):169–183 Association, Bethesda, MD, p 371
17. Özg€ ur A, Vu T, Erkan G, Radev DR (2008) 30. Faith JJ, Hayete B, Thaden JT, Mogno I,
Identifying gene-disease associations using Wierzbowski J, Cottarel G, Kasif S, Collins JJ,
centrality on a literature mined gene- Gardner TS (2007) Large-scale mapping and
interaction network. Bioinformatics 24(13): validation of escherichia coli transcriptional
i277–i285 regulation from a compendium of expression
18. Friedman N, Linial M, Nachman I, Pe’er D profiles. PLoS Biol 5(1):e8
(2000) Using Bayesian networks to analyze 31. Margolin AA, Nemenman I, Basso K, Wiggins
expression data. J Comput Biol 7 C, Stolovitzky G, Favera RD, Califano A
(3–4):601–620 (2006) Aracne: an algorithm for the recon-
19. Davidich MI, Bornholdt S (2008) Boolean struction of gene regulatory networks in a
network model predicts cell cycle sequence of mammalian cellular context. BMC Bioinfor-
fission yeast. PLoS One 3(2), e1672 matics 7(Suppl 1):S7
20. Sch€afer J, Strimmer K (2005) An empirical 32. Meyer PE, Kontos K, Lafitte F, Bontempi G
Bayes approach to inferring large-scale gene (2007) Information-theoretic inference of
association networks. Bioinformatics 21 large transcriptional regulatory networks.
(6):754–764 EURASIP J Bioinforma Syst Biol 2007:79879
21. Balov N (2013) A categorical network 33. Huynh-Thu VA, Irrthum A, Wehenkel L,
approach for discovering differentially Geurts P (2010) Inferring regulatory networks
expressed regulations in cancer. BMC Med from expression data using tree-based meth-
Genet 6(Suppl 3):S1 ods. PLoS One 5(9):e12776
22. Kwon AT, Hoos HH, Ng R (2003) Inference 34. Roy S, Bhattacharyya DK, Kalita JK (2014)
of transcriptional regulation relationships from Reconstruction of gene co-expression network
gene expression data. Bioinformatics 19 from microarray data using local expression pat-
(8):905–912 terns. BMC Bioinformatics 15(Suppl 7):S10
23. Sanguinetti G et al (2015) Combining tree- 35. Moult J, Fidelis K, Kryshtafovych A, Rost B,
based and dynamical systems for the inference Hubbard T, Tramontano A (2007) Critical
of gene regulatory networks. Bioinformatics 31 assessment of methods of protein structure
(10):1614–1622 prediction-round vii. Proteins 69(S8):3–9
24. Segal E, Taskar B, Gasch A, Friedman N, Koller 36. Mendes P, Sha W, Ye K (2003) Artificial gene
D (2001) Rich probabilistic models for gene networks for objective comparison of analysis
expression. Bioinformatics 17(Suppl 1): algorithms. Bioinformatics 19(Suppl 2):
S243–S252 ii122–ii129
25. Mitra S, Das R, Banka H, Mukhopadhyay S 37. Marbach D, Schaffter T, Mattiussi C, Floreano
(2009) Gene interaction–an evolutionary D (2009) Generating realistic in silico gene
biclustering approach. Information Fusion 10 networks for performance assessment of
(3):242–249 reverse engineering methods. J Comput Biol
26. Butte AJ, Kohane IS (2000) Mutual informa- 16(2):229–239
tion relevance networks: functional genomic 38. Stolovitzky G, Monroe D, Califano A (2007)
clustering using pairwise entropy measure- Dialogue on reverse-engineering assessment
ments, vol 5, Pacific symposium on biocom- and methods. Ann N Y Acad Sci 1115(1):1–22
puting. World Scientific, Singapore, pp 39. Siegenthaler C, Gunawan R (2014) Assessment
418–429 of network inference methods: how to cope
27. Eisen MB, Spellman PT, Brown PO, Botstein with an underdetermined problem. PLoS One
D (1998) Cluster analysis and display of 9(3):e90481
 Biological Network Inference from Microarray Data, Current Solutions, and Assessments 167

40. Gama-Castro S, Jiménez-Jacinto V, Peralta-Gil 46. Jupiter D, Chen H, VanBuren V (2009) Star-
M, Santos-Zavaleta A, Peñaloza-Spinola MI, net 2: a web-based tool for accelerating discov-
Contreras-Moreira B, Segura-Salazar J, ery of gene regulatory networks using
Muñiz-Rascado L, Martnez-Flores I, Salgado microarray co-expression data. BMC Bioinfor-
H et al (2008) RegulonDB (version 6.0): gene matics 10(1):332
regulation model of escherichia coli k-12 47. Tripathi S, Dehmer M, Emmert-Streib F
beyond transcription, active (experimental) (2014) Netbiov: an r package for visualizing
annotated promoters and textpresso naviga- large network data in biology and medicine.
tion. Nucleic Acids Res 36(suppl 1): Bioinformatics 30(19):2834–2836
D120–D124 48. Bozdag S, Li A, Wuchty S, Fine HA (2010)
41. Subramanian A, Tamayo P, Mootha VK, Fastmedusa: a parallelized tool to infer gene
Mukherjee S, Ebert BL, Gillette MA, Paulovich regulatory networks. Bioinformatics 26
A, Pomeroy SL, Golub TR, Lander ES et al (14):1792–1793
(2005) Gene set enrichment analysis: a 49. Smith VA, Yu J, Smulders TV, Hartemink AJ,
knowledge-based approach for interpreting Jarvis ED (2006) Computational inference of
genome-wide expression profiles. Proc Natl neural information flow networks. PLoS Com-
Acad Sci U S A 102(43):15545–15550 put Biol 2(11):e161, pp. 1436–1449
42. Kharumnuid G, Roy S (2015) Tools for in- 50. Wang M, Verdier J, Benedito VA, Tang Y,
silico reconstruction and visualization of gene Murray JD, Ge Y, Becker JD, Carvalho H,
regulatory networks (GRN). In 2nd IEEE Rogers C, Udvardi M et al (2013) Legumegrn:
international conference on advance comput- a gene regulatory network prediction server for
ing and communication engineering functional and comparative studies. PLoS One
(ICACCE’ 2015) 8(7):e67434
43. Schaffter T, Marbach D, Floreano D (2011) 51. Faisal FE, Meng L, Crawford J, Milenković T
Genenetweaver: in silico benchmark genera- (2015) The post-genomic era of biological net-
tion and performance profiling of network work alignment. EURASIP J Bioinforma Syst
inference methods. Bioinformatics 27 Biol 2015:3
(16):2263–2270
52. Ciriello G, Mina M, Guzzi PH, Cannataro M,
44. Smoot ME, Ono K, Ruscheinski J, Wang P-L, Guerra C (2012) AlignNemo: a local network
Ideker T (2011) Cytoscape 2.8: new features alignment method to integrate homology and
for data integration and network visualization. topology. PLoS One 7(6):e38107. doi:10.
Bioinformatics 27(3):431–432 1371/journal.pone.0038107
45. Baker C, Carpendale MT, Prusinkiewicz P, Sur- 53. Guzzi PH, Milano M, Roy S (2015) Towards
ette MG (2002) Genevis: visualization tools for the assessment of GRN algorithms based on
genetic regulatory network dynamics. In Pro- (disease) ontology. In: Proceedings of the
ceedings of the conference on Visualization’02. ACM conf on bioinformatics, computational
IEEE Computer Society, 2002, pp 243–250 biology and health informatics (BCB’15)
Methods in Molecular Biology (2016) 1375: 169–179
DOI 10.1007/7651_2015_248
© Springer Science+Business Media New York 2015
Published online: 24 May 2015

A Protocol to Collect Specific Mouse Skeletal Muscles

for Metabolomics Studies
Zhuohui Gan, Zhenxing Fu, Jennifer C. Stowe, Frank L. Powell,
and Andrew D. McCulloch

Abstract
Due to the highly sensitive nature of metabolic states, the quality of metabolomics data depends on the
suitability of the experimental procedure. Metabolism could be affected by factors such as the method of
euthanasia of the animals and the sample collection procedures. The effects of these factors on metabolites are
tissue-specific. Thus, it is important to select proper methods to sacrifice the animal and appropriate
procedures for collecting samples specific to the tissue of interest. Here, we present our protocol to collect
specific mouse skeletal muscles with different fiber types for metabolomics studies. We also provide a protocol
to measure lactate levels in tissue samples as a way to estimate the metabolic state in collected samples.

Keywords: Skeletal muscle, Dissection, Euthanasia, Metabolomics, Lactate, Amplex Red

1 Introduction

Skeletal muscle is the most abundant tissue in humans and verte-

brate animals, which suggests the possibility to carry across-species
metabolic studies. Additionally, skeletal muscle is also a relatively
feasible tissue to collect for metabolomics studies, especially for
healthy humans. Several types of skeletal muscles can be identified
by their composition (1). These skeletal muscles could have differ-
ent metabolisms. For example, soleus muscle is identified as an
oxidative muscle, while plantaris muscle is identified as a glycolytic
muscle, although the details can vary for a given muscle between
species and with training. Therefore, the dissection of specific skel-
etal muscles is necessary for some metabolic studies. The request of
specific skeletal muscle for metabolomics raises two issues. Firstly,
the collection of muscle requires the animal to be anesthetized or
euthanized in advance, which may impact the metabolic levels in
animals. Another issue involves muscle sample collection once the
animals are ready to use. Since metabolic status in muscle tissue can
be changed with the active enzymes very rapidly, it is important to
instantaneously stop the inherent enzymatic activity so that the
metabolic state can be reserved for further analysis (2, 3).

169
170 Zhuohui Gan et al.

Several methods are widely used for mouse euthanasia,

including: anesthesia, decapitation, cervical dislocation, and elec-
tric shock. Studies suggest that these methods themselves could
affect the metabolic state in tissue (4–6), for example, anes-
thetics, such as ketamine and isoflurane, are known to suppress
metabolic functions (7–13). Thus, the selection of a method to
sacrifice the animal is the first step. For this purpose, several
methods of euthanasia including: decapitation, cervical disloca-
tion, and pentobarbital sodium, were tested. We selected lactate
as an easy indicator of metabolic state. Lactate is a metabolite
sensitive to the metabolic state with a relatively stable tissue
turnover rate around 0.2 mM/min. The lactate levels in both
the dissected glycolytic muscles and oxidative muscles were
measured with an Amplex Red-based assay. Amplex Red is a
highly sensitive and selective probe of hydrogen peroxide
(H2O2). This assay converts lactate into H2O2 in the presence
of lactate oxidase and produces a resorufin signal by the reaction
between Amplex Red and the resulting H2O2 in the presence of
horseradish peroxidase.
lactate oxidase
O2 þ L  lactate 
! pyruvate þ H2 O2
horseradish peroxidase
H2 O2 þ Amplex Red 
! resorufin þ H2 O

The experimental results indicate that both decapitation and pen-

tobarbital sodium injection result in a higher lactate level in both
glycolytic plantaris muscles and oxidative soleus muscles com-
pared with cervical dislocation, as shown in Fig. 1a. The higher
lactate level usually suggests hypoxia in the tissue. This may be
caused by the depth of anesthesia with pentobarbital sodium,
which depresses ventilation, or by the involuntary contractions
of limbs after decapitation. In order to minimize the effect of the
euthanasia method on the metabolic state in mouse skeletal mus-
cle, cervical dislocation was selected as the preferred method for
mouse muscle collection. It is noteworthy that the optimized
method of euthanasia is tissue specific. For example, pentobarbital
sodium is regarded as a good anesthetic to collect tissues for the
purpose of cardiac physiology and mitochondrial functions (5,
14). However, some studies report that pentobarbital sodium
impacts cerebral metabolism (15). Hence, pentobarbital sodium
is not an optimized euthanasia method for brain-relevant meta-
bolic studies. Similarly, cervical dislocation is reported to change
lung platelet serotonin (16), but compared with the effects of
other euthanasia methods on mouse skeletal muscle, it is the
optimal choice. Therefore, it is important to take the tissue type
into consideration when selecting the euthanasia method for met-
abolic studies.
 A Protocol to Collect Specific Mouse Skeletal Muscles for Metabolomics Studies 171

a
2.00
plantaris

Lactate in tissue homogenate (mM)

soleus *
1.60 *

# #
1.20

0.80

0.40

0.00
Cervical Dislocation Decaptation Pentobarbital

b
2.00
plantaris
Lactate in tissue homogenate (mM)

soleus
1.60

1.20

*
0.80

#
0.40

0.00
Flash Freezing Tongs Freezing In-Situ Dissection

Fig. 1 (a) The comparison of euthanasia methods on lactate level in mouse skeletal muscles. Two types of
mouse skeletal muscles, plantaris and soleus, were dissected from the flash-frozen legs which were cut from
mice sacrificed by the assigned methods. The sample size for each data point is 3–4. The values are
mean  s.d. *: different from the lactate level in plantaris muscles from mice sacrificed by cervical
dislocation, p < 0.05. #: different from the lactate level in soleus muscles from mice sacrificed by cervical
dislocation, p < 0.05. (b) The comparison of collection methods on lactate level in mouse skeletal muscles.
Two types of mouse skeletal muscles, plantaris and soleus, were dissected from legs processed with the
assigned methods. The mice were sacrificed by cervical dislocation. The sample size for each data point is
3–4. The values are mean  s.d. *: different from the lactate level in plantaris muscles from flash-frozen legs,
p < 0.05. #: different from the lactate level in soleus muscles from flash-frozen legs, p < 0.05

Several methods have been developed to quickly inhibit

enzymatic activity, such as in situ clamping by cryogenic block
tongs or flash freezing in liquid nitrogen (2). Needle biopsy fol-
lowed by flash freezing is widely used to collect human muscle
samples for metabolic studies (17–19). With this approach, the
172 Zhuohui Gan et al.

whole collection process can be completed in seconds and hence

preserves the metabolic state in human muscles. However, the
volume of mouse skeletal muscle is quite small. This makes it
difficult to collect a specific mouse muscle by biopsy, especially
considering the relatively large amount of tissue required for meta-
bolic analyses. Thus, the dissection of the muscle becomes neces-
sary in order to collect enough specific mouse skeletal muscle for
metabolic studies. To dissect the muscle first or to freeze the muscle
first becomes a question. Methods to dissect fresh muscles exist
(20), but will require modification if the tissue is first frozen.
Aluminum-block tongs at liquid nitrogen temperatures are widely
used to freeze tissues (liver, lung) in situ for metabolic studies (21).
Whether the aluminum- block tongs are sufficient to freeze skeletal
muscles, which are located near the bone, is also a question.
To answer these questions, we measured the lactate levels of mus-
cles dissected from fresh mouse legs and then flash-frozen in liquid
nitrogen, muscles dissected from legs frozen by aluminum block
tongs in situ, and muscles dissected from mouse legs which had
been rapidly removed from the euthanized animal and flash-frozen
in liquid nitrogen. The dissection experiments show that it is feasi-
ble to dissect specific muscles from frozen-and-then-partially
thawed mouse legs. Figure 1b shows no statistical difference in
lactate levels between muscles dissected from aluminum-block-
tong frozen legs and muscles dissected from liquid nitrogen
flash-frozen legs, but both groups of muscles have a lower lactate
level compared with muscles dissected from fresh legs. Based on
this result, either in situ block tong freezing or whole-leg liquid
nitrogen flash freezing is appropriate. To simplify the experiment,
we collect the muscles by rapidly removing the mouse legs just after
euthanasia and flash freezing the legs completely in liquid nitrogen.
We then dissect the desired muscles from the partially thawed
legs on ice. The whole leg removal takes a couple of seconds.
The whole leg can be frozen completely in liquid nitrogen in twenty
seconds. The advantage of this method is that the low-temperature
in the barely thawed leg gives us a dissection time window.
In the following protocol, the process of preparing the mice
and the muscle collection steps are outlined. The assay that we used
to measure lactate level in collected muscles is also described.

2 Materials

2.1 Materials for 1. A dewar containing liquid nitrogen.

Muscle Dissection 2. A dissection tray containing ice.
3. Tissue forceps.
4. Surgical scissors.
 A Protocol to Collect Specific Mouse Skeletal Muscles for Metabolomics Studies 173

5. Iris scissor.
6. Hemostat (Halsted mosquito forceps).
7. Aluminum foil.
8. Indelible marker pen.
9. Disposable latex or nitrile gloves, gown or clean lab coat, and
eye protection.
10. Disposal container.
11. Cryotubes.

2.2 Materials 1. 1 phosphate buffered saline (PBS), PH 7.4 (no Ca2+, no Mg2+).
for Amplex Red-Based 2. L-lactate standard (Sigma, part #71718).
Lactate Assay Mix 10 μL of 100 mM lactate sample with 990 μL deionized
water to get a 1 mM lactate solution. Prepare lactate standard
samples using deionized water with lactate concentrations at 0,
20, 50, 100, 200, and 300 μM, respectively. Store at 20  C.

0 μM 20 μM 50 μM 100 μM 200 μM 300 μM

1 mM Lactate (μL) 0 20 50 100 200 300
H2O (μL) 1,000 980 950 900 800 700

3. Amplex Red stock (Sigma, part #90101).

Add 5 mg Amplex Red powder into 1.943 mL DiMethyl-
SulphOxide (DMSO) to get 10 mM Amplex Red stock.
Aliquot and store the stock at 20  C.
4. Horseradish peroxidase (HRP) stock (Sigma, part#P2088).
Add 2 mg HRP powder into 1 mL 1 PBS to get
500 U/mL HRP stock, and store on ice. Prepare fresh, and
do not store for further use.
5. Lactate oxidase (LOX) stock (Sigma, part#L0638).
Add 1 mL 1 PBS to 50 U lactate oxidase powder, aliquot as
100 μL each and store at 20  C.
6. 96-well white flat-bottom plate.
7. Pipettes and tips.
8. 0.5 or 1.5 mL Tubes.
9. Tube storage rack.
10. Ice bucket containing ice.
11. Marker pen.
12. A slotted metal spoon or other apparatus to remove samples
from liquid nitrogen.
13. Tissue homogenizer.
We use Kontes Glass DUALL® 20 tissue grinder to homoge-
nize mouse muscles. There are some alternative tools. A
174 Zhuohui Gan et al.

comparison is available at http://opsdiagnostics.com/notes/

mouseldh.htm.
14. Refrigerated Micro centrifuge (temperature setting to 4  C).
15. Disposal container.
16. Gloves, gown and eye protection.
CAUTION: Liquid nitrogen is very cold and can cause burns and
frost bite. Wear protective gloves when handling liquid nitrogen
frozen samples and proceed with extreme caution while working
with liquid nitrogen.

3 Methods

1. Obtain mice with the same age, sex, and strain.

2. House mice in the animal room for at least 1 day.
3. Withhold food from experimental mice for at least 4 h before
the sample collection to minimize metabolic variance caused by
digestion.
4. Prepare several pieces of aluminum foil 200  200 in size. Label
the foil on the center of the shinier side with sample name and
other identifying information.
5. Label 1.5–2 mL cryotubes. Keep on ice.
6. Cover the dissection tray with aluminum foil for dissection
surgery.
7. Move the mouse to the surgery room and sacrifice the mouse by
rapid cervical dislocation, making sure that the mouse being
euthanized is separated from any others to be studied, so they
are not stressed. The operator must be well trained (see Note 1).
8. Immediately cut off both of the mouse’s hind legs using large
sharp surgical scissors.
9. Wrap the leg in the aluminum foil, tissue touching the matte
(less shiny) side of the foil. Ensure limb is completely sealed in
foil and rapidly drop it into the dewar containing liquid nitro-
gen (see Note 2).
10. If the samples will not be processed on the same day, store the
frozen samples at 80  C or in liquid nitrogen. Otherwise, the
legs will be completely frozen in 20 s and the dissection process
can begin.
11. Remove a frozen leg from the liquid nitrogen using the slotted
metal spoon or other apparatus, open the aluminum foil and
place the leg on the aluminum foil of the dissection tray filled
with ice.
12. Allow the leg to thaw for approximately 2 min.
A Protocol to Collect Specific Mouse Skeletal Muscles for Metabolomics Studies 175

13. Make a small incision along the skin of the lower leg around the
“ankle,” and then tear off the entire skin covering the leg.
14. Remove the membrane surrounding the leg muscles
using forceps or cut the membrane using a dissection scissors
(see Note 3).
15. Gently separate the rear tendons from the bone, clamp the rear
tendons very distally underneath the heel using a hemostat, and
cut with Iris scissors.
16. Pull and separate the posterior muscles from the bone with the
hemostat. If it is still hard to move the posterior muscles,
wait for a few more minutes to thaw the leg a little more
(see Note 4).
17. Dissect soleus from the posteriors muscles by pulling the soleus
from its proximal tendon until the distal tendon and cut out the
soleus using a dissection scissors. Put the dissected soleus into a
cyrotube and freeze in liquid nitrogen or homogenize in assay
buffer immediately (see Note 5). The location of the soleus
muscle in a frozen-thawed leg is indicated in Fig. 2a.

Fig. 2 (a) The location of soleus muscle in a flash-frozen and then thawed mouse
leg. The soleus muscle is identifiable and dissectible in frozen-thawed mouse leg.
The integrity of soleus muscle keeps well. (b) The location of plantaris muscle in a
flash-frozen and then thawed mouse leg after the soleus muscle is removed. The
plantaris muscle is identifiable and dissectible with a good integrity
176 Zhuohui Gan et al.

18. Separate the two distal tendons and find the plantaris
muscle, cut its tendon. Pull the plantaris muscle by holding
the distal end of its tendon, and cut the proximal end (see Note
6). Put the dissected plantaris into a cytotube and freeze
in liquid nitrogen or homogenize in assay buffer immediately
(see Note 7). The location of plantaris muscle in a frozen-
thawed leg is shown in Fig. 2b.
19. Store the dissected muscles at 80  C if it won’t be used
immediately. Otherwise, proceed to homogenization.

3.1 Supplement: This assay was developed to detect lactate levels in dissected muscle
Amplex Red-Based samples. This is not a part of the muscle collection protocol, but an
Lactate Assay option for metabolic measurement. This protocol is sufficient to
run 20 lactate measurement reactions (see Note 8).
1. Label tubes for Amplex Red solution, HRP solution, LOX
solution and at least two tubes for each sample.
2. Take Amplex Red stock, HRP stock, lactate oxidase stock and
lactate standards out of the freezer to thaw.
3. Mix 100 μL Amplex Red stock and 900 μL 1 PBS to
get 1 mM Amplex Red solution, store on ice (this is AR-PBS)
(see Note 9).
4. Mix 20 μL HRP stock and 80 μL 1 PBS to get 100 U/mL
HRP solution, store on ice (this is HRP-PBS).
5. Mix 10 μL lactate oxidase stock and 90 μL 1 PBS to get
5 U/mL lactate oxidase solution (this is LOX-PBS).
6. Homogenize the muscle sample with the ratio of 1 mg per
30 μL 1 PBS, in ice bath.
7. Centrifuge the tube containing homogenate at 13,000  g for
10 min, 4  C to remove insoluble material.
8. Move the supernatant and transfer into clean tubes.
9. Dilute the supernatant with 1 PBS 1:10 (see Note 10), store
on ice.
10. Add 45 μL AR-PBS to each cell, the total number of cells is the
sum of the cells for sample measurements and the cells for the
lactate standard curve which requires six cells.
11. Add 5 μL HRP-PBS to each cell.
12. Measure the colorimetric absorbance at 571 nm as the back-
ground signal (abs1) at room temperature.
13. Add 45 μL diluted supernatant or lactate standards (0, 20,
50, 100, 200, and 300 μM) to each cell, mix by slightly shaking
(see Note 11).
14. Measure the colorimetric absorbance at 571 nm as the
measurement of H2O2 in supernatant samples (abs2).
 A Protocol to Collect Specific Mouse Skeletal Muscles for Metabolomics Studies 177

15. Add 5 μL LOX-PBS to each cell, mix by slightly shaking.

16. Allow reaction to occur for 15 min at room temperature, and
then measure the colorimetric absorbance at 571 nm as the
measurement of lactate in supernatant samples and standards
(abs3).
17. Calculate each sample’s lactate absorbance value, abs ¼ abs3-
abs2-abs1.
18. Determine the conversion equation between the concentra-
tions of lactate standard samples and its corresponding abs
values (see Note 12).
19. Convert abs of muscle samples to its lactate concentration by
the conversion equation.

4 Notes

1. An improper cervical dislocation may affect metabolic state.

The faster and cleaner the cervical dislocation, the better.
2. A thinner aluminum foil allows faster freezing of legs in the
liquid nitrogen. Other containers such as cryotubes might not
work as well as aluminum foil.
3. The removal of the outer membrane facilitates the muscle
dissection. Otherwise, the muscles would be bound together.
4. It is important to dissect the muscles at the appropriate tem-
perature. If the leg is thawed too much, the muscles will be soft
and might undergo metabolic degradation. When the leg is less
thawed, it is hard to dissect the intact soleus and plantaris
muscles since the muscles are still frozen together. The best
time point is when you can barely move the posterior muscles
with the hemostat.
5. The soleus is quite visible since its color is deep red as an
oxidative muscle. The mean quantity of an intact mouse soleus
muscle is ~6–9 mg.
6. The plantaris is bigger than the soleus, with a moderate red
color. Sometimes, you can see the tendon sticking to the plan-
taris. The mean quantity of an intact mouse plantaris muscle is
~11–16 mg.
7. A well-trained operator could finish the dissection of the soleus
and plantaris muscles in a couple of minutes. Faster is better.
8. This protocol has been well tested with the total 100 μL mix-
ture in each cell. It might be okay to adjust the total volume to a
smaller value such as 50 μL or less for 384-well plate. When you
adjust the total volume, please adjust the required volumes of
all relevant solutions proportionally. The final concentration
178 Zhuohui Gan et al.

of Amplex Red in each cell shall be 450 μM. The final

concentration of HRP in the cells shall be 5 U/mL. The
final concentration of LOX in the cell shall be 0.25 U/mL.
The concentrations of enzymes are fixed, while the concentra-
tion of Amplex Red is adjustable as long as the resulting signal is
in the linear range of the standard curve.
9. In this protocol, we use 100 μL Amplex Red stock for 20
measurements. To save Amplex Red, you can calculate and
adjust the required amount of Amplex Red stock based on
your desired number of measurements, such as 4 μL Amplex
Red stock/measurement.
10. For mouse skeletal muscle, we recommend 1:5 or 1:10 as the
dilution factor. For different tissues, the lactate level could be
different. You need to select a proper dilution ratio to ensure the
absorbance signal is in the linear range of the standard curve.
11. The maximum concentration of lactate standard can’t be
beyond 450 μM since the final concentration of Amplex Red
in the cell is 450 μM. You need to adjust the supernatant to
locate the absorbance value in the 0–450 μM. If the concentra-
tion of lactate in the measured supernatant is beyond 450 μM,
the extra H2O2 produced from lactate will react with Amplex
Red’s product, resorufin, and result in a decrease in the absor-
bance value. This is the reason why the original supernatant is
diluted.
12. Take abs ¼ a  [lactate] + b as the conversion equation,
deduce a, b using least square method or use the regression
function which is available in Excel, Matlab, R, and a lot of
other software.

Acknowledgements

Grant NIH/NHLBI 1P01HL098053 supported this manuscript

preparation.

References
1. Mizunoya W, Wakamatsu J, Tatsumi R et al 4. Evans CA, Kerkut GA (1981) Effect of nem-
(2008) Protocol for high-resolution separation butal anesthesia, electric shock, and shock
of rodent myosin heavy chain isoforms in a avoidance conditioning on acetylcholinesterase
mini-gel electrophoresis system. Anal Biochem activity and protein content in various regions
377(1):111–113 of the rat brain. Neurosci Behav Physiol 11
2. Fiehn O (2002) Metabolomics – the link (6):614–620
between genotypes and phenotypes. Plant 5. Marquez-Julio A, French IW (1967) The effect
Mol Biol 48(1–2):155–171 of ether, pentobarbital, and decapitation on
3. Noack S, Wiechert W (2014) Quantitative various metabolites of rat skeletal muscle. Can
metabolomics: a phantom? Trends Biotechnol J Biochem 45(9):1323–1327
32(5):238–244
 A Protocol to Collect Specific Mouse Skeletal Muscles for Metabolomics Studies 179

6. Pence HH, Pence S, Kurtul N et al (2003) The mitochondrial isolated from the hearts of
alterations in adenosine nucleotides and lactic anesthetized rats with high-dose pentobarbital
acid levels in striated muscles following death sodium. Jpn J Physiol 47(1):87–92
with cervical dislocation or electric shock. Soud 15. Du F, Zhang Y, Iltis I et al (2009) In vivo
Lek 48(1):8–11 proton MRS to quantify anesthetic effects of
7. Rezin GT, Goncalves CL, Daufenbach JF et al pentobarbital on cerebral metabolism and
(2009) Acute administration of ketamine brain activity in rat. Magn Reson Med 62
reverses the inhibition of mitochondrial respi- (6):1385–1393
ratory chain induced by chronic mild stress. 16. Yamamoto Y, Hasegawa H, Ikeda K et al
Brain Res Bull 79(6):418–421 (1988) Cervical dislocation of mice
8. Chang Y, Chen TL, Sheu JR et al (2005) Sup- induces rapid accumulation of platelet
pressive effects of ketamine on macrophage func- serotonin in the lung. Agents Actions 25
tions. Toxicol Appl Pharmacol 204(1):27–35 (1–2):48–56
9. de Oliveira L, Fraga DB, De Luca RD et al 17. Fischer JC, Ruitenbeek W, Stadhouders AM
(2011) Behavioral changes and mitochondrial et al (1985) Investigation of mitochondrial
dysfunction in a rat model of schizophrenia metabolism in small human skeletal muscle
induced by ketamine. Metab Brain Dis 26 biopsy specimens. Improvement of prepara-
(1):69–77 tion procedure. Clin Chim Acta 145
10. Pravdic D, Hirata N, Barber L et al (2012) (1):89–99
Complex I and ATP synthase mediate mem- 18. Boros-Hatfaludy S, Fekete G, Apor P (1986)
brane depolarization and matrix acidification Metabolic enzyme activity patterns in muscle
by isoflurane in mitochondria. Eur J Pharmacol biopsy samples in different athletes. Eur J Appl
690(1–3):149–157 Physiol Occup Physiol 55(3):334–338
11. Zhang Y, Xu Z, Wang H et al (2012) Anes- 19. Bergstrom J (1975) Percutaneous needle
thetics isoflurane and desflurane differently biopsy of skeletal muscle in physiological and
affect mitochondrial function, learning, and clinical research. Scand J Clin Lab Invest 35
memory. Ann Neurol 71(5):687–698 (7):609–616
12. Kohro S, Hogan QH, Nakae Y et al (2001) Anes- 20. Antal C, Teletin M, Wendling O et al (2007)
thetic effects on mitochondrial ATP-sensitive K Tissue collection for systematic phenotyping in
channel. Anesthesiology 95(6):1435–1440 the mouse. Curr Protoc Mol Biol Chapter 29:
13. Braun S, Gaza N, Werdehausen R et al (2010) Unit 29A 24
Ketamine induces apoptosis via the mitochon- 21. Winder WW, Fuller EO, Conlee RK (1983)
drial pathway in human lymphocytes and neu- Adrenal hormones and liver cAMP in exercis-
ronal cells. Br J Anaesth 105(3):347–354 ing rats – different modes of anesthesia. J Appl
14. Takaki M, Nakahara H, Kawatani Y et al (1997) Physiol Respir Environ Exerc Physiol 55
No suppression of respiratory function of (5):1634–1636
Methods in Molecular Biology (2016) 1375: 181–194
DOI 10.1007/7651_2015_250
© Springer Science+Business Media New York 2015
Published online: 14 May 2015

Functional Analysis of microRNA in Multiple Myeloma

Maria Teresa Di Martino, Nicola Amodio, Pierfrancesco Tassone,
and Pierosandro Tagliaferri

Abstract
MicroRNAs (miRNAs) are short non coding RNAs that regulate the gene expression and play a relevant
role in physiopathological mechanisms such as development, proliferation, death, and differentiation of
normal and cancer cells. Recently, abnormal expression of miRNAs has been reported in most of solid or
hematopoietic malignancies, including multiple myeloma (MM), where miRNAs have been found deeply
dysregulated and act as oncogenes or tumor suppressors. Presently, the most recognized approach for
definition of miRNA portraits is based on microarray profiling analysis. We here describe a workflow based
on the identification of dysregulated miRNAs in plasma cells from MM patients based on Affymetrix
technology. We describe how it is possible to search miRNA putative targets performing whole gene
expression profile on MM cell lines transfected with miRNA mimics or inhibitors followed by luciferase
reporter assay to analyze the specific targeting of the 30 untranslated region (UTR) sequence of a mRNA by
selected miRNAs. These technological approaches are suitable strategies for the identification of relevant
druggable targets in MM.

Keywords: microRNA, miRNA, Microarray profiling, miRNA replacement, miRNA inhibition,

Transfection

1 Introduction

Multiple myeloma (MM) is an incurable hematologic disorder in

which malignant plasma cells accumulate in the bone marrow. MM
is characterized by a range of genetic aberrations leading to clinical
heterogeneity in the patient population. The specific mechanisms
leading to the development of plasma cell disorders and progression
to symptomatic MM have not been fully elucidated. Treatment
options for patients with MM have significantly improved within
the past decade, resulting in enhanced response rates and survival;
treatment options outside of clinical trials are currently limited to
combinations of bortezomib, thalidomide and its analogs (IMiDs),
chemotherapy and steroids (dexamethasone, prednisone). High-
dose melphalan with stem cell transplantation is considered in
eligible patients. Although these agents are effective in the treat-
ment of a majority of MM patients, the clinical outcome is worst for
patients which can be defined as high-risk. Specifically, the

181
182 Maria Teresa Di Martino et al.

treatment results are clearly unsatisfactory for MM patients bearing

the t(4;14) translocation which represents therefore a substantial
challenge (1).
microRNAs (miRNAs) are small evolutionary conserved non-
coding RNAs that bind to the 30 -untranslated region (UTR) of
target mRNAs, resulting in translation repression or mRNA degra-
dation, and play important roles in cellular processes such as prolif-
eration, development, differentiation, and apoptosis. The
dysregulation of these tiny and important molecules in different
types of cancers including MM are widely described and make it
promising target for new therapeutic approach (2–9). Moreover,
recent promising findings are supporting the idea of miRNA-based
personalized therapeutic strategies in MM (2, 4, 10–19). Advances
in the oligonucleotide chemistry and formulation of miRNA
mimics or antisense oligonucleotides (miRNA inhibitors) have
added a new value in the enhancement of their therapeutic and
commercial potential (7, 20).
We recently reported that new molecular-based targeted
approaches, including the use of specific mimics or inhibitors,
have provided an example of personalized treatments of MM.
For instance, miR-29b replacement strategies are effective against
MM in preclinical models (10, 12, 21). Moreover, we described
that p53-mutated MM may have benefit by treatment with miR-
34a mimics (15) if delivered with appropriate nanovectors (13,
19). We also demonstrated that the use of miR-221 inhibitors may
provide benefit for patients with t(4;14) MM representing a
potential targeted and personalized therapeutic approach. Based
on our preclinical data, it is conceivable that these therapies may
benefit a subset of patients with the t(4;14) translocation, partic-
ularly those with increased expression of miR-221/222 (4, 14,
22). miR-21 (16) and miR-125a-5p (17) inhibitors were also
found to be promising anti-MM therapeutic agents when deliv-
ered in vitro and in vivo in MM models.
During last decade, microarray technologies have emerged as a
flexible method for analyzing large numbers of nucleic acid frag-
ments in parallel. Based on base-pairing rules microarrays provide a
medium for matching known and unknown DNA targets, as well
automating the process of identifying the unknowns. This tech-
nique accommodates parallel analysis for gene expression, as well as,
allows gene discovery and therefore provides simultaneously infor-
mation on several thousand of genes. Moreover, noncoding (nc)
RNA microarray platforms have been recently developed, revealing
differential expression patterns of microRNA, as well as of long
ncRNA.
Aberrant miRNA expression in MM as shown by microarray
analysis has been reported (23), indicating that miRNA dysregu-
lation characterizes the progression of the disease, modulates
important pathways involved in MM cell survival and reflects the
 Functional Analysis of microRNA in Multiple Myeloma 183

Fig. 1 The figure shows the workflow of procedures for miRNA functional analysis. First the detection of
dysregulated miRNAs by differential miRNA profiling. Then the replacement of downregulated miRNAs or
specific inhibition of upregulated miRNAs, and the validation of miRNA-specific targeting by (1) the luciferase
reporter assay, (2) the whole gene expression analysis by microarray, (3) the study of specific targets at mRNA
level by real-time PCR and (4) at protein level by Western Blot analysis

different MM genetic subtypes (23). Global miRNA expression

profiling data in MM further underlined the role of miRNA
dysregulation in cancer and their potential use in therapy,
and gives a solid rationale for the development of effective tools
for selective delivery of miRNAs and anti-miRNAs to myeloma
cells (9).
At this aim the Affymetrix platform represents a comprehensive
tool for microarray technology. The experimental workflow of
miRNA functional analysis for our approach is represented in
Fig. 1, that schematically shows our strategy for the detection of
dysregulated miRNAs in MM samples as compared to normal
plasma cells. We then briefly describe how to replace in MM cells
downregulated miRNAs or to use specific inhibitors for upregu-
lated miRNAs, which potentially act as tumor suppressors or as
oncomiRs, and how we validate a miRNA-specific targeting by
luciferase reporter assay. Our flow includes the evaluation of
miRNA replacement/inhibition on whole gene expression by
microarray analysis, the study of specific targets at mRNA level by
real-time PCR and at protein level by Western Blot analysis.
184 Maria Teresa Di Martino et al.

2 Materials

2.1 Screening by 1. CD138 MicroBeads human (Miltenyi Biotec, catalog no.

miRNA Profiling 130-051301).
2. RNeasy® Mini kit (Qiagen, catalog no. 74104).
3. TRIzol reagent (US Patent No. 5,346,994). Ready-to-use
reagent for the homogenization or cell lysis that maintains
the integrity of the RNA.
4. Affymetrix® FlashTag™ Biotin HSR RNA Labeling Kit (cata-
log no. 901910, sufficient for 10 reactions).
5. Affymetrix® miRNA 2.0, 3.0, or 4.0 Array (catalog no.
901753, 902017, or 902411, respectively, contains 2 arrays),
designed to interrogate a different miRBase Release up to
all mature miRNA sequences in miRBase Release 20, as for
4.0 Array.
Equipment
1. NanoDrop 1000 Spectrophotometer (www.nanodrop.com).
2. Microarray Affymetrix® workstation (hybridization oven, flu-
idic station, Scanner; www.affymetrix.com).

2.2 Synthetic miRNA 1. miRNA mimics (miRVana™ catalog no. 4464070, 5 nmol lyo-
Overexpression philized pellet) are small, chemically modified double-stranded
or Inhibition RNAs that mimic endogenous miRNAs and enable miRNA
functional analysis by upregulation of miRNA activity.
2.2.1 Transient
Transfection 2. miRNA inhibitors (miRVana™ catalog no. 4464066, 5 nmol
lyophilized pellet) are small, chemically modified single-
stranded RNA molecules designed to specifically bind to and
inhibit endogenous miRNA molecules and enable miRNA
functional analysis by downregulation of miRNA activity.
3. Negative control (Life Technologies).
4. Neon® Transfection System 100 μL kit (Invitrogen™), catalog
no. MPK10025. The Neon® Transfection System 100 μL Kit
includes 1 mL resuspension buffer R, 1 mL resuspension buffer
T, 75 mL E electrolytic buffer, 25 reaction delivery tips, five
electroporation tubes.
5. Exponentially growing MM cell lines.
6. Six-well plates.
7. RPMI-1640 medium.
8. Fetal bovine serum.

Equipment
1. Neon® Transfection System (Catalog Number MPK5000).
2. Neon® Pipette (Catalog Number MPP100).
 Functional Analysis of microRNA in Multiple Myeloma 185

3. Neon® Pipette Station (Catalog Number MPS100).

4. Equipment for cell culture: CO2 Incubator, laminar flow cabin.

2.2.2 miRNA Quantitative 1. tRNA isolation: TRIzol® Reagent (Life Technologies), chloro-
Analysis form, isopropanol, 75 % ethanol, and nuclease-free water.
2. TaqMan® MicroRNA Reverse Transcription Kit (Applied
Biosystems).
3. TaqMan® MicroRNA Assays (Applied Biosystems).
4. NanoDrop 1000 Spectrophotometer.
5. Optical 96-well reaction plates with barcode (Applied
Biosystems).
6. Viia7 Dx real-time PCR system (Applied Biosystems).

2.3 Analysis 1. RNeasy® Mini kit (Qiagen, catalog no. 74104).

of Target Modulation 2. TRIZOL reagent (US Patent No. 5,346,994). Ready-to-use
2.3.1 Gene Profiling reagent for the homogenization or cell lysis that maintains the
integrity of the RNA.
3. Affymetrix® GeneChip WT PLUS Reagent Kit (catalog no.
902309, sufficient for 10 reactions).
4. Affymetrix® GeneChip Hybridization, Wash, and Stain Kit
(catalog no. 900720, sufficient for 30 reactions).
5. Affymetrix® GeneChip Hybridization Control Kit (catalog no.
900454, sufficient for 30 reactions) including Control Oligo
B2, 3 nM (catalog no. 900301).
6. Affymetrix® GeneChip Human Transcriptome Array 2.0 (cata-
log no. 902233, contains 2 arrays), designed to empower next-
generation expression profiling studies, this array provides the
ability to go beyond gene-level expression profiling by
providing the coverage and accuracy required to detect all
known transcript isoforms produced by a gene. This
high-resolution array design contains an unprecedented >6.0
million probes covering coding transcripts and noncoding tran-
scripts. 70 % of the probes on this array cover exons for coding
transcripts, and the remaining 30 % of probes on the array cover
exon–exon splice junctions and noncoding transcripts. The
unparalleled coverage of this array provides the deepest insight
into all coding and noncoding transcripts available.
Equipment
1. NanoDrop 1000 Spectrophotometer (www.nanodrop.com).
2. Microarray Affymetrix® workstation (hybridization oven, flu-
idic station, Scanner, www.affymetrix.com).
186 Maria Teresa Di Martino et al.

Additionally support.
Library Files
GeneChip® Human Transcriptome Array 2.0 Analysis
(zip, 303 MB).
GeneChip® Human Transcriptome Array 2.0 AGCC Library File
Installer (zip, 109 KB).

2.3.2 Luciferase 1. Plasmid constructs (30 UTR of the gene of interest is cloned in
Reporter Assay for miRNA- pEZX-MT01 vector, Genecopoeia).
Target Validation (a) 400 mL LB liquid medium: 10 g/L tryptone; 5 g/L yeast
extract, 10 g NaCl.
(b) Kanamycin for bacterial selection.
(c) LB-Kanamycin bacterial plates: LB liquid media plus 8 g/L
agar. Autoclave, cool down to about 50  C in a water bath
and add Kanamycin to have the final concentration of
25 μg/mL; mix well and distribute 15–20 mL of medium
per 10 cm plate.
(d) Maxi prep kit (PureLink® HiPure Plasmid Maxiprep Kit,
Invitrogen, Life Technologies) for isolation of high purity
plasmid from bacteria.
2. MM cell transfection and 30 UTR luciferase reporter assay.
(a) Exponentially growing MM cells.
(b) miRVANA miRNA mimics (Applied Biosystems).
(c) Six-well plates.
(d) RPMI-1640 medium.
(e) Fetal bovine serum (FBS).
(f) Dual-Glo Luciferase Assay kit (Promega).

Equipment
1. Neon electroporation system (Life technologies).
2. Plate reader for luminesce detection.

3 Methods

3.1 Screening by Plasma cells from human peripheral blood mononuclear cells
miRNA Profiling (PBMCs) bone marrow are isolated at >90 %, purity as determined
by flow cytometry, by the use of CD138 magnetic beads sorting
according to the manufacturer’s instructions (www.miltenyibiotec.
com); subsequently total RNA (tRNA) including small
RNA fractions is extracted by a modified Qiagen protocol
(www.qiagen.com). Briefly, 5  104 cells are lysed by 250 μL of
TRIzol® solution, then the aqueous phase is loaded on the Qiagen
column. After two washings by high speed centrifugation at r.t. the
 Functional Analysis of microRNA in Multiple Myeloma 187

tRNA is eluted from the column by loading 15 μL of RNase-free

water directly on the top of the filter in the column followed by a
further 5 min incubation at r.t. The RNA solution is collected and
immediately stored at 80  C.
miRNA expression profiling is carried out according to the
Affymetrix (www.affymetrix.com) recommended protocol. Briefly,
after quantification of tRNA by NanoDrop spectrophotometry,
300 ng is processed using the FlashTag labeling kit according to
the manufacturer’s instructions (www.affymetrix.com), which is
based on a tailing reaction followed by ligation of the biotinylated
signal molecule to the target RNA sample. The 45-min assay, from
the extracted RNA sample to the labeled target, does not involve
complex reactions such as amplification or purification steps that
can reduce the yield or can bias the results by introducing false
positives or false negatives. The labeled RNA is then hybridized to
Affymetrix GeneChip microRNA arrays which is scanned by the
GeneChip Scanner 3000 7G.
Expression values for miRNAs are extracted from CEL files
using Affymetrix miRNA QC tool software (RMA normalized
and log2-transformed), for miRNA 2.0 and 3.0 Array, or Affyme-
trix® Expression Console™ Software (version 1.3.1 and higher) for
4.0 Array, including array quality control. 4.0 Array provides the
opportunity of data filtering by the use of analysis options by the
following 4 filtering options:
1. Analysis of all organisms.
2. Mouse only.
3. Human only.
4. Rat only.
By selecting either the human, mouse, or rat only analysis
option, only the probe sets for the selected organism(s) are reported
in the analysis output. In the previously miRNA Array Release (i.e.,
2.0 and 3.0) analyzed by QC Tool, the analysis for single organism
was performed by different tools. After generating the signal sum-
marization, differential expression analysis can be performed and
results can be visualized by Affymetrix® Transcriptome Analysis
Console (TAC) Software. Both Expression Console Software and
TAC Software can be free downloaded from www.affymetrix.com.
The analyses are performed on log2 transformed data. First,
hierarchical agglomerative clustering of the samples is performed
by Pearson’s correlation coefficient and average linkage as distance
and linkage metrics, respectively, on those probes whose average
change in expression levels varied at least twofold from the mean
across the dataset; P value threshold for sample enrichment is
default set at 0.5. Supervised analyses are carried out using
DChip software (www.hsph.harvard.edu/cli/complab/dchip).
The threshold for significance is determined by fold change analysis
applied on the expression values of the miRNAs.
188 Maria Teresa Di Martino et al.

3.2 Synthetic miRNA The following procedure refers to the transfection of RPMI-8226
Overexpression MM cells.
or Inhibition The day before the transfection, MM cells are seeded at
5.0  105 cells/mL in RPMI-1640 containing 10 % heath inacti-
3.2.1 Transient vated FBS and 1 % penicillin-streptomycin.
Transfection MM cells are transfect by the Neon transfection system (Life
Technologies): briefly, 1.0  106 cells are used for each transfec-
tion point, washed in PBS and resuspended in 100 μL of buffer R.
Add 2 μL of 100 μM miRNA mimics or inhibitors or negative
control (NC). The mix is resuspend and electroporated by the use
of the Neon pipette and 100 μL tips at the following electropora-
tion conditions: 1,050 V, 30 ms, 1 pulse. Transfected cells are
seeded into a six-well plate containing 2 mL of pre-warmed growth
medium without antibiotics. The plate is incubated into a 37  C/
5 % CO2 incubator and cells collected 24 and 48 h after transfection
and analyzed for miRNA and target expression.

3.2.2 miRNA Quantitative 0.5 mL TRIzol® Reagent is added to 1  106 harvested cells; cell
Analysis sample are lysed by pipetting the cells up and down several times.
The homogenized sample is incubated for 5 min at room tempera-
tRNA Isolation
ture then 0.2 mL of chloroform are added. After shaking vigorously
by hand for 15 s, the tube is incubated for 2 min at room tempera-
ture. The sample is then centrifuged at 12,000  g for 15 min at
4  C. The aqueous phase is now transferred into a fresh Eppendorf
tube avoiding to draw any of the interphase or organic layer into the
pipette. After adding of 0.5 mL of 100 % isopropanol to the
aqueous phase, mix well the sample and incubate for 10 min at
room temperature. After centrifugation at 12,000  g for 10 min
at 4  C the supernatant is discarded from the tube, and the RNA
pellet washed with 1 mL of 75 % ethanol. After centrifugation at
7,500  g for 5 min at 4  C, the supernatant is discarded and the
RNA pellet air-dried for 5–10 min. The RNA pellet is then resus-
pended in nuclease-free water by pipetting up and down several
times. The concentration is measured by NanoDrop
spectrophotometer.

cDNA Generation and qRT- To prepare the RT master mix using the TaqMan® MicroRNA
PCR Performance and Reverse Transcription Kit (Applied Biosystems) components, the
Analysis kit components are allowed to thaw on ice. Before use, the RT
primer tubes are vortexed. The RT master mix is prepared in a
polypropylene tube on ice, gently mixed and then centrifuged.
Note: RT master mix for each sample consists of:
– 100 mM dNTPs (with dTTP): 0.15 μL.
– MultiScribe™ Reverse Transcriptase, 50 U/μL: 1.0 μL.
– 10 Reverse Transcription Buffer: 1.50 μL.
– RNase Inhibitor, 20 U/μL: 0.19 μL.
 Functional Analysis of microRNA in Multiple Myeloma 189

– Nuclease-free water: 4.16 μL.

– Each 15-μL RT reaction consists of 7 μL master mix, 3 μL of 5
RT primer, and 5 μL RNA sample diluted at 10 ng/L.

qRT-PCR The stock cDNA is 15 times diluted to the working concentration

using nuclease-free water.
Then the following components are combined into one reac-
tion: 5 μL of diluted cDNA, 1 μL of 20 TaqMan probe, 4 μL of
nuclease free water, and 10 μL TaqMan Universal Master Mix
without UNG AmpErase. The reaction mixture is transferred to
each well of an optical 96-well PCR microplate which is then sealed
with an ultra clear sealing film. The plate is briefly centrifuged and
the RT-PCR run is performed on a real-time PCR system, such as
the Viia7 Dx instrument (Applied Biosystems) with the following
conditions: 95  C, 10 min (enzyme activation); 95  C, 15 s (dena-
turation) and 60  C, 1 min (annealing/ extension), for 40 cycles.
The results, analyzed by the comparative Ct method or 2-ΔΔCt
method (24) for miRNAs relative expression (corrected to a refer-
ence miRNA such as RNU44 or RNU6), can be reported as the
relative fold change versus the control sample.

3.3 Analysis of To identify mRNA targets which can contribute to explain the role
Target Modulation of miRNAs in MM, Affymetrix gene expression profiling is per-
formed. Target mRNAs are then evaluated as possible miRNA
targets by luciferase reporter assay by the use of the 30 UTR
mutated and wild type sequence of the miRNA-target gene cloned
in pEZX-MT01 vector (Genecopoeia).

3.3.1 Gene Profiling Total RNA (tRNA) including small RNA fractions, is extracted from
transfected MM cells using a modified protocol from Qiagen as
above described. tRNA samples consist of a combination of ribo-
somal RNA (rRNA), messenger RNA (mRNA), transfer RNA
(tRNA), and other small RNA species with the rRNA fraction con-
stituting the vast majority (non-rRNA depleted). Gene expression
profiling (GEP) is then carried out according to the Affymetrix
recommended protocol. Briefly, after quantification of tRNA by
NanoDrop Spectrophotometry, 100 ng is processed using the Gen-
eChip WT PLUS Reagent Kit according to manufacturer’s instruc-
tions (www.affymetrix.com), which uses a reverse transcription
priming method that primes the entire length of each RNA tran-
script, including both poly-A and non-poly-A mRNA to provide
complete transcriptome. By the use of 100 ng of tRNA without
rRNA depletion 10 μg of cRNA is obtained to be carried into the
second cycle. The cRNA concentration is measured by UV spectro-
photometry (NanoDrop) following beads purification. Total cRNA
yield is calculated by multiplying the measured concentration by the
190 Maria Teresa Di Martino et al.

estimated elution volume (10.5 μL). Generally total cRNA yield

generates sufficient cRNA for the following reaction. Ten micro-
grams of cRNA is in fact used to generate second cycle, first strand
cDNA. In cases where cRNA concentration is low, the volume of
cRNA could be reduced by SpeedVac to a maximum of 6.5 μL. The
second strand (ss) cDNA yield is, as previously, measured by UV
Spectrophotometry (NanoDrop). The total yield of single stranded
cDNA was calculated by multiplying the measured concentration by
the estimated elution volume (28 μL). The 100 ng of RNA input
starting amounts (without rRNA reduction) generated sufficient
single stranded cDNA in the second cycle to generate a hybridization
cocktail containing 5.5 μg of labeled target. The yield of single
stranded cDNA is fairly consistent across all samples as is expected
since we use a constant mass (10 μg) of cRNA as input into this step.
Target preparation was carried out exactly as described in the Whole
Transcript (WT) Sense Target Labeling Assay Manual. 5.5 μg of
single-stranded cDNA is then fragmented, labeled, and hybridized.
The kit generates amplified and biotinylated sense-stranded DNA
targets, coupled with GeneChip® WT Terminal Labeling and Con-
trols Kit (Affymetrix) is adequate for preparing hybridization-ready
targets. Hybridization cocktail is prepared according to manual
instruction (www.affymetrix.com) using 5.2 μg of labeled RNA. A
total of 200 μL were loaded on each Affymetrix Human Transcrip-
tome Array (HTA) 2.0. Hybridized arrays were labeled and washed
using the Hybridization, Wash and Stain Kit on a GeneChip® Fluid-
ics Station 450 with the appropriate fluidics scripts (see manual for
details; Expression Wash, Stain and Scan User Manual for Cartridge
Arrays). Arrays are subsequently scanned on a GCS3000 7G Scan-
ner. CEL file are generated after quality control of array scan.

Data Processing All resulting CEL files for each array type are processed as a single
and Analysis group via the Affymetrix Power Tools (APT) package using sketch
normalization and the RMA algorithm to summarize probeset
signal. QC metrics from the APT report file shows, the mean
perfect match (PM) and mean background intensity for all of the
samples included in the analysis.
Expression values for gene are extracted from CEL files using
Affymetrix® Transcriptome Analysis Console (TAC) software. Dif-
ferent GeneChip® Array have been developed in the last decade by
the company. For the older GeneChip Whole Transcriptome (WT)
Arrays, log2-transformed expression values are extracted from CEL
files and normalized using Transcript Cluster Annotations, and
robust multi-array average (RMA) procedure in Expression Con-
sole (EC) software (Affymetrix Inc.). The analyses are then per-
formed on log2 transformed data generated from EC. After
hierarchical clustering of the samples, enabled to group either
genes, specimens or both with similar expression patterns,
 Functional Analysis of microRNA in Multiple Myeloma 191

supervised analyses is carried out using DChip software (www.hsph.

harvard.edu/cli/complab/dchip). The threshold for significance is
determined by fold change analysis applied on the expression values
of the transcript.
These microarray experiment procedures offer the unique pos-
sibility to select a subset of differentially expressed genes between
two classes of samples (i.e., transfected versus control) and thus
providing a global picture of the modification induced by miRNA
mimics or inhibitors treatment on the expression pattern across the
entire genome of the myeloma cell.
New miRNA targets could be identified by GEP analysis and
then will validated by in vitro and in vivo experiments to delineate
the mechanism of action of miRNA mimics or inhibitors in MM.
In addition it could be suggested the development of novel
therapies by targeting molecules involved in cell survival, migra-
tion, as well as drug resistance pathways or for the use of miRNA
mimics or inhibitors in combination therapy regimen.
Then pathway analysis could be applied by the use of the fold
change genes list, for example, loading into Ingenuity Pathway
Analysis (IPA®) software. This tool will be of support to reveal
biological pathways modulated by miRNAs overexpression or
silencing. For example after miR-221/222 knockdown, we identi-
fied by IPA modulation of canonic pathways involved in cell prolif-
eration signals and activation of immune response (20).
The information derived from molecular profiling technologies
will translate into the definition of the whole transcriptome pertur-
bation induced by miRNA replacement or inhibition and pathway
analysis will define the scenario by integrating the experimental data
with previous stored knowledge.
This approach will offer the opportunity for the identification
of therapeutical targets as well as pathways involved in crucial
cellular function as angiogenesis or bone marrow drive survival or
drug resistance.
All these finding will be finally validated by protein analysis,
which at present mostly rely on flow cytometry, immunofluores-
cence or western blotting analysis.

3.3.2 Luciferase The following procedure is based on the use of 30 UTR sequence of
Reporter Assay for miRNA- the selected mRNA cloned in pEZX-MT01 vector, specifically
Target Validation designed by and purchased from Genecopoeia. The 30 UTR of
interest—or a deletion mutant lacking the predicted miRNA target
sequence(s)—is specifically cloned in such vector containing both
Firefly and Renilla luciferase reporters.

Amplification 50 ng of plasmid DNA containing the wild type 30 UTR sequence

of Plasmid DNA of the gene of interest cloned in pEZX-MT01 (Genecopoeia)—or a
deletion mutant lacking the miRNA-target sequence—is added to
50 μL of competent cells (DH5α, Life Technologies). After 20 min
192 Maria Teresa Di Martino et al.

of incubation on ice, the cells underwent heat shock at 42  C for

40 s and immediately placed on ice for 3 min. 250 μL of S.O.C.
medium are added to the cells and incubated at 37  C for an hour.
100 μL of the bacterial culture are then spread on a pre-warmed
(37  C) culture plate containing kanamycin and incubated over-
night at 37  C.
Culture bacterial colonies: one individual colony is touched with a
sterile pipette tip which is then released in a sterile culture tube
containing 3 mL of liquid LB medium + kanamycin. Place the
tubes with the bacterial culture in a 37  C incubator with agitation.
After 6 h, transfer 200 μL into 400 mL of LB broth containing
kanamycin and leave overnight. The isolation of plasmid DNA is
performed by PureLink® HiPure Plasmid Maxiprep Kit (Life Tech-
nologies) according to the manufacturer’s protocol.

MM Cells Transfection The following procedure refers to the transfection of RPMI-8226

MM cells.
The day before the transfection, MM cells are seeded at
5.0  105 cells/mL in RPMI-1640 containing 10 % FBS and 1 %
penicillin-streptomycin. The transfection of MM cells is obtained
by the Neon transfection system (Life Technologies): briefly, use
1.0  106 cells, washed in PBS and resuspended in 100 μL of
buffer R. 2 μL of 100 μM miRNA mimics and 2.5 μL of the wild
type 30 UTR plasmid concentrated at 1 μg/μL are then added. The
Neon pipette and 100 μL tips are used for electroporation at the
following conditions: 1,050 V, 30 ms, 1 pulse. The transfected cells
are seeded into a six-well plate containing 2 mL of pre-warmed
growth medium without antibiotics. The same procedures is per-
formed using a deletion mutant lacking the miRNA-target
sequence. The plate is incubated in 37  C/5 % CO2 incubator for
24 h.
Note: prepare the following controls:
(a) Mix 2.5 μg of empty vector (pEZX-MT01) with either 100 nM
miRNA mimic or negative control (NC).
(b) 2.5 μg of empty vector (pEZX-MT01) alone.
(c) 2.5 μg of the 30 UTR-pEZX-MT01 plasmid alone.
(d) 2.5 μg of the mutated 30 UTR-pEZX-MT01 plasmid alone.

30 UTR Luciferase 24 h after the transfection, MM cells are collected and lysed with
Reporter Assay 200 μL of passive lysis buffer contained into the Dual Luciferase
Reporter Assay System (Promega). After incubation at r.t. for
10 min the cells are centrifuged at 2,320 rcf for 5 min and the
superrnatants collected. The Luciferase Assay Reagent II and the
Stop & Glo buffer (Dual-Glo Luciferase Assay Kit) are thawed at
r.t. and the content of one bottle of Luciferase Assay Reagent II is
 Functional Analysis of microRNA in Multiple Myeloma 193

transfered to the bottle of Luciferase Assay Substrate to create the

Dual-Glo Luciferase Reagent. The solution is then mixed thor-
oughly by inversion.
Measuring Firefly luciferase: 30 μL of cell lysates are incubated
with 100 μL of the luciferase substrate for 10 min into a white 96-
well plate appropriate for luminescence and the Firefly lumines-
cence is measured by reader set up for 96-well plates (GloMax,
Promega).
Measuring Renilla luciferase activity: 100 μL of Dual-Glo Stop
& Glo Reagent are added to each well and, after mix, incubated for
at least 10 min at room temperature and the Renilla luminescence
measured in a plate reader.
The ratio of luminescence from the experimental reporter
(Firefly) to the control reporter (Renilla) is calculated for each
well, and the average of at least three replicates is determined.
The results are normalized for the specific miRNA mimic versus
the control treatment.
The final aim of the luciferase assay is to validate the mRNA as a
specific target of the miRNA under investigation. It is important to
make clear that the UTR microRNA target can be predicted by
several software like TargetScan. Such prediction needs molecular
validation by luciferase assays in order to formally prove the real
miRNA targeting, because the biological activity is not regulated by
the identified nucleotide sequence alone. It has also to be consid-
ered that miRNA luciferase assay measures the mRNA stability and
are different from the traditional luciferase assays aimed to measure
the transcriptional activity of a DNA sequence.

References
1. Anderson KC (2014) Multiple myeloma. 6. Misso G et al (2013) Emerging pathways as
Hematol Oncol Clin North Am 28:xi–xii. individualized therapeutic target of multiple
doi:10.1016/j.hoc.2014.08.001 myeloma. Expert Opin Biol Ther 13(Suppl
2. Tagliaferri P et al (2012) Promises and chal- 1):S95–S109. doi:10.1517/14712598.2013.
lenges of microRNA-based treatment of multi- 807338
ple myeloma. Current Cancer Drug Targets 7. Misso G et al (2014) Mir-34: a new weapon
12:838–846 against cancer? Mol Ther Nucleic Acids 3:
3. Tassone P, Tagliaferri P (2012) Editorial: new e194. doi:10.1038/mtna.2014.47
approaches in the treatment of multiple mye- 8. Lionetti M et al (2013) Biological and clinical
loma: from target-based agents to the new era relevance of miRNA expression signatures in
of microRNAs (dedicated to the memory of primary plasma cell leukemia. Clin Cancer Res
Prof. Salvatore Venuta). Curr Cancer Drug 19:3130–3142. doi:10.1158/1078-0432.
Targets 12:741–742 CCR-12-2043
4. Rossi M et al (2013) From target therapy to 9. Lionetti M, Agnelli L, Lombardi L, Tassone P,
miRNA therapeutics of human multiple mye- Neri A (2012) MicroRNAs in the pathobiology
loma: theoretical and technological issues in of multiple myeloma. Curr Cancer Drug Tar-
the evolving scenario. Curr Drug Targets gets 12:823–837
14:1144–1149 10. Amodio N et al (2013) miR-29b induces
5. Rossi M et al (2014) MicroRNA and multiple SOCS-1 expression by promoter demethyla-
myeloma: from laboratory findings to transla- tion and negatively regulates migration of mul-
tional therapeutic approaches. Curr Pharm tiple myeloma and endothelial cells. Cell Cycle
Biotechnol 15:459–467 12:3650–3662. doi:10.4161/cc.26585
194 Maria Teresa Di Martino et al.

11. Amodio N, Di Martino MT, Neri A, Tagliaferri 18. Rossi M et al (2013) miR-29b negatively reg-
P, Tassone P (2013) Non-coding RNA: a novel ulates human osteoclastic cell differentiation
opportunity for the personalized treatment of and function: implications for the treatment
multiple myeloma. Expert Opin Biol Ther 13 of multiple myeloma-related bone disease. J
(Suppl 1):S125–S137. doi:10.1517/ Cell Physiol 228:1506–1515. doi:10.1002/
14712598.2013.796356 jcp.24306
12. Amodio N et al (2012) DNA-demethylating 19. Scognamiglio I et al (2014) Transferrin-
and anti-tumor activity of synthetic miR-29b conjugated SNALPs encapsulating 20 -O-
mimics in multiple myeloma. Oncotarget methylated miR-34a for the treatment of mul-
3:1246–1258 tiple myeloma. Biomed Res Int 2014:217365.
13. Di Martino MT et al (2014) In vivo activity of doi:10.1155/2014/217365
miR-34a mimics delivered by stable nucleic 20. Monroig PD, Chen L, Zhang S, Calin GA
acid lipid particles (SNALPs) against multiple (2014) Small molecule compounds targeting
myeloma. PloS One 9:e90005. doi:10.1371/ miRNAs for cancer therapy. Adv Drug Deliv
journal.pone.0090005 Rev. doi:10.1016/j.addr.2014.09.002
14. Di Martino MT et al (2014) In vitro and in vivo 21. Amodio N et al (2012) miR-29b sensitizes mul-
activity of a novel locked nucleic acid (LNA)- tiple myeloma cells to bortezomib-induced apo-
inhibitor-miR-221 against multiple myeloma ptosis through the activation of a feedback loop
cells. PloS One 9:e89659. doi:10.1371/jour with the transcription factor Sp1. Cell Death Dis
nal.pone.0089659 3:e436. doi:10.1038/cddis.2012.175
15. Di Martino MT et al (2012) Synthetic miR-34a 22. Di Martino MT et al (2013) In vitro and in vivo
mimics as a novel therapeutic agent for multi- anti-tumor activity of miR-221/222 inhibitors
ple myeloma: in vitro and in vivo evidence. Clin in multiple myeloma. Oncotarget 4:242–255
Cancer Res 18:6260–6270. doi:10.1158/ 23. Lionetti M et al (2009) Identification of micro-
1078-0432.CCR-12-1708 RNA expression patterns and definition of a
16. Leone E et al (2013) Targeting miR-21 inhi- microRNA/mRNA regulatory network in dis-
bits in vitro and in vivo multiple myeloma cell tinct molecular groups of multiple myeloma.
growth. Clin Cancer Res 19:2096–2106. Blood 114:e20–e26. doi:10.1182/blood-
doi:10.1158/1078-0432.CCR-12-3325 2009-08-237495
17. Leotta M et al (2014) A p53-dependent tumor 24. Livak KJ, Schmittgen TD (2001) Analysis of
suppressor network is induced by selective relative gene expression data using real-time
miR-125a-5p inhibition in multiple myeloma quantitative PCR and the 2(-Delta Delta C
cells. J Cell Physiol 229:2106–2116. doi:10. (T)) method. Methods 25:402–408. doi:10.
1002/jcp.24669 1006/meth.2001.1262
Methods in Molecular Biology (2016) 1375: 195–206
DOI 10.1007/7651_2015_245
© Springer Science+Business Media New York 2015
Published online: 26 June 2015

Microarray Analysis in Glioblastomas

Kaumudi M. Bhawe and Manish K. Aghi

Abstract
Microarray analysis in glioblastomas is done using either cell lines or patient samples as starting material. A
survey of the current literature points to transcript-based microarrays and immunohistochemistry (IHC)-
based tissue microarrays as being the preferred methods of choice in cancers of neurological origin.
Microarray analysis may be carried out for various purposes including the following:
i. To correlate gene expression signatures of glioblastoma cell lines or tumors with response to chemo-
therapy (DeLay et al., Clin Cancer Res 18(10):2930–2942, 2012)
ii. To correlate gene expression patterns with biological features like proliferation or invasiveness of the
glioblastoma cells (Jiang et al., PLoS One 8(6):e66008, 2013)
iii. To discover new tumor classificatory systems based on gene expression signature, and to correlate
therapeutic response and prognosis with these signatures (Huse et al., Annu Rev Med 64(1):59–70,
2013; Verhaak et al., Cancer Cell 17(1):98–110, 2010)
While investigators can sometimes use archived tumor gene expression data available from repositories
such as the NCBI Gene Expression Omnibus to answer their questions, new arrays must often be run to
adequately answer specific questions. Here, we provide a detailed description of microarray methodologies,
how to select the appropriate methodology for a given question, and analytical strategies that can be used.
Experimental methodology for protein microarrays is outside the scope of this chapter, but basic sample
preparation techniques for transcript-based microarrays are included here.

Keywords: Microarray, Glioblastoma, Gene-expression

1 Introduction

Glioblastoma, the most common malignant primary brain tumor,

carries an invariably poor prognosis (3, 5, 6). Targeting underlying
biological foundations of the disease will be crucial to developing
more effective treatment strategies (3, 5, 6). Transcriptional
profiling through microarray analysis and protein expression
profiling through immunohistochemistry (IHC)-based microar-
rays represent vital resources for researchers seeking to accomplish
these goals (3, 5, 6).
Here, we first describe protocols for gathering gene expression
data with transcript-based microarrays. Next, we review the various
methods of data analysis and clustering along with merits and

195
196 Kaumudi M. Bhawe and Manish K. Aghi

demerits of each approach. Finally, we highlight important consid-

erations to keep in mind while selecting the optimal approach to
test your particular hypothesis.

2 Materials

2.1 Sample Types 1. Cells.

and Associated a. Glioblastoma cells.
Culture Media and
i. Types of cells.
Equipment
1. Glioblastoma-derived tumor-initiating stem cells (7).
2. Glioma cell lines such as A172, CCF-SSTG1,T98G,
U373MG, U178MG, TP365MG, U118MG,
U251MG, GL15, U105MG, U251MG, U343MG,
U373MG, and SF767 (8, 9).
3. Primary glioblastoma cultures.
ii. Normal control cells for comparison.
1. Neural stem cells such as CB541 and CB660 (10).
2. Peripheral blood collected in blue-top monoject
tubes (3.5 % sodium citrate anticoagulant, Teromo
Corp., Japan) (10).
b. Medium.
i. Stem-cell medium (8) 9 made of DMEM/F-12 contain-
ing 20 % bovine serum albumin, insulin and transferrin
(BIT)-serum-free supplement, and basic fibroblast and
epidermal growth factors (Provitro, 20 ng/mL each) (8).
ii. DMEM, containing 10 % fetal bovine serum (FBS).
iii. RPMI-1640 Medium (Sigma-Aldrich Sweden AB,
Stockholm, Sweden) (9).
2. Tissue.
a. Patient glioblastoma specimens.
i. Flash-frozen paraffin-embedded (FFPE) tumor sections
[11)—poorer quality RNA from paraffin sections requires
special preparatory protocols and stringent purity criteria.
ii. Frozen tumor pieces
b. Frozen pieces from subcutaneous or intracranial xenografts
treated with vehicle versus drug of interest.

2.2 Materials RNA Isolation

for Transcript-Based
1. RecoverAll Total Nucleic Acid Isolation Kit (Ambion, Inc.) (1).
Microarrays
2. RNeasy kit (Qiagen).
3. TotalPrep RNA Amplification kit (Illumina).
4. Blood and Cell Culture Kit (Qiagen).
5. DNaseI (Invitrogen).
 Microarray Analysis in Glioblastomas 197

6. Cesium chloride column.

7. Ultracentrifuge.
Assessment of RNA Quality
8. ND-1000 spectrophotometer (NanoDrop Technologies,
Wilmington, DE) (9).
9. Agilent 2100 bioanalyzer (Agilent).

Hybridization-Ready Sample Preparation

10. SuperscriptII (Invitrogen).
11. Reference total RNA obtained from nonneoplastic human
brain tissue samples of five individuals (Bio-Chain) (8).
12. HybBag mixing system with 1 OneArray Hybridization
Buffer (Phalanx Biotech) (11).
13. Salmon sperm DNA (Promega) (11).
14. Molecular Dynamics™ Axon 4100A scanner (11).
15. ABI PRISM 7900 (Applied Biosystems) (8) for RT-PCR for
validating the transcript-based microarray data.
16. Absolute SYBR Green ROX Mix (ABgene).
17. Biotin-16-UTP.
18. Cy5 NHS ester (GE Healthcare Life Sciences) (11).
Microarrays and Signal Detection
19. Illumina Human whole-genome Sentrix-6V2 BeadChip array.
20. Affymetrix GeneChip expression arrays (Human Genome
U133 Plus 2.0 Array) (9).
21. Whole-Genome DASL Assay with HumanRef-8 BeadChips
(Illumina, Inc.; San Diego, CA) (1).
22. Whole Human Genome Oligo Microarray 4x44K (Agilent) (1).
23. Human HT-12 v4 Expression BeadChip Kits (Illumina; San
Diego, CA) (1).
24. Human Whole Genome OneArray v2 (Phalanx Biotech) (11).
25. GeneChip Expression Analysis Technical Manual (Rev. 5, Affy-
metrix Inc., Santa Clara, CA) (9).
26. Fluidics Station 450 (Affymetrix Inc.) for washing and staining
microarrays.
27. 45  C incubator, capable of rotation up to 60 rpm.
28. Bead station array scanner.
29. GeneChip® Scanner 3000 7G (Affymetrix Inc.) (9).

2.3 Ways of 1. Length of probe—arrays can be classified into “complementary

Classifying Transcript- DNA (cDNA) arrays,” which use long probes of hundreds or
Based Microarrays thousands of base pairs (bps), and “oligonucleotide arrays,”
198 Kaumudi M. Bhawe and Manish K. Aghi

which use short probes (usually 50 bps or less). Manufacturing

methods include “deposition” of previously synthesized
sequences and “in situ synthesis.”
2. Manufacturing technique—Usually, cDNA arrays are manufac-
tured using deposition, while oligonucleotide arrays are
manufactured using in situ technologies. In situ technologies
include “photolithography” (e.g., Affymetrix, Santa Clara, CA),
“ink-jet printing” (e.g., Agilent, Palo Alto, CA), and “electro-
chemical synthesis” (e.g., Combimatrix, Mukilteo, WA) (12).
3. Number of samples— “Single-channel arrays” analyze a single
sample at a time, whereas “multiple-channel arrays” can ana-
lyze two or more samples simultaneously. An example of an
oligonucleotide, single-channel array is the Affymetrix Gene-
Chip (12).

2.4 Types of Protein 1. Analytical/capture microarrays, where a library of antibodies,

Microarrays aptamers, or affibodies arrayed on the support surface act as
capture molecules since each binds specifically to a particular
protein: Samples such as cell lysates can be then applied to the
array, and a variety of detection methods can be used to deter-
mine the relative levels of array proteins found in the sample
solution (13).
2. Functional protein microarrays/target protein microarrays,
which are constructed by immobilizing large numbers of pur-
ified proteins and are used to identify protein–protein, pro-
tein–DNA, protein–RNA, protein–phospholipid, and
protein–small-molecule interactions, to assay enzymatic activity
and to detect antibodies and demonstrate their specificity. They
differ from analytical arrays in that they contain full-length
functional proteins or protein domains and can in some cases
be used to study the biochemical activities of the entire prote-
ome in a single experiment (13).
3. Reverse-phase protein arrays (RPPA), so called because in this
case, the sample, which can be cell lysate or complex tissue
lysate, is applied to the microarray, and then probed with anti-
bodies against the target proteins of interest. Methods of detec-
tion are usually chemiluminescence, fluorescence, or
colorimetry. Reference peptides are printed on the arrays to
allow for protein quantification of the sample lysates (13).
4. Tissue microarrays (TMA) probed using IHC protocols are
where laser-capture-microdissected tissue may be spotted in
an array format, and then assayed with a variety of antibodies
towards expressed proteins. The added benefit of IHC-based
arrays is the fact that expression and tissue localization of
proteins can be simultaneously studied. A significant drawback
is the lack of molecular weight verification of identified
 Microarray Analysis in Glioblastomas 199

proteins, which means that the detection antibodies must be

thoroughly validated using western blotting prior to use in the
IHC technique.

2.5 Typical Workflow Illustrated in Fig. 1 is the typical workflow of microarray-based

of Microarray-Based experiments. Note that although protein-based microarrays are
Experiments outside the scope of this chapter, the analysis methodologies
described here can be applied irrespective of whether expression
levels are measured based on transcript or protein.

2.6 Examples 1. Bioconductor packages.

of Software a. DESeq (normalizing tag counts for transcriptome tag
and Databases Used sequencing) (7).
for Microarray Data
b. Signaling pathway impact analysis (SPIA) (7).
Clustering and
Analysis
c. Affy.
d. Org.Hs., e.g., (tests for enrichment of gene ontology terms) (7).
2.6.1 Software
e. DNAcopy (7).

Fig. 1 A representative microarray-based experimental workflow. Shown are the

typical steps taken in microarray analysis from sample processing to data
analysis
200 Kaumudi M. Bhawe and Manish K. Aghi

f. CGHcall.
g. CGHnormaliter (correction for intensity dependence).
h. Bead array R package (svn release 1.7.0) (8).
i. Lumi R package (release 1.1.0) for variance stabilizing and
spline normalizing (8).
2. Recount program (to correct for potential sequencing errors
during transcriptome tag sequencing) (7).
3. TagDust (7).
4. Bowtie short read aligner (to remove tags coming from mito-
chondrial RNA or rRNA) (7).
5. limma (comparing between microarrays) (7).
6. Ingenuity Pathways Knowledge Base and Analysis Software
(www.ingenuity.com) (8).
7. BLAT (Kent 2002) (14).
8. AltAnalyze (15, 16) for quintile normalization to look at dif-
ferential gene expression (14).
9. Partek genomic suite (http://www.partek.com/) for analysis of
the microarray data (14).
10. Significance Analysis of Microarrays (SAM) 3.0 (Stanford Uni-
versity) for statistical analyses (17).
11. Imagene 6.0 data extraction software (BioDiscovery Inc.) (17).
12. AROMA (18).

2.6.2 Databases 13. Gene Ontology.

14. Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway
database.
15. The Cancer Genome Atlas (TCGA) dataset consisting of 397
glioblastoma cases.
16. NCI-60 expression data from CellMiner (9).
17. Database for Annotation, Visualization and Integrated Discov-
ery (DAVID) Bioinformatics Resources 6.7 (http://david.abcc.
ncifcrf.gov/home.jsp) for Gene Set Enrichment Analysis
(GSEA) (9).
18. Biocarta.
19. PANTHER.
20. SPSS 16.0 (SPSS Inc., Chicago) (17).
21. ArrayExpress database (accession no. E-MEXP-3296) (1).
22. GenePixPro™ Software (11).
23. C5.BP.V3.0 (gene ontology: biological processes) and C2.CP.
V3.0 (canonical pathways) MSigDB gene sets for GSEA (11).
24. Chinese glioma genome atlas (2).
 Microarray Analysis in Glioblastomas 201

3 Methods for Transcript-Based Microarrays

1. Based on your initial sample, use the appropriate reagents for

isolating total RNA (items 1–4, Section 2.2). In most cases, the
specific instructions are given by the kit manufacturers.
2. Determine RNA quality and integrity utilizing an Agilent 2100
Bioanalyzer (Agilent Technologies) and absorbance at A260/
A280. Only high-quality RNA, having a RIN of >7.0 and an
A260/280 absorbance ratio of >1.8, should be utilized for
further experimentation. This step is particularly important
for RNA derived from paraffin-embedded tissue, whose purity
may be limited, often requiring re-purification (11).
3. Most microarrays require 2 μg of high-quality total RNA from
each sample. Most microarrays require conversion of RNA to
biotinylated fragmented complementary RNA (cRNA). cRNA
is necessary because the oligonucleotides are in the sense direc-
tion and so one has to use antisense RNA. Amplification is
necessary since most microarrays require about 25–100 μg of
total RNA to be hybridized (14). Microarrays are generally
classified into two broad categories based on their method of
synthesis. The two categories are spotted microarrays and oli-
gonucleotide microarrays. In the case of spotted microarrays,
the probes can be oligonucleotides, cDNA, or small fragments
of PCR products corresponding to mRNAs, and they are
synthesized prior to deposition on the array surface and are
then “spotted” onto glass. For such spotted arrays one can use
either mRNA, cDNA, or cRNA because both strands are used
as probes on the microarray (13). In the case of oligonucleotide
microarrays, probes can be either produced by piezoelectric
deposition with full-length oligonucleotides or in situ synthe-
sis. While spotted microarrays are more amenable to in-house
printing for custom-made arrays, oligonucleotide microarrays
have higher probe density and also higher reproducibility from
one array to another in terms of experimental results. Biotinyla-
tion of the test sample is necessary when the microarray has
streptavidin to capture the RNA.
4. On any given microarray, once the capture probe is immobi-
lized to the substrate, it is important to perform two additional
steps prior to using the microarray. If a covalent chemistry was
used for immobilization, any residual reactive groups on the
surface should be removed. This is commonly called quenching
the surface. Under certain conditions, this is also referred to as
capping. For example, residual epoxide (EP) groups can be
reacted with an amine compound such as ethanolamine,
whereas aldehyde groups can be reduced to alcohols
using sodium borohydride. The second process is commonly
202 Kaumudi M. Bhawe and Manish K. Aghi

called blocking. Once residual reactive groups are destroyed,

the issue of nonspecific adsorption will need to be addressed.
What you choose to block with depends on several factors such
as the treated surface, hybridization cocktail, and sample
matrix. Common blocking agents include detergents such as
Tween 20, salmon sperm DNA, tRNA, or proteins such as
bovine serum albumin (BSA) (19).
5. The key physicochemical process involved in microarrays is
hybridization. Samples are typically hybridized overnight
(12–16 h) at a temperature between 42 and 45  C.
6. The arrays hybridized with sample must then be washed,
stained, and scanned with methods appropriate for the micro-
array of choice. For example, for Affymetrix microarrays, you
can use the Fluidics Station 450 (Affymetrix Inc.) and scan with
the GeneChip Scanner 3000 7G (Affymetrix Inc.) (9).
7. The first step in any analysis is to apply a background correction
which accounts for the percent of intensity coming from non-
specific binding to the microarray. Background correction can
be applied using the intensity levels in the vicinity of spots in the
case of spotted arrays. In the case of high-density arrays, mis-
match probes can be used to estimate the amount of signal
coming from nonspecific binding.
8. After background correction, the data is generally log-
transformed. The log transformation improves the characteristics
of the data distribution and allows the use of classical parametric
statistics for analysis. With two-channel arrays, the intensity values
of the two competing samples are expressed as ratios and then log-
transformed. In contrast, with single-channel technology (e.g.,
Affymetrix), the “absolute” expression level of the genes is log-
transformed. Logarithmic transformation also converts multipli-
cative error into additive error (12).
9. Normalization of the raw data is a subsequent necessary step so
that the final data can be compared across platforms. The aim of
normalization is to remove any systematic biases that may be
causing artifactual intensity variance between samples on
account of inherent differences in dye characteristics, array
manufacturing, and spatial location of the sample on a given
array. Some examples of freely available normalization tools are
Bioconductor packages such as MAS 5.0, Robust Microarray
Average (RMA), and GC-RMA33 for single-channel arrays,
and LOESS normalization for two-channel arrays (12).
10. Once the data is normalized, it can be used for hypothesis
testing. Analysis methods described from this point on can
also be used for meta-analysis of existing expression data in
databases such as KEGG, TCGA, DAVID, ArrayExpress data-
base, and others.
 Microarray Analysis in Glioblastomas 203

11. Currently, there are three major types of applications of

transcript-based microarrays in medicine. The first involves
finding differences in expression levels between predefined
groups of samples. This is called a “class comparison” experi-
ment. A second application, “class prediction,” involves identi-
fying the class membership of a sample based on its gene
expression profile. This requires the construction of a classifier
(a mathematical model) able to analyze the gene expression
profile of a sample and predict its class membership. The classi-
fier is constructed based on a representative set of samples with
known class membership. This classifier will then be used to
assess the likelihood of developing glioblastoma in patients not
included in construction of the classifier. The third type of
application involves analyzing a given set of gene expression
profiles with the goal of discovering subgroups that share com-
mon features. This application is known as “class discovery.”
For example, the expression profiles of a large number of
patients with glioblastoma will be measured with the goal of
identifying subgroups of patients who have a similar gene
expression profile. This effort is conducted to generate a
molecular taxonomy of disease. In other words, how many
molecular types of glioblastoma are in a sample of patients
affected by the disease? (12).
12. An unsupervised clustering analysis can be carried out in order
to search for obvious patterns. Clusters identified in such a
manner can then be further validated (1). These clusters can
be graphically illustrated in the form of “heatmaps” showing
upregulated and downregulated gene sets from one sample to
the next (Fig. 2).
13. If there is no clustering detected then an unbiased gene selec-
tion approach may be used, where the samples to be compared
are clustered on high-variance probes (top 98th percentile and
above) (1), and then examined for correlations with any classes
established through other means such as histopathology or
imaging.
14. In class comparison and class discovery studies, the expression
characterization of the groups (e.g., health vs. disease) is often
followed by “functional profiling.” The purpose of this task is
to gain insight into the biological processes that are altered in
disease. GSEA is currently the most widely used method of
functional profiling. GSEA is a computational method that
determines whether an a priori-defined set of genes shows
statistically significant, concordant differences between two
biological states (e.g., phenotypes) (19). When comparing
two distinct biological phenotypes, there are some major lim-
itations to the simple approach of identifying the genes that
show the largest expression differences across the phenotypes
in question. The limitations are as follows:
204 Kaumudi M. Bhawe and Manish K. Aghi

Fig. 2 A representative heatmap of gene expression obtained by microarray analysis. Shown is an unpublished
heatmap showing differentially expressed genes in a glioblastoma cell engineered to express shRNA targeting
autophagy gene ATG7

i. No individual gene may meet the threshold for statistical

significance, due to a small signal-to-noise ratio.
ii. In case of a long list of statistically significant genes without
known biological connections between them, it becomes
difficult to interpret the data meaningfully.
iii. Since cellular processes typically involve a large number of
genes acting in concert, seemingly minor expression
changes in a set of related genes may be more interesting
to follow up on as compared to a small set of unrelated
genes that show largely statistically significant differences in
expression levels between the groups compared.
iv. When different groups study the same biological system,
the list of statistically significant genes from the two studies
may show very little overlap while there may be identical
genetic pathways being affected that remain undetected
because of a limitation in the analysis methodology.
GSEA is a computational protocol that seeks to get around the
limitations listed above (20).
 Microarray Analysis in Glioblastomas 205

4 Notes

1. The sample preparation technique greatly limits the range of

microarrays that can be used for a given study. Most transla-
tional studies begin with FFPE samples, while studies aimed at
deciphering underlying molecular pathways might use cell lines
as beginning material. Cell line-derived samples are invariably
of higher quality than frozen tissue-derived samples, which in
turn are of significantly higher quality than FFPE samples.
While cell line-derived samples or frozen tissue-derived samples
can be used directly as starting material for most commercially
available arrays, FFPE material, on the other hand, suffers from
having degraded and low-quality RNA. As a result, specialized
microarray assays such as the cDNA-mediated annealing, selec-
tion, extension, and ligation (DASL) assay by Illumina must be
used when working with FFPE samples. The DASL assay uses
random priming in the cDNA synthesis, and therefore does not
depend on an intact poly(A) tail for T7-oligo-d(T) priming. In
addition, the assay requires a relatively short target sequence of
about 50 nucleotides for query oligonucleotide annealing,
allowing the assay to perform well with significantly degraded
RNAs (21).
2. The subsequent algorithms and software packages used for
analysis are usually linked to the particular microarray of choice.
However the underlying analysis strategies are common across
software packages, and need to be chosen based on the type of
statistical analysis deemed necessary to answer the questions
posed by the researchers. Here we have presented in detail a
prototypical microarray experimental workflow. Depending on
sample type, microarray choice, and software used, readers
must draw parallels or make choices based on their own unique
research goals.

References

1. DeLay M, Jahangiri A, Carbonell WS, Hu YL, implications. Annu Rev Med 64(1):59–70.
Tsao S, Tom MW, Paquette J, Tokuyasu TA, doi:10.1146/annurev-med-100711-143028
Aghi MK (2012) Microarray analysis verifies 4. Verhaak RG, Hoadley KA, Purdom E, Wang V,
two distinct phenotypes of glioblastomas resis- Qi Y, Wilkerson MD, Miller CR, Ding L,
tant to antiangiogenic therapy. Clin Cancer Res Golub T, Mesirov JP, Alexe G, Lawrence M,
18(10):2930–2942. doi:10.1158/1078- O’Kelly M, Tamayo P, Weir BA, Gabriel S,
0432.ccr-11-2390 Winckler W, Gupta S, Jakkula L, Feiler HS,
2. Jiang T, Tie X, Han S, Meng L, Wang Y, Wu A Hodgson JG, James CD, Sarkaria JN, Brennan
(2013) NFAT1 Is highly expressed in, and reg- C, Kahn A, Spellman PT, Wilson RK, Speed
ulates the invasion of, glioblastoma multiforme TP, Gray JW, Meyerson M, Getz G, Perou CM,
cells. PLoS One 8(6):e66008. doi:10.1371/ Hayes DN, Cancer Genome Atlas Research
journal.pone.0066008 Network (2010) Integrated genomic analysis
3. Huse JT, Holland E, DeAngelis LM (2013) identifies clinically relevant subtypes of glio-
Glioblastoma: molecular analysis and clinical blastoma characterized by abnormalities in
206 Kaumudi M. Bhawe and Manish K. Aghi

PDGFRA, IDH1, EGFR, and NF1. Cancer 12. Tarca ALRRDS (2006) Analysis of microarray
Cell 17(1):98–110. doi:10.1016/j.ccr.2009. experiments of gene expression profiling. Am J
12.020 Obstet Gynaecol 192(2):15
5. Bao ZS, Zhang CB, Wang HJ, Yan W, Liu YW, 13. Hartmann M, Roeraade J, Stoll D, Templin
Li MY, Zhang W (2013) Whole-genome MF, Joos TO (2009) Protein microarrays for
mRNA expression profiling identifies func- diagnostic assays. Anal Bioanal Chem 393
tional and prognostic signatures in patients (5):1407–1416. doi:10.1007/s00216-008-
with mesenchymal glioblastoma multiforme. 2379-z
CNS Neurosci Ther 19(9):714–720. doi:10. 14. Solomon O, Oren S, Safran M, Deshet-Unger
1111/cns.12118 N, Akiva P, Jacob-Hirsch J, Cesarkas K, Kabesa
6. Tivnan A, McDonald KL (2013) Current R, Amariglio N, Unger R, Rechavi G, Eyal E
progress for the use of miRNAs in glioblastoma (2013) Global regulation of alternative splicing
treatment. Mol Neurobiol. doi:10.1007/ by adenosine deaminase acting on RNA
s12035-013-8464-0 (ADAR). RNA 19(5):591–604. doi:10.1261/
7. Engstrom PG, Tommei D, Stricker SH, Ender rna.038042.112
C, Pollard SM, Bertone P (2012) Digital tran- 15. Emig D, Salomonis N, Baumbach J, Lengauer
scriptome profiling of normal and T, Conklin BR, Albrecht M (2010) AltAnalyze
glioblastoma-derived neural stem cells identi- and DomainGraph: Analyzing and visualizing
fies genes associated with patient survival. exon expression data. Nucleic Acids Res 38:
Genome Med 4(10):76. doi:10.1186/gm377 W755–W762
8. Ernst A, Hofmann S, Ahmadi R, Becker N, 16. Salomonis N, Schlieve CR, Pereira L, Wahl-
Korshunov A, Engel F, Hartmann C, Felsberg quist C, Colas A, Zambon AC, Vranizan K,
J, Sabel M, Peterziel H, Durchdewald M, Hess Spindler MJ, Pico AR, Cline MS, et al. (2010)
J, Barbus S, Campos B, Starzinski-Powitz A, Alternative splicing regulates mouse embryonic
Unterberg A, Reifenberger G, Lichter P, stem cell pluripotency and differentiation. Proc
Herold-Mende C, Radlwimmer B (2009) Natl Acad Sci 107:10514–10519
Genomic and expression profiling of glioblas-
17. Lin Y, Zhang G, Zhang J, Gao G, Li M, Chen
toma stem cell-like spheroid cultures identifies
Y, Wang J, Li G, Song S-W, Qiu X, Wang Y,
novel tumor-relevant genes associated with sur-
vival. Clin Cancer Res 15(21):6541–6550. Jiang T (2013) A panel of four cytokines pre-
doi:10.1158/1078-0432.ccr-09-0695 dicts the prognosis of patients with malignant
gliomas. J Neuro-Oncol 114(2):199–208.
9. Sooman L, Ekman S, Andersson C, Kultima doi:10.1007/s11060-013-1171-x
HG, Isaksson A, Johansson F, Bergqvist M,
Blomquist E, Lennartsson J, Gullbo J (2013) 18. Godoy PR, Mello SS, Magalhaes DA, Donaires
Synergistic interactions between camptothecin FS, Nicolucci P, Donadi EA, Passos GA,
and EGFR or RAC1 inhibitors and between Sakamoto-Hojo ET (2013) Ionizing
imatinib and Notch signaling or RAC1 inhibi- radiation-induced gene expression changes in
tors in glioblastoma cell lines. Cancer Che- TP53 proficient and deficient glioblastoma cell
mother Pharmacol 72(2):329–340. doi:10. lines. Mutat Res 756(1–2):46–55. doi:10.
1007/s00280-013-2197-7 1016/j.mrgentox.2013.06.010
10. Zeeberg BR, Kohn KW, Kahn A, Larionov V, 19. Matson RS, Wadia PP, Miklos DB, Song Y,
Weinstein JN, Reinhold W, Pommier Y (2012) Wang D, Yamada M, Martinsky T (2009)
Concordance of gene expression and functional Microarray methods and protocols. CRC
correlation patterns across the NCI-60 cell lines Press, Boca Raton, FL
and the cancer genome atlas glioblastoma sam- 20. Subramanian A, Tamayo P, Mootha VK,
ples. PLoS One 7(7):e40062, doi: 10.1371/ Mukherjee S, Ebert BL, Gillette MA, Paulovich
journal.pone.0040062.g001. 10.1371/jour- A, Pomeroy SL, Golub TR, Lander ES,
nal.pone.0040062.t001. 10.1371/ journal. Mesirov JP (2005) Gene set enrichment analy-
pone.0040062.t002 sis: a knowledge-based approach for interpret-
11. Quann K, Gonzales DM, Mercier I, Wang C, ing genome-wide expression profiles. Proc Natl
Sotgia F, Pestell RG, Lisanti MP, Jasmin J-F Acad Sci U S A 102(43):15545–15550.
(2013) Caveolin-1 is a negative regulator of doi:10.1073/pnas.0506580102
tumor growth in glioblastoma and modulates 21. Doerks T, Copley RR, Schultz J, Ponting CP,
chemosensitivity to temozolomide. Cell Bork P (2002) Systematic identification of
Cycle 12(10):1510–1520. doi:10.4161/cc. novel protein domain families associated with
24497 nuclear functions. Genome Res 12(1):47–56,
10.1101/
Methods in Molecular Biology (2016) 1375: 207–221
DOI 10.1007/7651_2015_247
© Springer Science+Business Media New York 2015
Published online: 14 May 2015

Analysis of microRNA Microarrays in Cardiogenesis

Diego Franco, Fernando Bonet, Francisco Hernandez-Torres,
Estefania Lozano-Velasco, Francisco J. Esteban,
and Amelia E. Aranega

Abstract
microRNAs are a subclass of noncoding RNAs which have been demonstrated to play pivotal roles in
multiple cellular mechanisms. microRNAs are small RNA molecules of 22–24 nt in length capable of
modulating protein translation and/or RNA stability by base-priming with complementary sequences of
the mRNAs, normally at the 30 untranslated region. To date, over 2,000 microRNAs have been already
identified in humans, and orthologous microRNAs have been also identified in distinct animals and plants
ranging a wide vast of species. High-throughput analyses by microarrays have become a gold standard to
analyze the changes on microRNA expression in normal and pathological cellular or tissue conditions. In
this chapter, we provide insights into the usage of this uprising technology in the context of cardiac
development and disease.

Keywords: microRNA, Microarrays, Cardiac development, Meta-analyses

1 Background

Cardiac development is a complex process in which multiple cell types

are involved (1–3). From the early stages of cardiogenic specification
soon after gastrulation, the heart progressively acquires a tubular
shape formed by an inner endothelial lining and an outer myocardium
layer (4). The heart is the first organ to display left-right asymmetry
(5, 6) and more importantly, it is the first organ to be functional
during organogenesis (2). Soon after the heart start to pump, atrial
and ventricular chambers are progressively configured and valve pri-
mordia are formed in the intertwining areas, i.e., outflow tract, atrio-
ventricular canal, and inflow tract (2). Subsequently, the heart is
separated and each of these structures becomes divided into left and
right parts (7, 8). Over the last decades, we have gained crucial
insights into the molecular mechanisms that govern cardiac morpho-
genesis. Given the complexity of cardiac morphogenesis, multiple
pathways and transcriptional factors are involved during cardiogenesis
at different stages, as recently reviewed (9, 10). In the last years, a
novel layer of complexity is emerging in the cardiovascular

207
208 Diego Franco et al.

development field, namely the post-transcriptional regulatory

networks driven by noncoding RNAs. Among noncoding RNAs,
microRNAs have emerged as a pivotal mechanism is cardiovascular
development, since targeted deletion of the microRNAs processing
ribonuclease Dicer resulted in severe cardiovascular defects. More-
over, targeted deletion of single microRNAs, such as miR-1 and
miR-126, respectively, also resulted in severe cardiovascular develop-
ment impairment (11, 12). In view of the important contribution of
distinct microRNAs in cardiovascular field, great efforts have been
devoted to map the microRNA microarray fingerprints of distinct
normal and abnormal cardiovascular contexts (see for a recent review
(13–15)). In this chapter, we provide insights into the usage of
microRNA microarray analyses in the field of cardiovascular develop-
ment, as well as insights as how to proceed further beyond the classical
microarray approach, such as microarray meta-analyses.

2 Materials and Methods

Initial steps to ensure appropriate microRNA microarray analyses

start already on the experimental design. Considering a simple
experiment in which a control and an experimental condition will
be analyzed, triplicates of each biological assay should be per-
formed. After experimentation, each condition should be processed
for RNA isolation, array hybridization, data acquisition and nor-
malization, as detailed in the following subheadings. In addition,
several other steps are also recommended such as independent
validation and assessment of predicted functional roles. Examples
of the latter are also illustrated below.

2.1 Isolation of RNA Purification and preparation of total RNA that includes small RNAs
for microRNA (<200 nt) from a biological samples is the first critical step for a
Microarrays successful expression profiling analysis of microRNAs. Therefore,
the method used for RNA simple preparation is critical to the
success of the experiment. An important limitation is that naked
RNA is extremely susceptible to degradation by endogenous ribo-
nucleases (RNases) that are present in all living cells. Thus, the key
to successful isolation of high-quality RNA is to ensure that neither
endogenous nor exogenous RNases are introduced during the
extraction procedure. We normally used a TRIzol-based isolation
protocol to isolate total RNA, without any special requirements for
small RNA enrichment, as detailed below.

2.1.1 Tissue Homogenize 100 mg of tissue or cell pellet in 2 ml of TRIzol

Homogenization reagent using an Ultra Turrax tissue homogenizer, or cell pellet in
0.2 ml of TRIzol reagent using a Pellet Pestle Cordless Motor, and
incubate them at room temperature (20–30  C) for 2–3 min. Add
0.6 ml of chloroform or 60 μl in the case of cell pellet, and vortex
 Analysis of microRNA Microarrays in Cardiogenesis 209

samples vigorously for 15 s. Centrifuge the samples at 5,100  g

for 15 min at 4  C. Following centrifugation, the mixture separates
into lower Green phenol phase, an interphase, and a colorless upper
aqueous phase. RNA remains exclusively in the aqueous phase.
Transfer upper aqueous phase carefully without disturbing the
interphase into fresh tube.

2.1.2 RNA Precipitation Measure the volume of the aqueous phase and add isopropyl alco-
hol at 1:1 proportion. Incubate samples at room temperature
(20–30  C) for 10 min and centrifuge 6,300  g for 15 min at
4  C. Remove the supernatant completely. The RNA precipitate,
often invisible before centrifugation, forms a pellet on the side and
at bottom of the tube.

2.1.3 RNA Wash Wash the RNA pellet once adding 0.2 ml of filter-sterilized 75 %
ethanol and centrifuge at 5,500  g for 5 min at 4  C. Remove all
leftover ethanol. It is important avoid completely drying the RNA
pellet as this greatly will decrease its solubility. Redissolve RNA
pellet in 35–50 μl of Milli-Q water RNAse-free by passing solution
a few times through a pipette tip. Measure the samples in Nano-
Drop 2000c and keep it in the freezer (80  C) until further use.

2.1.4 DNAse Treatment 1. Prepare the following mixture:

Individual
Component reaction
Total RNA 10–50 μg
10 Incubation Buffer 5 μl
DNase I recombinant, RNase-free 2.5–10 units
Optionally: RNaseOUT Recombinant Ribonuclease 10 units
Inhibitor
Milli-Q water, RNase-free Up to 50 μl
Incubate at 25 to 37  C for 15–20 min.
2. Stop the reaction by adding 2 μl of 0.2 M EDTA (pH 8.0) to a
final concentration of 8 mM and heating to 75  C for 10 min.
The concentration of EDTA has to be taken into account for all
subsequent applications.
3. Add 50 μl of Phenol RNA (pH 4.7) and 50 μl of chloroform
and vortex samples vigorously for 15 s. Centrifuge the samples
at 16,000  g for 10 min at 4  C.
4. Following centrifugation, the mixture separates into a lower
phenol phase and an upper aqueous phase. RNA remains exclu-
sively in the aqueous phase. Transfer upper aqueous phase
carefully into fresh tube.
210 Diego Franco et al.

5. Measure the volume of the aqueous phase and add the

following reagents as stated below:
(a) 1/10 Volume of 3 M sodium acetate, pH 4.5.
(b) Two volumes of ice-cold absolute ethanol.
Mix and freeze (20  C) at least for 30 min.
6. Centrifuge the samples at 16,000  g for 10 min at 4  C.
Remove the supernatant completely. The RNA precipitate,
often invisible before centrifugation, forms a pellet on the
side and bottom of the tube.
7. Finally repeat step 2.1.3 (RNA wash).

2.2 Required 1. Milli-Q water, nuclease-free.

Reagents 2. TriPure Isolation Reagent (TRIzol) (Roche).
3. Chloroform.
4. Isopropyl alcohol.
5. 70 % ethanol.
6. DNase I recombinant, RNase-free (Roche).
7. RNaseOUT™ Recombinant Ribonuclease Inhibitor
(Invitrogen).
8. 0.2 M EDTA pH 8.0.
9. Phenol RNA extraction pH 4.7 (Sigma).
10. Ice-cold absolute ethanol.
11. Sodium acetate 3 M, pH 4.5.

2.3 Equipment 1. Gas extraction hood.

2. Powder-free gloves.
3. Refrigerated centrifuge
4. Micropipettes.
5. Vortex mixer.
6. Eppendorf 1.5 ml microcentrifuge tubes.
7. Ultra Turrax tissue homogenizer.

2.4 Protocol Tips The following precautions should be taken to prevent RNase con-
tamination and degradation of the RNA sample and reagents:
l Always use gloves.
l Use nuclease-free, low nucleic acid binding plasticware and
filter barrier pipette tips.
l Keep tubes capped whenever possible.
l Make sure your equipment and solution are RNAse-free.
 Analysis of microRNA Microarrays in Cardiogenesis 211

l Keep samples on ice as much as possible. For long-time storage,

RNA may be stored at 80  C. Avoid repeated freezing and
thawing cycles.
l Avoid contact with skin or clothing.
l Safety glasses are highly recommended.
l Avoid vapor breathing.
l Work in a chemical safety hood.

3 microRNA Microarray Platform Selection

As previously said, biological triplicates are needed in order to

perform appropriate microRNA microarray analyses, in analogy
with strategies of mRNA microarray studies. However, in contrast
to mRNA microarrays, in which, on average a single copy of each
gene is spotted into each chip, on the microRNA microarrays, on
average triplicates or quadruplicates of each microRNA is spotted
into each chip. Thus, this question has opened the possibility of
using single or duplicate replicas of each microRNA sets. In our
experience, we have used several strategies, starting from doing the
conventional approach, i.e., three biological experiments each of
the analyzed on distinct chips, to a more risky but much cheaper
approach, pooled biological samples (>3 per condition) into a
single quadruplicate containing chip per condition. Within the
first condition, 9 reads per microRNA were obtained while within
the second approach 4 reads per microRNA were generated. In
both cases, appropriate statistical analyses can be performed (>3
reads per condition). Since distinct biological questions were asked
on each experiment it is difficult to assess which of the approach was
most adequately developed. However, validation analyses of a rep-
resentative set of those microRNAs identified by microRNA micro-
array analyses revealed a rather similar validation rate (>80 %).
Thus, it is rather likely that running one single microarray per
condition leads to similarly robust differentially expressed micro-
RNA identification.
We have used two distinct microRNA microarray platforms to
dissect differentially expressed microRNAs in distinct biological
contexts, mainly cardiac and skeletal muscle development. On the
one hand we used N-code miRVana arrays (Life Technology) and on
the other hand Agilent arrays. Generating microRNA analyses with
mirVana arrays required that each biological sample is hybridized to
a single array and all arrays (6 in the case of a 2 conditions, 3 replica
analyses) are run in parallel. On the other hand, Agilent arrays
contained 8 arrays within a single glass and therefore hybridization,
probe clearance and signal scanning were always run simultaneously.
In our experience, both platforms were successfully used and
212 Diego Franco et al.

identification of differentially expressed microRNAs was achieved in

both cases. However, it is important to highlight that miRVana
arrays displayed larger data variations as compared to Aligent arrays,
yet miRVana arrays were more versatile for experimentation design.
Our data are in line within a recent report by Callari et al. (16) since
in several cases, discordant results are obtained within similar phys-
iopathological conditions, suggesting that a large variability on data
acquisition and analyses. These authors compared four distinct
microarray platforms (Agilent, Exiqon, Illumina, Miltenyi) within
the same biological context, colon cancer tissues. They found a poor
overlap among differentially expressed genes. Interestingly, those
differentially expressed microRNA with high concordant correla-
tion among distinct platform where equally validated by qRT-PCR.
Thus, these data suggest that independent of the selected platform,
validation by qRT-PCR is compulsory.
In any case, after choosing the appropriate platform and
experimental design, arrays hybridization, clearance and input
screening was externalized to a national microarray analyses plat-
form (Genoma España, Madrid) and more recently to a commercial
SME (Bioarrays, SL, Alicante). Hybridization conditions, clearance
and input screening were performed by specialized staff and raw
data were obtained. To our point of view this is rather convenient
for small to medium academic institutions in which acquiring and
using microarray analyses platform is rather unaffordable.

4 microRNA Data Analysis, Normalization and Representation

In microRNA profiling experiments, using microarray technology,

an adequate analysis has to be achieved in order to avoid incorrect
conclusions. As in mRNA gene expression microarray procedures,
the data analysis pipeline usually includes preprocessing, normali-
zation, parametric or nonparametric statistical analysis to detect
those microRNAs differentially expressed in our experimental
model, multivariate data exploration and gene enrichment func-
tional analyses. We also follow this well established pipeline in our
microRNA profiling studies in cardiogenesis, which were carried
out using free software as described below. In general, microRNA
microarray preprocessing and statistical analyses are performed call-
ing Bioconductor functions (bioconductor.org/) in R software
(r-project.org).
1. Once the raw data are obtained (after background correction,
which depends on platform), we usually impute those densi-
tometry values <1 using the KNN algorithm implemented in
the Bioconductor impute package (bioconductor.org/
packages/release/bioc/html/impute.html). Then, all data
are transformed to the logarithmic scale (log2).
 Analysis of microRNA Microarrays in Cardiogenesis 213

2. As described elsewhere (17), choosing an optimal normalization

method is a particular and important aspect to correct systematic
and technical (nonbiological) variability among arrays. Because
quantile normalization has been confirmed as one of the most
robust methods, data is then normalized using the quartiles
normalization function implemented in the Bioconductor
limma package (bioconductor.org/packages/release/bioc/
html/limma.html).
3. Statistically significant differences between groups are generally
identified using the t-test and multiple hypothesis correction
(false discovery rate, FDR) implemented in the multtest pack-
age (bioconductor.org/packages/release/bioc/html/mult-
test.html). Applying a Student’s t-test with a limited number
of samples (for example, four in each experimental group) is
inappropriate as the obtained statistical significance is not
robust; in this situation the mean and the standard deviation
could be easily biased by outliers. Thus, a nonparametric statis-
tical test can be used as a rough filter to narrow down the list of
most relevant microRNAs. The Rank Product method, imple-
mented in the RankProd package, (bioconductor.org/
packages/release/bioc/html/RankProd.html) has proven to
be superior, in our experience, to other statistical methods for
microarray data analysis. Moreover, this approach includes a
multiple hypothesis test, for raw p-value correction, to ascertain
a false positive rate similar to false discovery rate correction.
These, and others, statistical approaches are also implemented
in the user-friendly TM4 microarray software suite (tm4.org/).
4. Microarray data is generally represented using multivariate
procedures. Heat maps of hierarchical clustering, the unsuper-
vised way of grouping samples based only on their gene expres-
sion similarities, may be obtained using free software Cluster
and TreeView (rana.lbl.gov/EisenSoftware.htm) or the TM4
microarray software suite (tm4.org/); principal component
analysis graphs can be easily represent with this last platform.
5. When necessary, Venn diagrams can be easily drawn with
VENNY (bioinfogp.cnb.csic.es/tools/venny/).

5 microRNA Validation

microRNA arrays are powerful tools for studying the regulatory

mechanisms mediated by miRNAs during heart development and
their importance in heart diseases. Currently, miRNA array technology
is very useful in establishing broad patterns of miR genes expression
and in screening for differential gene expression during these pro-
cesses. In addition, miRNA studies can also facilitate the discovery of
biomarkers and disease signatures. However, array results can be
214 Diego Franco et al.

influenced by each step of the complex assay, from array manufacturing

to sample preparation (extraction, labeling, hybridization and image
analysis (18–20)). For that reason, validation of expression differences
must be accomplished with an alternate method to ensure a robust
data quality for publication. In this sense, real-time PCR is currently
the most accurate and reproducible approach to gene quantification
and, hence, for miRNA array validation.
In our lab we use the miRCURY LNA™ Universal RT micro-
RNA PCR system (Exiqon) for microRNA arrays validation. This is
a microRNA-specific, LNA-based system designed for sensitive and
accurate detection of microRNA by quantitative real-time PCR
using SYBR® Green. The method is based on universal reverse
transcription (RT) followed by real-time PCR amplification with
LNA enhanced primers.

5.1 First-Strand 1. Adjust each of the template RNA samples to a concentration of

cDNA Synthesis (RT) 5 ng/μl. Dilute template RNA using nuclease-free water.
2. Gently thaw the 5 Reaction buffer and nuclease-free water,
and immediately place on ice. Mix by vortexing. Immediately
before use, remove the Enzyme mix from the freezer, mix by
flicking the tubes and place on ice. Spin down all reagents.
3. If performing first-strand cDNA synthesis on multiple RNA
samples, it is recommended to prepare an RT working solution
of the 5 Reaction buffer, water and Enzyme mix as follow:

Reagent Volume (μl)

5 Reaction buffer 2
Nuclease-free water 5
Enzyme mix 1
Template total RNA (5 ng/μl) 2
Total volume 10

4. Mix the reaction by very gentle vortexing or pipetting to ensure

that all reagents are thoroughly mixed. After mixing, spin down.
5. Perform retrotranscription reaction following the next steps
(a) Incubate for 60 min at 42  C.
(b) Heat-inactivate the reverse transcriptase for 5 min at 95  C.
(c) Immediately cool to 4  C.
(d) Store at 4  C or freeze.
6. Immediately before use, dilute only the amount of cDNA
template needed for the planned real-time PCR reactions
80 in nuclease-free water (e.g., add 395 μl nuclease-free
water to each 5 μl of reaction. It is not recommended to store
the 1:80 dilution of cDNA.
 Analysis of microRNA Microarrays in Cardiogenesis 215

5.2 qPCR Protocol Although Exiqon has optimized qPCR experiments by using the
miRCURY LNA™ ExiLENT SYBR® Green master mix, our expe-
rience shows how LNA™ PCR primer sets can be used with others
SYBR® Green master mixes such as GoTaq® qPCR Master Mix
(Promega) with similar results. Whatever the master mix is used
proceed as follow:
1. Place cDNA (from previous step), nuclease-free water, and
PCR Master mix on ice and thaw for 15–20 min. Protect the
PCR Master mix vials from light. Immediately before use, mix
the PCR Master mix by pipetting up and down. The rest of the
reagents are mixed by vortexing and spun down.
2. When multiple real-time PCR reactions are performed with the
same microRNA primer set, it is recommended to prepare a
primer master mix working-solution of the PCR primers and
the PCR Master mix as follow:

Reagent Volume (μl)

PCR Master mix 5
PCR primer mix 1
Diluted cDNA template 4
Total volume 10

3. Mix the reaction by gentle pipetting to ensure that all reagents

are mixed thoroughly. After mixing cap tubes or strips, seal the
plate with optical sealing as recommended by the manufac-
turer. Spin down in a centrifuge (1,500  g for 1 min). The
experiment can be paused at this point. Store the reactions
protected from light at 4  C for up to 24 h.
4. Perform real-time PCR amplification followed by melting
curve analysis according to the user’s qPCR instrument; optical
read as well as melting curve analyses depends on qPCR instru-
ment used by the costumer. The relative level of expression of
each miR gene can be calculated through Livak analysis method
(21) by using 5S and 6U as internal controls.

6 Functional GO Analyses of Differentially Expressed microRNAs

microRNA microarray analyses often results in a long list of

microRNAs differentially upregulated or downregulated in experi-
mental versus control status or within different developmental
stages (22–27). As an example, we compared differential expression
profile of microRNAs in the developing ventricular chamber at
three distinct stages by microarrays which resulted in the
216 Diego Franco et al.

identification of 66/486 (~13 %) microRNAs upregulated 24/486

(~5 %) between E12.5 and E15.5 ventricular chambers, and 41/
486 (~8 %) microRNAs upregulated between E15.5 and E18.5
stages (27). Finding out the functional relevance of these differen-
tially expressed microRNAs is a triple challenge. On the one hand, a
single microRNA can target tens to hundreds distinct mRNAs
(TargetScan; Miranda on-line predictive software), on the other
hand, microRNAs normally decreased mRNA stability or protein
translation (28), but in other cases, it might exclusively alter protein
expression without modification of mRNA levels (29) or even
enhance mRNA/protein expression (29–32). Thirdly, most micro-
RNA microarray assays are performed using tissue homogenates,
such as our own case (27). Therefore correlation between micro-
RNA and mRNA expression levels is not enough to provide puta-
tive cause-effect relationships.
We have used different approaches to trim down the list of
putative microRNA–mRNA interaction in the cardiovascular con-
text as detailed below. Our first approach to untangle such gene
regulatory network was to collect of all putative gene targets of all
upregulated (107 microRNAs; ~60,000 putative target genes) and
downregulated (61 microRNAs; ~35,000 putative target genes)
microRNAs during ventricular development, respectively and do
gene ontology analyses. Estimates of putative target genes were
generated using TargetScan, yet other predictive on-line tools are
also available. We used TargetScan because a multi-species 30 UTR
comparative algorism is used, which we think is the most appropri-
ate method to date to minimize false positive. Gene ontology
analyses can be currently performed using multiple tools (www.
geneontology.org), such as Bioconductor (www.biocondutor.
org), DAVID (http://david.abcc.ncifcrf.gov), or GeneTools
(http://www.genetools.microarray.ntnu.no) We used DAVID to
search for putative signaling pathways involved in differentially
expressed microRNAs during ventricular development. Tgf-beta
and Mapk signaling were among the most statistically significant
GO terms identified. In line with these predictions, several micro-
RNAs displayed increasing expression levels during ventricular
development, such as miR-24, miR-125 and miR-143, have been
recently reported to regulate mapk signaling in different biological
contexts (33–36).
1. Identify all putative mRNA targets of all differentially expressed
microRNAs. Generate an Excel file.
2. Get on-line with DAVID (http://david.abcc.ncifcrf.gov).
3. Select functional annotation.
4. Upload the list of all putative mRNA targets.
5. Select the appropriate gene nomenclature identifier (if collected
from TargetScan it will be the Official Gene Symbol identifier).
 Analysis of microRNA Microarrays in Cardiogenesis 217

6. Run the application and select Functional Categories, Gene

Ontology Terms or Pathways.
7. Using Pathways and selecting thereafter KEGG pathways, it
provides a nice signaling flowchart in which relevant genes
from the input list are marked.
CALCIUM SIGNALING PATHWAY

Na+
NCX MCU NCX
Na+ Mitochondrion VDAC ANT
PMCA Cyp-D
ER/SR
+P
Neurotransmitter, GPCR Gs ADCY PKA PLN
autacoid
cAMP
TnC
ORAI STIM
Contraction
Depletion MLCK
of Ca2+ stores
PHK Metabolism
VOCs CaV1 SERCA
Membrane CaV2 MAPK
depolarization RYR signalng pathway
CaV3
Apoptosis
Neurotransmitter, CaN Proliferation
ROC CALM
autacoid Fertilization
Ca2+ CAMK Learning and memory
Neurotransmitter, Long term potentiation
hormone, GPCR Gq
autacoid
PLCδ NOS Long term depression
Growth factor PTK ER/SR
PLCβ
IP3R ADCY
PLCγ IP3
TCR PDE1
Antigen PLCε
BCR Other signaling pathways
cAMP
Sperm PLCζ FAK2
DAG
? ATP Phosphatidylinositol
IP3 3K signaling pathway
cADPR
CD38 NAADPR PKC
NAADP
NADH
? SPHK ? Exocytosis
S1P Secretion

04020 4/15/14

8. Identified targets are cross-checked back to provide micro-

RNA–mRNA putative interactions. (For example, PDE1 is
searched into TargetScan to provide the list of putative target-
ing microRNAs. miR-18, originally in our differentially
expressed input list is now identified).
9. Select putative target gene (e.g., PDE1 in calcium handling
signaling pathway).
10. Search putative microRNAs binding sites in Target Scan (all
putative miRNAs).
11. Identify putative microRNA–mRNA interaction (e.g., miR-18-
original list, miR-128 was not in the original list).
12. Design biochemical assays to provide direct functional evidence
of mRNA–microRNA interactions (e.g., luciferase-based assays
as reported by Chinchilla et al. (27).
While such an approach provides a general overview of the gene
regulatory networks that are putatively regulated by those differen-
tially expressed microRNAs, it is rather tedious to search for indi-
vidual microRNA-mRNA partnership and moreover large number
of false positive might arise since tissue mismatch (e.g., mRNA in
218 Diego Franco et al.

myocardial cell but microRNA in cardiac fibroblasts) is not

accounted for.
We therefore have envisioned novel in silico approaches to sort
out the limitations previously reported. On the one hand, a
literature-based approach in conjunction with curate analyses of
target mRNA expression databases has been envisaged. On the
other hand, microRNA microarray meta-analyses have been devel-
oped. In the following subheadings we provide background infor-
mation on the application range and their limitations.

7 Literature-Based Functional Analyses of Differentially Expressed microRNAs

Taken as a paradigm the list of upregulated and downregulated

microRNA within a cellular context, for instance cardiac develop-
ment, and taken GO functional analyses as a first step, a short list of
putative microRNAs are selected to target a discrete gene regu-
latory network. Thus, all components of these regulatory networks
are then scrutinized by PubMed search for previously identified
involvement within the cardiovascular development field. Genes
with a previous track record of involvement in cardiovascular
research are further carried on to scrutinize into tissue expression
databases such as Genepaint (www.genepaint-org). Double positive
genes are more likely to be bona fide target of differentially
expressed microRNAs within such a biological context. Our
proof-of-principle on this approach came from searching for novel
targeted by a triplet of microRNAs that were capable of interfering
with endocardial cushion EMT. Conjugation of common target
screening, followed by literature-based selection and gene expres-
sion database comparison trimmed down ~1,200 putative genes
into 14 candidate genes. Gain-of-function analyses within muscle
cells, provided evidence that 12/14 (85 %) were functionally vali-
dated (Bonet et al. in preparation).

8 Meta-analyses of Differentially Expressed microRNAs in the Cardiovascular

Context

An alternative way to delimit the bone fide microRNAs involved in

a discrete cellular pathway or morphogenetic event might come by
comparing our own results with those already published in similar
related conditions or processes within public database such as GEO
(http://www.ncbi.nlm.nih.gov/geo/) or ArrayExpress (http://
www.ebi.ac.uk/arrayexpress/). We have compared the microRNA
expression profile in different models of cardiogenesis: ventricular
chamber development (27), induced pluripotent stem cell (iPS)-
derived cardiomyocytes (GSE32935) from Gene Expression
 Analysis of microRNA Microarrays in Cardiogenesis 219

Omnibus (ncbi.nlm.nih.gov/geo/) and the aging heart

(GSE35672). Data were obtained from different platforms (Mir-
Vana v2.0–miRBase version 8.0, Illumina Human v2 MicroRNA
expression beadchip datasets, and Exiqon Mouse microRNA v11.0,
respectively). To achieve an appropriate comparison, for each
group:
1. Data is first independently log2 transformed.
2. An intra-normalization, using the quantiles function, is carried
out as described above. Such function was implemented in the
Bioconductor limma package (bioconductor.org) run in R
software (www.r-project.org).
3. Probes are Z-scored ððx i  x Þ=SDÞ.
4. As an unsupervised way of grouping samples, based only on
their gene expression similarities, a hierarchical clustering is
carried out. Hierarchical clustering (euclidean distance and
complete linkage), an unsupervised way of grouping samples
based only on their gene expression similarities, was carried out
using TM4 software suite (tm4.org).
Using such an approach we reported the value that microRNA
microarrays to understand the role of microRNAs in cardiovascular
development (14). Meta-analyses comparing microRNA signature
of the developing ventricular chambers with two additional condi-
tions revealed similar microRNA signatures in the developing car-
diac chambers and the differentiating and maturing cardiomyocytes
derived from induced pluripotent stem cells (17/35; 48 %) which
are not altered in the adult and aging heart (0/10). Furthermore,
such a proof-of-principle microRNA microarray meta-analyses pro-
vide also novel hints, as decoding a subset of microRNAs that
behave in opposite pattern during in vitro (iPS-derived) and
in vivo (chamber maturation) cardiogenesis (7/35; 20 %), opening
new avenues to dissect the functional role of these microRNAs in
the cardiovascular setting. Moreover, future meta-analyses studies,
including not only healthy conditions, but emerging microRNA
signatures of diseased status, such as atrial fibrillation, cardiac
hypertrophy or ischemia will certainly increase our understanding
of microRNA biology in the normal and diseased heart. In addi-
tion, microRNA microarrays also provided novel insights into the
intricate biology of microRNA transcriptional regulation and puta-
tive target recognition, broadening thus the spectrum of their
applicability.
220 Diego Franco et al.

References
1. Kelly RG, Buckingham ME (2002) The 13. Espinoza-Lewis RA, Wang DZ (2012)
anterior heart-forming field: voyage to the MicroRNAs in heart development. Curr Top
arterial pole of the heart. Trends Genet Dev Biol 100:279–317
18:210–216 14. Bonet F, Hernandez-Torres F, Esteban FJ,
2. Moorman AF, Christoffels VM, Anderson RH, Aranega A, Franco D (2013) Comparative ana-
van den Hoff MJ (2007) The heart-forming lyses of microRNA microarrays during cardio-
fields: one or multiple? Philos Trans R Soc genesis: functional perspectives. Microarrays
Lond B Biol Sci 362:1257–1265 2:81–96. doi:10.3390/microarrays2020081
3. Kelly RG (2012) The second heart field. Curr 15. Bonet F, Hernandez-Torres F, Franco D
Top Dev Biol 100:33–65 (2014) Towards the therapeutic usage of
4. López-Sánchez C, Garcı́a-Martı́nez V (2011) microRNAs in cardiac disease and regenera-
Molecular determinants of cardiac specifica- tion. Exp Clin Cardiol 20:720–756
tion. Cardiovasc Res 91:185–195 16. Callari M, Dugo M, Musella V, Marchesi E,
5. de Castro Mdel P, Acosta L, Domı́nguez JN, Chiorino G, Grand MM, Pierotti MA, Dai-
Aránega A, Franco D (2003) Molecular diver- done MG, Canevari S, De Cecco L (2012)
sity of the developing and adult myocardium: Comparison of microarray platforms for mea-
implications for tissue targeting. Curr Drug suring differential microRNA expression in
Targets Cardiovasc Haematol Disord paired normal/cancer colon tissues. PLoS
3:227–239 One 7(9):e45105
6. Campione M, Ros MA, Icardo JM, Piedra E, 17. Meyer SU, Pfaffl MW, Ulbrich SE (2010) Nor-
Christoffels VM, Schweickert A, Blum M, malization strategies for microRNA profiling
Franco D, Moorman AF (2001) Pitx2 expres- experiments: a “normal” way to a hidden layer
sion defines a left cardiac lineage of cells: evi- of complexity? Biotechnol Lett 32
dence for atrial and ventricular molecular (12):1777–1788
isomerism in the iv/iv mice. Dev Biol 231 18. Der SD, Zhou A, Williams BR, Silverman RH
(1):252–264 (1998) Identification of genes differentially
7. Franco D, Campione M, Kelly R, Zammit PS, regulated by interferon alpha, beta, or gamma
Buckingham M, Lamers WH, Moorman AF using oligonucleotide arrays. Proc Natl Acad
(2000) Multiple transcriptional domains, with Sci U S A 95:15623–15628
distinct left and right components, in the atrial 19. Eisen M, Brown P (1999) DNA arrays for
chambers of the developing heart. Circ Res 87 analysis of gene expression. Meth Enzymol
(11):984–991 303:179–205
8. Franco D, Lamers WH, Moorman AF (1998) 20. Winzeler EA, Schena M, Davis RW (1999)
Patterns of expression in the developing myo- Fluorescence-based expression monitoring
cardium: towards a morphologically integrated using microarrays. Meth Enzymol 306:3–18
transcriptional model. Cardiovasc Res 21. Livak K, Schmittgen T (2001) Analysis of rela-
38:25–53 tive gene expression data using real-time quan-
9. Chinchilla A, Franco D (2006) Regulatory titative PCR and the 2ΔΔCT method.
mechanisms of cardiac development and repair. Methods 25:402–408
Cardiovasc Hematol Disord Drug Targets 22. Cheng Y, Ji R, Yue J, Yang J, Liu X, Chen H,
6:101–112 Dean DB, Zhang C (2007) MicroRNAs are
10. Franco D, Chinchilla A, Aránega AE (2012) aberrantly expressed in hypertrophic heart: do
Transgenic insights linking pitx2 and atrial they play a role in cardiac hypertrophy? Am J
arrhythmias. Front Physiol 3:206 Pathol 170(6):1831–1840
11. Zhao Y, Ransom JF, Li A, Vedantham V, von 23. Wang Y, Weng T, Gou D, Chen Z, Chintagari
Drehle M, Muth AN, Tsuchihashi T, McManus NR, Liu L (2007) Identification of rat lung-
MT, Schwartz RJ, Srivastava D (2007) Dysre- specific microRNAs by microRNA microarray:
gulation of cardiogenesis, cardiac conduction, valuable discoveries for the facilitation of lung
and cell cycle in mice lacking miRNA-1-2. Cell research. BMC Genomics 8:29
129:303–317 24. Wang J, Xu R, Lin F, Zhang S, Zhang G, Hu S,
12. Fish JE, Santoro MM, Morton SU, Yu S, Yeh Zheng Z (2009) MicroRNA: novel regulators
RF, Wythe JD, Ivey KN, Bruneau BG, Stainier involved in the remodeling and reverse remo-
DY, Srivastava D (2008) miR-126 regulates deling of the heart. Cardiology 113(2):81–88
angiogenic signaling and vascular integrity. 25. Matkovich SJ, Van Booven DJ, Youker KA,
Dev Cell 15:272–284 Torre-Amione G, Diwan A, Eschenbacher
 Analysis of microRNA Microarrays in Cardiogenesis 221

WH, Dorn LE, Watson MA, Margulies KB, 31. Letonqueze O, Lee J, Vasudevan S (2012)
Dorn GW 2nd (2009) Reciprocal regulation MicroRNA-mediated posttranscriptional
of myocardial microRNAs and messenger mechanisms of gene expression in proliferating
RNA in human cardiomyopathy and reversal and quiescent cancer cells. RNA Biol 9
of the microRNA signature by biomechanical (6):871–880
support. Circulation 119(9):1263–1271 32. Lee S, Vasudevan S (2013) Post-transcriptional
26. Naga Prasad SV, Duan ZH, Gupta MK, Sur- stimulation of gene expression by microRNAs.
ampudi VS, Volinia S, Calin GA, Liu CG, Kot- Adv Exp Med Biol 768:97–126
wal A, Moravec CS, Starling RC, Perez DM, 33. Papadopoulos GL, Reczko M, Simossis VA,
Sen S, Wu Q, Plow EF, Croce CM, Karnik S Sethupathy P, Hatzigeorgiou AG (2009) The
(2009) Unique microRNA profile in end-stage database of experimentally supported targets: a
heart failure indicates alterations in specific car- functional update of TarBase. Nucleic Acids
diovascular signaling networks. J Biol Chem Res 37:D155–D158
284(40):27487–27499 34. Fiedler J, Jazbutyte V, Kirchmaier BC,
27. Chinchilla A, Lozano E, Daimi H, Esteban FJ, Gupta SK, Lorenzen J, Hartmann D,
Crist C, Aranega AE, Franco D (2011) Micro- Galuppo P, Kneitz S, Pena JT, Sohn-Lee C
RNA profiling during mouse ventricular matu- et al (2011) MicroRNA-24 regulates vascular-
ration: a role for miR-27 modulating Mef2c ity after myocardial infarction. Circulation
expression. Cardiovasc Res 89(1):98–108 124:720–730
28. Condorelli G, Latronico MV (2014) Cavar- 35. Mayorga ME, Penn MS (2012) miR-145 is
retta E. microRNAs in cardiovascular diseases: differentially regulated by TGF-β1 and ischae-
current knowledge and the road ahead. J Am mia and targets disabled-2 expression and wnt/
Coll Cardiol 63(21):2177–2187 β-catenin activity. J Cell Mol Med
29. Vasudevan S (2012) Posttranscriptional upre- 16:1106–1113
gulation by microRNAs. Wiley Interdiscip Rev 36. Li DF, Tian J, Guo X, Huang LM, Xu Y, Wang
RNA 3(3):311–330 CC, Wang JF, Ren AJ, Yuan WJ, Lin L (2013)
30. Steitz JA, Vasudevan S (2009) miRNPs: versa- Induction of microRNA-24 by HIF-1 protects
tile regulators of gene expression in vertebrate against ischemic injury in rat cardiomyocytes.
cells. Biochem Soc Trans 37(Pt 5):931–935 Physiol Res 61:555–565
Methods in Molecular Biology (2016) 1375: 223
DOI 10.1007/7651_2015_256
© Springer Science+Business Media New York 2015
Published online: 27 November 2015

Erratum to: Classification and Clustering on Microarray

Data for Gene Functional Prediction Using R
Liliana López Kleine, Rosa Montaño, and Francisco Torres-Avilés

Erratum to: Methods in Molecular Biology

DOI 10.1007/7651_2015_240

There is an error in given name and family name of the author Liliana López Kleine.
The correct name should read as Liliana López-Kleine (given name: Liliana and family
name: López-Kleine)

The online version of the original chapter can be found under

http://dx.doi.org/10.1007/7651_2015_240

223
Methods in Molecular Biology (2016) 1375: 225–226
DOI 10.1007/7651_2015
© Springer Science+Business Media New York 2016

INDEX

A G
Affymetrix ......................... 1–9, 183, 184, 185, 187, 189, Gene expression .........................................................2, 11,
190, 197, 198, 202 12, 16, 17, 29, 30, 31, 41, 43, 52, 53, 55–72, 91,
Amplex red ..........................................170, 173, 176, 178 92, 93, 94, 95, 96, 99, 100, 101, 137, 138, 140,
Assessment................ 128, 138, 147, 155–165, 197, 208 141, 143, 144, 146, 147, 148, 149, 150, 151,
152, 156, 159, 160, 163, 182, 183, 189, 195,
B 200, 203, 204, 212, 213, 218, 219
Biclustering................................. 18, 55–72, 91–101, 159 Gene ontology......................................................... 17, 20,
Bioinformatics .................................................... 13, 15, 18 21, 29, 95, 97, 106, 110, 113, 118, 119, 125,
163, 200, 216, 217
Biological network inference ............................... 155–165
Gene regulatory networks (GRNs)..............96, 137–153,
C 155, 156, 157, 158, 159, 160, 161, 162, 163,
164, 216, 217, 218
Cancer samples .............................................1–9, 111, 112 Gene target ................................................ 11, 13, 19, 216
Cardiac development ........................................... 207, 218 Glioblastoma ........................................................ 195–205
Classification..................................................... 41–54, 146 Gold standard...................................................... 156, 160,
Cloud computing.............. 26, 27, 29, 30, 31, 32, 33, 35 161, 162, 163, 164, 165
Clustering ................................18, 21, 27, 35, 41–53, 56, Graph theory ............................................... 140, 155, 156
57, 59, 64, 70, 79, 96, 97, 98, 99, 108, 111, 112,
113, 114, 115, 159, 187, 190, 195, 199, 203, H
213, 219
Correlation ................ 19, 29, 56, 57, 58, 59, 62, 63, 64, Hierarchical clustering .................. 44, 46, 111, 114, 190,
213, 219
65, 66, 68, 70, 71, 100, 124, 125, 126, 129, 131,
133, 134, 140, 159, 160, 187, 212, 216
L
D Lactate .......................................126, 170, 171, 172, 173,
Data analysis .................................. 11–22, 25–38, 41, 57, 176, 177, 178
64, 95, 96, 97, 141, 147, 148, 195, 199,
M
212–213
Databases ..............................................17, 18, 19, 56, 77, Measure ...................................................... 19, 59, 62, 63,
78, 84, 109, 110, 113, 114, 118, 127, 146, 200, 66, 70, 71, 99, 100, 101, 105, 106, 107, 108,
202, 218 114, 120, 128, 131, 137, 161, 162, 163, 172,
Data mining................................78, 83, 87, 95, 101, 144 176, 177, 193, 209, 210
Dissection .......................... 169, 171, 172, 174, 175, 177 Meta-analyses ...............................................208, 218–219
Metabolomics ...............................................124, 169–178
E MetaMirClust ............................................................75–88
Euthanasia ................................................... 170, 171, 172 Microarray data analysis ..............................12–13, 25–38,
105–115, 141, 212–213
Expression microarray.......................................... 108, 212
Expression patterns ........................................92, 161, 191 Microarray profiling ............................................. 208, 212
Microarrays ..................................................... 1–9, 12–13,
F 15, 20, 21, 25–38, 41–54, 55, 56, 57–59, 64, 66,
70, 71, 91, 92, 93, 96, 105–115, 117–120, 126,
Functional pathways .............................................. 21, 162 137–153, 155–165, 182, 183, 184, 185, 191,
Functional prediction................................................41–54 195–205, 207–219

225
226 M ICROARRAY DATA ANALYSIS: METHODS
Index
AND APPLICATIONS

microRNA ............................. 1–9, 11–22, 32, 34, 75–88, Prediction ...................13, 17, 18, 19, 20, 22, 29, 41–54,
181–193, 207–213, 214, 219 128, 129, 133, 138, 140, 144, 160, 193, 203, 216
cluster................................................................... 75–88 Proteome .............................................110, 125, 126, 198
inhibition ....................................................... 183, 184,
188, 191 S
microarray................................................................ 1–9 Semantic similarity ............................ 106, 107, 108, 114,
replacement .................................................... 183, 191 115, 118, 119–120
Multivariate data analysis ...............................43, 212, 213 Sequence features .............. 125, 126, 131, 132, 133, 134
Skeletal muscle .............................................169–178, 211
N Synteny ......................................................................77, 78
Next-generation sequencing (NGS) ................ 12, 13–15, Systems biology................. 117, 137, 138, 152, 155, 164
16, 17, 18, 20, 21, 25, 28, 31, 32, 78, 84
Normalization .................................................1–9, 13, 17,
T
31, 43, 60, 61, 126, 190, 200, 202, 208, Transcriptome ......................................34, 162, 185, 186,
212–213, 219 187, 189, 190, 191, 199, 200
Transfection.................................. 76, 184, 186, 188, 192
O Translation......................................... 124, 125, 130, 132,
Ontologies .............. 20, 21, 29, 106, 107, 118, 119, 165 133, 134, 146, 182, 216

P U
Pathways ....................................................................1, 20, Undetected proteins ................................... 124, 126, 132
21, 56, 75, 77, 91, 105, 123, 127, 128, 129,
Z
138, 139, 140, 146, 149, 150, 151, 162, 182,
191, 199, 200, 204, 205, 207, 216, 217, 218 Zero-inflated Poisson regression.................................. 127