RNA Bioinformatics

Methods in
Molecular Biology 1269
Ernesto Picardi Editor
RNA
Bioinformatics
METHODS IN M O L E C U L A R B I O LO G Y
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
RNA Bioinformatics
Edited by
Ernesto Picardi
Department of Biosciences, Biotechnology and Biopharmaceutics,
University of Bari, Bari, Italy
Editor
Ernesto Picardi
Department of Biosciences
Biotechnology and Biopharmaceutics
University of Bari
Bari, Italy
Institute of Biomembranes and Bioenergetics
National Research Council
Bari, Italy
National Institute of Biostructures and Biosystems (INBB)
Rome, Italy
ISSN 1064-3745 ISSN 1940-6029 (electronic)

ISBN 978-1-4939-2290-1 ISBN 978-1-4939-2291-8 (eBook)
DOI 10.1007/978-1-4939-2291-8
Springer New York Heidelberg Dordrecht London
Library of Congress Control Number: 2014958476
© Springer Science+Business Media New York 2015

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. Exempted from this
legal reservation are brief excerpts in connection with reviews or scholarly analysis or material supplied specifically for
the purpose of being entered and executed on a computer system, for exclusive use by the purchaser of the work.
Duplication of this publication or parts thereof is permitted only under the provisions of the Copyright Law of the
Publisher’s location, in its current version, and permission for use must always be obtained from Springer. Permissions
for use may be obtained through RightsLink at the Copyright Clearance Center. Violations are liable to prosecution
under the respective Copyright Law.
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.
While the advice and information in this book are believed to be true and accurate at the date of publication, neither
the authors nor the editors nor the publisher can accept any legal responsibility for any errors or omissions that may be
made. The publisher makes no warranty, express or implied, with respect to the material contained herein.
Printed on acid-free paper
Humana Press is a brand of Springer

Springer is part of Springer Science+Business Media (www.springer.com)
Preface
RNA is a versatile nucleic acid polymer with a structure analogous to single-stranded DNA
even though it has a backbone composed of ribose sugar and the organic base thymine (T)
is replaced by uracil (U). In contrast to DNA, RNA molecules are less stable and character-
ized by secondary and tertiary structures, which underline their different function.
According to the central dogma of molecular biology, RNA is directly involved in the flux
of genetic information from the DNA to the proteins. However, recent developments in
molecular biology indicate that RNA molecules have a plethora of functional roles and are
indispensable for living organisms and cell homeostasis. Indeed, on the basis of their func-
tions, RNA molecules can be divided into two groups, coding and noncoding. While the
coding fraction is represented by messenger RNAs (mRNAs), noncoding RNAs (ncRNAs)
include at least two main classes: (1) structural ncRNAs as transfer RNAs (tRNAs), ribo-
somal RNAs (rRNAs), and small nucleolar RNAs (snoRNAs); (2) regulatory ncRNAs as
micro RNAs (miRNAs), piwi-interacting RNAs (piwiRNAs), and long noncoding RNAs
(lncRNAs).
A large fraction of eukaryotic transcriptomes consists of ncRNAs that play relevant
biological roles as the regulation of gene expression in normal as well as pathological condi-
tions. NcRNA functions are generally mediated by interactions with other RNA molecules
or DNA regions or proteins. In all cases, RNA secondary and tertiary structures are basic
for a correct function and interaction. However, the characterization of RNA molecules is
a challenging task and, thus, during past decades a variety of bioinformatics tools have been
developed. Nowadays, RNA bioinformatics represents one of the most active fields of bio-
informatics and computational biology.
The interest towards RNA bioinformatics has increased rapidly thanks to the advent of
recent high-throughput sequencing technologies that enable the investigation of complete
transcriptomes at single nucleotide resolution.
The present book has been conceived with the aim of providing an overview of RNA
bioinformatics methodologies, starting from “classical” technologies to predict secondary
and tertiary structures, to novel strategies and algorithms based on massive RNA sequenc-
ing. Indeed, the content of the book is organized as follows:
– Part I—RNA secondary and tertiary structures. This section includes chapters devoted
to main computational algorithms to predict, draw, and edit secondary and tertiary
RNA structures or infer RNA-RNA interactions.
– Part II—Analysis of high-throughput RNA sequencing data. The aim of this section
is to provide a global overview of current methodologies to handle and analyze large
sequencing dataset generated by next-generation sequencing (NGS) technologies.
Indeed, the section includes chapters about quality control of RNA sequencing data
or the mapping of RNA-Seq reads on complete reference genomes or the gene
expression profiling. In addition, methodologies to investigate posttranscriptional
events as alternative splicing or RNA editing and entire meta-transcriptomes are also
described in detail.
v
vi Preface
– Part III—Web resources for RNA data analysis. Finally, this section provides chapters
about several available web resources to work with RNA data without specific computer
requirements (hardware and software) or specialized bioinformatics skills.
Hoping that the book content meets the reader expectations, I would like to acknowl-
edge those who helped make this book possible: all chapter authors for their work and
excellent contributions; the Series Editor, John Walker, for his constant support and sug-
gestions; my wife Angela and daughter Adele for their patience and encouragement.
A special thanks is addressed to my mentor Graziano Pesole for his always indispensable
paternal suggestions.
Finally, I would to dedicate this effort to my parents since they have always believed in
me and to the memory of my first supervisor Carla Quagliariello who introduced me for the
first time to the wonderful and fascinating world of RNA bioinformatics.
Bari, Italy Ernesto Picardi

Contents
Preface. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
Contributors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
PART I RNA SECONDARY AND TERTIARY STRUCTURES

1 Free Energy Minimization to Predict RNA Secondary Structures
and Computational RNA Design. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Alexander Churkin, Lina Weinbrand, and Danny Barash
2 RNA Secondary Structure Prediction from Multi-aligned Sequences . . . . . . . . 17
Michiaki Hamada
3 A Simple Protocol for the Inference of RNA Global Pairwise Alignments . . . . 39
Eugenio Mattei, Manuela Helmer-Citterich, and Fabrizio Ferrè
4 De Novo Secondary Structure Motif Discovery Using RNAProfile . . . . . . . . . 49
Federico Zambelli and Giulio Pavesi
5 Drawing and Editing the Secondary Structure(s) of RNA . . . . . . . . . . . . . . . . 63
Yann Ponty and Fabrice Leclerc
6 Modeling and Predicting RNA Three-Dimensional Structures. . . . . . . . . . . . . 101
Jérôme Waldispühl and Vladimir Reinharz
7 Fast Prediction of RNA–RNA Interaction Using Heuristic Algorithm . . . . . . . 123
Soheila Montaseri
PART II ANALYSIS OF HIGH-THROUGHPUT RNA SEQUENCING DATA

8 Quality Control of RNA-Seq Experiments. . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
Xing Li, Asha Nair, Shengqin Wang, and Liguo Wang
9 Accurate Mapping of RNA-Seq Data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
Kin Fai Au
10 Quantifying Entire Transcriptomes by Aligned RNA-Seq Data . . . . . . . . . . . . 163
Raffaele A. Calogero and Francesca Zolezzi
11 Transcriptome Assembly and Alternative Splicing Analysis . . . . . . . . . . . . . . . . 173
Paola Bonizzoni, Gianluca Della Vedova, Graziano Pesole,
Ernesto Picardi, Yuri Pirola, and Raffaella Rizzi
12 Detection of Post-Transcriptional RNA Editing Events . . . . . . . . . . . . . . . . . . 189
Ernesto Picardi, Anna Maria D’Erchia, Angela Gallo,
and Graziano Pesole
13 Prediction of miRNA Targets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
Anastasis Oulas, Nestoras Karathanasis, Annita Louloupi,
Georgios A. Pavlopoulos, Panayiota Poirazi, Kriton Kalantidis,
and Ioannis Iliopoulos
vii
viii Contents
14 Using Deep Sequencing Data for Identification of Editing Sites

in Mature miRNAs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
Shahar Alon and Eli Eisenberg
15 NGS-Trex: An Automatic Analysis Workflow for RNA-Seq Data . . . . . . . . . . . 243
Ilenia Boria, Lara Boatti, Igor Saggese, and Flavio Mignone
16 e-DNA Meta-Barcoding: From NGS Raw Data to Taxonomic Profiling . . . . . 257
Fosso Bruno, Marzano Marinella, and Monica Santamaria
17 Deciphering Metatranscriptomic Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
Evguenia Kopylova, Laurent Noé, Corinne Da Silva, Jean-Frédéric Berthelot,
Adriana Alberti, Jean-Marc Aury, and Hélène Touzet
18 RIP-Seq Data Analysis to Determine RNA–Protein Associations . . . . . . . . . . . 293
PART III WEB RESOURCES FOR RNA DATA ANALYSIS

19 The ViennaRNA Web Services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
Andreas R. Gruber, Stephan H. Bernhart, and Ronny Lorenz
20 Exploring the RNA Editing Potential of RNA-Seq Data by ExpEdit . . . . . . . . 327
Mattia D’Antonio, Ernesto Picardi, Tiziana Castrignanò,
Anna Maria D’Erchia, and Graziano Pesole
21 A Guideline for the Annotation of UTR Regulatory Elements
in the UTRsite Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 339
Matteo Giulietti, Giorgio Grillo, Sabino Liuni, and Graziano Pesole
22 Rfam: Annotating Families of Non-Coding RNA Sequences . . . . . . . . . . . . . . 349
Jennifer Daub, Ruth Y. Eberhardt, John G. Tate, and Sarah W. Burge
23 ASPicDB: A Database Web Tool for Alternative Splicing Analysis . . . . . . . . . . 365
Mattia D’Antonio, Tiziana Castrgnanò, Matteo Pallocca,
Anna Maria D’Erchia, Ernesto Picardi, and Graziano Pesole
24 Analysis of Alternative Splicing Events in Custom Gene Datasets
by AStalavista . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
Sylvain Foissac and Michael Sammeth
25 Computational Design of Artificial RNA Molecules for Gene Regulation. . . . . 393
Alessandro Laganà, Dario Veneziano, Francesco Russo,
Alfredo Pulvirenti, Rosalba Giugno, Carlo Maria Croce,
and Alfredo Ferro
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413
Contributors
ADRIANA ALBERTI • Genoscope—National Sequencing Center, Evry, France

SHAHAR ALON • Department of Neurobiology, George S. Wise Faculty of Life Sciences,
Tel-Aviv University, Tel-Aviv, Israel; Sagol School of Neuroscience, Tel-Aviv University,
Tel-Aviv, Israel
KIN FAI AU • Department of Internal Medicine, University of Iowa, Iowa City, IA, USA;
Department of Biostatistics, University of Iowa, Iowa City, IA, USA
JEAN-MARC AURY • Genoscope—National Sequencing Center, Evry, France
DANNY BARASH • Department of Computer Science, Ben-Gurion University, Been-Sheva,
Israel
STEPHAN H. BERNHART • Department of Bioinformatics, University of Leipzig, Leipzig,
Germany
JEAN-FRÉDÉRIC BERTHELOT • Inria Lille Nord-Europe, Villeneuve d’Ascq, France
LARA BOATTI • Department of Sciences and Technological Innovation (DiSIT),
University of Piemonte Orientale “A. Avogadro”, Alessandria, Italy
PAOLA BONIZZONI • Department of Informatics, Systems and Communication (DISCo),
University of Milano-Bicocca, Milano, Italy
ILENIA BORIA • Department of Chemistry, University of Milano, Milano, Italy
FOSSO BRUNO • Department of Biosciences, Biotechnology and Biopharmaceutics,
University of Bari, Bari, Italy
SARAH W. BURGE • Wellcome Trust Sanger Institute, Cambridgeshire, UK; European
Molecular Biology Laboratory, European Bioinformatics Institute, Cambridgeshire, UK
RAFFAELE A. CALOGERO • Department of Molecular Biotechnology and Health Sciences,
University of Torino, Torino, Italy
TIZIANA CASTRIGNANÒ • Consorzio Interuniversitario CINECA, Rome, Italy
ALEXANDER CHURKIN • Department of Computer Science, Ben-Gurion University,
Beer-Sheva, Israel
CARLO MARIA CROCE • Department of Molecular Virology, Immunology and Medical
Genetics, Comprehensive Cancer Center, The Ohio State University, Columbus, OH, USA
MATTIA D’ANTONIO • Consorzio Interuniversitario CINECA, Rome, Italy
ANNA MARIA D’ERCHIA • Department of Biosciences, Biotechnology and Biopharmaceutics,
University of Bari, Bari, Italy; Institute of Biomembranes and Bioenergetics, National
Research Council, Bari, Italy
JENNIFER DAUB • Wellcome Trust Sanger Institute, Cambridgeshire, UK; European
RUTH Y. EBERHARDT • Wellcome Trust Sanger Institute, Cambridgeshire, UK; European
ELI EISENBERG • Raymond and Beverly Sackler School of Physics and Astronomy,
Tel-Aviv University, Tel-Aviv, Israel; Sagol School of Neuroscience, Tel-Aviv University,
Tel-Aviv, Israel
FABRIZIO FERRÈ • Center for Molecular Bioinformatics (CBM), Department of Biology,
University of Rome Tor Vergata, Rome, Italy
ix
x Contributors
ALFREDO FERRO • Department of Clinical and Molecular Biomedicine, University

of Catania, Catania, Italy
SYLVAIN FOISSAC • UMR1388 GenPhySE, French National Institute
for Agricultural Research (INRA), Castanet Tolosan, France
ANGELA GALLO • RNA Editing Laboratory, Oncohaematology Department,
IRCCS Ospedale Pediatrico “Bambino Gesù”, Rome, Italy
ROSALBA GIUGNO • Department of Clinical and Molecular Biomedicine, University
of Catania, Catania, Italy
MATTEO GIULIETTI • Institute of Biomembranes and Bioenergetics, National Research
Council, Bari, Italy
GIORGIO GRILLO • Institute of Biomedical Technologies, National Research Council,
Bari, Italy
ANDREAS R. GRUBER • Biozentrum, University of Basel, Basel, Switzerland; Swiss Institute
of Bioinformatics, Basel, Switzerland
MICHIAKI HAMADA • Faculty of Science and Engineering, Waseda University, Tokyo, Japan;
Computational Biology Research Center, National Institute of Advanced Industrial
Science and Technology (AIST), Tokyo, Japan
MANUELA HELMER-CITTERICH • Center for Molecular Bioinformatics (CBM),
Department of Biology, University of Rome Tor Vergata, Rome, Italy
IOANNIS ILIOPOULOS • Division of Basic Sciences, University of Crete Medical School,
Heraklion, Greece
KRITON KALANTIDIS • Institute of Molecular Biology and Biotechnology, FORTH, Heraklion,
Greece; Department of Biology, University of Crete, Heraklion, Greece
NESTORAS KARATHANASIS • Institute of Molecular Biology and Biotechnology, FORTH,
Heraklion, Greece; Department of Biology, University of Crete, Heraklion, Greece
EVGUENIA KOPYLOVA • Laboratoire d’Informatique Fondamentale de Lille (LIFL), UMR
CNRS 8022, Villeneuve d’Ascq Cédex, France; Inria Lille Nord-Europe, Villeneuve
d’Ascq Cédex, France
ALESSANDRO LAGANÀ • Department of Molecular Virology, Immunology and Medical
Genetics, Comprehensive Cancer Center, The Ohio State University, Columbus, OH, USA
FABRICE LECLERC • Institut de G’en’etique et Microbiologie (IGM) UMR 8621,
Universit’e Paris Sud, Orsay, France
XING LI • Division of Biomedical Statistics and Informatics, Department of Health Sciences
Research, Mayo Clinic, Minneapolis, MN, USA
SABINO LIUNI • Institute of Biomedical Technologies, National Research Council, Bari, Italy
RONNY LORENZ • Institute for Theoretical Chemistry, University of Vienna, Vienna, Austria
ANNITA LOULOUPI • Biomedical Science-Medical Biology, University of Amsterdam,
Amsterdam, The Netherlands
MARZANO MARINELLA • Institute of Biomembranes and Bioenergetics, CNR, Bari, Italy
EUGENIO MATTEI • Center for Molecular Bioinformatics (CBM), Department of Biology,
University of Rome Tor Vergata, Rome, Italy
FLAVIO MIGNONE • Department of Sciences and Technological Innovation (DiSIT),
University of Piemonte Orientale “A. Avogadro”, Alessandria, Italy
SOHEILA MONTASERI • Department of Mathematics, Statistics and Computer Science,
University of Tehran, Tehran, Iran
ASHA NAIR • Division of Biomedical Statistics and Informatics, Department of Health
Sciences Research, Mayo Clinic, Minneapolis, MN, USA
Contributors xi
LAURENT NOÉ • Laboratoire d’Informatique Fondamentale de Lille (LIFL), UMR CNRS

8022, Villeneuve d’Ascq Cédex, France; Inria Lille Nord-Europe, Villeneuve d’Ascq Cédex,
France
ANASTASIS OULAS • Institute of Marine Biology, Biotechnology and Aquaculture, HCMR,
Heraklion, Greece
MATTEO PALLOCCA • Translational Oncogenomics Unit, Italian National Cancer Institute
“Regina Elena”, Rome, Italy
GIULIO PAVESI • Dipartimento di Bioscienze, Università di Milano, Milano, Italy
GEORGIOS A. PAVLOPOULOS • Division of Basic Sciences, University of Crete Medical School,
Heraklion, Greece
GRAZIANO PESOLE • Department of Biosciences, Biotechnology and Biopharmaceutics,
Research Council, Bari, Italy; National Institute of Biostructures and Biosystems (INBB),
Rome, Italy
ERNESTO PICARDI • Department of Biosciences, Biotechnology and Biopharmaceutics,
Research Council, Bari, Italy; National Institute of Biostructures and Biosystems (INBB),
Rome, Italy
YURI PIROLA • Department of Informatics, Systems and Communication (DISCo),
PANAYIOTA POIRAZI • Institute of Molecular Biology and Biotechnology, FORTH,
Heraklion, Greece
YANN PONTY • Laboratoire d’Informatique de ‘X (LIX) UMR 7161, CNRS and INRIA
AMIB, Ecole Polytechnique, Palaiseau, France
ALFREDO PULVIRENTI • Department of Clinical and Molecular Biomedicine, University of
Catania, Catania, Italy
VLADIMIR REINHARZ • School of Computer Science, McGill University, Montreal, QC, Canada
RAFFAELLA RIZZI • Department of Informatics, Systems and Communication (DISCo),
FRANCESCO RUSSO • Department of Clinical and Molecular Biomedicine, University
of Catania, Catania, Italy; Laboratory for Integrative System Medicine (LISM),
Institute of Informatics and Telematics (IIT) and Institute of Clinical Physiology (IFC),
National Research Council (CNR), Pisa, Italy
IGOR SAGGESE • Department of Sciences and Technological Innovation (DiSIT), University
of Piemonte Orientale “A. Avogadro”, Alessandria, Italy
MICHAEL SAMMETH • Bioinformatics, National Laboratory of Cientific Computing
(LNCC), Rio de Janeiro, Brazil
MONICA SANTAMARIA • Institute of Biomembranes and Bioenergetics, CNR, Bari, Italy
CORINNE DA SILVA • Genoscope—National Sequencing Center, Evry, France
JOHN G. TATE • Wellcome Trust Sanger Institute, Cambridgeshire, UK; European
HÉLÈNE TOUZET • Laboratoire d’Informatique Fondamentale de Lille (LIFL), UMR
CNRS 8022, Villeneuve d’Ascq Cédex, France; Inria Lille Nord-Europe, Villeneuve
d’Ascq Cédex, France
GIANLUCA DELLA VEDOVA • Department of Informatics, Systems and Communication
(DISCo), University of Milano-Bicocca, Milano, Italy
xii Contributors
DARIO VENEZIANO • Department of Molecular Virology, Immunology and Medical Genetics,

Comprehensive Cancer Center, The Ohio State University, Columbus, OH, USA;
Department of Clinical and Molecular Biomedicine, University of Catania, Catania, Italy
JÉRÔME WALDISPÜHL • School of Computer Science, McGill University, Montreal, QC, Canada
SHENGQIN WANG • School of Biological Science and Medical Engineering,
Southeast University, Nanjing, Jiangsu, PR, China
LIGUO WANG • Division of Biomedical Statistics and Informatics, Department of Health
Sciences Research, Mayo Clinic, Minneapolis, MN, USA
LINA WEINBRAND • Department of Computer Science, Ben-Gurion University, Beer-Sheva, Israel
FEDERICO ZAMBELLI • Istituto di Biomembrane e Bioenergetica, Consiglio Nazionale delle
Ricerche - CNR, Bari, Italy
FRANCESCA ZOLEZZI • Singapore Immunology Network (SIgN), Agency of Science,
Technology and Research (A*STAR), Biopolis, Singapore, Singapore
Part I
RNA Secondary and Tertiary Structures

Chapter 1
Free Energy Minimization to Predict RNA Secondary

Structures and Computational RNA Design
Alexander Churkin, Lina Weinbrand, and Danny Barash
Abstract
Determining the RNA secondary structure from sequence data by computational predictions is a long-
standing problem. Its solution has been approached in two distinctive ways. If a multiple sequence align-
ment of a collection of homologous sequences is available, the comparative method uses phylogeny to
determine conserved base pairs that are more likely to form as a result of billions of years of evolution than
by chance. In the case of single sequences, recursive algorithms that compute free energy structures by
using empirically derived energy parameters have been developed. This latter approach of RNA folding
prediction by energy minimization is widely used to predict RNA secondary structure from sequence. For
a significant number of RNA molecules, the secondary structure of the RNA molecule is indicative of its
function and its computational prediction by minimizing its free energy is important for its functional
analysis. A general method for free energy minimization to predict RNA secondary structures is dynamic
programming, although other optimization methods have been developed as well along with empirically
derived energy parameters. In this chapter, we introduce and illustrate by examples the approach of free
energy minimization to predict RNA secondary structures.
Key words RNA folding prediction, Free energy minimization methods, RNA free energy parameters,
RNA bioinformatics, RNA secondary structure prediction
1 Introduction
The RNA molecule, once considered as an intermediate step

between DNA and protein, has become a central subject of research
in recent years. Although the functional role of RNAs is often
related to their 3-D structure, the RNA secondary structure is
experimentally accessible and contains much information to shed
light on the relationship between structure and function. In gen-
eral, RNA folding is thought to be hierarchical in nature [1, 2]
where a stable secondary structure forms first and subsequently
there is a refinement to the tertiary fold. Thus, predicting RNA
secondary structures from sequence is an important problem in
computational biology. The secondary structure of an RNA
molecule is a representation of the pattern, given an initial RNA
Ernesto Picardi (ed.), RNA Bioinformatics, Methods in Molecular Biology, vol. 1269,
DOI 10.1007/978-1-4939-2291-8_1, © Springer Science+Business Media New York 2015
3
4 Alexander Churkin et al.
sequence, of complementary base-pairings that are formed between

the nucleic acids. The sequence, represented as a string of four let-
ters, is a single strand consisting of the nucleotides A, C, G, and U,
which are generally assumed to pair to form a secondary structure
with minimum free energy. Therefore, aside from the comparative
method in case a multiple sequence alignment of homologous
sequences is available, if a practitioner would like to predict sec-
ondary structure from an RNA sequence that has no homologs or
that belongs to a poorly annotated family in a comprehensive RNA
database such as Rfam [3], then free energy minimization should
be the method of choice.
The folding prediction problem of the secondary structure
from a single RNA sequence has been an area of active research
since the late 1970s. Dynamic programming methods for this prob-
lem were initially developed in [4–6]. Energy minimization meth-
ods by dynamic programming using auxiliary information [7, 8]
have led to Zuker’s mfold prediction server [9] and the Vienna
RNA package [10, 11]. The predictive accuracy of these pack-
ages was improved by incorporating expanded energy rules [12],
derived from an independent set of experiments, into the folding
prediction algorithm. It should be noted that other methods aside
from dynamic programming have been employed in the context of
free energy minimization, mainly stochastic ones such as in
MPGAfold [13], but the ones that are mostly used are variants of
dynamic programming. For example, RNAfold in the Vienna RNA
package is slightly different from mfold in using the partition func-
tion algorithm of McCaskill [14], but both are based on dynamic
programming and their results are in most cases very similar.
An important aspect of energy minimization software from the
practical standpoint is the ability to predict suboptimal solutions in
addition to the optimal one [15, 16]. The “mfold style” suboptimal
solutions are computed along with a filtering step that ensures that
suboptimal solutions differ to avoid redundancy [15]. A different
way of calculating the suboptimal solutions was carried out in [16]
for the Vienna RNA package. It calculates all suboptimal solutions
within an energy range above the minimum free energy without a
pre-prescribed filtering step. It is then possible to apply a post-
prescribed filtering step, such as with RNAshapes [17]. We will
illustrate in the examples the merits of both approaches, namely
the “mfold style” and the “Vienna style” ways of calculating the
suboptimal solutions. Finally, there are many advanced applica-
tions that utilize free energy minimization, such as in a genome-
scale structure-based clustering of RNAs called LocARNA [18]
and de novo prediction of structured RNAs from genomic
sequences [19], to name but a few. Basic tasks that accompany
structure prediction by energy minimization are available in the soft-
ware packages for the new mfold package called UNAFold [20]
and the most recent version of the Vienna RNA package [21], both
being actively used and regularly maintained. Herein, we provide a
Free Energy Minimization to Predict RNA Secondary Structures 5
few examples that illustrate the use of free energy minimization to

predict RNA structures and utilize the programs included in the
software prediction packages.
2 Materials
2.1 Biological We obtained the RNA sequences for our examples from the study
Sequences published by You et al. [22] and Krol et al. [23]. This is by no
means inclusive as there are many different ways to obtain biologi-
cal sequences, e.g., from a sequencing experiment to investigate a
certain organism or from a known database such as Rfam [3], to
name but a few. Let us briefly describe these biological sequences
and their significance. The first sequence used for illustration, the
5BSL3.2 HCV taken from [22], is a functionally important stem-
loop structure for virus replication that appears in the coding
region of the hepatitis C virus RNA-dependent RNA polymerase,
NS5B. The 5BSL3.2 is about 50 bases in length and is part of a
larger predicted cruciform structure (5BSL3). It was confirmed
experimentally that the 5BSL3.2 consists of an 8-bp lower helix, a
6-bp upper helix, a 12-base terminal loop, and an 8-bp internal
loop. Mutagenesis experiments were performed to investigate the
relationship between its structure and function. In addition, the
size of the sequence is amenable to energy minimization prediction
methods (see Note 1). Thus, the sequence of the 5BSL3.2 HCV
provides a good test bed for energy minimization prediction meth-
ods since the structure of both the wildtype sequence and several
of its mutants was already investigated by structure probing experi-
ments. The second sequence used for illustration, an miRNA pre-
cursor taken from [23], is a hairpin loop structure of about 70
bases in length. It is important as a participant in gene regulation
and its structure was verified experimentally, thus providing another
good test bed for energy minimization prediction methods.
For convenience, the sequence of the 5BSL3.2 HCV is shown
below:
AGCGGGGGAGACAUAUAUCACAGCCUGUCUCGUGCCC
GACCCCGC
Its two-point mutants that were experimentally tested in [22] are
clearly specified in the text herein and the relevant figure captions.
2.2 Computational The most widely used energy minimization software for predicting
Hardware RNA secondary structures from sequence can be accessed using
webservers, e.g., [9, 11], requiring no special computational
resources. With the exception of some specific large-scale tasks that
may require whole genome scans, or specific cases of a software
that may require a massively parallel computational platform [13],
almost all tasks that utilize energy minimization prediction are
approached with a standard PC. If one prefers to download a
software package instead of using a webserver directly or for more

involved tasks, the most straightforward hardware platforms to
use are Linux and Unix-based computers, although Windows
and Mac OS X and possibly other operating systems are in general
supported as well.
2.3 Computational When downloading the most widely used energy minimization soft-
Software ware packages such as UNAFold/mfold or the Vienna RNA pack-
age, some common requirements are to have a Perl interface such as
Perl 5.6.1 or greater and an application programming interface (API)
such as OpenMP and OpenGL. Specifications can be found on the
instructions page that is available for each individual package.
3 Methods
Free energy minimization to predict RNA secondary structures is

an advanced procedure from the algorithmic side, which has been
gradually developed and improved over the years, but conceptually
simple from the practical side. The user submits a sequence as
input and can either choose to accompany the sequence with
default parameters or to change the value of the parameters accord-
ing to the problem at hand. The most basic free energy minimiza-
tion problem is to predict structure from sequence, but there are
several other related problems that use the basic problem repeat-
edly and are therefore relying on free energy minimization. Thus,
we chose to illustrate free energy minimization by representative
examples. First, the basic problem of predicting structure from
sequence will be exemplified. Second, a related problem of predict-
ing which sequence mutations will cause a dramatic change in
structure will be shown. Finally, the third example will illustrate the
problem of sequence design, which is the inverse problem
(sequence from structure) relative to the direct problem (structure
from sequence). It uses the direct problem iteratively, after starting
from an initial guess for the sequence and mutating the sequence
in each step in order to optimize a certain objective function.
The basic structure for the user to follow is common in all cases:
1. Access the relevant webserver and/or download the program.
2. Insert an input sequence and change parameter values (default
values are given).
3. Run the program and observe the output predicted structure,
mostly given in graphical forms (e.g., RNA secondary struc-
ture drawings, dot plots) along with textual information.
3.1 Predicting RNA We start by describing the basic procedure of predicting RNA sec-
Secondary Structures ondary structures by energy minimization given an initial sequence.
The format for the input sequence is similar in almost all software
packages that have been developed to solve this problem. The user
can either select to manually insert a sequence consisting of the
letters A,C,G,U or use the FASTA format. For example, in the

mfold webserver [9], the user starts by filling in the RNA Folding
Form, inserting the 5BSL3.2 HCV sequence that was given
towards the end of Subheading 2, including its name. Default
parameters can remain with the values given in the various fields
unless there is a clear justification to insert different values. Upon
pressing “Fold RNA” found at the end of the RNA Folding Form,
mfold performs the free energy minimization prediction and out-
puts several ways to display the results. Selecting “Structure 1”
with displaying it in one of the different formats available, such as
pdf, will result in Fig. 1. Going back to the RNA Folding Form and
increasing the percent of suboptimality from the default of 5–30,
repeating the same procedure above, will result in four different
structures instead of one. Opting for several suboptimal solutions
instead of a single optimal solution can be important (see Note 2).
The default parameter is usually enough when using the “mfold
style” suboptimal solutions [15], but in our case we increase the
percent of suboptimality because we are interested in checking if
one of the weak suboptimals has the potential to become dominant
when introducing a change to the sequence. All structures can now
be viewed using an energy dot plot by selecting all four structures
in the “Compare selected foldings” field and pressing “Do the
Fig. 1 Free energy minimization for the secondary structure prediction of wild-
type 5BSL3.2 HCV. UNAFold version 3.8 with mfold utils version 4.6 was used to
predict and generate the secondary structure drawing
Fig. 2 Energy dot plot for the secondary structure prediction of wildtype 5BSL3.2 HCV. Using mfold utils version
4.6, the percent of suboptimality was increased to 30 to capture the fourth suboptimal solution, for an illustra-
tive reason that will become clear after generating and examining Fig. 3
comparison”. This will result in Fig. 2, where in the upper triangu-

lar of the dot plot all four structures can be viewed while the lower
triangular contains only the optimal structure. Going back to the
RNA Folding Form and slightly modifying the sequence by insert-
ing a two-point mutation C31A-U33G, repeating the same proce-
dure as above with default parameters (percent of suboptimality is
now back to 5) will result in Fig. 3. Notice that the stable optimal
structure for the mutant resembles the fourth suboptimal structure
for the wildtype predicted structure. Such an observation makes
sense because often times the mutant structure of a point mutation
Fig. 3 Free energy minimization for the secondary structure prediction of a spe-
cific two-point mutant 5BSL3.2 HCV (C31A-U33G). UNAFold version 3.8 with
mfold utils version 4.6 was used to predict and generate the secondary structure
drawing. One can observe the similarity in shape between the structure obtained
and the fourth suboptimal solution displayed in the dot plot of Fig. 2
that is in the near-neighborhood of the wildtype will originate from

a suboptimal solution of the wildtype that becomes the optimal
solution upon introducing the mutation.
3.2 Predicting RNA The first example above served to illustrate the most basic use of free
Deleterious Mutations energy minimization to predict RNA structures. Having also
observed in this example the relationship between suboptimal solu-
tions of the wildtype and their possible appearance as an optimal
solution of a near-neighbor mutant, we now focus on a second
example that utilizes this concept to efficiently predict deleterious
mutations in the sequence that may cause a dramatic change in
structure. As detailed in a review on this topic [24], there are several
ways to approach this problem [25, 26]. The method described in
[25] relies on Vienna’s RNAfold [10] for the direct problem of
RNA structure prediction, Vienna’s RNAsubopt [16] for calculat-
ing all suboptimal solutions, and the concept described above of a
suboptimal solution becoming an optimal one. To experiment, the
user can now access the RNAmute webserver [27] and insert the
same example sequence of the 5BSL3.2 HCV that was used as input
in the first two figures. Changing the number of mutations to 2 in

the appropriate field and pressing “Submit Form” will result in a
table of mutations that are ranked according to the distance of their
free energy minimization predicted structure relative to the wildtype
predicted structure. One of the few two-point mutations appearing
distinctively with the largest distance is the one shown in Fig. 4,
where the labeled changes appearing in red, corresponding to the
example depicted in Fig. 3. This well illustrates the fact that compo-
nents from free energy minimization to predict RNA secondary
structures, such as RNAfold and RNAsubopt from the Vienna RNA
package, may be used as the backbone of other related problems
such as RNA deleterious mutation prediction (see Note 3).
3.3 Computational The final example is taken from the field of RNA design. The
RNA Design inverse RNA folding problem for designing sequences that
fold into a given RNA secondary structure was introduced in [10].
The approach to solve it by stochastic optimization relies on the
solution of the direct problem. Initially, a seed sequence is chosen,
after which a local search strategy (or global, as in [28]) is used to
mutate the seed and apply repeatedly the direct problem of RNA
Fig. 4 Predicted rearranging mutation in the 5BSL3.2 wildtype. The RNAmute webserver was used, relying
on RNAfold, RNAsubopt, and other programs from the Vienna RNA package. The output two-point mutant
C31A-U33G happens to coincide with the experimental result that was illustrated in Fig. 3
folding prediction by energy minimization. Recently, an extension

to the problem was developed that allows designing sequences that
fold into a prescribed shape, leaving some flexibility in the second-
ary structure of RNA motifs that do not necessarily possess a
known functional role. The shape of the RNA can be represented
as a tree-graph [29] and its relation to RNA secondary structures can
then be clarified by examining Fig. 5. The program that implements
Fig. 5 A tree-graph illustration for a coarse grain representation [29] of RNA secondary structures
this type of sequence design, put forth in [30], relies on pro-

grams from the Vienna RNA package such as RNAfold,
RNAinverse, and RNAdistance [10]. This time, the user down-
loads a program called RNAfbinv [31] (RNA fragment-based
design), runs the application, and in the input screen inserts a sec-
ondary structure as input in dot-bracket representation. Upon
pressing “Fragment”, a new window opens where the user can
choose a fragment constraint using a combo-box to select a desired
motif that should be preserved. Then, in the initial input screen,
upon pressing “Process” a subsequent window is displayed where
certain physical parameters appear with default numerical values.
Upon pressing “Result”, the program computes several designed
sequences ranked by their base-pair distance from the input struc-
ture. Examples for such a solution and the various stages in the
GUI that demonstrate the user experience are available in Figs. 6,
7, and 8. Figure 7 shows an example of the design results of an
artificial structure while in Fig. 8, a design result of an miRNA
precursor (in compliance with Note 4) taken from [23] is shown.
Fig. 6 Sequence design with a motif constraint by using repetitively free-energy minimization of RNA second-
ary structures. (a) Initial input screen in the RNAfbinv application for designing sequences. (b) Motif selection
screen in the RNAfbinv application for designing sequences
Fig. 7 Sequence design of an artificial example by energy minimization. Best result and 20th result obtained
with RNAfbinv according to base-pair distance from the wildtype structure are displayed
In the future, the uses for such free energy minimization applications
that involve computational RNA design are expected to grow con-
siderably, with the increase of experimental evidence regarding the
functional importance of certain RNA motifs. Also, these applica-
tions can be used to detect known noncoding RNAs in new loca-
tions when the query RNA pattern is suspected to possess more
conservation in structure and less conservation in sequence. Results
of the flexible computational RNA design procedure are sequences
that can be searched for efficiently by using sequence-based
methods.
Fig. 8 Sequence design of an miRNA precursor by energy minimization. Best result obtained with RNAfbinv
according to the base-pair distance from the wildtype structure is shown with the depicted motif constraint
4 Notes
1. All free energy minimization methods rely on thermodynamic

parameters that are empirically derived for small combinations
of nucleotides. Therefore, the upper range estimate for the
sequence length that these programs are useful for is ~150 nt.
Predictions made with longer sequences tend to become inac-
curate and possess artifacts. If the user needs to make predictions
for longer sequences, it should be done locally by a moving
window approach, for example with Vienna’s RNALfold [21].
This way, energy minimization predictions are performed
inside a window whose size is no longer than 150 nt.
2. An important breakthrough in the field of RNA folding pre-
diction by energy minimization was the ability to calculate
meaningful suboptimal solutions. This was first devised in ref.
15, known as the “mfold style” approach that performs an
automatic filtering to the suboptimal solutions. It is still being
widely used at present and it differs from the “Vienna style”
approach that calculates all suboptimal solutions within an
energy range above the minimum free energy without a

pre-prescribed filtering step [16]. Whether one should favor
one approach over the other depends on the application, and
there are also alternative approaches to calculate suboptimal
solutions [17]. In any case, it is recommended to take advan-
tage of the ability to calculate suboptimal solutions in free
energy minimization to predict secondary structures.
3. As observed in the illustrations for Example 2 and Example 3,
one can solve more specific problems such as RNA deleterious
mutation prediction and computational RNA design using
components taken from the latest versions of the most widely
used RNA software prediction packages [20, 21].
4. RNA folding [1, 2] is an involved process that depends on
many factors in the environment, such as ion concentration
and small molecules that can bind to the RNA and influence
the folding in ways that are impossible to predict. These limi-
tations should be taken into account and one needs to be care-
ful from making predictions about the folding of RNAs that
are known to reside in such environments. Nevertheless, there
are many cases of significant interest where folding prediction
by energy minimization is indispensable as an approach to
analyze RNA structure and its relation to function.
Acknowledgments
The authors would like to thank Idan Gabdank and Assaf Avihoo
for their assistance in this study. This work was partially supported
by the Kreitman Foundation at Ben-Gurion University.
References
1. Brion P, Westhof E (1997) Hierarchy and 7. Zuker M, Stiegler P (1981) Optimal computer
dynamics of RNA folding. Annu Rev Biophys folding of large RNA sequences using thermo-
Biomol Struct 26:113–137 dynamics and auxiliary information. Nucleic
2. Tinoco I, Bustamante C (1999) How RNA Acids Res 9:133–148
folds. J Mol Biol 293:271–281 8. Zuker M, Sankoff D (1984) RNA secondary
3. Griffiths-Jones S, Bateman A, Marshall M, structures and their prediction. Bull Math Biol
Khanna A, Eddy SR (2003) Rfam: an RNA 46:591–621
family database. Nucleic Acids Res 31: 9. Zuker M (2003) Mfold web server for nucleic
439–441 acid folding and hybridization prediction.
4. Nussinov R, Pieczenik G, Grigg JR, Kleitman Nucleic Acids Res 31:3406–3415
DJ (1978) Algorithms for loop matchings. 10. Hofacker IL, Fontana W, Stadler PF,
SIAM J Appl Math 35:68–82 Bonhoeffer LS, Tacker M, Schuster P (1994)
5. Waterman MS, Smith TF (1978) RNA sec- Fast folding and comparison of RNA second-
ondary structure: a complete mathematical ary structures. Monatsh Chem 125:167–188
analysis. Math Biosci 42:257–266 11. Hofacker IL (2003) Vienna RNA second-
6. Nussinov R, Jacobson AB (1980) Fast algo- ary structure server. Nucleic Acids Res 31:
rithm for predicting the secondary structure 3429–3431
of single stranded RNA. Proc Natl Acad Sci 12. Mathews DH, Sabina J, Zuker M, Turner DH
U S A 77(11):6309–6313 (1999) Expanded sequence dependence of
thermodynamic parameters improves predic- RNA polymerase is required for hepatitis c

tion of RNA secondary structure. J Mol Biol virus RNA replication. J Virol 78(3):
288:911–940 1352–1366
13. Shapiro BA, Wu J-C, Bengali D, Potts MJ 23. Krol J, Sobczak K, Wilczynska U, Drath M,
(2001) The massively parallel genetic algo- Janiska A, Kaczynska D, Krzyzosiak WJ (2004)
rithm for RNA folding: MIMD implementa- Structural features of microRNA (miRNA)
tion and population variation. Bioinformatics precursors and their relevance to miRNA
17:137–148 biogenesis and small interfering RNA/short
14. McCaskill JS (1990) The equilibrium partition hairpin RNA design. J Biol Chem 279:
function and base pair binding probabilities for 42230–42239
RNA secondary structure. Biopolymers 29: 24. Barash D, Churkin A (2011) Mutational anal-
1105–1119 ysis in RNAs: comparing programs for RN del-
15. Zuker M (1989) On finding all suboptimal eterious mutation prediction. Brief Bioinform
foldings of an RNA molecule. Science 244: 12(2):104–114
48–52 25. Churkin A, Barash D (2008) An efficient
16. Wuchty S, Fontana W, Hofacker IL, Schuster P method for the prediction of deleterious
(1999) Complete suboptimal folding of RNA multiple-point mutations in the secondary
and the stability of secondary structures. structure of RNAs using suboptimal folding
Biopolymers 49:145–165 solutions. BMC Bioinformatics 9:222
17. Steffen B, Voss B, Rehmsmeier M, Reeder J, 26. Waldispühl J, Devadas S, Berger B, Clote P
Giegerich R (2006) RNAshapes: an integrated (2008) Efficient algorithms for probing the
RNA analysis package based on abstract shapes. RNA mutational landscape. PLoS Comput
Bioinformatics 22(4):500–503 Biol 4:e1000124
18. Will S, Reiche K, Hofacker IL, Stadler PF, 27. Churkin A, Gabdank I, Barash D (2011) The
Backofen R (2007) Inferring non-coding RNA RNAmute web server for the mutational anal-
families and classes by means of genome-scale ysis of RNA secondary structures. Nucleic
structure-based clustering. PLoS Comput Biol Acids Res 39:W92–W99
3(4):e65 28. Levin A, Lis M, Ponty Y, O’Donnell CW,
19. Gorodkin J, Hofacker IL, Torarinsson E, Yao Devadas S, Berger B, Waldispühl J (2012) A
Z, Havgaard JH, Ruzzo WL (2010) De novo global sampling approach to designing and
prediction of structures RNAs from genomic reengineering RNA secondary structures.
sequences. Trends Biotechnol 28(1):9–20 Nucleic Acids Res 40(20):10041–10052
20. Markham NR, Zuker M (2008) Software for 29. Shapiro BA (1988) An algorithm for compar-
nucleic acid folding and hybridization. ing RNA secondary structures. Comput Appl
Methods Mol Biol 453:3–31 Biosci 4:387–393
21. Lorenz R, Lorenz R, Bernhart SH, Höner zu 30. Avihoo A, Churkin A, Barash D (2011)
Siederdissen C, Siederdissen C, Tafer H, RNAexinv: an extended inverse RNA folding
Flamm C, Stadler PF, Hofacker IL (2011) from shape and physical attributes to
ViennaRNA package 2.0. algorithms. Mol Biol sequences. BMC Bioinformatics 12(319):24
6:26 31. Weinbrand L, Avihoo A, Barash D (2013)
22. You S, Stump DD, Branch AD, Rice CM RNAfbinv: an interactive Java application for
(2004) A cis-acting replication element in the fragment-based design of RNA sequences.
sequence encoding the NS5B RNA-dependent Bioinformatics 29(22):2938–2940
Chapter 2
RNA Secondary Structure Prediction

from Multi-Aligned Sequences
Michiaki Hamada
Abstract
It has been well accepted that the RNA secondary structures of most functional non-coding RNAs
(ncRNAs) are closely related to their functions and are conserved during evolution. Hence, prediction of
conserved secondary structures from evolutionarily related sequences is one important task in RNA bioin-
formatics; the methods are useful not only to further functional analyses of ncRNAs but also to improve
the accuracy of secondary structure predictions and to find novel functional RNAs from the genome.
In this review, I focus on common secondary structure prediction from a given aligned RNA sequence, in
which one secondary structure whose length is equal to that of the input alignment is predicted. I system-
atically review and classify existing tools and algorithms for the problem, by utilizing the information
employed in the tools and by adopting a unified viewpoint based on maximum expected gain (MEG)
estimators. I believe that this classification will allow a deeper understanding of each tool and provide users
with useful information for selecting tools for common secondary structure predictions.
Key words Common/consensus secondary structures, Comparative methods, Multiple sequence

alignment, Covariation, Mutual information, Phylogenetic tree, Energy model, Probabilistic model,
Maximum expected gain (MEG) estimators
1 Introduction
Functional non-coding RNAs (ncRNAs) play essential roles in

various biological processes, such as transcription and translation
regulation [11, 45, 49]. Not only nucleotide sequences but also
secondary structures are closely related to the functions of ncRNAs,
so secondary structures are conserved during evolution. Hence,
the prediction of these conserved secondary structures (called
“common secondary structure prediction” throughout this review)
from evolutionarily related RNA sequences is among the most
important tasks in RNA bioinformatics, because it provides useful
information for further functional analysis of the targeted RNAs
[4, 6, 28, 41, 46]. It should be emphasized that common second-
ary structure predictions are also useful to improve the accuracy of
17
18 Michiaki Hamada
secondary structure prediction [13, 48] or to find functional RNAs

from the genome [15, 66].
Approaches to common secondary structure prediction are
divided into two categories with respect to the input: (1) unaligned
RNA sequences and (2) aligned RNA sequences. Common sec-
ondary structure predictions from unaligned RNA sequences can
be solved by utilizing the Sankoff algorithm [54]. However, it is
known that the Sankoff algorithm has huge computational costs:
O(L3n) and O(L2n) for time and space, respectively, where L is the
length of RNA sequences and n is the number of input sequences.
For this reason, the second approach, whose input is aligned RNA
sequences, is often used in actual analyses. The problem focused
on this review is formulated as follows.
Problem 1 (RNA Common Secondary Structure Prediction of

Aligned Sequences)
Given an input multiple sequence alignment A, predict a secondary
structure y whose length is equal to the length of the input alignment.
The secondary structure y is called the common (or consensus) second-
ary structure of the multiple alignment A.
In contrast to conventional RNA secondary structure predic-
tion (see Fig. 1a), the input of common secondary structure predic-
tion is a multiple sequence alignment of RNA sequences (see
Fig. 1b) and the output is a secondary structure with the same
length as the alignment. In general, the predicted common RNA
secondary structure is expected to represent a secondary structure
that commonly appears in the input alignments or is conserved
during evolution.
To develop tools (or algorithms) for solving this problem, it
should be made clear what a better common secondary structure is.
However, the evaluation method for predicted common RNA sec-
ondary structures is not trivial. A predicted common secondary
structure depends on not only RNA sequences in the alignment
but also multiple alignments of input sequences, and in general, no
reference (correct) common secondary structures are available (see
Note 1). A predicted common secondary structure is therefore
evaluated, based on reference RNA secondary structures for each
individual RNA sequence in the alignment as follows.
Evaluation Procedure 1 (For Problem 1)

Given reference secondary structures for each RNA sequence, evalu-
ate the predicted common secondary structure y as follows. First, map
y onto each RNA sequence in the alignment A (see the column
“Mapped structure with gaps” in Fig. 2, for example). Second, remove
all gaps in each sequence and the corresponding base-pairs in the
mapped secondary structure in order to maintain the consistency of
the secondary structures (see the column “Mapped structure without
gaps” in Fig. 2, for example). Third, calculate the quantities TP,
RNA Secondary Structure Prediction from Multi-Aligned Sequences 19
Fig. 1 (a) Conventional RNA secondary structure prediction, in which the input is an individual RNA sequence
and the output is an RNA secondary structure of the sequence. (b) Common (or consensus) RNA secondary
structure prediction in which the input is a multiple sequence alignment of RNA sequences and the output is
an RNA secondary structure whose length is equal to the length of the alignment. The secondary structure is
called the common (or consensus) secondary structure
20 Michiaki Hamada
Fig. 2 An evaluation procedure for the predicted common RNA secondary structure of an input alignment, in
which the reference secondary structure of each RNA sequence in the alignment is given. This procedure is
based on the idea that a common secondary structure should reflect as many of the secondary structures of
each RNA sequence in the input alignment as possible. Mapped RNA secondary structures without gaps are
computed by getting rid of base-pairs that correspond to gaps. Note that it is difficult to compare a predicted
common secondary structure with a reference common secondary structure, because the reference common
RNA secondary structure for an arbitrary input alignment is not available in general. In most studies of com-
mon secondary structure prediction, evaluation is conducted by using this procedure or a variant. See [23] for
a more detailed discussion of evaluation procedures for Problem 1
TN, FP, and FN for each mapped secondary structure y(map) with
respect to the reference secondary structure (TP, TN, FP, and FN are
the respective numbers of true positive, true-negative, false-positive
and false-negative base-pairs of a predicted secondary structure with
respect to the reference structures). Finally, calculate the evaluation
measures sensitivity (SEN), positive predictive value (PPV), and
Matthew correlation coefficient (MCC) for the sum of TP, TN, FP,
and FN over all the RNA sequences in the alignment A.
Figure 2 shows an illustrative example of Evaluation
Procedure 1. Note that there exist a few variants of this procedure
(e.g., [5]).
In this review, I aim to classify the existing tools (or algorithms)
for Problem 1; These tools are summarized in Table 1, which
includes all the tools for common secondary structure prediction
as of 17th June 2013. To achieve this aim, I describe the informa-
tion that is often utilized in common secondary structure
predictions, and classify tools from a unified viewpoint based on
maximum expected gain (MEG) estimators. I also explain the rela-
tions between the MEG estimators and Evaluation Procedure 1.
This review is organized as follows. In Subheading 2, I sum-
marize the information that is commonly utilized when designing
algorithms for common secondary structure prediction. In
Subheading 3, several concepts to be utilized in the classification of
tools are presented, and the currently available tools are classified
within this framework.
Table 1
List of tools for common secondary structure prediction from aligned sequences
Tool References Description

(Without pseudoknot)
CentroidAlifold [19, 23] Achieved superior performance in a recent benchmark (CompaRNA)
[48]. Using an MEG estimator with the γ-centroid-type gain
function (Subheading 3.3.2) in combination with a mixture
probability distribution, including several types of information
(Subheading 3.2.4)
ConStruct [37, 72] A semi-automatic, graphical tool based on mutual information
(Subheading 2.2)
KNetFold [6] Computes a consensus RNA secondary structure from an RNA
sequence alignment based on machine learning (Bayesian networks)
McCaskill-MEA [32] Adopting majority rule with the McCaskill energy model [40], leading
to an algorithm that is robust to alignment errors
PETfold [59] Considers both phylogenetic information and thermodynamic stability
by extending Pfold, in combination with an MEA estimation
Pfold [34, 35] Uses phylogenetic tree information with simple SCFG
PhyloRNAalifold [14] Incorporates the number of co-varying mutations on the phylogenetic
tree of the aligned sequences into the covariance scoring of
RNAalifold
PPfold [63] A multi-threaded implementation of the Pfold algorithm, which is
extended to evolutionary analysis with a flexible probabilistic model
for incorporating auxiliary data, such as data from structure probing
experiments
RNAalifold [5, 26, 27] Considers both thermodynamic stability and co-variation in
combination with RIBOSUM-like scoring matrices [33]
RSpredict [62] Takes into account sequence covariation and employs effective
heuristics for accuracy improvement
(With pseudoknot)
hxmatch [73] Computes consensus structures including pseudoknots based on
alignments of a few sequences. The algorithm combines
thermodynamic and covariation information to assign scores to all
possible base-pairs, the base-pairs are chosen with the help of the
maximum weighted matching algorithm
ILM [52] Uses mutual information and helix plot in combination with heuristic
optimization
IPKnot [57] Uses MEG estimators with γ-centroid gains and heuristic probability
distribution of RNA interactions together with integer linear
programming to compute a decoded RNA secondary structure
MIfold [12] A MATLAB(R) toolbox that employs mutual information, or a related
covariation measure, to display and predict common RNA
secondary structure
To the best of my knowledge, this is a complete list of tools for the problem as of 17 June 2013. Note that tools for
common secondary structure prediction from unaligned RNA sequences are not included in this list. See Table 2 for
further details of the listed tools
22 Michiaki Hamada
Table 2
Comparison of tools (in Table 1) for common secondary structure predictions from aligned sequences
Softwarea Used informationb Gainc Prob. dist.d
Name SA WS TS ML CV PI MI MR EI
(Without pseudoknot)
CentroidAlifold ✓ ✓ ✓ ✓ ✓ ✓ ✓ e
γ-cent Anyf
ConStruct ✓ ✓ ✓ ✓ ✓ na na
KNetFold ✓ ✓ ✓ na na
McCaskill-MEA ✓ ✓ Contra Av(Mc)
PETfold ✓ ✓ ✓ Contra Pf+Av(Mc)
Pfold ✓ ✓ ✓ Delta Pf
PhyloRNAalifold ✓ ✓ ✓ Delta Ra
PPfold ✓ ✓ ✓ ✓ Delta Pf
RNAalifold ✓ ✓ ✓ ✓ Delta Ra
RSpredict ✓ ✓ ✓ ✓ na na
(With pseudoknot)
hxmatch ✓ ✓ ✓ na na
ILM ✓ ✓ ✓ ✓ na na
IPKnot ✓ ✓ ✓ ✓ ✓ ✓ γ-cent Anyf
MIfold ✓g ✓ na na
a
Type of software available. SA stand alone, WS web server, TS thermodynamic stability (Subheading 2.1.1)
b
In the “Information used” columns, ML machine learning (Subheading 2.1.2); CV covariation (Subheading 2.3); PI
phylogenetic (evolutionary) information(Subheading 2.4); MI mutual information (Subheading 2.2); MR majorityrule
(Subheading 2.5); EI experimental information (Subheading 2.1.3)
c
In the column“Gain,” γ-cent:γ-centroid-type gain function (Subheading 3.3.2); contra: CONTRAfold-type gain func-
tion (Subheading 3.3.3)
d
In the column “Prob. dist.,” Pf Pfold model (Subheading 3.2.2); Ra RNAalipffold model (Subheading 3.2.1);
Av(Mc) averaged probability distribution with McCaskill model (Subheading 3.2.3); “+” indicates a mixture distribu-
tion (of several models). “na” means “Not available” due to no use of probabilistic models
e
If the method proposed in [16] is used, experimental information derived from SHAPE [36] and PARS [31] is easily
incorporated in CentroidAlign
f
CentroidAlifold and IPKnot can employ a mixed distribution given by an arbitrary combination of RNAalipffold,
Pfold, and an averaged probability distribution based on the McCaskill or CONTRAfold modelsgMATLAB codes
are available
2 Materials
Several pieces of information are generally utilized in tools and

algorithms for predicting common RNA ondary structures. These
will now be briefly summarized.
2.1 Fitness to Each The common RNA secondary structure should be a representative
Sequence in the Input secondary structure among RNA sequences in the alignment.
Alignment Therefore, the fitness of a predicted common secondary structure
to each RNA sequence in the alignment is useful information. In
particular, in Evaluation Procedure 1, the fitness of a predicted
common secondary structure to each RNA sequence is evaluated.
This fitness is based on probabilistic models for RNA second-
ary structures of individual RNA sequence, such as the energy-
based and machine learning models shown in Subheadings 2.1.1
and 2.1.2, respectively. These models provide a probability distri-
bution of secondary structures of a given RNA sequence. (p(θ | x)
denotes a probability distribution of RNA secondary structures for
a given RNA sequence x.)
2.1.1 Thermodynamic Turner’s energy model [38] is an energy-based model, which con-
Stability: Energy-Based siders the thermodynamic stability of RNA secondary structures.
Models This model is widely utilized in RNA secondary structure predic-
tions, in which experimentally determined energy parameters
[38, 39, 75] are employed. In the model, structures with a lower
free energy are more stable than those with a higher free energy.
Note that Turner’s energy model leads to a probabilistic model for
RNA sequences, providing a probability distribution of secondary
structures, which is called the McCaskill model [40].
2.1.2 Machine Learning In addition to the energy-based models described in the previous
(ML) Models subsection, probabilistic models for RNA secondary structures
based on machine learning (ML) approaches have been proposed.
In contrast to the energy-based models, machine learning models
can automatically learn parameters from training data (i.e., a set of
RNA sequences with secondary structures). There are several mod-
els based on machine learning which adopt different approaches:
(1) Stochastic context free grammar (SCFG) models [10]; (2) the
CONTRAfold model [9] (a conditional random field model); (3)
the Boltzmann Likelihood (BL) model [1–3]; and (4) non-
parametric Bayesian models [56].
See Rivas et al. [51] for detailed comparisons of probabilistic
models for RNA secondary structures.
2.1.3 Experimental Recently, experimental techniques to probe RNA structure by

Information high-throughput sequencing (SHAPE [36]; PARS [31]; FragSeq
[64]) have enabled genome-wide measurements of RNA struc-
ture. Those experimental techniques stochastically estimate the
flexibility of an RNA strand, which can be considered as a kind of
loop probability for every nucleotide in an RNA sequence.
Remarkably, secondary structures of long RNA sequences, such as
HIV-1 [69], HCV (hepatitis C virus) [44], and large intergenic
ncRNA (the steroid receptor RNA activator) [43], have been
recently determined by combining those experimental techniques
24 Michiaki Hamada
with computational approaches. If available, such experimental

information is useful in common secondary structure predictions,
because they provide reliable secondary structures for each RNA
sequence.
2.2 Mutual The mutual information of the ith and jth columns in the input
Information alignment is defined by
f ij (XY )
M ij = åf
X ,Y
ij (XY )log
f i (X ) f j (Y )
= KL ( f ij (XY ) || f i (X ) f j (Y ))) (1)

where fi(X) is the frequency of base X at alignment position i;
fij(XY ) is the joint frequency of finding X in the ith column and Y
in the jth column, and KL(⋅ | | ⋅ ) denotes the Kullback–Liebler dis-
tance between two probability distributions. As a result, the com-
plete set of mutual information can be represented as an upper
triangular matrix: {Mij}i < j.
Note that the mutual information score makes no use of base-
pairing rules of RNA secondary structure. In particular, mutual
information does not account for consistent non-compensatory
mutations at all, although information about them would be useful
when predicting common secondary structures as described in the
next subsection.
2.3 Sequence Because secondary structures of ncRNAs are related to their func-
Covariation tions, mutations that preserve base-pairs (i.e., covariations of a
of Base-Pairs base-pair) often occur during evolution. Figure 3 shows an exam-
ple of covariation of base-pairs of tRNA sequences, in which many
covariations of base-pairs are found, especially in the stem parts in
the tRNA structure.
The covariation of the ith and jth columns in the input align-
ment is evaluated by the averaged number of compensatory muta-
tions defined by
2
Cij = å dijx , y Pijx Pijy
N (N - 1) (x , y )ÎA
(2)

where N is the number of sequences in the input alignment A. For
an RNA sequence x in A, Pijx = 1 if xi and xj form a base-pair and
Pijx = 0 otherwise, and
dijx , y = 2 - d (xi , y i ) - d (x j , y j ) (3)

where δ is the delta function: δ(a, b) = 1 only if a = b, and δ(a, b) = 0
otherwise.
For instance, RNAalifold [65] uses the information of covaria-
tion in combination with the thermodynamic stability of common
secondary structures.
a
(((((((..((((.........)))).(((((.......))))).....(((((......
Seq1/1-74 GGGCCUGUAGCUCAGAGGAUUAGAGCACGUGGCUACGAACCACGGUGUCGGGGGUUCGAA 60
Seq2/1-74 GGGCUAUUAGCUCAGUUGGUUAGAGCGCACCCCUGAUAAGGGUGAGGUCGCUGAUUCGAA 60
Seq3/1-72 GGCGCCGUGGCGCAGUGGA--AGCGCGCAGGGCUCAUAACCCUGAUGUCCUCGGAUCGAA 58
Seq5/1-72 GCGUUGGUGGUAUAGUGGUG-AGCAUAGCUGCCUUCCAAGCA-GUUGACCCGGGUUCGAU 58
Seq4/1-68 ACUCCCUUAGUAUAAUU----AAUAUAACUGACUUCCAAUUA-GUAGAUUCUGAAU-AAA 54
.........10........20........30........40........50.........
.)))))))))))).
Seq1/1-74 UCCCUCCUCGCCCA 74 b A
Seq2/1-74 UUCAGCAUAGCCCA 74 GC
Seq3/1-72 ACCGAGCGGCGCUA 72 GC
Seq5/1-72 UCCCGGCCAACGCA 72 GC
Seq4/1-68 CCCAGAAGAGAGUA 68 CG
.........70... CA
CG UAA
U GA U GC GACC G
G A C
C A C C C G GG C
G G U U
A GAGC C
U
_ _ A A G
C GAU
A _
UA
GC
C
C CA
U A
U C
C
Fig. 3 An example of covariation of base-pairs in an alignment of tRNA sequences: (a) a multiple alignment of
tRNA sequences and (b) a predicted common secondary structure of the alignment. The figures are taken from
the output of an example on the RNAalifold [5] Web Server (http://rna.tbi.univie.ac.at/cgi-bin/RNAalifold.cgi)
2.4 Phylogenetic Because most secondary structures of functional ncRNAs are

(Evolutionary) conserved in evolution, the phylogenetic (evolutionary) informa-
Information tion with respect to the input alignment is useful for predicting
common secondary structures, and is employed by several tools.
Pfold [34, 35] incorporates this information into probabilistic
model for common secondary structures (Subheading 3.2.2) for
the first time.
2.5 Majority Rule Kiryu et al. [32] proposed the use of the majority rule of base-
of Base-Pairs pairs in predictions of common secondary structures. This rule
states that base-pairs supported by many RNA sequences should
be included in a predicted common secondary structure.
Specifically, Kiryu et al. utilized an averaged probability distri-
bution of secondary structures among RNA sequences to pre-
dict a common RNA secondary structure from a given alignment
(Subheading 3.2.3).
The aim of this approach is to mitigate alignment errors,
because the effects of a minor alignment errors can be disregarded
in the prediction of common secondary structures.
26 Michiaki Hamada
3 Methods
3.1 MEG Estimators In this section, I classify the existing algorithms for common RNA
secondary structure prediction (shown in Table 1) from a unified
viewpoint, based on a previous study [23] in which the following
type of estimator [18, 24] was employed.
yˆ = argmax y ÎS (A ) å G (q , y ) p (q | A) (4)
q ÎY
where S (A) denotes a set of possible secondary structures with
length | A | (the length of the alignment A), G(θ, y) is called a “gain
function” and returns a measure of the similarity between two com-
mon secondary structures, and p(θ | A) is a probabilistic distribution
on S (A) . This type of estimator is an MEG estimator; When the
gain function is designed according to accuracy measures for target
problems, the MEG estimator is often called a “maximum expected
accuracy (MEA) estimator” [18] (see Note 2).
In the following, a common secondary structure q Î S (A) is
represented as an upper triangular matrix q = qij { }
. In this
1£i < j £|A |
matrix θij = 1 if the ith column in A forms a base-pair with the jth
column in A, and θij = 0 otherwise.
The choices of p(θ | A) and G(θ, y) are described in
Subheadings 3.2 and 3.3, respectively.
3.2 Choice The probabilities p(θ | A) provide a distribution of common RNA

of Probabilistic Models secondary structures given a multiple sequence alignment A. This
p(θ | A) distribution is given by the following models.
3.2.1 RNAalipffold Model The RNAalipffold model is a probabilistic version of RNAalifold

[27], which provides a probability distribution of common RNA
secondary structures given a multiple sequence alignment. The
distribution on S (A) is defined by
1 æ -E (q , A) ö
p (RNAalipffold) (q | A) = exp ç + Cov (q , A) ÷ (5)
Z (T ) è kT ø
where E(θ, A) is the averaged free energy of RNA sequences in the

alignment with respect to the common secondary structure θ in A
and Cov(θ, A) is the base covariation (cf. Subheading 2.3) with
respect to the common secondary structure θ. The ML estimate of
this distribution is equivalent to the prediction of RNAalifold.
Note that the negative part of the exponent in Eq. 5 is called
the pseudo energy and it plays an essential role in finding ncRNAs
from multiple alignments [15, 67].
3.2.2 Pfold Model The Pfold model [34, 35] incorporates phylogenetic (evolution-
ary) information about the input alignments into a probabilistic
distribution of common secondary structures:
p (A | q ,T ) p (q | M )
p (pfold) (q | A) = p (pfold) (q | A,T , M ) = (6)
p (A | T , M )
where T is a phylogenetic tree, A is the input data (i.e., an align-
ment), M is a prior model for secondary structures (based on
SCFGs; see Note 3). Unless the original phylogenetic tree T is
obtained, T is taken to be the ML estimate of the tree, TML, given
the model M and the alignment A.
3.2.3 Averaged Predictions based on RNAalipffold and Pfold models tend to be

Probability Distribution affected by alignment errors in the input alignment. To address
of Each RNA Sequence this, averaged probability distributions of RNA sequences involved
in the input alignment were introduced by Kiryu et al. [32]. This
leads to an MEG estimator with the probability distribution
1
p (ave) (q | A) = å p(q | x)
n xÎA
(7)

where p(θ | x) is a probabilistic model for RNA secondary struc-
tures, for example, the McCaskill model [40], the CONTRAfold
model [9], the BL model [1–3], and others [56]. Note that nei-
ther covariation nor phylogenetic information about alignments is
considered in this probability distribution.
In [32], the authors utilized the McCaskill model as a proba-
bilistic model for individual RNA sequences (i.e., they used p(θ | x)
in Eq. 14), and showed that their method was more robust with
respect to alignment errors than RNAalifold and Pfold. Using
averaged probability distributions for RNA sequences in an input
alignment is compatible with Evaluation Procedure 1. See Hamada
et al. [23] for a detailed discussion.
3.2.4 Mixture of Several Hamada et al. [23] pointed out that arbitrary information can be
Distributions incorporated into common secondary structure predictions by uti-
lizing a mixture of probability distributions. For example, the
probability distribution
p (q | A) = w1 × p (pfold) (q | A) + w2 × p (alifold) (q | A) + w3 × p (ave) (q | A), (8)
where w1, w2, and w3 are positive values that satisfy w1 + w2 + w3 = 1 ,

includes covariation (Subheading 2.3), phylogenetic tree (Sub
heading 2.4), and majority rule (Subheading 2.5) information.
In CentroidAlifold [23], users can employ a mixed distribution
given by an arbitrary combination of RNAalipffold (Subheading 3.2.1),
Pfold (Subheading 3.2.2), and an averaged probability distribution
28 Michiaki Hamada
Fig. 4 Example of a predicted common secondary structure from the Centroid

Alifold Web Server [55] (http://www.ncrna.org/centroidfold). The input is a mul-
tiple sequence alignment of the traJ 5′ UTR. The color of a base-pair indicates
the averaged base-pairing probabilities (among RNA sequences in the align-
ment) of the base-pair, where warmer colors represent higher probabilities
(Subheading 3.2.3) based on the McCaskill or CONTRAfold model.

An example of a result from the CentroidAlifold Web Server is shown
in Fig. 4, in which colors of base-pairs indicate the marginal base-
pairing probability with respect to this mixture distribution. Hamada
et al. also showed that computation of MEG estimators with a mix-
ture distribution can be easily conducted by utilizing base-pairing
probability matrices (see also the next section), when certain gain
functions are employed. Moreover, computational experiments indi-
cated that CentroidAlifold with a mixture model is significantly bet-
ter than RNAalifold, Pfold, or McCaskill-MEA.
3.3 Choice of Gain A choice of the gain function in MEG estimators corresponds to the
Functions G(θ, y) decoding method (i.e., prediction of one final common secondary
structure from the distribution) of common RNA secondary struc-
tures, given a probabilistic model of common secondary structures.
3.3.1 The Kronecker A straightforward choice of the gain function is the Kronecker
Delta Function delta function:
ì1 if y is the exactly same as q

G (delta) (q , y ) ( = d (q , y ) ) = í (9)
î0 otherwise

The MEG estimator with this gain function is equivalent to the
“maximum likelihood estimator” (ML estimator) with respect to a
given probabilistic model for RNA common secondary structures
(cf. Subheading 3.2), which predicts the secondary structure with
the highest probability.
3.3.2 The γ-Centroid- It is known that the probability of the ML estimation is extremely
Type Gain Function [19] small, due to the immense number of secondary structures that
could be predicted; this fact is known as the “uncertainty” of the
solution, and often leads to issues in bioinformatics [17]. Because
the MEG estimator with the delta function considers only the solu-
tion with the highest probability, it is affected by this uncertainty.
A choice of gain function that partially overcomes this uncertainty
of solutions is the γ-centroid-type gain function [19]:
Gg(centroid) (q , y ) = å éëg I (qij = 1)I (y ij = 1) + I (qij = 0)I (y ij = 0)ùû (10)

i, j

where γ > 0 is a parameter that adjusts the relative importance of
the SEN and PPV of base-pairs in a predicted structure (i.e., a
larger γ produces more base-pairs in a predicted secondary struc-
ture). This gain function is motivated by the concept that more
true base-pairs (TP and TN) and fewer false base-pairs (FP and
FN) should be predicted [19] when the entire distribution of sec-
ondary structures is considered. Note that, when γ = 1, the gain
function is equivalent to that of the centroid estimator [8].
3.3.3 The CONTRAfold- In conventional RNA secondary structure predictions, another

Type Gain Function [9] gain function has been proposed (see Note 4), which is based on
the number of accurate predictions of every single position (not
base-pair) in an RNA sequence (see Note 5).The CONTRAfold-
type gain function is
|A | é ù
Gg( contra) (q , y ) = å ê g å I (qij* = 1)I (y ij* = 1) + ÕI (qij* = 0)I (y ij* = 0) ú (11)
i =1 ê
ë j : j ¹i j : j ¹i úû
where θ∗ and y∗ are symmetric extensions of θ and y, respectively

(i.e., θij∗ = θij for i < j and θij∗ = θji for j < i). This gain function is also
applicable to MEG estimators for common secondary structure
predictions. It should be emphasized that MEG estimators based
on the CONTRAfold-type gain function (for any γ > 0) do not
include centroid estimators, while the γ-centroid-type gain func-
tion includes the centroid estimator as a special case (i.e., γ = 1).
30 Michiaki Hamada
3.3.4 Remarks About From a theoretical viewpoint, the γ-centroid-type gain function is
Choice of Gain Function more appropriate for Evaluation Procedure 1 than either the delta
function or the CONTRAfold-type gain function (see Note 6),
which is also supported by several empirical (computational)
experiments. See Hamada et al. [23] for a detailed discussion.
Another choice of the gain function is MCC (or F-score),
which takes a balance between SEN and PPV of base-pairs, and the
MEG estimator with this gain function leads to an algorithm that
maximizes pseudo-expected accuracy [22]. In addition, the estima-
tor with this gain function includes only one parameter for predict-
ing secondary structure (see Note 7).
3.4 Computation The MEG estimator with the delta function (that predicts the
of Common Secondary common secondary structure with the highest probability with
Structure Through respect to a given probabilistic model) can be computed by employ-
MEG Estimators ing a CYK (Cocke–Younger–Kasami)-type algorithm. For exam-
ple, see [65] for details.
3.4.1 MEG Estimator
with Delta Function
3.4.2 MEG Estimator The MEG estimator with the γ-centroid-type (or CONTRAfold-
with γ-Centroid (or type) gain function is computed based on “base-pairing probability
CONTRAfold) Type Gain matrices” (BPPMs) (see Note 8) and Nussinov-style dynamic pro-
Function gramming (DP) [9, 23]:
ì M i +1, j
ï M i , j -1
ï
M i, j = max í M i +1, j -1 + Sij (12)
ï
ïmax éM i ,k + M k +1, j ù
î k ë û

where Mi, j is the optimal score of the subsequence xi⋯j and Sij is a
score computed from the BPPM(s). For instance, for the γ-centroid
estimator with the RNAalipffold model, the score Sij is equal to
Sij = (g + 1) pij(alipffold) - 1 where pij(alipffold) is the base-pairing probabil-
ity with respect to the RNAalipffold model. This DP algorithm
maximizes the sum of (base-pairing) probabilities pij(alipffold) which
are larger than 1 / (g + 1) , and requires O( | A | 3) time.
3.4.3 MEG Estimators The MEG estimator with an averaged probability distribution
with Averaged Probability (Subheading 3.2.3) can be computed by using averaged base-
Distributions pairing probabilities, {pij}i < j:
1
pij(ave) = å pij(x )
n xÎA
(13)

where
ì å I (qt (i )t ( j ) = 1) p (q | x ¢ ) if bothxi andx j are not gaaps

(x ) ï
p ij = íq ÎS (x ) (14)
ïî 0 otherwise

In the above, x′ is the RNA sequence given by removing gaps from
x and n is the number of sequences in the alignment A. The
function τ(i) returns the position in x′ corresponding to the posi-
tion i in x.
The common secondary structure of MEG estimator with the
γ-centroid gain function and the averaged probability distribution
are computed by using the DP recursion in Eq. 12 where
Sij = (g + 1) pij(ave) - 1 . This procedure has a time complexity of
O(n | A | 3) where n is the number of sequences in the alignment.
3.4.4 MEG Estimators The MEG estimator with a mixture of distribution (Subheading
with a Mixture Distribution 3.2.4) and the delta function (Subheading 3.3.1) cannot be com-
puted efficiently. However, if the γ-centroid-type (or CONTRAfold-
type) gain function is utilized, the prediction can be conducted
using a similar DP recursion to that in Eq. 12. For instance, the DP
recursion of the γ-centroid-type gain function with respect to Eq. 8
is equivalent to the one in Eq. 12 with Sij = (g + 1) pij* - 1 where
w3
pij* = w1 × pij(pfold) + w2 × pij(alipffold ) +
n
åp (x )
ij . (15)
x ÎA
In the above, pij(pfold) and pij(alipffold) are base-pairing probabilities
for the Pfold and RNAalipffold models, respectively, and {pij(x)} is a
base-pairing probability matrix with respect to a probabilistic
model for secondary structures of single RNA sequence x
(McCaskill or CONTRAfold model). Note that the total compu-
tational time of CentroidAlifold with a mixture of distributions still
remains O(n | A | 3).
3.4.5 MEG Estimators Using probability distributions of secondary structures with pseu-
with Probability doknots in MEG estimators generally has higher computational
Distribution Including cost [42]. To overcome this, for example, IPKnot [57] utilizes an
Pseudoknots approximated method for determining the probability distribution
as along with integer linear programming for predicting a final
common secondary structure.
3.5 A Classification Table 1 shows a comprehensive list of tools for common secondary
of Tools for Problem 1 structure prediction from aligned RNA sequences (in alphabetical
order within groups that do or do not consider pseudoknots
(see Note 9)). To the best of my knowledge, Table 1 is a complete
list of tools for Problem 1 as of 17 June 2013.
32 Michiaki Hamada
In Table 2, the tools in Table 1 are classified based on the

considerations of Subheadings 2 and 3. The classification leads
to much useful information: (1) the pros and cons of each tool;
(2) the similarity (or dissimilarity) among tools; (3) which tools are
more suited to Evaluation Procedure 1; and (4) a unified frame-
work within which to design algorithms for Problem 1. I believe
that the classification will bring a deeper understanding of each
tool, although several tools (which are not based on probabilistic
models and depend fundamentally on heuristic approaches) cannot
be classified in terms of MEG estimators.
3.6 Discussion Predicting a multiple sequence alignment (point estimation) from

unaligned sequences is not reliable because the probability of the
3.6.1 Multiple Sequence
alignment becomes extremely small. This is called the “uncer-
Alignment of RNA
tainty” of alignments which raises serious issues in bioinformatics
Sequences
[17]. In one Science paper [74], for instance, the authors argued
that the uncertainty of multiple sequence alignment greatly influ-
ences phylogenetic topology estimations: phylogenetic topologies
estimated from multiple alignments predicted by five widely used
aligners are different from one another. Similarly, point estimation
of multiple sequence alignment will greatly affect consensus sec-
ondary structure prediction.
In Problem 1, because the quality of the multiple sequence
alignment influences the prediction of common secondary struc-
ture, the input multiple alignment should be given by a multiple
aligner which is designed specifically for RNA sequences. Although
strict algorithms for multiple alignments taking into account sec-
ondary structures are equivalent to the Sankoff algorithm [54] and
have huge computational costs, several multiple aligners which are
fast enough to align long RNA sequences are available: these are
CentroidAlign [20, 76], R-coffee [71], PicXXA-R [53], DAFS
[58], and MAFFT [30]. In those multiple aligners, not only nucle-
otide sequences but also secondary structures are considered in the
alignment, and they are, therefore, suitable for generating input
multiple alignments for Problem 1.
Because the common secondary structure depends on mul-
tiple alignment, an approach adopted in RNAG [70] also seems
promising. This approach iteratively samples from the condi-
tional probability distributions P(Structure | Alignment) and
P(Alignment | Structure). Note, however, that RNAG does not
solve Problem 1 directly.
3.6.2 Improvement Although several studies have been conducted for RNA secondary
of RNA Secondary structure predictions for a single RNA sequence [9, 19, 38, 47],
Structure Predictions Using the accuracy is still limited, especially for long RNA sequences. By
Common Secondary employing comparative approaches using homologous sequence
Structure information, the accuracy of RNA secondary structure prediction
will be improved. In many cases, homologous RNA sequences of
the target RNA sequence are obtained, and someone would like to
know the common secondary structure of those sequences.
Gardner and Giegerich [13] introduced three approaches for
comparative analysis of RNA sequences, and common secondary
structure prediction is essentially utilized in the first of these.
However, if the aim is to improve the accuracy of secondary struc-
ture predictions, common secondary structure prediction is not
always the best solution, because it is not designed to predict the
optimal secondary structure of a specific target RNA sequence. If
you have a target RNA sequence for which the secondary structure
is to be predicted, the approach adopted by the CentroidHomfold
[21, 25] software is more appropriate than a method based on
common RNA secondary structure prediction.
3.6.3 How to Incorporate As shown in this review, there are two ways to incorporate several
Several Pieces pieces of information into an algorithm for common secondary
of Information structure prediction. The first approach is to modify the (internal)
in Algorithms algorithm itself in order to handle the additional information. For
example, PhyloRNAalifold [14] incorporates phylogenetic infor-
mation into the RNAalifold algorithm by modifying the internal
algorithm and PPfold [63] modifies the Pfold algorithm to handle
experimental information. The drawbacks of this approach are the
relatively large implementation cost and the heuristic combination
of the information.
On the other hand, another approach adopted in
CentroidAlifold [23] is promising because it can easily incorporate
many pieces of information into predictions if a base-pairing prob-
ability matrix is available. Because the approach depends on only
base-pairing probability matrices, and does not depend on the
detailed design of the algorithm, it is easy to implement an algo-
rithm using a mixture of distributions.
Moreover, a method to update a base-pairing probability
matrix (computed using sequence information only) which incor-
porates experimental information [16] has recently been proposed.
The method is independent of the probabilistic models of RNA
secondary structures, and is suitable for incorporating experimen-
tal information into common RNA secondary structure predic-
tion. A more sophisticated method by Washietl et al. [68] can also
be used to incorporate experimental information into common
secondary structure predictions, because it produces a BPPM that
takes experimental information into account.
3.6.4 A Problem that Is Problem 1, which is considered in this paper, can be extended to
Mathematically Related predictions of RNA–RNA interactions, another important task in
to Problem 1 RNA bioinformatics (e.g., [29, 50]).
Problem 2 (Common Joint Structure Predictions of Two Aligned

RNA Sequences)
34 Michiaki Hamada
Given two multiple alignments A1 and A2 of RNA sequences, then

predict a joint secondary structure between A1 and A2.
Because the mathematical structure of Problem 2 is similar to
that of Problem 1, the ideas utilized in designing the algorithms
for common secondary structure predictions can be adopted in the
development of methods for this new problem. In fact, Seemann
et al. [60, 61] have employed similar idea, adapting the PETfold
algorithm to Problem 2 (implemented in the PETcofold software).
Note that the problem of (pairwise) alignment between two mul-
tiple alignments (cf. see [24] for the details) has a similar mathe-
matical structure to Problem 1.
3.7 Conclusion In this review, I focused on RNA secondary structure predictions

from aligned RNA sequences, in which a secondary structure
whose length is equal to the length of the input alignment is pre-
dicted. A predicted common secondary structure is useful not only
for further functional analyses of the ncRNAs being studied but
also for improving RNA secondary structure predictions and for
finding ncRNAs in genomes. In this review, I systematically classi-
fied existing algorithms on the basis of (1) the information utilized
in the algorithms and (2) the corresponding MEG estimators,
which consist of a gain function and a probability distribution of
common secondary structures. This classification will provide a
deeper understanding of each algorithm.
4 Notes
1. Reference common secondary structures are available for only

reference multiple sequence alignments in the Rfam database
[7] (http://rfam.sanger.ac.uk/).
2. The gain G(θ, y) is equal to the accuracy measure Acc(θ, y) for
a prediction and references, so the MEG estimator maximizes
the expected accuracy under a given probabilistic distribution.
3. The SCFG is based on the rules S → LS | L, F → dFd | LS,
L → s | dFd.
4. Historically, the CONTRAfold-type gain function was pro-
posed earlier than the γ-centroid-type gain function.
5. This is not consistent with Evaluation Procedure 1, because
accurate predictions of base-pairs with respect to reference
structures are evaluated in it.
6. The CONTRAfold-type gain function has a bias toward accu-
rate predictions of base-pairs, compared to the γ-centroid-type
gain function.
7. The γ-centroid-type and CONTRAfold-type gain functions
contain a parameter adjusting the ratio of SEN and PPV for a
predicted secondary structure.
8. The BPPM is a probability matrix {pij} in which pij is the

marginal probability that the ith base xi and the jth base xj
form a base-pair with respect to a given probabilistic distribu-
tion of secondary structures. For many probabilistic models,
including the McCaskill model and the CONTRAfold model,
the BPPM for a given sequence can be computed efficiently by
utilizing inside–outside algorithms. See [40] for the details.
9. Tools for predicting common secondary structures without
pseudoknots are much faster than those for predicting second-
ary structures with pseudoknots.
Acknowledgement
This work was supported in part by MEXT KAKENHI (Grant-in-

Aid for Young Scientists (A): 24680031; Grant-in-Aid for Scientific
Research (A): 25240044).
References
1. Andronescu M, Condon A, Hoos HH, 9. Do CB, Woods DA, Batzoglou S (2006)
Mathews DH, Murphy KP (2007) Efficient CONTRAfold: RNA secondary structure pre-
parameter estimation for RNA secondary struc- diction without physics-based models.
ture prediction. Bioinformatics 23(13):19–28 Bioinformatics 22(14):e90–e98
2. Andronescu M, Condon A, Hoos HH, Mathews 10. Dowell RD, Eddy SR (2004) Evaluation of
DH, Murphy KP (2010) Computational several lightweight stochastic context-free
approaches for RNA energy parameter estima- grammars for RNA secondary structure predic-
tion. RNA 16(12):2304–2318 tion. BMC Bioinform 5:71
3. Andronescu MS, Pop C, Condon AE (2010) 11. Esteller M (2011) Non-coding RNAs in human
Improved free energy parameters for RNA disease. Nat Rev Genet 12(12):861–874
pseudoknotted secondary structure prediction. 12. Freyhult E, Moulton V, Gardner P (2005)
RNA 16(1):26–42 Predicting RNA structure using mutual infor-
4. Balik A, Penn AC, Nemoda Z, Greger IH mation. Appl Bioinform 4(1):53–59
(2013) Activity-regulated RNA editing in 13. Gardner PP, Giegerich R (2004) A comprehen-
select neuronal subfields in hippocampus. Nucl sive comparison of comparative RNA structure
Acids Res 41(2):1124–1134 prediction approaches. BMC Bioinform 5:140
5. Bernhart SH, Hofacker IL, Will S, Gruber AR, 14. Ge P, Zhang S (2013) Incorporating
Stadler PF (2008) RNAalifold: improved con- phylogenetic-based covarying mutations into
sensus structure prediction for RNA align- RNAalifold for RNA consensus structure pre-
ments. BMC Bioinform 9:474 diction. BMC Bioinform 14(1):142
6. Bindewald E, Shapiro BA (2006) RNA second- 15. Gruber AR, Findeiss S, Washietl S, Hofacker
ary structure prediction from sequence align- IL, Stadler PF (2010) Rnaz 2.0: improved
ments using a network of k-nearest neighbor noncoding RNA detection. Pac Symp
classifiers. RNA 12(3):342–352 Biocomput 15:69–79
7. Burge SW, Daub J, Eberhardt R, Tate J, 16. Hamada M (2012) Direct updating of an RNA
Barquist L, Nawrocki EP, Eddy SR, Gardner base-pairing probability matrix with marginal
PP, Bateman A (2013) Rfam 11.0: 10 years of probability constraints. J Comput Biol 19(12):
RNA families. Nucl Acids Res 41(Database 1265–1276
issue):D226–D232 17. Hamada M (2014) Fighting against uncer-
8. Carvalho LE, Lawrence CE (2008) Centroid tainty: an essential issue in bioinformatics.
estimation in discrete high-dimensional spaces Briefings Bioinform 15(5):748–767
with applications in biology. Proc Natl Acad Sci 18. Hamada M, Asai K (2012) A classification of
USA 105(9):3209–3214 bioinformatics algorithms from the viewpoint
36 Michiaki Hamada
of maximizing expected accuracy (MEA). J 32. Kiryu H, Kin T, Asai K (2007) Robust predic-
Comput Biol 19(5):532–549 tion of consensus secondary structures using
19. Hamada M, Kiryu H, Sato K, Mituyama T, averaged base-pairing probability matrices.
Asai K (2009) Prediction of RNA secondary Bioinformatics 23(4):434–441
structure using generalized centroid estima- 33. Klein RJ, Eddy SR (2003) RSEARCH: finding
tors. Bioinformatics 25(4):465–473 homologs of single structured RNA sequences.
20. Hamada M, Sato K, Kiryu H, Mituyama T, BMC Bioinform 4:44
Asai K (2009) CentroidAlign: fast and accurate 34. Knudsen B, Hein J (1999) RNA secondary
aligner for structured RNAs by maximizing structure prediction using stochastic context-
expected sum-of-pairs score. Bioinformatics free grammars and evolutionary history.
25(24):3236–3243 Bioinformatics 15(6):446–454
21. Hamada M, Sato K, Kiryu H, Mituyama T, 35. Knudsen B, Hein J (2003) Pfold: RNA sec-
Asai K (2009) Predictions of RNA secondary ondary structure prediction using stochastic
structure by combining homologous sequence context-free grammars. Nucl Acids Res 31(13):
information. Bioinformatics 25(12):i330–i338 3423–3428
22. Hamada M, Sato K, Asai K (2010) Prediction 36. Low JT, Weeks KM (2010) SHAPE-directed
of RNA secondary structure by maximizing RNA secondary structure prediction. Methods
pseudo-expected accuracy. BMC Bioinform 52(2):150–158
11:586 37. Luck R, Graf S, Steger G (1999) ConStruct: a
23. Hamada M, Sato K, Asai K (2011) Improving tool for thermodynamic controlled prediction
the accuracy of predicting secondary structure of conserved secondary structure. Nucl Acids
for aligned RNA sequences. Nucl Acids Res Res 27(21):4208–4217
39(2):393–402 38. Mathews DH, Sabina J, Zuker M, Turner DH
24. Hamada M, Kiryu H, Iwasaki W, Asai K (2011) (1999) Expanded sequence dependence of
Generalized centroid estimators in bioinfor- thermodynamic parameters improves predic-
matics. PLoS ONE 6(2):e16450 tion of RNA secondary structure. J Mol Biol
25. Hamada M, Yamada K, Sato K, Frith MC, Asai 288(5):911–940
K (2011) CentroidHomfold-LAST: accurate 39. Mathews DH, Disney MD, Childs JL,
prediction of RNA secondary structure using Schroeder SJ, Zuker M, Turner DH (2004)
automatically collected homologous sequences. Incorporating chemical modification con-
Nucl Acids Res 39(Web Server issue): straints into a dynamic programming algorithm
W100–W106 for prediction of RNA secondary structure.
26. Hofacker IL (2007) RNA consensus structure Proc Natl Acad Sci USA 101(19):7287–7292
prediction with RNAalifold. Methods Mol Biol 40. McCaskill JS (1990) The equilibrium partition
395:527–544 function and base-pair binding probabilities for
27. Hofacker IL, Fekete M, Stadler PF (2002) RNA secondary structure. Biopolymers 29(6–7):
Secondary structure prediction for aligned RNA 1105–1119
sequences. J Mol Biol 319(5):1059–1066 41. Meer EJ, Wang DO, Kim S, Barr I, Guo F,
28. Jager D, Pernitzsch SR, Richter AS, Backofen Martin KC (2012) Identification of a cis-acting
R, Sharma CM, Schmitz RA (2012) An element that localizes mrna to synapses. Proc
archaeal sRNA targeting cis- and trans-encoded Natl Acad Sci 109(12):4639–4644
mRNAs via two distinct domains. Nucl Acids 42. Nebel ME, Weinberg F (2012) Algebraic and
Res 40(21):10964–10979 combinatorial properties of common RNA
29. Kato Y, Sato K, Hamada M, Watanabe Y, Asai pseudoknot classes with applications. J Comput
K, Akutsu T (2010) RactIP: fast and accurate Biol 19(10):1134–1150
prediction of RNA-RNA interaction using 43. Novikova IV, Hennelly SP, Sanbonmatsu KY
integer programming. Bioinformatics 26(18): (2012) Structural architecture of the human
i460–i466 long non-coding RNA, steroid receptor RNA
30. Katoh K, Toh H (2008) Improved accuracy of activator. Nucl Acids Res 40(11):5034–5051
multiple ncRNA alignment by incorporating 44. Pang PS, Elazar M, Pham EA, Glenn JS (2011)
structural information into a MAFFT-based Simplified RNA secondary structure mapping
framework. BMC Bioinform 9:212 by automation of SHAPE data analysis. Nucl
31. Kertesz M, Wan Y, Mazor E, Rinn JL, Nutter Acids Res 39(22):e151
RC, Chang HY, Segal E (2010) Genome-wide 45. Pauli A, Rinn JL, Schier AF (2011) Non-
measurement of RNA secondary structure in coding RNAs as regulators of embryogenesis.
yeast. Nature 467(7311):103–107 Nat Rev Genet 12(2):136–149
46. Penn AC, Balik A, Greger IH (2013) Steric information for RNA folding of multiple align-
antisense inhibition of AMPA receptor Q/R ments. Nucl Acids Res 36(20):6355–6362
editing reveals tight coupling to intronic edit- 60. Seemann SE, Menzel P, Backofen R, Gorodkin
ing sites and splicing. Nucl Acids Res 41(2): J (2011) The PETfold and PETcofold web
1113–1123 servers for intra- and intermolecular structures
47. Proctor JR, Meyer IM (2013) COFOLD: an of multiple RNA sequences. Nucl Acids Res
RNA secondary structure prediction method 39(Web Server issue):W107–W111
that takes co-transcriptional folding into 61. Seemann SE, Richter AS, Gesell T, Backofen
account. Nucl Acids Res 41(9):e102 R, Gorodkin J (2011) PETcofold: predicting
48. Puton T, Kozlowski LP, Rother KM, Bujnicki conserved interactions and structures of two
JM (2013) CompaRNA: a server for continu- multiple alignments of RNA sequences.
ous benchmarking of automated methods for Bioinformatics 27(2):211–219
RNA secondary structure prediction. Nucl 62. Spirollari J, Wang JT, Zhang K, Bellofatto V, Park
Acids Res 41(7):4307–4323 Y, Shapiro BA (2009) Predicting consensus
49. Qureshi IA, Mehler MF (2012) Emerging structures for RNA alignments via pseudo-energy
roles of non-coding RNAs in brain evolution, minimization. Bioinform Biol Insights 3:51–69
development, plasticity and disease. Nat Rev 63. Sukosd Z, Knudsen B, Kjems J, Pedersen CN
Neurosci 13(8):528–541 (2012) PPfold 3.0: fast RNA secondary struc-
50. Richter AS, Backofen R (2012) Accessibility ture prediction using phylogeny and auxiliary
and conservation: general features of bacterial data. Bioinformatics 28(20):2691–2692
small RNA–mRNA interactions? RNA Biol 64. Underwood JG, Uzilov AV, Katzman S,
9(7):954–965 Onodera CS, Mainzer JE, Mathews DH, Lowe
51. Rivas E, Lang R, Eddy SR (2012) A range of TM, Salama SR, Haussler D (2010) FragSeq:
complex probabilistic models for RNA secondary transcriptome-wide RNA structure probing
structure prediction that includes the nearest- using high-throughput sequencing. Nat
neighbor model and more. RNA 18(2):193–212 Methods 7(12):995–1001
52. Ruan J, Stormo GD, Zhang W (2004) An iter- 65. Washietl S, Hofacker IL (2004) Consensus
ated loop matching approach to the prediction folding of aligned sequences as a new measure
of RNA secondary structures with pseu- for the detection of functional RNAs by com-
doknots. Bioinformatics 20(1):58–66 parative genomics. J Mol Biol 342(1):19–30
53. Sahraeian SM, Yoon BJ (2011) PicXAA-R: effi- 66. Washietl S, Hofacker IL, Lukasser M,
cient structural alignment of multiple RNA Huttenhofer A, Stadler PF (2005) Mapping of
sequences using a greedy approach. BMC conserved RNA secondary structures predicts
Bioinform 12(Suppl 1):S38 thousands of functional noncoding RNAs in
54. Sankoff D (1985) Simultaneous solution of the the human genome. Nat Biotechnol 23(11):
RNA folding alignment and protosequence 1383–1390
problems. SIAM J Appl Math 45:810–825 67. Washietl S, Hofacker IL, Stadler PF (2005) Fast
55. Sato K, Hamada M, Asai K, Mituyama T and reliable prediction of noncoding RNAs.
(2009) CENTROIDFOLD: a web server for Proc Natl Acad Sci USA 102(7):2454–2459
RNA secondary structure prediction. Nucl 68. Washietl S, Hofacker IL, Stadler PF, Kellis M
Acids Res 37(Web Server issue):W277–W280 (2012) RNA folding with soft constraints: rec-
56. Sato K, Hamada M, Mituyama T, Asai K, onciliation of probing data and thermodynamic
Sakakibara Y (2010) A non-parametric bayesian secondary structure prediction. Nucl Acids Res
approach for predicting rna secondary struc- 40(10):4261–4272
tures. J Bioinform Comput Biol 8(4):727–742 69. Watts JM, Dang KK, Gorelick RJ, Leonard
57. Sato K, Kato Y, Hamada M, Akutsu T, Asai K CW, Bess JW, Swanstrom R, Burch CL, Weeks
(2011) IPknot: fast and accurate prediction of KM (2009) Architecture and secondary struc-
RNA secondary structures with pseudoknots ture of an entire HIV-1 RNA genome. Nature
using integer programming. Bioinformatics 460(7256):711–716
27(13):85–93 70. Wei D, Alpert LV, Lawrence CE (2011)
58. Sato K, Kato Y, Akutsu T, Asai K, Sakakibara Y RNAG: a new Gibbs sampler for predicting
(2012) DAFS: simultaneous aligning and fold- RNA secondary structure for unaligned
ing of RNA sequences via dual decomposition. sequences. Bioinformatics 27(18):2486–2493
Bioinformatics 28(24):3218–3224 71. Wilm A, Higgins DG, Notredame C (2008)
59. Seemann SE, Gorodkin J, Backofen R (2008) R-Coffee: a method for multiple alignment of
Unifying evolutionary and thermodynamic non-coding RNA. Nucl Acids Res 36(9):e52
38 Michiaki Hamada

72. Wilm A, Linnenbrink K, Steger G (2008) 75. Xia T, SantaLucia J, Burkard ME, Kierzek R,
ConStruct: improved construction of RNA Schroeder SJ, Jiao X, Cox C, Turner DH
consensus structures. BMC Bioinform 9:219 (1998) Thermodynamic parameters for an

73. Witwer C, Hofacker IL, Stadler PF (2004) expanded nearest-neighbor model for forma-
Prediction of consensus RNA secondary struction of RNA duplexes with Watson–Crick base-
tures including pseudoknots. IEEE/ACM pairs. Biochemistry 37(42):14719–14735
Trans Comput Biol Bioinform 1(2):66–77 76. Yonemoto H, Asai K, Hamada M (2013)

74. Wong KM, Suchard MA, Huelsenbeck JP CentroidAlign-Web: a fast and accurate multi-
(2008) Alignment uncertainty and genomic ple aligner for long non-coding RNAs. Int J
analysis. Science 319(5862):473–476 Mol Sci 14(3):6144–6156
Chapter 3
A Simple Protocol for the Inference of RNA Global

Pairwise Alignments
Eugenio Mattei, Manuela Helmer-Citterich, and Fabrizio Ferrè
Abstract
RNA alignment is an important step in the annotation and characterization of unknown RNAs, and several
methods have been developed to meet the need of fast and accurate alignments. Being the performances
of the aligning methods affected by the input RNA features, finding the most suitable method is not trivial.
Indeed, no available method clearly outperforms the others. Here we present a simple workflow to help
choosing the more suitable method for RNA pairwise alignment. We tested the performances of six algo-
rithms, based on different approaches, on datasets created by merging publicly available datasets of known
or curated RNA secondary structure annotations with datasets of curated RNA alignments. Then, we
simulated the frequent case where the secondary structure is unknown by using the same alignment data-
sets but ignoring the known structure and instead predicting it. In conclusion, the proposed workflow for
pairwise RNA alignment depends on the input RNA primary sequence identity and the availability of reli-
able secondary structures.
Key words RNA alignment, RNA structure comparison, RNA sequence–structure relationship, RNA
functional annotation, Computational biology
1 Introduction
RNAs are complex molecules that can fold, by nucleotide pairing,

into intricate secondary and tertiary structures, which play a lead-
ing role in influencing their function. Accordingly, the need for
instruments for a structure-based functional characterization is
becoming more pressing. In recent years, the development of new
sequencing technologies [1] led to the discovery of several novel
RNA transcript variants and entirely new classes of ncRNAs [2, 3].
The characterization and annotation of these molecules is in gen-
eral facilitated by the search for similarities among annotated
RNAs. Yet, RNAs are more subject to structure rather than
sequence constraints in order to preserve their secondary structure
upon genomic variations [4], and therefore sequence-only
comparison can be often misleading when comparing homologous
39
40 Eugenio Mattei et al.
but divergent RNA molecules. It was demonstrated, indeed, that

the inclusion of secondary structure information is mandatory to
improve comparison results when sequence identity drops under
50–60 % [5]. Hence, the search for evolutionary structural signa-
tures inspired the development of many comparison tools [6].
Structure comparison plays an important role not only in RNA
alignment and classification but also in ncRNA annotation, motifs
finding and phylogenetic inference. Despite its importance, the
structural alignment of RNAs is still an open issue. Several align-
ment tools have been developed using different kinds of approaches.
Sankoff’s dynamic programming algorithm [7] represents the pri-
mary reference for structural RNA alignment but its high compu-
tational complexity (O(N6) in time and O(N4) in space) motivated
the search for more efficient algorithms. Lower complexity has
been reached by different methods using for example dynamic pro-
gramming [8–10], formal grammars [11], genetic algorithms [12,
13], and tree-based encodings [14, 15].
Given two RNAs to be aligned, the selection of the best proce-
dure is not trivial, given also the absence in the current literature of
a thorough benchmark of the available algorithms. Herein, we
propose a simple procedure that can guide the selection of the
most reliable approach, which is motivated by testing several popu-
lar algorithms on different datasets and conditions. As shown by
our results, the discriminating factors are the sequence identity of
the two input RNAs and the availability of reliable secondary
structures.
2 Materials
2.1 Computational LocARNA [10] uses a folding and aligning strategy to perform pair-
Methods for RNA wise alignments. The latest version (v. 1.7.10) can be downloaded
Alignment Tested from http://www.bioinf.uni-freiburg.de/Software/LocARNA/
in This Work and requires the Vienna package [16] that can be downloaded from
http://www.tbi.univie.ac.at/~ronny/RNA/index.html. LocARNA
can be also accessed through a Web interface (http://rna.informa-
tik.uni-freiburg.de:8080/LocARNA/Input.jsp). Detailed installa-
tion instructions are included in the package (see Note 1).
LocARNA uses a two-step procedure to align RNAs. In the first
place, a base pair probability matrix for each input sequence is com-
puted using RNAfold (included in the Vienna package). Then the
two matrices are used as a guide to find the optimal alignment.
RNA StrAT [15] employs a tree-based strategy to perform
pairwise structural alignments. The latest version (v. 7.1) is
available upon request; instructions and contact information can
be found at http://www-lbit.iro.umontreal.ca/rnastrat/. RNA
StrAT uses a three-step procedure to align RNA secondary struc-
tures. Firstly, the two input secondary structures are broken down
into stems and stem-loop structures and then these substructures
RNA Global Pairwise Alignment 41
are compared using a variant of the tree edit distance. The scores
obtained from the previous calculation are then used to guide the
alignment of the stem and stem-loop structures. During the align-
ment, also the nucleotidic sequence is taken into account to
improve the alignment performances.
needle from the EMBOSS package [17] performs RNA
sequence alignments using an implementation of the Needleman–
Wunsch algorithm with affine gaps. needle can be downloaded
from http://emboss.open-bio.org/html/adm/ch01s01.html
(see Note 2); a detailed guide is included in the Web page.
2.2 Test Datasets We built three datasets to test the alignment performances of the
six selected alignment methods. We retrieved curated secondary
structures from the RNA STRAND [18] and RNAspa [19] datas-
ets. RNA STRAND integrates information about known RNA sec-
ondary structure of any type and from different organisms,
retrieved from several public databases. Instead, the RNAspa data-
set is a collection of curated secondary structures from Rfam.
Datasets of curated sequence alignments were retrieved from
RNASTAR [20] and bralibase II [5]. Bralibase II is a collection of
RNA alignment datasets proposed for benchmarking of alignment
algorithms. Among the available datasets supplied by bralibase, we
selected the dataset 2 including pairwise tRNA alignments.
RNASTAR includes refined Rfam alignments that were manually
curated using structural information from the Protein Data Bank,
PDB [21]. Since these curated datasets of alignments do not pro-
vide secondary structure annotation, we combined together struc-
tural datasets and secondary structure datasets to obtain a collection
of curated alignments of known RNA structures. More specifically,
we used RNA STRAND secondary structure for RNASTAR align-
ments, RNAspa secondary structure for bralibase alignments, and
finally the remaining RNAs in RNAspa for Rfam alignments.
The sum-of-pairs (SPS) score [5] was employed as measure to
evaluate the performances of the alignment methods. SPS is
defined as the number of correct pairs (aligned nucleotide pairs
found in the reference alignment, i.e., the correctly aligned posi-
tions) over the total number of predicted pairs, and it can be con-
sidered as a measure of the sensitivity of a method. An SPS score of
0 indicates two completely different alignments (i.e., the alignment
reconstructed from the algorithm is completely wrong); conversely,
a score of 1 indicates that the reconstructed alignment is identical
to the reference alignment.
2.3 Hardware No specific requirements are needed to handle and align RNA
Requirements sequences and structures. The proposed methods work well on
normal desktop computers running Linux. Nevertheless all the
methods described in this paper have a Web-based interface that
can be used instead of the command line version.
3 Methods
In order to select the best alignment tool to use according to the

sequence identity of the two input RNAs and the availability of
their secondary structure, we tested six different aligners using
three different datasets. As a result, we propose a workflow that
can be used as a guideline for the alignment of two given RNAs
(Fig. 1).
3.1 Comparing We run six different RNA alignment tools on the datasets pre-
Algorithms sented in Subheading 2.2, namely needle, LocARNA, RNA StrAT,
Performance RNAforester (included in Vienna package), RNAdistance (also
included in Vienna package), and gardenia [14]. These algorithms
can be divided into three classes according to their approach:
sequence-based (needle), folding and aligning (LocARNA),
tree-based (all the others). We used needle as a measure of how
much the secondary structure is important for the alignment of
two RNAs; the less the sequence-based alignment is correct, the
more the secondary structure can give a positive contribution in
guiding the alignment.
Figure 2 shows the results of the structure-based aligners
at different levels of sequence identity, measured as SPS (see
Subheading 2.2). Specifically, when sequence identity is lower than
Fig. 1 The proposed workflow, showing the required steps to find the best align-
ment approach according to the features of the input RNA sequences
Fig. 2 Alignment performances of the six tested methods on datasets composed by curated structures and
alignments. Employed methods are needle, LocARNA, RNA StrAT, gardenia, RNAforester, RNAdistance.
Alignment accuracy is evaluated as the sum-of-pairs (SPS, described in Subheading 2.2), at different intervals
of sequence identity as measured using needle
70–80 % RNA StrAT shows the best performances, while above

70–80 % LocARNA is the most accurate (see Note 3). Moreover,
we run the same test using predicted secondary structures instead
of curated ones in order to evaluate the performances in a case of
unknown secondary structure, which the most frequent setting
(Fig. 3). In this test, the reference alignment is used as before, but
the input structures passed to the alignment algorithms are pre-
dicted by RNA fold (included in Vienna package) instead of using
the real or curated ones. Even if databases of RNA known struc-
tures are growing in size, in the majority of cases the user has only
access to their sequences. Our results show that, when the struc-
ture is predicted, structural information leads to alignments that
are worst than the ones obtained with a simple sequence align-
ment. LocARNA represents an exception because rather than using
directly the predicted secondary structure, we run it using only
sequence information and let it compute secondary structures
while aligning. In conclusion, when the secondary structure is
known we suggest using LocARNA for RNAs sharing more
than 70–80 % sequence identity, or otherwise use RNA StrAT.
Fig. 3 Alignment performances of the six methods using predicted secondary structures obtained using
RNAfold
For the more frequent cases where structural annotations are not
available, LocARNA is the tool of choice, and sequence alignment
can be also sufficient depending on the user needs.
3.2 Finding The command to perform a sequence alignment with needle is:
Sequence Identity
needle -asequence rna1 -bsequence rna2
with Needle
where rna1 and rna2 are two plain text files containing the input
sequences. The two input files can be in different formats; below is
shown the widely used fasta format:
>S1
ACCAGGUGAAAUUCCUGGACCGACGGUUAAAGUCCGG
The output is printed on screen but it can be saved into a file
(in different formats) using the following command:
needle -asequence rna1 -bsequence rna2 -outfile out
3.3 Aligning Using The command to perform an alignment with LocARNA is:
LocARNA
locarna rna1 rna2
where rna1 and rna2 are two plain text files containing the input
sequences (and, optionally, their secondary structure, or a set of
structural constraints, as described later). The two input files must
be formatted as shown below:
S1 ACCAGGUGAAAUUCCUGGACCGACGGUUAAAGUCCGG
Secondary structure annotation can be supplied in the stan-
dard dot-bracket notation as a new line. The input files must then
be modified as shown below:
S1 ACCAGGUGAAAUUCCUGGACCGACGGUUAAAGUCCGG
#S .(((((.......))))).(((((.......)).)))
The secondary structure is used by LocARNA to compute a
base pair probability matrix satisfying structural constraints.
Among the parameters that can help in tuning the alignment,
LocARNA allows the user to choose the substitution matrix with
the switch ‘--ribosum-file=<path_to_the_matrix>’; default is the
RIBOSUM85-60 matrix. Another useful parameter is ‘--clustal=
<file>’ that save the output of the program using ClustalW [22]
format in the file ‘my_out’. By default the output is printed on
screen.
3.4 Aligning Using The command to run RNA StrAT is:

RNA StrAT
rnastrap rna1 rna2
where input plain text files rna1 and rna2 must have the format
reported below:
>S1
ACCAGGUGAAAUUCCUGGACCGACGGUUAAAGUCCGG
.(((((.......))))).(((((.......)).)))
No further parameters are available in the package.
4 Notes
1. LocARNA needs a Vienna installation to work. By using the

‘—enable-librna’ switch along with the ‘./configure’
LocARNA will search the Vienna installation in the default
path. If your Vienna installation is not in the default path you
should use the ‘—enable-librna PATH_TO_VIENNA’ switch.
For the sake of simplicity we suggest to install Vienna
package by adding the Linux repository.
2. EMBOSS package in UBUNTU system is already included in
the repository.
3. Table 1 shows the average computational time per sequence for
each dataset and method. The computational time generally
depends on the sequence length, which is different in the three
Table 1
Mean running time (in seconds) per sequence
bralibase RNAspa RNAstrand

needle 0.01 0.01 0.01
gardenia 0.01 0.01 0.02
RNA StrAT 0.02 0.32 0.46
LocARNA 0.02 0.08 0.02
RNAdistance 0.01 0.01 0.01
RNAforester 0.03 0.73 0.81
datasets (bralibase: average length 76 nucleotides; RNAspa:

145 nucleotides; RNAstrand: 119 nucleotides). Other input
RNA characteristics can also influence the running time of
some algorithms, for example their structural complexity
(i.e., how many secondary sub-structures the RNA contains).
Acknowledgements
This work was supported by the EPIGEN flagship project and

PRIN 2010 (prot. 20108XYHJS_006) to MHC.
References
1. Mercer TR, Gerhardt DJ, Dinger ME et al programs upon structural RNAs. Nucleic Acids
(2012) Targeted RNA sequencing reveals the Res 33:2433–2439. doi:10.1093/nar/gki541
deep complexity of the human transcriptome. 6. Wan Y, Kertesz M, Spitale RC et al (2011)
Nat Biotechnol 30:99–104. doi:10.1038/ Understanding the transcriptome through
nbt.2024 RNA structure. Nat Rev Genet 12:641–655.
2. Cabili MN, Trapnell C, Goff L et al (2011) doi:10.1038/nrg3049
Integrative annotation of human large inter- 7. Sankoff D (1985) Simultaneous solution of the
genic noncoding RNAs reveals global proper- RNA folding, alignment and protosequence
ties and specific subclasses. Genes Dev problems. SIAM J Appl Math 45:810–825
25:1915–1927. doi:10.1101/gad.17446611 8. Havgaard JH, Torarinsson E, Gorodkin J
3. Baker M (2011) Long noncoding RNAs: the (2007) Fast pairwise structural RNA alignments
search for function. Nat Methods 8:379–383. by pruning of the dynamical programming
doi:10.1038/nmeth0511-379 matrix. PLoS Comput Biol 3:1896–1908.
4. Burge SW, Daub J, Eberhardt R et al (2013) doi:10.1371/journal.pcbi.0030193
Rfam 11.0: 10 years of RNA families. Nucleic 9. Harmanci AO, Sharma G, Mathews DH
Acids Res 41:D226–D232. doi:10.1093/nar/ (2007) Efficient pairwise RNA structure pre-
gks1005 diction using probabilistic alignment con-
5. Gardner PP, Wilm A, Washietl S (2005) A straints in Dynalign. BMC Bioinformatics
benchmark of multiple sequence alignment 8:130. doi:10.1186/1471-2105-8-130
10. Will S, Reiche K, Hofacker IL et al (2007) 17. Rice P, Longden I, Bleasby A (2000)
Inferring noncoding RNA families and classes EMBOSS: the European molecular biology
by means of genome-scale structure-based open software suite. Trends Genet 16:
clustering. PLoS Comput Biol 3:e65. 276–277
doi:10.1371/journal.pcbi.0030065 18. Andronescu M, Bereg V, Hoos HH, Condon
11. Dowell RD, Eddy SR (2006) Efficient pair- A (2008) RNA STRAND: the RNA secondary
wise RNA structure prediction and align- structure and statistical analysis database.
ment using sequence alignment constraints. BMC Bioinformatics 9:340. doi:10.1186/1471-
BMC Bioinformatics 7:400. doi:10.1186/ 2105-9-340
1471-2105-7-400 19. Horesh Y, Doniger T, Michaeli S, Unger R
12. Taneda A (2010) Multi-objective pairwise RNA (2007) RNAspa: a shortest path approach for
sequence alignment. Bioinformatics 26:2383– comparative prediction of the secondary structure
2390. doi:10.1093/bioinformatics/btq439 of ncRNA molecules. BMC Bioinformatics 8:
13. Notredame C, Higgins DG (1996) SAGA: 366. doi:10.1186/1471-2105-8-366
sequence alignment by genetic algorithm. 20. Widmann J, Stombaugh J, McDonald D et al
Nucleic Acids Res 24:1515–1524 (2012) RNASTAR: an RNA STructural
14. Blin G, Denise A, Dulucq S et al (2007) Alignment Repository that provides insight
Alignments of RNA structures. IEEE ACM into the evolution of natural and artificial
Trans Comput Biol Bioinformatics 7:309– RNAs. RNA 18:1319–1327. doi:10.1261/
322. doi:10.1109/TCBB.2008.28 rna.032052.111
15. Guignon V, Chauve C, Hamel S (2005) An 21. Berman HM, Kleywegt GJ, Nakamura H,
edit distance between RNA stem-loops. In: Markley JL (2013) The future of the protein
Consens MP, Navarro G (eds) SPIRE. Springer, data bank. Biopolymers 99:218–222.
Heidelberg, pp 335–347 doi:10.1002/bip.22132
16. Lorenz R, Bernhart SH, Höner Zu Siederdissen 22. Larkin MA, Blackshields G, Brown NP et al
C et al (2011) ViennaRNA Package 2.0. (2007) Clustal W and Clustal X version 2.0.
Algorithms Mol Biol 6:26. doi:10.1186/ Bioinformatics 23:2947–2948. doi:10.1093/
1748-7188-6-26 bioinformatics/btm404
Chapter 4
De Novo Secondary Structure Motif Discovery

Using RNAProfile
Abstract
RNA secondary structure plays critical roles in several biological processes. For example, many trans-acting
noncoding RNA genes and cis-acting RNA regulatory elements present functional motifs, conserved both
in structure and sequence, that can be hardly detected by primary sequence analysis alone. We describe
here how conserved secondary structure motifs shared by functionally related RNA sequences can be
detected through the software tool RNAProfile. RNAProfile takes as input a set of unaligned RNA
sequences expected to share a common motif, and outputs the regions that are most conserved through-
out the sequences, according to a similarity measure that takes into account both the sequence of the
regions and the secondary structure they can form according to base-pairing and thermodynamic rules.
The method is split into two parts. First, it identifies candidate regions within the input sequences,
and associates with each region a locally optimal secondary structure. Then, it compares candidate regions
to one another, both at sequence and structure level, and builds motifs exploring the search space through
a greedy heuristic. We provide a detailed guide to the different parameters that can be employed, and usage
examples showing the different software capabilities.
Key words RNAProfile, RNA untranslated sequences, RNA secondary structure, UTR, Modtools,
Posttranscriptional regulation
1 Introduction
RNA molecules play a large number of different roles in eukaryotic

cells, other from the transfer of coding information performed by
mRNAs and tRNAs. Examples are the posttranscriptional regula-
tory activity of miRNAs, the catalytic activity of ribozymes, and
several others that have been identified only recently, like the abil-
ity to influence epigenetic states [1, 2]. Such great versatility is
mainly due to the fact that, unlike DNA, RNA is a single stranded
nucleic acid, and has the ability to fold into a three-dimensional
structure through the formation of intramolecular base pairings.
RNA structure can therefore act both in cis and in trans. While
functional elements that depend from RNA structure cannot be
identified by considering sequence information alone, in several
49
50 Federico Zambelli and Giulio Pavesi
cases they can be however described by considering RNA secondary

structure. Examples of these motifs are the secondary structure
signals present in the untranslated regions (UTRs) of mRNA,
responsible for posttranscriptional regulation of the gene expres-
sion. Several other classes of RNAs, such as rRNAs, tRNAs,
microRNAs, and other noncoding small RNAs, as well as viral
mRNAs, contain similar motifs with regulatory activity.
Thus, in the last few years, several bioinformatic algorithms
and tools have been introduced to predict and compare RNA sec-
ondary structures (e.g., [3–6]). Also, several databases have been
established and maintained to organize and exploit the wealth of
information available on different class of RNAs and their associ-
ated secondary structures (e.g., [7–9]).
The problem of predicting shared structural features of RNA
molecules, likely to correspond to functional elements, can be split
into two separate problems. First of all, reliable models of the (sec-
ondary) structure of the RNAs should be built. Then, a set of
RNAs sharing the same function should then present similarity at
the sequence and most important, at the structure level, at least in
some regions.
The prediction of the structure of a RNA sequence is usually
based on the minimization of a given potential energy function
(e.g., [10, 11]). If the prediction is limited to secondary structure,
than the optimal secondary structure associated with the global
minimum of the energy function can be obtained in feasible time.
On the other hand, RNA folding is a dynamic process, that usually
involves many intermediate states corresponding to local minima
of energy. As a consequence, the actual in vivo structure of RNA
molecules may not correspond to the most thermodynamically
stable one. Thus, while single sequence energy-based prediction
algorithms can exhaustively explore the space of possible secondary
structure conformations, even if a very reliable energy function is
employed the structure corresponding to a global minimum is not
guaranteed to match the actual one.
On the other hand, the prediction of RNA secondary structure
is more reliable when applied to predicting a “consensus structure”
for a set of sequences sharing a similar structure [12]. However, in
a set of functionally related RNAs, relevant structural similarity
usually covers only a fraction of the sequences themselves, often
just a single hairpin, whereas the rest of the overall structure exhib-
its more variability. Motifs found in mRNA-untranslated regions
(UTRs) are typical instances of this behavior [13]. Thus, the prob-
lem of finding these functional motifs can be recast as motif finding
aimed at discovering similar elements in sequences with related
function, where similarity cannot be limited to primary sequence
alone, but has to include also structural features (Fig. 1).
Herein we describe RNAProfile [3], a bioinformatic tool for
finding conserved secondary structure motifs shared by a set of
RNAProfile 51
Fig. 1 A secondary structure hairpin, composed by a stem made by intramolecu-

lar base pairings and a loop with unpaired bases. Bases composing the loop have
usually a functional role, and thus are generally more conserved in functional
motifs than the paired ones
functionally related RNA sequences. RNAProfile bypasses the need

of predicting the overall secondary structure of the sequences stud-
ied, and can identify similar regions within the sequences themselves,
where similarity is measured by considering both sequence similarity
and the optimal secondary structure associated with the regions.
2 Materials
RNAProfile is available at http://www.beaconlab.it/modtools/

(current version is 2.2). It is possible to download either the source
code, or precompiled binary files for Unix-like operating systems
(Unix, Linux, BSD, Mac OS X) and for Windows. In order to

compile the source code, the C compiler “gcc” and the “make”
utility should be available. For secondary structure prediction,
RNAProfile includes functions for free energy calculation and sec-
ondary structure prediction taken from the Vienna RNA package
[11]. The software is memory and time efficient, and can be run
on any current PC configuration.
3 Methods
3.1 The Algorithm We present here a summary of the main steps performed by the
algorithm, while a more detailed description can be found in [3]
and relative Supplementary Materials.
3.1.1 Candidate Regions Given a set of RNA sequences expected to share a structural motif,
Selection the first step is to identify a set of candidate regions from each
sequence, containing potential secondary structure motifs. Since
RNA secondary structure can be decomposed into hairpins, the
only requested parameter is the number of hairpins expected to
form the motif, with a single hairpin as default. No other con-
straints like size and number of features like loops, stacks, and con-
necting elements or any threshold for the folding energy are
required. The prediction of the secondary structure of the input
sequences is performed only locally. That is, given as input the
number h of hairpins that have to be contained in the secondary
structure of the motif, the algorithm selects from each input
sequence the regions whose predicted optimal secondary structure
contains exactly h hairpins. If this parameter is also not available,
the analysis can be simply iterated starting from a single hairpin,
and increasing the number h at each run. In the absence of further
constraints, the algorithm examines all the possible substrings of
each input sequence. The general idea is that a structural motif
should correspond to a local free energy optimum for the region
forming it, and its formation thus should depend solely on local
interactions among the nucleotides of the region and not on the
presence of other structures elsewhere in the sequence.
Moreover, searching for a fixed number of hairpins within each
region of a RNA sequence of length n allows the reduction of com-
putational complexity from the exponential time required to enu-
merate all the potential secondary structures that a RNA sequence
can form [14] to a polynomial one. Notice also that the regions
length is not predetermined, since the algorithm processes every
possible substring of suitable size of the sequence in order to check
if it contains the required number of hairpins. Finally, this approach
takes also advantage from the fact that folding parameters used by
energy-based RNA secondary structure methods are usually more
reliable when applied to small regions of limited size (10–50 nt).
RNAProfile 53
The algorithm does not need a priori free energy thresholds, so

that each region with an associated optimal secondary structure
with free energy lower than zero (lower than the unfolded state) is
accepted as a candidate motif instance. Candidate regions from the
same sequence may overlap one another.
3.1.2 Building Motifs After being selected from the input sequences, together with their
secondary structure, candidate regions are compared to one
another, in order to build groups of “most similar” regions likely to
form a conserved motif. These groups will contain exactly one
region from each input sequence. Comparisons are made by com-
puting pairwise alignment of the regions, with a scoring function
that takes into account at the same time similarity at sequence and
structure level. Given k input sequences of size n, the selection step
returns O(n) candidate regions per sequence, with O(nk) possible
candidate motifs that can be built by selecting one region per
sequence. Thus, exhaustive enumeration of all possible region com-
binations would be computationally unfeasible, and a greedy heu-
ristic has been introduced in order to explore the solution space.
Briefly, the heuristic works as follows. Given a set of k sequences
S = {S1, S2, …, Sk} and their relative sets of associated regions R = {R1,
R2, …, Rk}, RNAProfile first computes all the pairwise alignments
between the regions from R1 with those from R2. All the align-
ments are scored and ranked, and the p highest scoring alignments
are kept. Each of the alignments is described with an alignment
profile, that summarizes the sequence-structure similarity of the
two aligned regions. At the second step, regions from R3 are then
aligned with the p profiles that were built at the previous step. The
resulting alignments, that now contain three regions, are again
scored, ranked, and the best p alignments are kept. This step is iter-
ated until the last set of regions Rk has been reached. The align-
ments built at the last step will thus contain exactly one region per
input sequence, and will correspond to the motifs output by the
algorithm. Also, a fitness value is associated with each region
included in an alignment, estimating how well the region fits the
motif profile defined by the alignment itself. A positive fitness value
is an indicator that the region is likely to represent an instance of
the motif described by the profile, vice versa for negative fitness
values. The rationale is that, since there are no guarantees that all
the sequences of S will share an instance of the motif, a motif profile
could include regions that do not represent instance of the motif,
and hence have little similarity to the others building the motif.
When the motif is expected to appear only in a small subset
(less than half) of the input sequences, the algorithm can be run
employing a different method for the construction of motif pro-
files. For each i from 0 to k − 1 the selected regions Ri of a sequence
Si are aligned to all the regions Rj of the other sequences with j > i,
and again the p highest scoring profiles are kept. Thus, instead of
saving the best alignments of regions from two sequences, the

algorithm saves the highest scoring pairwise alignments obtained
by comparing all possible pairs of candidate regions, with the sole
condition that aligned regions come from different sequences.
In the second iteration each of the highest scoring profiles is
aligned with regions from the sequences that have not been used
to build the profile itself, and so on until each profile contains a
minimum number q of regions that can be set as input parameter.
However, it is possible to output the highest scoring profiles
obtained at each iteration, in order to inspect them: this strategy is
suitable when the motifs appear in a few q of the k sequences, with
q < k/2. The advantage of this method is that it is more likely to
find motifs shared by a small subset of the input sequences. The
main drawback is that it is prone to fall in local optima, if some
input sequences are highly similar to one another, and that the
computational complexity rises from O(kn2) to O(k2n2) where
again k is the number of input sequences and n is their length.
3.2 Running The command for using RNAProfile with default settings is:
RNAPRofile
> ./rnaprofile –f <filename>
where <filename> is the input file.
3.2.1 Input Input files can be of two types. The first and most common type is
a multifasta file containing the RNA sequences to be analyzed. A
careful selection of the sequences is important and may have a
strong impact on the output. When possible, it is useful to avoid
including long sequences (i.e., full mRNAs). Rather, it is advisable
to include only the portion that is more likely to contain a func-
tional motif. For example, when motifs appear within mRNA
untranslated regions, it can be a good strategy to omit the coding
portion of the transcripts, in order to exclude the bias induced by
the higher sequence conservation usually found in coding sequences
(see Note).
The second kind of input file accepted by RNAProfile contains
preprocessed sequences or regions, for which a secondary structure
is already available. In this case, the secondary structure prediction
and region selection steps of the algorithm are bypassed. In the
latter case, the file format is as follows, with for each input sequence
the sequence itself in FASTA format, followed by the secondary
structure associated with its regions in bracket notation.
>Sequence1
CAGTCAGTACGTCTGACAGTCAGTACATGCTCGATGGTACGTATGCATGCGTGT
CATCAGTCTGAGTCAGTACTGACGTAGTCAGTCTGACTGACGTATGCAGTCTGA
!Sequence1
GTTCGTCCTCAGTGCAGGGCAAC
(((.(((((......))))))))
RNAProfile 55
AACTTCAGCTACAGTGTTAGCTAAGTT
(((((.(((((......))))))))))
CCACAGGCTCAGTGTGGTCTTGG
(((.(((((......))))))))
GCCTTCTGCACCAGTGTGTGTAAAGGC
(((((.(((((......))))))))))
GCCTTCTGCGCCAGTGTGTGTAAAGGC
(((((.(((((......))))))))))
TAATTGCAAACGCAGTGCCGTTTCAATTG
((((((.(((((......)))))))))))
>Sequence2
CTACTGACAGTCAGTCATGCGTACAGTGTCAGTCATGCAGTCAGTACCGTACGTA
CATGACGTCATGCATGCATGCAGTCAGTCATGCAGTCATGCATGCATGCAGTCAG
!Sequence2
CCACAGGCTCAGTGTGGTCTTGG
(((.(((((......))))))))
GCCTTCTGCACCAGTGTGTGTAAAGGC
(((((.(((((......))))))))))
GCCTTCTGCGCCAGTGTGTGTAAAGGC
(((((.(((((......))))))))))
TAATTGCAAACGCAGTGCCGTTTCAATTG
((((((.(((((......)))))))))))
The regions of sequence seqname (defined in the FASTA

header) must be introduced with a line beginning with ! (exclama-
tion mark) followed by the same seqname as in the above example.
The order in which regions appear in the file is unimportant and it
is possible to put first all the sequences and then all their respective
regions. If for one or more input sequences no regions are pro-
vided, RNAProfile will apply the secondary structure and region
selection steps only on those sequences.
3.2.2 Parameters Default settings have been determined by running several tests,
with the aim of speeding up the computation obtaining at the same
time reliable results. All input parameters can anyway be modified
and fine tuned by users as follows.
-A <int>: run the program in “all versus all” mode, useful when
only a subset of the input sequences is expected to share a
motif as discussed in Subheading 3.1.2, or when the default
run did not yield significant results. The algorithm stops when
profiles containing <int> regions have been generated. The
best profiles for each iteration are anyway output.
-H <int>: sets the number of hairpins that the secondary structure
associated to each candidate region must contain. Default is a
single hairpin. A good strategy can be to run RNAProfile on
the same set of sequences increasing by one the value of H at
each run, and stopping when the score for the best profile
found given n hairpins is smaller than the one for the best pro-
file with n − 1 hairpins.
-v: verbose mode.

-r yes: process the sequences in random order. The default is to
process the sequences in the same order with which they appear
in the input file. This option is incompatible with –A. Different
permutations of the input order are likely to produce different
results, but, on the other hand, obtaining the same highest
scoring motifs on different permutations is an indicator of the
convergence of the method towards a set of highly conserved
motifs.
-r <seed>: use <seed> (integer number) as seed for the random selec-
tion of sequences. Runs with the same seed will have identical
output.
-P <int>: number of profiles to be kept at each step. The default is
100. Increasing this number allows for a more thorough explo-
ration of the solution space, at the expense of computational
time.
-l <int>: minimum length of the candidate regions. Default is 20
for single hairpin motifs and 10 more bases are added for each
additional hairpin. It should be increased in case of complex
structures and reduced when looking for small hairpins.
-L <int>: maximum length for candidate regions. Default is 40 for
one hairpin and 20 more bases are added for each additional
hairpin.
-o <filename>: the program performs only the first step, that is,
secondary structure prediction and the identification of candi-
date regions, which are saved in file <filename> together with
the predicted secondary structure.
-s <int>: do not compare regions differing in length for more than
<int>. This can speed up the program by avoiding the align-
ment of regions of largely different size, and thus very unlikely
to produce good results. Default is 10.
-i <int>: run the program consecutively for <int> times. To use in
conjunction with –r yes in order to process the sequences in a
different order at each run. Once again, finding the same high-
est scoring motif(s) in all or most of the runs is an indicator of
the presence of a highly conserved motif in the sequences.
-e <float>: set an energy threshold for the secondary structure asso-
ciated to candidate regions. Default is 0, and <float> should be
lower than 0.
-p <int>: at each step, no more than <int> of the best profiles can
be generated by using the same starting profile. Used to avoid
premature convergence. Default is 10.
RNAProfile 57
3.2.3 Output Results are output into a file, whose name is shown on the screen
at the end of the run. A result file has the following format:
ALIGNMENT RESULTS
Input file: mouse+human.ferritin.fna
Number of profiles saved at each step: 100 Max number of
profiles originating from the same profile (or region): 10
Region minimum length: 20 Region maximum length: 40
Energy threshold: 0
Max difference in length between regions: 10
Random alignment: yes (Seed used: 1065538283)
Best profiles:
Profile 1. Score: 2.66
(profile data here)
>gi|33859501|ref|NM_009653.1| Mus musculus aminolev-

ulinic acid synthase 2, mRNA
gGTTcGTCCTcagtgcAGGGCAACa
.(((.(((((......)))))))). (E: -7.2 Fitness: 4.6)
>gi|6753913|ref|NM_010240.1| Mus musculus ferritin
light chain 1 (Ftl1), mRNA
cTTGcTTCAAcagtgtTTGAACGGa
.(((.(((((......)))))))). (E: -1.8 Fitness: 9.0)
>gi|6753911|ref|NM_010239.1| Mus musculus ferritin
heavy chain (Fth), mRNA
cCTGcTTCAAcagtgcTTGAACGGa
.(((.(((((......)))))))). (E: -4.4 Fitness: 8.1)
>gi|20149497|ref|NM_000146.2| Homo sapiens ferritin,
light polypeptide (FTL), mRNA
cTTGcTTCAAcagtgtTTGGACGGa
.(((.(((((......)))))))). (E: -1.3 Fitness: 9.8)
>gi|4557298|ref|NM_000032.1| Homo sapiens aminolevu-
linate, delta-, synthase 2
cGTTcGTCCTcagtgcAGGGCAACa
.(((.(((((......)))))))). (E: -7.5 Fitness: 6.3)
>gi|507251|gb|L20941.1|HUMFERRITH Human ferritin
heavy chain mRNA, complete cds
cCTGcTTCAAcagtgcTTGGACGGa
.(((.(((((......)))))))). (E: -3.9 Fitness: 8.8)
>gi|806340:c862-1 H.sapiens (24) Ferritin H pseudogene
aATTTCTTTatttGAAGGAATg
.((((((((....)))))))). (E: -4.5 Fitness: -3.6)
>gi|182512|gb|J04755.1|HUMFERHX Human ferritin H
processed pseudogene, complete cds
ctTAGTCATTgccatGATGACTGca
..((((((((.....)))))))).. (E: -7.5 Fitness: -39.7)
>gi|806342|emb|X80336.1|HS5FERHPE H.sapiens (5)
Ferritin H pseudogene
TTCTTCaCCaaTCtcatGAGGaGAGGGA
((((((.((..((....)))).)))))) (E: -5.8 Fitness: -321.6)
The input parameters are reported at the beginning of the file.

Then, the highest scoring profiles are listed ranked by score. The
output for each profile includes the frequency of each nucleotide in
each column of the alignment, together with the list of regions
used to build it, their secondary structure and folding energy (E).
As described in Subheading 3.1.2, the fitness value can be used to
discriminate regions that are true instances of the motif described
by the profile (positive fitness value) from those less likely. The
more negative the fitness, the more unlikely it is for the correspon-
dent region to be a true instance of the associated motif profile.
3.3 Usage Examples In these examples the algorithm has been run iteratively with
default parameters, unless otherwise specified, starting with h = 1
(h being the number of hairpins in candidate regions) and increas-
ing h by 1 until the best profile found with h hairpins had a score
lower than the best profile reported using h − 1 hairpins.
3.3.1 Iron Responsive The iron responsive element (IRE) is a well-known mRNA struc-
Element tural element (e.g., [15–17]) that has been also associated with
neurodegenerative diseases (e.g., [18]). It is composed by a con-
served hairpin structure present in the UTRs of transcripts coding
for proteins involved in cellular iron metabolism. IREs usually
present an unpaired nucleotide (cytosine) at the 5′ of the stem, or
a cytosine nucleotide and two additional bases at the 5′ opposing
one free 3′ nucleotide. The iron responsive element is then a good
benchmark for testing the functionality of RNAProfile. A dataset
with the full mRNA sequences of human and mouse ferritin (light
and heavy chain) and aminolevulinate synthase 2 was prepared.
To add experimental noise, sequences from three human ferritin
pseudogenes, hence not likely to contain an IRE motif, were also
included in the dataset. As it can be seen in Fig. 2 the algorithm is
able to identify correctly the IREs as building the highest scoring
motif. Candidate motif regions from pseudogenes have instead a
negative fitness value, strongly hinting that they are not actual
instances of the IRE motif described by the profile.
3.3.2 Atypical IREs Not all the IREs fall in the two types described in the previous
example. A new dataset with sequences including atypical IREs
found in the 5′-UTR of mRNAs from fruitfly, bullfrog, starfish and
crayfish was prepared. Eight 5′-UTRs from plant ferritin mRNAs,
lacking the IRE, were also added to the dataset to test the “all ver-
sus all” functionality of the algorithm. In fact, only one third of the
sequences of this dataset actually contain an IRE. The algorithm
was thus run with the –A option active, and looking for motifs con-
taining up to six regions (appearing in at most half of the sequences).
As shown in Fig. 3, four regions containing four IRE instances
form the highest scoring motif. The bottom helix has different
length in the four instances, while the starfish and crayfish IREs
share a base pair missing at the bottom of the topmost helix.
RNAProfile 59
Fig. 2 IRE motif occurrences reported by RNAProfile, with respective energy and fitness values. Notice the
negative fitness values for the instances reported in pseudogenes
The crayfish element does not present the unpaired cytosine usually
considered the distinctive signature of IREs. However, all these ele-
ments have been experimentally validated as functional IREs in [19].
3.3.3 Translation Control The Nanos protein determines the correct anterior/posterior pat-
Element terning in fruitfly embryos [20]. The translation of its mRNA is
repressed in the bulk cytoplasm and activated only in the posterior
domain. Translational control is dependent on a RNA secondary
structure feature called translation control element (TCE) found in
the 3′-UTR of Nanos mRNA. This structure is Y shaped and is
bound by the protein Smaug leading to translational repression
[21]. For this test RNAProfile was run on the 3′-UTR sequences
of Nanos mRNA homologs from D. melanogaster, D. virilis and D.
simulans. The score of the highest scoring profile remained stable
when searching for motifs with one or two hairpins and the best
profile found with one hairpin was contained within the best pro-
file found with two hairpins. The latter captures the two hairpins
forming the Y shape as seen in Fig. 4.
Fig. 3 Four IRE instances found in the “unconventional IRE” dataset building the highest scoring motif. Allowing
the algorithm to reach the imposed limit of six regions per motif lowered the score of the motif itself by includ-
ing two unlikely instances, as also shown by the low fitness of the two additional sequences
Fig. 4 The Nanos TCE element identified by RNAProfile, composed of two hairpins building a Y-shaped secondary
structure
RNAProfile 61
3.4 Single Sequence In some cases, it might be helpful to have a tool for secondary
RNA Secondary structure prediction available to work together with motif finding
Structure Prediction tools like RNAProfile. In fact, while functional RNA secondary
structures may be different from the ones predicted according to
minimal energy or other criteria, tools like mfold [22] can be con-
strained to predict structures that include fixed elements in some
regions, like those identified by RNAProfile. In this way the pre-
dicted structure for the whole molecule might represent a more
accurate representation of the possible in vivo secondary structure.
3.5 Secondary RNAProfile output consists of a text file where the structure motifs
Structure are described using the conventional “dot-and-brackets” notation,
Visualization Tools where round brackets correspond to paired bases, and dots to the
unpaired ones. While this method is practical, a graphical represen-
tation of the RNA motifs should be more suitable when producing
figures for articles or oral presentations. Tools like PseudoViewer
[23] and Jviz.Rna [24] can be very useful for this task, since they
permit to draw a model of the secondary structure starting from
the dot-bracket notation.
4 Note
As discussed in Subheading 3.2.1, a careful selection of the

sequences to be analyzed is of the utmost importance for the cor-
rect identification of RNA motifs, for example by removing unnec-
essary portion of the sequences. Another point to be considered
when preparing an input where most of the sequences are expected
to contain a functional motif is to arrange them in phylogenetic
order following a sort of decreasing similarity gradient. In fact, as
described in Subheading 3.1.2, the greedy heuristic for the com-
parisons of regions proceeds by following the same order of the
input sequences. Therefore, it may happen that processing very
dissimilar sequences at the beginning could mask the presence of
motifs that would be otherwise detected when the profiles are built
more gradually. On the other hand, including highly similar
sequences in the input should be avoided, since in this case
sequence similarity will have the effect of “shadowing” similarity
mainly due to secondary structure. In case of doubt, it might be
advisable to use the –r yes option in conjunction with a suitable
value for –i, in order to perform multiple tests on different permu-
tations of the sequences. Highest scoring motifs appearing in all or
most of the results are then the most likely candidates to corre-
spond to actual functional motifs shared by the input sequences.
References
1. Sabin LR, Delás MJ, Hannon GJ (2013) 13. Mignone F, Gissi C, Liuni S et al (2002)
Dogma derailed: the many influences of RNA Untranslated regions of mRNAs. Genome Biol
on the genome. Mol Cell 49(5):783–794 3(3):REVIEWS0004
2. Dieterich C, Stadler PF (2012) Computational 14. Waterman MS (1995) Introduction to compu-
biology of RNA interactions. Wiley Interdiscip tational biology. Chapman & Hall, London
Rev RNA 4(1):107–120 15. Hentze MW, Muckenthaler MU, Galy B et al
3. Pavesi G, Mauri G, Stefani M et al (2004) (2010) Two to tango: regulation of mamma-
RNAProfile: an algorithm for finding con- lian iron metabolism. Cell 142(1):24–38
served secondary structure motifs in unaligned 16. Tandara L, Salamunic I (2012) Iron metabo-
RNA sequences. Nucleic Acids Res 32(10): lism: current facts and future directions.
3258–3269 Biochem Med 22(3):311–328
4. Rabani M, Kertesz M, Segal E (2008) 17. Ma J, Haldar S, Khan MA et al (2012) Fe2+
Computational prediction of RNA structural binds iron responsive element-RNA, selec-
motifs involved in posttranscriptional regula- tively changing protein-binding affinities and
tory processes. Proc Natl Acad Sci U S A regulating mRNA repression and activation.
105(39):14885–14890 Proc Natl Acad Sci U S A 109(22):
5. Hiller M, Pudimat R, Busch A et al (2006) 8417–8422
Using RNA secondary structures to guide 18. Cahill CM, Lahiri DK, Huang X et al (2009)
sequence motif finding towards single-stranded Amyloid precursor protein and alpha synuclein
regions. Nucleic Acids Res 34(17):e117 translation, implications for iron and inflam-
6. Bafna V, Tang H, Zhang S (2006) Consensus mation in neurodegenerative diseases. Biochim
folding of unaligned RNA sequences revisited. Biophys Acta 1790(7):615–628
J Comput Biol 13(2):283–295 19. Huang TS, Melefors O, Lind MI et al (1999)
7. Mokrejs M, Vopálenský V, Kolenaty O et al An atypical iron-responsive element (IRE)
(2006) IRESite: the database of experimentally within crayfish ferritin mRNA and an iron reg-
verified IRES structures (www.iresite.org). ulatory protein 1 (IRP1)-like protein from
Nucleic Acids Res 34(Database issue): crayfish hepatopancreas. Insect Biochem Mol
D125–D130 Biol 29(1):1–9
8. Burge SW, Daub J, Eberhardt R et al (2013) 20. Lehmann R, Nüsslein-Volhard C (1991) The
Rfam 11.0: 10 years of RNA families. Nucleic maternal gene nanos has a central role in pos-
Acids Res 41(Database issue):D226–D232 terior pattern formation of the Drosophila
9. Grillo G, Turi A, Licciulli F et al (2010) embryo. Development 112(3):679–691
UTRdb and UTRsite (RELEASE 2010): a col- 21. Crucs S, Chatterjee S, Gavis ER (2000)
lection of sequences and regulatory motifs of Overlapping but distinct RNA elements con-
the untranslated regions of eukaryotic mRNAs. trol repression and activation of nanos transla-
Nucleic Acids Res 38(Database issue): tion. Mol Cell 5(3):457–467
D75–D80 22. Zuker M (2003) Mfold web server for nucleic
10. Reuter JS, Mathews DH (2010) RNAstructure: acid folding and hybridization prediction.
software for RNA secondary structure predic- Nucleic Acids Res 31(13):3406–3415
tion and analysis. BMC Bioinformatics 11:129 23. Byun Y, Han K (2009) PseudoViewer3: gener-
11. Lorenz R, Bernhart SH, Höner Zu Siederdissen ating planar drawings of large-scale RNA
C et al (2011) ViennaRNA Package 2.0. structures with pseudoknots. Bioinformatics
Algorithms Mol Biol 6:26 25(11):1435–1437
12. Witwer C, Hofacker IL, Stadler PF (2004) 24. Wiese KC, Glen E, Vasudevan A (2005) JViz.
Prediction of consensus RNA secondary struc- Rna—a Java tool for RNA secondary structure
tures including pseudoknots. IEEE/ACM visualization. IEEE Trans Nanobioscience
Trans Comput Biol Bioinformatics 1(2):66–77 4(3):212–218
Chapter 5
Drawing and Editing the Secondary Structure(s) of RNA

Abstract
Secondary structure diagrams are essential, in RNA biology, to communicate functional hypotheses and
summarize structural data, and communicate them visually as drafts or finalized publication-ready figures. While
many tools are currently available to automate the production of such diagrams, their capacities are usually
partial, making it hard for a user to decide which to use in a given context. In this chapter, we guide the reader
through the steps involved in the production of expressive publication-quality illustrations featuring the RNA
secondary structure. We present major existing representations and layouts, and give precise instructions to
produce them using available free software, including jViz.RNA, the PseudoViewer, RILogo, R-chie,
RNAplot, R2R, and VARNA. We describe the file formats and structural descriptions accepted by popular
RNA visualization tools. We also provide command lines and Python scripts to ease the user’s access to
advanced features. Finally, we discuss and illustrate alternative approaches to visualize the secondary structure
in the presence of probing data, pseudoknots, RNA–RNA interactions, and comparative data.
Key words RNA visualization, Secondary structure, Graph drawing, Pseudoknots, Non-canonical
motifs, Structure-informed multiple sequence alignments
1 Introduction
1.1 Context The secondary structure of RNA represents a discrete abstraction,

and Definitions or projection, of a three-dimensional structure restricted to (a sub-
set of) its interacting positions. As such, it is naturally used within
schematics depicting the general architecture of a given RNA con-
formation. It provides a context for various sequential information,
including evolutionary conservation, accessibility to chemical/
enzymatic probes, or predicted single-strandedness. In structural
studies, it is also used as a scaffold to perform an interactive three-
dimensional modeling by homology. Such schematics are finally
used to illustrate functional scenario involving RNA, possibly
interacting with another molecule or ligand, raising the need for
the production of publication-quality diagrams.
Electronic Supplementary Material: The online version of this chapter (doi: 10.1007/978-1-4939-
2291-8_5) contains supplementary material, which is available to authorized users
63
64 Yann Ponty and Fabrice Leclerc
Formally, the secondary structure of an RNA is simply a set

of base-pairs connecting some of the positions in an RNA sequence.
Any base-pair will be denoted by a pair (i, j), indicating that the
nucleotide at position i in the sequence is paired with the nucleo-
tide at position j. In its strictest definition, this set of base-pairs
obeys additional constraints:
● Only canonical interactions: Any base-pair must be canonical
(A–U, G–C, or G–U bases, interacting on Watson-Crick/
Watson Crick edges in cis orientation);
● Non-crossing base-pairs: Any pair of base-pairs (i, j) and
(i′, j′) should be free of crossing interactions (i < i′ < j < j′ or
i′ < i < j′ < j ). This property forbids the representation of com-
plex structural features like pseudoknots;
● Single partner: Any position may interact with at most a sin-
gle distinct partner.
These restrictions lead to tree-like objects which can be drawn
on the 2D plane, and are easy to analyze visually. They also ensure
the computational tractability of the underlying algorithmic process-
ing, while retaining the capacity to suggest the general architecture
(helices, junctions…) adopted within a given conformation.
However, enforcing these restrictions leads to a loss of infor-
mation which could deteriorate the quality of subsequent analyses.
In such cases, a more detailed view of the base–base interaction
network must be adopted. Some or all of the above constraints can
be lifted, leading to the extended secondary structure, a repre-
sentation which typically supports additional annotations, such as
stacked bases, or interacting edges and orientation for base-pairs.
Crossing helices, usually referred to as pseudoknots, using a topo-
logical analogy, are also supported by some tools, but their planar
layout is usually challenging, and even intrinsically unfeasible for
some RNA architectures.
1.2 Objectives The secondary structure can be naturally represented as a graph whose
of RNA Visualization vertices are individual nucleotides. In this representation, edges are
expected to connect consecutive nucleotides in the sugar-phosphate back-
bone, or pairs of positions involved in a base-pair mediated by hydrogen
bonds. Besides these formal requirements, additional properties have
been identified as desirable by earlier work [25]:
Modularity Inspection of the drawing should suggest a decomposi-
tion of the structure into functional domains. For
instance, helices should be easy to identify.
Robustness Structural similarity should be revealed by similar-looking

representations. In particular, the introduction of few
evolutionary events (insertions/deletions of single or
paired nucleotides) should not lead to significant changes.
Clarity The resulting drawing should lend itself to an easy visual
analysis. In particular, overlapping nucleotides and crossing
RNA 2D Visualization 65
lines should be avoided, and a constant distance should

separate backbone-consecutive nucleotides and base-
pairs, respectively.
Realism Drawings should ideally constitute a projection of the
three-dimensional structure. In this respect, relative dis-
tances between consecutive nucleotides and base-pairing
partners should be proportional to that observed in
three-dimensional structures.
Aesthetics Arguably the most subjective aspect of RNA visualization
encompasses a large variety of criteria ranging from stan-
dard color schemes to a strict adherence to historical rep-
resentations for classic families of RNAs (e.g., cloverleaf
shape for transfer RNAs).
Unfortunately, satisfying combinations of these properties may

be computationally intractable, or simply even impossible for intrinsic
reasons. Consequently, many alternative representations and layout
strategies have been proposed over the years, each arising as a tradeoff.
1.3 Existing Tools Many tools are now available for visualizing the secondary structure
of RNA. These tools differ on many aspects, depending on their
intended application and functionalities. Among the distinguishing
features of available tools, one denotes the presence of a graphical
user interface, allowing for a convenient post-processing before
resorting to a time-consuming general-purpose editor. Moreover,
while certain formats may be converted into others in a lossless
manner, most conversion may lose some format-specific data, and a
support for a rich variety of input format is therefore desirable. The
output format of a software is also of great importance, as generated
figures typically require further post-processing before meeting
the quality criteria expected by publishers. Therefore, vector for-
mats such as Portable Document Format (PDF), or Scalable Vector
Graphics (SVG) should always be preferred to bitmap formats.
Table 1 summarizes the foremost tools available for visualizing
the secondary structure. Which, of the available alternatives, should
be preferred will typically depend on the intended application. We
hope that this overview of existing tools and representations will
assist the reader in his choice of the right visualization, and ulti-
mately ease the production of more informative pictures.
1.4 Outline In this chapter, we focus on the necessary steps involved in the
production of publication-quality illustrations involving the RNA
secondary structure. After introducing the main data sources and file
formats, we describe a preferred workflow, and stress on the advan-
tages of vector graphics manipulation. Then, we describe major
proposed representations and layouts for the secondary structure.
We also provide simple command lines to invoke main tools.
We also describe how to address the specific visualization needs
arising from exemplary use-cases: chemical probing experiments, RNA
pseudoknots and interactions, and multiple-sequence alignments.
66
Table 1
Main features of selected tools offering a visualization of the RNA secondary structure
Name Platform(s) [BWMLa] Ease of use Layouts Interactivity Pseudoknots Ext. sec. str.b Outputc References
jViz.RNA •••• ++ Circular + + • EPS PNG [29, 32]
Linear
Graph
PseudoViewer •• ∘∘ ++ Graph ++ ∘ EPS SVG PNG [5]
GIF
RNAMovies •••• ++ Graph + ∘ SVG PNG [17]
JPEG GIF
RILogo •••• + Linear ∘ SVG [23]

R-chie •••• ++ Linear + ∘ PDF PNG [18]
RNAplot •d••• + Graph ∘ PS SVG GML [13]
RNAView ∘••• + Graph + • PS [34]
e
RNAViz ∘••• + Graph ++ + ∘ PDF [27]
R2R ∘••• - Graph + ∘ PDF SVG [31]
S2S/Assemble ∘••• + Linear Graph ++ + • SVG [15, 16]
VARNA •••• ++ Circular ++ + • EPS SVG PNG [7]
Linear JPEG GIF
Graph
xRNA ∘••• + Graph ++ + ∘ EPS –
a
Indicates support (•) or lack of support (∘) for Browser-based, Windows, Mac and Linux executions
b
Indicates support (•) or lack of support (∘) for an extended secondary structure
c
Bold output formats indicate vector graphics, allowing for convenient post-processing
d
Only available from the Vienna RNA webserver, available at http://rna.tbi.univie.ac.at/
e
Export accessible through a print option, using a custom printer driver such as Adobe Distiller
2 Materials
Reflecting the diversity of functions and cellular processes involv-

ing non-coding RNAs, secondary structure data is organized
within a maze of domain-specific databases, and encoded in a vari-
ety of file formats. We briefly mention these sources, and describe
in further details the syntax of major file formats.
2.1 Data Sources At the primary structure level, including sequences and alignments,
the RFAM database [12] is the authoritative source of RNA data.
Unfortunately this centralized authority is without a counterpart on
the secondary structure level, leading to a wealth of specialized repos-
itories focusing on RNAs of specific functions. The reason of this
situation is mostly due to the young history of RNA computational
biology, coupled with the fact that, up until recently, RNA structural
data was relatively scarce compared to proteins. Consequently, data-
bases have typically arisen from a variety of domain-specific efforts,
which have not currently undergone standardization.
Of notable interest at the secondary structure level is the
STRAND database [1], a collection of secondary structures found
in, or inferred from, a variety of databases. However, the purpose
of this database, which was initially assembled to train new energy
models from existing sequence/structure data and benchmark
computational prediction tools, conflicts with the goal of com-
pleteness, and secondary structure data must currently be sought
within a variety of specialized databases. One should however note
that recent initiatives, such as the RNA ontology consortium [21]
or the RNA Central project [2], have led to propositions for more
rational organizations of RNA structural data, and a solution to
these issues may be found in the near future (Fig. 1).
(Extended) Secondary Structure File Formats

Stockhlom Vienna/DBN BPSeq Connect RNAML
Tools jViz.RNA PseudoViewer RNAMovies RILogo R-chie RNAplot R2R S2S VARNA
PDF EPS/PS SVG PNG JPG GIF
Vector Graphics Formats Bitmap Graphics Formats
Fig. 1 Main input and output formats supported by the major visualization tools for the RNA secondary
structure
Fig. 2 Stockholm formatted file fragment, excerpted from RFAM seed alignment for miRNA mir-22. RFAM
ID: RF00653
2.2 RNA Secondary Reflecting the diversity of biological contexts and computational
Structure File Formats use-cases involving RNA secondary structure, many formats were
proposed over time to support its description. Aside from the
ubiquitous FASTA format, extended by the popular Vienna RNA
software package [14] to include a dot-bracket encoding of the
secondary structure, the STOCKHOLM format is used within RFAM
to present a WUSS-formatted consensus secondary structure, pos-
sibly including pseudoknots. The BPSEQ and CONNECT formats
represent an adjacency list, indicating the partner (if any) of each
position.
Notably, the RNAML format [30] represents a unifying effort of
the community to represent virtually any sort of RNA-related data.
While this format is not widely used to represent the secondary
structure because of its intrinsic verbosity, it is clearly the format of
choice to represent and manipulate the extended secondary struc-
ture, including non-canonical base-pairs.
2.2.1 STOCKHOLM STOCKHOLM format files are arguably the preferred representation
Format for RNA multiple sequence alignments. As illustrated in Fig. 2, a
STOCKHOLM file breaks the global alignment of a set of sequences
into portions of bounded width. Unique identifying prefixes
(Organism name/Accession ID) are associated with each portion.
Additionally a set of mark-up lines, recognizable by their prefix
#  , specify various additional information, including accession
identifiers, experimental parameters, bibliographical references etc.
Fig. 3 Dot-parentheses (aka dot-bracket) notation for the minimal free-energy structure of the Rat Alanine
tRNA, predicted using RNAFold
Two mark-up line types are especially relevant to RNA bioinfor-

matics, the # = GC RF lines, which indicate a sequence consen-
sus, and the # = GC SS_cons , which indicates the secondary
structure consensus. The latter uses a parenthesis scheme to repre-
sent base-pairing positions using corresponding parentheses.
Pseudoknots can also be represented using an additional alphabeti-
cal system ( A matching with a , B with b …).
2.2.2 Vienna RNA In this format, the RNA sequence is coupled with a dot-parenthe-
Dot-Parentheses (aka sis string where matching pairs of opening and closing parentheses
dot-bracket) and identify base-pairs. This expression is well parenthesized, meaning
parenthesis that any opening parenthesis can be unambiguously associated
and Pseudobase Notations with a closing parenthesis, inducing a set of non-crossing base-
pairs. For instance, in Fig. 3, the bases at first and ante-penultimate
positions are base-paired.
Since matching pairs of parenthesis cannot unambiguously
identify crossing base-pairs, pseudoknots cannot be represented
strictly within this format. Therefore this format was extended by
the PseudoBase to include support for multiple parentheses sys-
tems, allowing crossing interactions to be represented. In Fig. 4,
two crossing helices, forming a H-type pseudoknot, are initiated
by base-pairs at positions (1604, 1623) and (1615, 1630),
respectively.
2.2.3 BPSEQ Format The BPSeq format is an alternative to the CT format introduced by the
Comparative RNA Web site (CRW, hosted at the University of Texas
Austin by the Robin Gutell Lab). It essentially consists in a simplified ver-
sion of the CT format. As illustrated in Fig. 5 any BPSeq file starts with
four self-explanatory lines identifying the data source and content, and is
followed by the structure, specified as a sequence of space-separated trip-
lets (x, y, z), where: x
x Position;
y IUPAC code for base;
z Base-pairing partner position (0 if unpaired).
2.2.4 CONNECT This format was introduced by Mfold [37], the historical tool for the ab-
(CT) Format initio prediction of RNA secondary structure, and is still used to date by
several prediction tools. After an initial header consisting of the sequence
Fig. 4 Pseudobase notation for the Gag/pro ribosomal frameshift site of Bovine Leukemia Virus. Source: pseu-
dobase, entry# PKB1
Fig. 5 Fragment of a BPSEQ formatted 5 s rRNA inferred using comparative modeling. Source: comparative
RNA web site
length followed by a comment (see Fig. 6), each position is represented by

six fields (a, b, c, d, e, f), each encoded in a fixed-width (eight characters)
column:
a Position;
b IUPAC code for base;
c Position of previous base in the backbone (5′–3′ order, 0 is used if
first position);
d Position of next base in the backbone (5′–3′ order, 0 is used if last
position);
e Base-pairing partner position (0 if unpaired);
Fig. 6 CONNECT format for the Mfold 3.7 [37] predicted structure of the human let-7 pre-miRNA
Fig. 7 Fragment of a three-dimensional model for the Oceanobacillus iheyensis Group II intron (PDBID:3IGI)
f Position (duplicated).
This format can be gently abused to allow for the description of
pseudoknots, although such motif cannot be predicted by Mfold.
2.2.5 PDB File Format The PDB format is a comprehensive text-formatted representation
of macromolecules used with the authoritative eponym repository
of experimentally derived 3D models [3]. Originally introduced to
represent protein structure, it has been enriched over the years to
include fine details of the experimental protocol used for any struc-
ture derivation. However, it does not support detailed information
regarding base-pairing position, and is mostly used by RNA soft-
ware as a raw geometrical descriptor of a 3D model, as illustrated
in Fig. 7.
Fig. 8 RNAML formatted fragment of an all-atom 3D model for the Oceanobacillus iheyensis Group II intron
(PDBID: 3IGI, also shown in Fig. 7), produced by the RNAView software. The base-pair XML sections are
automatically annotated geometrically from the PDB file, and represent base-pairing positions. Such base-pairs
may be non-canonical, i.e. they may involve non-standard pairs of nucleotides, interacting edges, or orientation.
Here, positions 2 and 260 form a canonical base-pair (U–A, both Watson-Crick edges, cis orientation), while
positions 4 and 259 form a non-canonical base-pair (U–U, Sugar/Watson-Crick edges, cis orientation)
2.2.6 RNAML Format RNAML [30] is an XML format, introduced to address the dual need
to unify data representations related to RNA, and to represent
novel important features of its structure (e.g. non-canonical base-
pairs and motifs). Although still challenged in its former goal by
less structured, domain-specific, formats (arguably because of the
intrinsic verbosity arising from its ambitious goals), RNAML has
established itself as the format of choice for an enriched symbolic
description of the tertiary structure, as illustrated in Fig. 8.
Accordingly, it is currently supported by most automated methods
capturing non-canonical interactions and motifs.
3 Methods
3.1 Producing The production of publication-quality illustrations can be greatly

Publication-Quality eased by the choice of adequate tools. Indeed, while one could in
Pictures of RNA principle produce illustrative pictures of RNA by relying only on
vector graphics software such as Inkscape or Adobe®;
Illustrator®;, such a task would quickly be extremely time-
consuming as soon as RNAs of moderate lengths are considered.
In particular, this scenario would require spending a huge amount
of time on trivial tasks, such as the layout of individual bases, or the
reorientation of helices, and would not necessarily yield homoge-
neous output pictures.
For these reasons, current workflows attempt to keep using
domain-specific tools as far as possible. As illustrated in Fig. 9, a vari-
ety of file formats (see Subheading 2.2) can be used to represent the
secondary structure, possibly resulting from an automated annota-
tion of a 3D all-atom model. Such files are then used, possibly in
a Raw Data All-atoms 3D models

RNAML PDB
RNAView
Annotation MC-Annotate
FR3D
(Extended) Secondary Structure File Formats
Stockhlom Vienna/DBN BPSeq Connect RNAML
b RNA-Aware Tools
Interactive Editors Command-line tools Web-based tools
VARNA R-chie PseudoViewer
Manual xrna RILogo R-chie

Parameterization
Refinement
rnaviz RNAplot RILogo Rerun
+ Annotations
S2S/Assemble R2R RNAplot
PseudoViewer VARNA RNAMovies
c Post-Processing
Vector Graphics Editors
Inkscape Adobe
R
Illustrator
R
d Rasterization (Optional) Raster Graphics Editors

Gimp Adobe
R
Photoshop
R
Fig. 9 Typical workflow for the production of publication-quality diagrams. The preferred execution (blue
arrows, dashed arrow indicates a functionality which is not widely available) will start with a scripted produc-
tion of a data-rich initial draft, followed by an interactive refinement within a specialized GUI, and conclude
with a limited post-processing session using a general-purpose vector graphics editor
combination with additional data and annotations, as input for a

large number of software which use some algorithm to produce an
initial layout. This layout may then be further edited within a dedi-
cated GUI, e.g. to refine the automated layout, or to add further
annotations to the drawing. Finally, the output may be exported as
a picture, encoded using a general purpose picture format. Vector
format pictures may then be manually edited using tools such as
Inkscape or Adobe®; Illustrator®; , typically to add legends
or merge several pictures into panels.
One should always prefer vector formats (SVG, PDF, EPS/PS)
over bitmap graphics (PNG, JPG). Indeed, the latter, having lost
track of the underlying geometry, only allows for very basic transla-
tions and color adjustment while the former, by retaining a sym-
bolic description of the picture, enables a convenient post-processing
including rotations and grouping/dissociation of objects, and the
possibility of correcting factual errors such as erroneous nucleo-
tides. Finally, bitmap pictures of virtually any resolution can be
generated from a vector graphics, while the production of a bitmap
picture amounts to being blocked in a given resolution. Such a
commitment is arguably dangerous before the final zoom level of
the picture and the requirements of the publisher are known.
Working with bitmap picture may lead users to create files of
extreme resolutions, which are subsequently difficult to handle,
and sometimes even store. Therefore, the rasterization of RNA
secondary structure pictures should be postponed as much as pos-
sible in the figure-creation process.
3.2 Representations, Many representations have been proposed to display the secondary
Layouts, and Basic structure of RNA, each with its strengths and limitations. In this
Usage section, we give an overview of the existing propositions and, for
each representation and tool, we provide minimal command lines
and/or python scripts to obtain them.
3.2.1 Linear Layout The linear layout represents base-pairs as arcs drawn over and/or
under a linearly drawn RNA sequence. It arguably constitutes one
of the easiest ways to represent the secondary structure, and can be
thought as a genomic perspective over the secondary structure.
Among the strengths of this representation, one counts an easy
identification of helices as sets of nested arcs, and a convenient
(unbiased) representation of pseudoknots. This representation may
also be easily adapted into a side-by-side visual comparison of two
or several candidate structures for a given RNA, provided an align-
ment is available at the sequence level. In such a joint representa-
tion, shared base-pairs are easy to identify, as illustrated in Fig. 10.
However, the linear representation suffers from a poor density
of information. Indeed, the height of arcs associated with base-pairs
typically grows proportionally to the distance between its associated
Fig. 10 ON (above) and OFF (below) states of the pbuE adenine riboswitch [19], jointly represented using a linear
layout created using VARNA (see actual command in Note 1) [7], followed by minimal post-processing. Shared
structural elements (blue arcs) are easily identified in this representation, a property which is not satisfied by their
typical graph layout (gray boxes)
positions, leading to diagrams whose total area scales quadratically

with the sequence length. Consequently, one must either accept to
lose the fine details of the sequence/structure or resort to interac-
tive visualization strategies based on zoom/pan navigation opera-
tions. Such strategies, however, will impede the visualization of
long-range interactions, by forbidding the simultaneous visualiza-
tion of—sequentially distant—interacting positions.
Such diagrams can be generated, for instance starting from a
DBN file XXX.txt, using VARNA [7]:
# Running VARNA ( Linear layout / SVG output )

java - cp VARNAvx - y . jar fr . orsay . lri . varna . applications . VARNAcmd
-i XXXX . yyy -o out . svg -algorithm line
or the RChie [18] software:
# Running R - chie ( Linear layout / PDF output )

Rscript rchie . R - -format1 " vienna " - -pdf - -output = out . pdf XXX . yyy
jviz.RNA [32] may also be used to generate such a represen-

tation in an interactive way.
3.2.2 Circular Layout The circular layout constitutes a modified version of the linear lay-
out, in which the sequence is drawn along a circle. In this represen-
tation, base-pairs are drawn either as arcs or as chords of the circle,
as illustrated in Fig. 11.
This representation is usually more compact than its linear
counterpart (albeit only by a constant factor), and puts more
Fig. 11 Left: RFAM consensus sequence/structure for the Hepatitis delta virus ribozyme family (RFAM id:
RF00094), drawn using VARNA [7]; Right: comparative model for the 16s ribosomal RNA in Thermus ther-
mophilus (source: comparative RNA Web site [6]), drawn using jviz.RNA [32]
emphasis on long distance interactions. However, one must remain

cautious in a visual analysis of such a diagram, as the height of an
arc is no longer strictly proportional to its distance. This phenom-
enon has a particularly strong impact on the representation of heli-
ces involving both ends of an RNA. In this case, helices are
represented by chords/arcs which are in close proximity to the
circular support (see helix supported by (19, 93) in Fig. 11, Left),
and may erroneously be interpreted as a local structural feature.
This representation may also be of limited help to build one’s
intuition regarding the similarity of secondary structures. For
instance, drawing two structures simultaneously using both
inward and outward base-pair will not even assign similar-looking
patterns to perfectly conserved helices. Furthermore, it does not
easily allow for a visual realignment of homologous structures, as
minor changes in the structure will lead to helices having different
orientations. For these reasons, the circular layout is typically
reserved for an automated broad overview of the structural orga-
nization in larger RNAs, mixing long- and short-range
interactions.
Preferred software for producing this representation includes
jViz.RNA [32] (using the GUI + cautious post-processing, see
Note 2), or VARNA [7], invoked as:
# Running VARNA ( Circular layout / P ostScript output )

-i XXXX . yyy -o out . eps -algorithm circular
a c
Fig. 12 Various layout strategies for the display of (extended/consensus) RNA secondary structures as (outer)
planar graphs, also known as squiggle plots. Source: three-dimensional model of the TPP riboswitch (PDB
id:2HOM), and associated RFAM family (RFAM: RF00059)
3.2.3 Squiggle Plots: This representation attempts to draw the secondary structure as a
Planar Graph schematics of its three-dimensional structure. Helices are almost
Representations universally represented as straight, ladder-like, segments. Besides
this norm, layout strategies largely differ, especially with respect to
the layout of multiples loops (3-way junctions and more), as sum-
marized in Fig. 12. This representation is usually preferred to illus-
trate functional scenario.
RNAView (Fig. 12a) draws the secondary structure of an RNA
(PDB model) as a direct 2D projection of the three-dimensional
model, chosen to maximize the spread of the diagram. This strat-
egy results in drawings which, despite being usually self-overlap-
ping, are often indicative of the relative orientation of helices for
small RNAs, as illustrated in Fig. 12a. The software is also men-
tioned in Subheading 2.2 (Fig. 8) for its capacity to annotate the
base-pairs of a 3D model. After downloading/installing the soft-
ware from http://ndbserver.rutgers.edu, a PostScript projec-
tion can be produced from an input RNA (PDB format file XXXX.
pdb), by simply running the command:
# Running RNAView on RNA 3 D model ( PDB format )

rnaview -p XXXX . pdb
Note that, in addition to the output PostScript file, this

command also creates an RNAML file XXXX.pdb.xml, which con-
tains the extended secondary structure for this 3D model.
VARNA (Fig. 12b) offers two types of layout strategies to ren-
der squiggle plots. The first one, named the radial strategy, aims at
a consistent spacing of backbone-consecutive and paired nucleo-
tides. To that purpose, unpaired nucleotides are positioned along
an imaginary circle, whose radius is computed so that consecutive
elements are always distant by a predefined distance. After down-
loading the latest version of VARNA as a JAR archive VARNAvx-y.
jar from http://varna.lri.fr, the default radial strategy can be
used to draw an extended secondary structure (denoted by an
RNAML file, including non-canonical base-pairs, pseudoknots and
multiple interactions per position, and output as an SVG file,
through the single-line command:
# Running VARNA on the ext . sec . str . ( RNAML -> Vector graphics )
java - cp VARNAvX - Y . jar fr . orsay . lri . varna . applications . VARNAcmd
-i XXXX . pdb . xml -o YYYY . svg
The resulting layout, illustrated in Fig. 12b, does not guaran-

tee non-overlapping diagrams. The result may therefore require
further manual post-processing for larger RNAs, a task which can
be performed within VARNA’s dedicated GUI, provided that the
specific file format varna is chosen for the output:
# Running \ Software { VARNA } ( RNAML -> VARNA session file )

-i XXXX . pdb . xml -o YYYY . varna
Alternatively, the NAView algorithm [4] is implemented in

VARNA, and usually ensures a non-overlapping layout, resulting in
layouts which are similar to Fig. 12c. It can be invoked by setting
the algorithm command-line option:
# Running VARNA ( NAView algorithm -> P os tS cr ip t output )

-i XXXX . pdb . xml -o YYYY . eps -algorithm naview
A skeleton view can also be produced, as shown in both subpan-

els of Fig. 10 (see Note 1 for precise commands), by using the
combination of options below (replacing XXXX with your input
file, YYYY with the output file, including an extension which will
decide its type, and ZZZ with the length of the RNA):
# Running VARNA ( NAView algorithm -> P os tS cr ip t output )

java - cp VARNAvx - y . jar fr . orsay . lri . varna . applications . VARNAcmd -i
XXXX -o YYYY -flat true -highlightRegion " 1 - ZZZ : fill =# A0A0A0 ,
outline =# A0A0A0 , radius =10 " -drawBases false -backbone " # A0A0A0 "
-fillBases " # A0A0A0 " -bpStyle simple -numPeriod 50
RNAPlot (Fig. 12c), a software which is part of the Vienna

RNA package, also uses the NAView algorithm for its default lay-
out. As can be seen from Fig. 12c, a peculiarity of this layout is its
unique orientation, going counterclockwise in the 5′ → 3′ direc-
tion. After successful installation of the package, downloaded from
http://www.tbi.univie.ac.at/RNA/, it is invoked on a simple DBN
file (see Fig. 3 for an example) xxxx.yyy, by running the
command:
# Running RNAplot ( NAView a lgorithm )

RNAplot < xxxx . yyy
The resulting PostScript drawing is output to a file zzz_

ss.ps, where zzz is the first token found on the comment line of
the input file (defaults to rna.ps if absent). A radial layout can
also be specified through a dedicated option:
# Running RNAplot ( Radial a lgorithm )

RNAplot - -layout -type =0 < xxxx . yyy
The PseudoViewer (Fig. 12d) uses an original strategy for

choosing the relative orientations of helices. As illustrated in
Fig. 12d, pairs of consecutive helices, connected by a bulge or an
interior loops, are drawn along the same axis. The resulting struc-
tures appear simpler, possibly at the cost of uneven distances along
the backbone. Furthermore, multiple junctions/loops are stretched
and oriented to completely forbid overlaps. Finally, this advanced
layout strategy satisfactorily supports a very large class of pseu-
doknots. The software can be used online, as a web server or a web
service, at http://pseudoviewer.inha.ac.kr/. Graphical User
Interfaces (GUI) are provided for the Windows platform but, to
the best of our knowledge, no command-line version is currently
available.
jViz.RNA (Fig. 12e) implements an interactive layout strat-
egy for squiggle plots, which relies on a spring model. Both back-
bone bonds and base-pairs are modeled as springs, each with a
preferred length and tension. The layout is iteratively refined, sim-
ulating the springs reactions until an equilibrium is found. As illus-
trated in Fig. 12e, helices are not necessarily drawn in rectangular
boxes, and are not even necessarily straight, leading longer helices
to exhibit a wavy shape. However, a user may intervene during the
process to disentangle the structure, or modify the spring tensions
to obtain non-overlapping diagrams. Like the PseudoViewer,
this software does not offer offline command-line capabilities, and
a dedicated GUI, freely available at http://jviz.cs.sfu.ca, must be

used to produce drawings.
The R2R (Fig. 12f) software also uses a radial layout as its
default algorithm. However, more sophisticated layout strategies
can be specified, for instance in order to force angles to fall within
a restricted list (typically k    8, k  [0, 15] ). As illustrated in
Fig. 12f, another interesting feature is the flat drawing of unpaired
nucleotides along the exterior loop. Another interesting aspect is
the display of comparative information regarding a structural fam-
ily, represented as a STOCKHOLM-formatted multiple sequence
alignment including a consensus structure. The installation and
usage of R2R, described in detail in Subheading 3.5, is perhaps a
bit more involved than its competitors, but the quality of the
resulting picture is clearly worth the effort.
3.2.4 Tree Layout In the absence of crossing interactions such as pseudoknots, RNA
secondary structures can be unambiguously decomposed in a vari-
ety of ways, leading to tree-like objects. Such a decomposition of
the conformation space is at the core of any dynamic programming
scheme, a strategy for solving combinatorial optimization prob-
lems which is especially popular in RNA bioinformatics. For
instance, atomic contributions in the Turner energy model [22]
are entirely determined by the joint knowledge of internal nodes (a
base-pair) in combination with their immediate children.
Consequently, this representation may be useful to illustrate the
principles underlying RNA algorithms.
Unfortunately, there are currently few available options to pro-
duce such a representation, and the most realistic option consists
in using the general-purpose graph visualization software
GraphViz [11], which unfortunately requires a DOT-formatted
file as input. Consequently, a secondary structure, typically denoted
by a dot-parenthesis expression (see Subheading 2.2.2), will require
some conversion. We provide in Note 3 a minimal python (2.x)
script to convert, and dump to the standard output as a DOT-
formatted file, a secondary structure.
Invoking this code for a given sequence/structure, while captur-
ing the output through standard I/O redirection, one obtains a file
which can then be fed to the dot utility of GraphViz (See Fig. 13).
This utility can then be configured to produce quality pictures in virtu-
ally any format, including many vector formats such as PDF and SVG.
3.3 Mapping Probing Footprinting techniques are widely applied to map the 2D and 3D
Data onto Secondary structures of RNA using both chemical and/or enzymatic probes.
Structure Models They provide useful information about the relative accessibility of
the RNA to a given probe at the nucleotide resolution. Depending
on the physico-chemical conditions or the presence of different
molecular partners (RNA, proteins, antibiotics, etc.), the chemical
or enzymatic probing data give a snapshot of the RNA structure
Fig. 13 Planar graph layout using VARNA [7] and tree representation using the default options of the dot [11]
algorithm of GraphViz [10] of the secondary structure of a transfer RNA (PDB id: 1TN2:A, annotated using
RNAView [34])
under specific conditions. They are used to study the conforma-

tional changes during RNA folding or induced upon protein or
ligand binding [33]; they also nicely complement the structural
data obtained on the conformational changes associated with RNA
folding [28] or RNA/protein interactions [9].
With the development of high-throughput techniques using
Selective 2′-Hydroxyl acylation Analyzed by Primer Extension
(SHAPE) chemistry (SHAPE-seq [24]), UV-crosslinking or
immunoprecipitation (HITS-CLIP [36]), emerging computa-
tional methods attempt to integrate such data into RNA structure
predictions [20, 26, 35]. However, these predictions are never
entirely accurate, and must be confronted with the experimental
data in an iterative refinement. Visualization is therefore an impor-
tant aspect in the process of generating primary 2D representations
and models. Probing data can be represented using graphical
annotations (symbols, color maps and marks, labels, etc.) mapped

onto a 2D structure.
In this use-case, we focus on probing data obtained on the H/
ACA guide RNAs from the snoR9 family (RFAM: RF00065) in
Pyrococcus abyssi [8]. Here, one may map chemical probing data
onto a secondary structure by generating different representations
for the same RNA molecule, corresponding to chemical probing
data obtained under different conditions. Since the structure is
used to provide a context for the interpretation of probing data,
one will typically use a template for the secondary structure
(sequence/data and layout) onto which the molecular properties
are displayed. Chemical probing data can be represented in differ-
ent ways:
● Symbols may be used to indicate the degree of accessibility to
some chemical or enzymatic probe (Fig. 14b);
● A color map can also be used to superimpose the accessibility
directly onto the secondary structure (Fig. 14c);
● The chemical probing data associated with conformational
changes of the RNA (folding or interactions with ligands) can
be displayed in different panels to provide a graphical indica-
tion of the condition-dependent changes, affecting the tertiary
and secondary structures.
VARNA offers multiple features to ease the production of such
diagrams. Specifically, an extensive array of command-line options
can be used to produce a collection of pictures that share a general
orientation. Three types of annotations are of particular interest:
● Color maps: A color map associates numerical values to each
position in an RNA (colorMap option), and uses a (multiple)
color gradient (colorMapStyle option) to associate colors
to each position, based on their associated value. Minimal and
maximal values can be manually input (colorMapMax and
colorMapMin options), otherwise they are simply guessed
from the values. For color maps, the tedious task of manually
inputting the values (e.g., chemical reactivities), on a
nucleotide-per-nucleotide basis, can be avoided by loading an
external file containing all the data at once from within the
GUI. Alternatively, an ad hoc script may format the command-
line such that those values are directly provided to the soft-
ware. This usage is illustrated by the code listed in Note 4
whose execution produces Fig. 14c.
● Glyphs: Arrows, pins, or additional symbols are sometimes
used to indicate significantly accessible (backbone) regions, or
a difference in accessibility. Position-specific markers can be
specified using the chemProb options in VARNA, as shown in
Listings 4, resulting in the corresponding panels of Fig. 14
(Panels b–f).
a b c
d e f
Fig. 14 2D structure representations integrating chemical probing of the Pab21 HACA sRNP from P. abyssi
(RFAM: RF00065), as generated by VARNA. (a) 2D structure representation of the HACA RNA motif highlighting
the key structural features. (b) Free RNA Chemical probing indicating the accessible positions for Pb cleavage
on a scale from 1 to 3 in the intensity of the reaction. (c) Alternate representation showing the accessibility
using a color map. (d–f) Modifications of the RNA probing data due to the successive binding of L7Ae, CBF5
and NOP10, and the RNA substrate. Green pins: initial probing on RNA; red pins: increase in reactivity; blue
pins: decrease in reactivity; violet regions: newly protected positions
● Highlighted Regions (a.k.a. sausages): Regions highlights

may be used to emphasize a change of reactivity observed on a
portion of the backbone, possibly hinting towards a local con-
formational change. To produce such annotations, the high-
lightRegion option in VARNA may be used, and allows to
specify the width and color of the region, as shown in panels d,
e, and f of Fig. 14 (produced as shown in Note 4).
RNAplot can also be used to display accessibility values as a
color map, using the --pre and --post options to decorate the
PostScript output. Some basic examples can be found in the listing
below:
# Running RNAplot from a DBN file ’ XXX . txt ’
# Ma r k i n g n u c l e o t i d e 10
RNAplot - -post " 10 cmark " < XXX . txt
# Drawing b ackbone for region (10 ,15) using red color ( r =1 , g = b =0) and line thickness 2
RNAplot - -pre " 10 15 2 1. 0. 0. omark " < XXX . txt
# Defining a custom PS macro and using it to fill base number 10 in blue ( r = g =0 , b =1)
RNAplot - -pre " / cfmark { setrgbcolor newpath 1 sub coor exch get aload pop fsize 2 div 0 360
arc fill } bind def 10 0. 0. 1. cfmark " < XXX . txt
As can be seen in this last example, achieving even simple oper-

ations may require an extensive knowledge of the PostScript
language, whose suffix-order for operations induces a steep learn-
ing curve. Furthermore, formatting the command line manually
would quickly become tedious. Therefore, we propose in
Listing 5.1 a minimal Python wrapper which, given a DBN format-
ted RNA and a list of values, invokes RNAplot to produce a color
map-decorated output similar to that of Fig. 14c:
Listing 5.1 Using RNAplot to display probing values as a color

map
import os , sys
def getColor ( val , minval , maxval , col1 =(1. ,1. ,1.) , col2 =(0.0 ,0.8 ,0.0) ) :
( r1 , g1 , b1 ) ,( r2 , g2 , b2 ) = col1 , col2
span = float ( maxval ) - float ( minval )
k = ( float ( val ) - float ( minval ) ) /( span if span !=0 else 1.0)
l = 1. - k
return ( r1 * l + r2 *k , g1 * l + g2 *k , b1 * l + b2 * k )
def formatValuesAsPS ( values ) :
minval , maxval = min ( values ) , max ( values )
valtab = [ " % s % s cfmark " %( i +1 , " %.2 f %.2 f %.2 f " % getColor (v , minval , maxval ) ) for i , v in
enumerate ( values ) ]
return " " . join ( valtab )
# Invokes RNAplot to display a color map
# Arg1 : path to DBN file
# Arg2 : list of comma - s e par ate d values ( eg "1 ,2 ,3")
if __name__ == " __main__ " :
inputFile = sys . argv [1]
values = sys . argv [2]. split ( " ," )
extraMacro = " / cfmark { setrgbcolor newpath 1 sub coor exch get aload pop fsize 2 div 0 360
arc fill } bind def "
os . system ( " RNAplot -- pre \" % s % s \" < % s " %( extraMacro , f o r m a t V a l u e sAsPS ( values ) , inputFile ) )
3.4 3Displaying Pseudoknots and RNA–RNA interactions depart from the tree-like
Pseudoknots paradigm, and therefore constitute a challenge to computational
and Interactions methods. Most prediction algorithms address the problem by
restricting the search space. However, visualization tools cannot
arbitrarily enforce such restrictions. This narrows down the choice
of suitable layouts and tools to few options.
3.4.1 Pseudoknots Basic representations, such as the linear and circular representations,
do not require a complicated layout, and therefore remain largely
unaffected by the presence of pseudoknot. From a software designer
perspective, supporting pseudoknot for these representations is then
mostly a matter of being able to parse suitable input formats. jViz.
RNA (Fig. 15b), R-chie and VARNA (Fig. 15d) may be used with-
out much additional effort. In particular, the latter will also support
more complicated structures, such as those featuring multiple part-
ners for a given nucleotide, as illustrated by Fig. 15d.
However, a large subset of existing pseudoknots can be repre-
sented planarly as squiggle plots, i.e. in such a way that base-pairs and
backbone are non-crossing. To create such representations, the
PseudoViewer is the only fully satisfactory option. Pseudoknots
are laid out within dedicated boxes, that are embedded in a tree-like
general layout. As shown in Fig. 15a, the result is aesthetically pleas-
ing even in the absence of user intervention, its only drawback being
a consequent backbone distortion. Alternatively, the spring layout of
jViz.RNA can be used to obtain decent squiggle plots, as illustrated
in Fig. 15c usually after some manual disentangling from the user.
3.4.2 Interactions The need for automated representations of RNA–RNA interactions

is relatively recent, and has arisen from a recent increase of interest
in the computational prediction of RNA–RNA interactions.
RILogo is currently the only automated tool that produces
representations of RNA–RNA interactions. The web version
requires as input either one or two multiple sequence alignments
(Stockholm or a modified version of FASTA). In the case of a
single file, the consensus secondary structure will use a special char-
acter & to indicate the end of the first RNA and the beginning of
the second. The base-pairs which are split by the separator then be
involved in the hybridization of the two RNAs. Alternatively, two
Stockholm files may be specified, using one or several parentheses
systems ( A / a , B / b …) to indicate the subset of base-pairs
involved in the interaction as follows:
Listing 5.2 Fasta-like input file XXXX.txt expected by RILogo
> organism1
GCGGGGUGCGC & AGGACCCACUCCU
> organism2
CGGCCCGGCCG & AGCGGGCCGCGCU
> structure
((( AABBB ) ) ) &(((( aabbb ) ) ) )
a b c
Fig. 15 Pseudoknotted secondary structure of the Hepatitis Delta virus ribozyme (PseudoBase entry PKB75),
rendered as a planar squiggle plot by the PseudoViewer (a), as a circular diagram and as a force-driven spring
layout by jViz.RNA (b and c). Extended secondary structure of a variant (PDB id: 1SJ3), drawn linearly by
VARNA (d)
Listing 5.3 Invoking RILogo
rilogo XXXX . txt > output . svg
Feeding such a file to the RILogo webserver or software

(both available at http://rth.dk/resources/rilogo) as illustrated
in Listing 5.3 one obtains arc-annotated (linear) representations of
the interacting families, similar to the ones above and shown in
Fig. 16a.
a b
Fig. 16 Available representations for RNA–RNA interactions: consensus secondary structures of the H/ACA
guide and its target sequence from the 16S rRNA (RFAM: RF00065) automatically generated by RILogo (Panel
a), or as a manually edited pseudoknoted structure using R2R (Panel b) and VARNA (Panel c)
Neither R2R nor VARNA offers a native support for RNA–RNA

interactions. However, it is possible to use some tricks to represent
consensus sequence and secondary structures for RNA–RNA inter-
actions. Within R2R, using templates and options designed for
pseudoknots, one may generate some representations where both
interacting RNAs are merged into the same alignment using some
random sequence spacer. Then, the sequence spacer and its associ-
ated annotations may be manually removed using appropriate soft-
ware (Inkscape or commercial alternatives) to display a clean
representation of consensus sequences and secondary structures of
two interacting RNAs (see Fig. 16b). VARNA may also be used in its
linear representation to represent the interaction as a pseudoknot,
possibly in combination with annotations, and manually post-pro-
cess the output to obtain result similar to Fig. 16c.
3.5 Visualizing The identification of novel families of structured ncRNAs relies heav-
Structure-Informed ily on the implementation of a comparative approach. This approach
RNA Alignments starts with a multiple sequence alignments, from which a draft consen-
sus secondary structure is identified. This draft is then used to refine
the alignment, and retrieve novel homologs, upon which the second-
ary structure can be further refined. Iterating these steps leads to a
final structure-informed multiple sequence alignment, whose quality
must be assessed before concluding on the existence of a new func-
tional family. To that purpose, a global visualization of the conserva-
tion/covariation levels in the context of the structure is essential.
The R2R software is particularly suited to generate representa-
tions of structure-informed RNA alignments. Beside allowing to
decorate a squiggle plot using conservation and covariation levels,
R2R also allows to define subfamilies of sequences, possibly associ-

ated with diverging secondary structures. Another unique feature
of R2R is its definition of a complete set of directives to express
preferences. These preferences may then be used to produce sets of
similar-looking representations for individual sequences, helping in
the visual identification of differences. Figure 17, a panel of sec-
ondary structure representations produced for the RF00065
RFAM family (see Note 5 for precise commands), illustrates such
capabilities. Here, the consensus structure of the RF00065 family
is displayed using variable-length regions and color-shaded back-
bone regions for sequence elements that vary in size associated
with the internal loop, terminal loop, or 3′-end regions of the
RNAs. Each member of the family is also displayed with the full
sequence to view the details of each member in the family. Two
subfamilies corresponding to different sizes in the 5′ guide
sequence (7 or 8 nucleotides) of the internal loop are shown to
zoom on the variations in the loop size.
A typical usage of R2R includes the following steps:
Step 1 (Initial Alignment) Start from a Stockholm file XXXX.
sto featuring a multiple sequence alignment, along with a
consensus secondary structure denoted in WUSS notation.
The software will refuse a blocked alignment (in which
sequences are broken into blocks of fixed lengths), and we
provide in Listing 5.4 a minimal Python script to trans-
form a blocked Stockholm file into a linear one.
Listing 5.4 Transforming an Stockholm-formatted RFAM mul-

tiple sequence alignment into an unblocked alignment accepted by
R2R as input
import os , sys , string

def unblockStockholm ( inpath , outpath ) :
outfile = open ( outpath , " w " )
( seqIDs , seqContents ) = ([] ,{})
for l in open ( inpath ) :
if len ( l ) >12 and (( not l . startswith ( " # " ) ) or ( l [:12]
in [ " #= GC SS_cons " ," #= GC RF " ]) ) :
( id , content ) = ( l [0:37]. strip () ,l [37: -1]. strip
() )
if id not in seqContents :
seqIDs . append ( id )
seqContents [ id ] = " "
seqContents [ id ] += content
elif len ( l ) >1:
for id in seqIDs :
outfile . write ( string . ljust ( id , 37) +
seqContents [ id ]+ " \ n " )
seqIDs , seqContents = [] ,{}
outfile . write ( l )
outfile . close ()
# T r a n s f o r m s a ’ b l o c k e d ’ S T O C K H O L M file ( eg RFAM a l i g n m e n t ) into a
linear one
# Arg1 : path to ’ blocked ’ S TOCKHOLM file
# Arg2 : path to ’ l i n e a r i z e d ’ S T O C K H O L M output file
if __name__ == " __main__ " :
( inStoc , outFile ) = ( sys . argv [1] , sys . argv [2])
unblockStockholm ( inStoc , outFile )
Fig. 17 Consensus sequence and secondary structures of the H/ACA guide RNA from the snoR9 family (RFAM:
RF00065) generated by R2R (see Note 5 for a sequence of commands) from the RFAM seed alignment. Four
individual secondary structures from the seed alignment are annotated to display the common structural
features (internal loop, K-loop, ANA box) and the subtle differences in sequence and loop size. The modular
sub-structures correspond to two subfamilies: ILOOP75 and ILOOP85 where the 5′ guide sequence of the
internal loop contains either 7 or 8 nucleotides, respectively
Step 2 (Computing statistics) Conservation and covariation lev-

els must be computed, and appended to the alignment into
a file XXXX.cons.sto by running the command
# Compute c o nse r v a t i o n levels

r2r - -GSC -weighted -consensus XXXX . sto XXXX . cons . sto 3 0.97 0.9 0.75 4
0.97 0.9 0.75 0.5 0.1
The numerical values occurring in the above command

correspond to the recommended thresholds and parameter
values, and we direct to the manual any user with specific
needs and expectations in this respect.
Step 3 (User preferences) The produced file may then be manu-
ally edited to specify preferences regarding the layout and
annotations. To that purpose, generic feature lines, starting
with #GF = can be added appended to the file prior to the
terminal nn characters. A huge collection of such features
is available, whose complete listing would perhaps require
a dedicated chapter. We illustrate some of the most salient
features in Note 6 and 7.
Step 4 (Batch metafile) Once these preferences have been
expressed and stored within a Stockholm file XXXX.
user.sto, a meta file must be defined. In a minimal set-
ting, it may simply contain the name of the processed
Stockholm input file. More complex examples, as shown
in Note 8, may include several lines to produce multiple
figures. These figures may either represent the content of
multiple Stockholm files, focus on specific sequences/
organisms (see oneseq keyword), or produce a coarse-
grained skeleton view (see skeleton-with-pairbonds
keyword). For instance, the following meta file produces:
1. a consensus drawing;
2. a drawing for Pyrococcus furiosus, adhering to the same lay-
out rules;
3. a skeleton view of the consensus.
XXXX . user . sto

XXXX . user . sto oneseq Pyrococcus _ furiosus
XXXX . user . sto skeleton -with -pairbonds
Step 5 (Running R2R) Once the meta file XXXX.meta is ready,

the software can simply be run through either of the fol-
lowing commands:
r2r XXXX . meta XXXX . pdf

r2r XXXX . meta XXXX . svg
Fig. 18 Structure-informed representation of the RFAM multiple-sequence alignment of the snoR9 family
(RFAM: RF00065), produced using R-Chie
to produce either a PDF or SVG output, amenable to further

vector post-processing.
The representations offered by R2R allow for a critical inspection
of a proposed consensus structure. However, they do not necessar-
ily inform the viewer regarding the quality of the underlying align-
ment. Therefore, we advocate complementing this representation
with sequence-focused alternatives produced by R-chie, an R
package which uses a linear layout to draw one or several struc-
tures, stacked on top of a multiple sequence alignment. This
sequence alignment is typically color-encoded to indicate conser-
vation, insertion, compatibility, or contradiction with the consen-
sus structure. R-chie can be invoked online, or downloaded at
http://www.e-rna.org/r-chie. Once the necessary R dependencies
are installed, it can be executed on a gapped FASTA multiple
sequence alignment XXXX.txt and a Vienna/DBN consensus
structure YYYY.txt via the following command:
Rscript rchie . R - -msafile XXXX . txt - -colour1 " #4 DAF4A " - -msacol " #00
A651 ,#0072 BC ,#00 B9F2 ,# F15A22 ,#231 F20 ,# AAAAAA ,# DA6FAB " - -pdf
- -output = " out . pdf " - -format1 " vienna " YYYY . txt
Running the command produces a PDF file named out.pdf

similar to that shown in Fig. 18.
4 Notes
1. Execution of the command-line standalone version of VARNA

to generate Arc representations (see Fig. 10 for graphical
rendering).
# Top
java - cp VARNAvx - y . jar fr . orsay . lri . varna . applications . VARNAcmd \
- sequenceDBN " G G G C C C G G C U C C C G C C C U C U C C G G G G A A U C G U G A A C C G G G G G U U C C G G C C G G G C C U A C A " \
- structureDBN " ( ( ( ( ( ( ( ( ( . . . . . . . ( . ( ( ( ( ( ( ( ( ( . . . . . ) ) ) ) ) ) ) ) ) ) .....) ) ) ) ) ) ) ) ) ... " \
- highlightRegion " 10 -16: radius =10 , fill =# acf269 , outline =# acf269 ;43 -47: radius =10 , fill =# acf269 ,
outline =# acf269 " \
-o RF00065_testarc1 . svg - algorithm line - drawBases False - s paceBetweenBases 0.6 - bpStyle line
# Bottom
- structureDBN " ( ( ( ( ( ( ( ( ( . . . . . . . ( . ( ( ( ( ( ( ( ( ( . . . . . ) ) ) ) ) ) ) ) ) ) .....) ) ) ) ) ) ) ) ) ... " \
- chemProb " 10 -11: glyph = pin , dir = out , intensity =3 , color =# f90000 ;17 -18: glyph = pin , dir = out , intensity
=1 , color =# f90000 ; " \
- highlightRegion " 26 -27: radius =10 , fill =# d2ccf4 , outline =# d2ccf4 ;31 -37: radius =10 , fill =# d2ccf4 ,
outline =# d2ccf4 " \
-o RF00065_testarc2 . svg - algorithm line - drawBases False - s paceBetweenBases 0.6 - bpStyle line
2. jViz.RNA tends to create extremely large EPS files, which

cannot be easily manipulated within existing vector graphics
editors. This problem apparently originates from the way char-
acters are output. Indeed, texts are apparently not exported as
PostScript strings, but as splines defined using a large
number of points. Besides increasing the file size, this means
that the nucleotides and annotations cannot be easily manipu-
lated within dedicated editors. Therefore, unless a bitmap pic-
ture is eventually sought and very limited post-processing is
involved, we generally do not recommend using jViz.RNA to
draw RNAs of length greater than 300 nts.
3. Minimal python script to convert a secondary structure,
denoted by a dot-parenthesis notation, into a Graphviz input
file (DOT format) amenable to a tree-like layout.
import sys
# Get custom styles by changing these lines ( cf GraphViz manual )
STYLE_DEFAULT = " shape =\" rectangle \" , style = filled , margin =\"0 ,0\" , fontsize =20 , color = grey40 ,
fontcolor = grey20 , fillcolor = grey90 , fontname =\" Helvetica \" "
STYLE_UNPAIRED = " shape =\" circle \" , color = blue , fillcolor = aliceblue "
STYLE_PAIRED = " shape =\" hexagon \" "
STYLE_EDGES = " color = grey50 "
# C o n v e r t s an RNA s e q u e n c e / s e c o n d a r y s t r u c t u r e ( dot - p a r e n t h e s i s n o t a t i o n ) into a DOT - f o r m a t t e d
GraphViz input , written into a previously opened file f
def drawAsDOT ( seq , secstr , f ) :
print >> f , " digraph rna {\ n node [% s ];\ n edge [% s ];\ n -1 [ label =\" Root \"]; " %(
STYLE_DEFAULT , STYLE_EDGES )
stack = [ -1]
for i in range ( len ( secstr ) ) :
k = stack [ -1]
if secstr [ i ]== " ( " :
stack . append ( i )
print >> f , " % s -> % s ; " %( k , i )
elif secstr [ i ]== " ) " :
stack . pop ()
print >> f , " % s [ label =\"% s \" ,% s ]; " %( k , seq [ k ]+ seq [ i ] , STYLE_PAIRED )
else :
print >> f , " % s -> % s ; " %( k , i )
print >> f , " % s [ label =\"% s \" ,% s ]; " %( i , seq [ i ] , STYLE_UNPAIRED )
print >> f , " } "
# Typical usage of this script ( RNAToDOT . py ) :

# python RNAToDOT . py (((...) ) ) GGGAUACCC > rnaTree . dot
# dot - Tpdf -o rnaTree . pdf rnaTree . dot
if __name__ == " __main__ " :
struct , seq = sys . argv [1] , sys . argv [2]
drawAsDOT ( seq , struct , sys . stdout )
4. Execution of the command-line standalone version of VARNA for a

series of representations (see Fig. 14 for the graphical rendering).
# panel A
- structureDBN " ( ( ( ( ( ( ( ( ( . . . . . . . ( . ( ( ( ( ( ( ( ( ( . . . . . ) ) ) ) ) ) ) ) ) ) .....) ) ) ) ) ) ) ) ) ... " \
- highlightRegion " 10 -16: radius =10 , fill =# acf269 , outline =# acf269 ;43 -47: radius =10 , fill =# acf269 ,
outline =# acf269 " \
-o RF00065_panelA . svg - alg orithm radiate - flat true - drawBases False - spaceBetweenBases 0.6 \
- title " RF00065 : Pyrococcus abyssi GE " - titleSize 12
# panel B
- structureDBN " ( ( ( ( ( ( ( ( ( . . . . . . . ( . ( ( ( ( ( ( ( ( ( . . . . . ) ) ) ) ) ) ) ) ) ) .....) ) ) ) ) ) ) ) ) ... " \
- chemProb " 9 -10: glyph = pin , dir = out , intensity =1 , color =#3 e844b ;10 -11: glyph = pin , dir = out , intensity
=1 , color =#3 e844b ; " \
-o RF00065_panelB . svg - alg orithm radiate - flat true - drawBases False - spaceBetweenBases 0.6
# panel C
- structureDBN " ( ( ( ( ( ( ( ( ( . . . . . . . ( . ( ( ( ( ( ( ( ( ( . . . . . ) ) ) ) ) ) ) ) ) ) .....) ) ) ) ) ) ) ) ) ... " \
- colorMap " 0 ; 0 ; 0 ; 0 ; 0 ; 0 ; 0 ; 0 ; 0 ; 1 ; 1 ; 2 ; 3 ; 2 ; 1 ; 1 ; 1 ; 1 ; 3 ; 3 ; 1 ; 1 ; 0 ; 0 ; 0 ; 2 ; 2 ; 2 ; 3 ; 3 ; 1 ; 3 ; 2 ; \
1;1;1;1;0;0;0;0;0;1;3;3;3;1;0;0;0;0;0;0;0;0;0;1;1;3 " \
- colorMapMax 3 - colorMapMin 0 - colorMapStyle green \
-o RF00065_panelC . svg - alg orithm radiate - flat true - drawBases True - spaceBetweenBases 0.6
# panels D , E , F
- structureDBN " ( ( ( ( ( ( ( ( ( . . . . . . . ( . ( ( ( ( ( ( ( ( ( . . . . . ) ) ) ) ) ) ) ) ) ) .....) ) ) ) ) ) ) ) ) ... " \
- chemProb " 9 -10: glyph = pin , dir = out , intensity =1 , color =#3 e844b ;10 -11: glyph = pin , dir = out , intensity
=1 , color =#3 e844b ; " \
- highlightRegion " 26 -27: radius =10 , fill =# d2ccf4 , outline =# d2ccf4 ;31 -37: radius =10 , fill =# d2ccf4 ,
outline =# d2ccf4 " \
-o RF00065_panelD . svg - alg orithm radiate - flat true - drawBases False - spaceBetweenBases 0.6
5. Shell script to execute the R2R program:

# !/ bin / tcsh -f
#
set R2RPATH = " "

if ( $ R2RPATH == " " ) then
echo " Usage : please set the R2RPATH variable "
exit
endif
set r2r = " $ R2RPATH / src / r2r "

set subfam = " perl $ R2RPATH / src / S e l e c t S u b F a m i l y F r o m S t o c k h o l m . pl "
set GSCparams = ’3 0.97 0.9 0.75 4 0.97 0.9 0.75 0.5 0.1 ’
$ r2r -- GSC - weighted - consensus RF00065_seed . sto RF00065_seed . cons . sto $ GSCparams
$ subfam RF00065_seed . cons . sto ILOOP75 > RF00065_seed - ILOOP75 . sto

$ r2r -- GSC - weighted - consens us RF00065_seed - ILOOP75 . sto RF00065_seed - ILOOP75 . cons . sto $
GSCparams
$ subfam RF00065_seed . cons . sto ILOOP85 > RF00065_seed - ILOOP85 . sto
$ r2r -- GSC - weighted - consens us RF00065_seed - ILOOP85 . sto RF00065_seed - ILOOP85 . cons . sto $
GSCparams
setenv R2R_DUMP_FILE RF00065_seed . txt

$ r2r RF00065_seed . meta RF0 0065_seed . pdf
$ r2r RF00065_seed . meta RF0 0065_seed . svg
$ r2r -- GSC - weighted - consensus R F 0 0 0 6 5 _ g u i d e _ t a r g e t . sto R F 0 0 0 6 5 _ g u i d e _ t a r g e t . cons . sto $

GSCparams
$ subfam R F 0 0 0 6 5 _ g u i d e _ t a r g e t . cons . sto hybrid > RF00065_guide_target - hybrid . sto

$ r2r -- GSC - weighted - consensus RF00065_guide_target - hybrid . sto RF00065_guide_target - hybrid . cons
. sto $ GSCparams
setenv R2R_ D U M P _ F I L E R F 0 0 0 6 5 _ g u i d e _ t a r g e t . txt

$ r2r R F 0 0 0 6 5 _ g u i d e _ t a r g e t . meta R F 0 0 0 6 5 _ g u i d e _ t a r g e t . pdf
$ r2r R F 0 0 0 6 5 _ g u i d e _ t a r g e t . meta R F 0 0 0 6 5 _ g u i d e _ t a r g e t . svg
6. Stockholm seed alignment of the snoR9 family (RFAM ID:

RF00065) decorated with R2R labels to generate a series of
consensus and secondary structures for RNA–RNA interactions
(see Fig. 16 for the graphical rendering).
# STOCKHOLM 1 . 0
Pa21 - S892 G G G C C C G G C . . . . . U C C C G C C C U C U C C G G G . . . GA . . . . . . . . . . . . . . AUC . . . . . . . . . . . . .

. . . . GUGA . ACCG . GG . GGUUCCG . . G . CCGGGCCU . . ACA . U U U U U U U U U U U U U U A G G A A U U G G C G G G G G N N N N N
Ph21 - S880 G G G C C C G G U . . . . . U C C C G C C C U C U C C G G G . . . GA . . . . . . . . . . . . . . AUC . . . . . . . . . . . . .
. . . . GUGA . ACCG . GG . GGUUCCG . . A . CCGGGCCG . . ACA . U U U U U U U U U U U U U U A G G A A U U G G C G G G G G N N N N N
Pf1 - S880 G G G C C C G G U . . . . . U C C C G C C C U C U C C G G G . . . GA . . . . . . . . . . . . . . AUC . . . . . . . . . . . . .
. . . . GUGA . ACCG . GG . GGUUCCG . . A . CCGGGCCC . . ACA . U U U U U U U U U U U U U U A G G A A U U G G C G G G G G N N N N N
#= GC S S _ c o n s < <<<<<<<<........... <<<... <<<<... <<............. _________ ........
. . . -. . > >. >>>> - >> - >. . . . . . . - > - > > > > > > > > . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
#= GC SS_cons_1 . . . . . . . . . . . . . . < <<<<< . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . <<<<< . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . > >>>> . . >>>>>> . . . . . .
#= GC R2R_LABEL q . . . . . . . . . . . . 1 LLlLLL . . . e . f . . . . . . . . . . . . 4 . p . . . . . . . . . . . . . . . . . p . . . . . .
. . . . 5 KHH . HHiH . H H B H L L L L L j 7 8 H 9 H H H h H H H r . . . . a . C C C C C 2 0 C C C C C C C . GG G G G m o n G G G G G g . . . . .
#= GC S U B F A M _ h y b r i d _ R 2 R _ L A B E L - - - - - - - - - - - - - - AAaAAAbf - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
- - - - - - - - - - - - - - -g . chHHHH - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - DDDDDmonEEEEi - - - - - -
#= GC R 2 R _ X L A B E L _ K L O O P .................................
KKKKKKKKKKKKKKKKKKKKKKKKKKKKKKKKKKKKKKKK ................................................
....................
#= GC R 2 R _ X L A B E L _ K L O O P L ..................................................k..............
............................................................................
#= GC R 2 R _ X L A B E L _ I L O O P . . . . . . . . . . . . . . PQXXXX . . W . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . YYYYYRSU . V . . . . . . . . . . . . . . . . . . . . . . . . . . . . . YYYYY . . PPPPPP . . . . . .
#= GC R 2 R _ X L A B E L _ I L O O P L .................x....w..........................................
. . . . . . . . . . . . . . . . . . . . . y . . su . v . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
#= GC R 2 R _ X L A B E L _ A N A .................................................................
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . CCC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
#= GF R2R i g n o r e _ s s _ e x c e p t _ f o r _ p a i r s _1 o u t l i n e _ o n l y _ b p
#= GF R2R place_explicit 2 2 - - - 45 1 0 0 0 - 90
#= GF R2R place_explicit n n - - - 45 1 0 0 0 - 90
#= GF R2R place_explicit m m - - - 45 1 0 0 0 - 90
#= GF R2R delcols 0
#= GF R2R tick_label l guide
#= GF R2R tick_label K K - Loop
#= GF R2R tick_label a ANA box
#= GF R2R tick_label g target
#= GF R2R keep q r
#= GF R2R var_backbone_range 4 5
#= GF R2R outline_nuc L
#= GF R2R outline_nuc l
#= GF R2R outline_nuc G
#= GF R2R outline_nuc n
#= GF R2R outline_nuc j
#= GF R2R circle_nuc 2 rgb :0 ,0 ,0
#= GF R2R circle_nuc C rgb :0 ,0 ,0
#= GF R2R nuc_color 2 rgb :0 ,0 ,0
#= GF R2R nuc_color C rgb :0 ,0 ,0
#= GF R2R var_backbone_range_size_fake_nucs 3 e f
#= GF R2R s h a d e _ a l o n g _ b a c k b o n e KLOOP : K rgb :0 ,129 ,255
#= GF R2R s h a d e _ a l o n g _ b a c k b o n e ANA : C rgb :200 ,200 ,200
#= GF R2R o u t l i n e _ a l o n g _ b a c k b o n e ILOOP : Y rgb :255 ,228 ,196
#= GF R2R o u t l i n e _ a l o n g _ b a c k b o n e ILOOP : P rgb :193 ,255 ,193
#= GF R2R o u t l i n e _ a l o n g _ b a c k b o n e ILOOP : Q rgb :193 ,255 ,193
#= GF R2R o u t l i n e _ a l o n g _ b a c k b o n e ILOOP : X rgb :193 ,255 ,193
#= GF R2R s h a d e _ a l o n g _ b a c k b o n e ILOOP : Y rgb :255 ,228 ,196
#= GF R2R s h a d e _ a l o n g _ b a c k b o n e ILOOP : P rgb :193 ,255 ,193
#= GF R2R s h a d e _ a l o n g _ b a c k b o n e ILOOP : Q rgb :193 ,255 ,193
#= GF R2R s h a d e _ a l o n g _ b a c k b o n e ILOOP : X rgb :193 ,255 ,193
#= GF SUBFAM_PERL_PRED hybrid return 1;

#= GF SUBFAM_hybrid_R2R subst_ ss _1 primary outl ine_only_bp
#= GF S U B F A M _ h y b r i d _ R 2 R box_nuc b rgb :200 ,0 ,0
#= GF S U B F A M _ h y b r i d _ R 2 R box_nuc f rgb :200 ,0 ,0
#= GF S U B F A M _ h y b r i d _ R 2 R box_nuc g rgb :200 ,0 ,0
#= GF S U B F A M _ h y b r i d _ R 2 R box_nuc c rgb :200 ,0 ,0
#= GF SUBFAM_hybrid_R2R outline_nuc E
#= GF SUBFAM_hybrid_R2R outline_nuc D
#= GF SUBFAM_hybrid_R2R outline_nuc m
#= GF SUBFAM_hybrid_R2R outline_nuc n
#= GF SUBFAM_hybrid_R2R outline_nuc i
#= GF S U B F A M _ h y b r i d _ R 2 R set_dir pos0 - 90
#= GF SUBFAM_hybrid_R2R place_explicit b b - - +45 3 0 0 0 0
#= GF SUBFAM_hybrid_R2R place_explicit g g - - +90 1 . 5 0 0 0 90
#= GF SUBFAM_hybrid_R2R place_explicit c c - - 90 1 0 0 0 0
#= GF SUBFAM_hybrid_R2R place_explicit h h - - +45 3 0 0 0 90
#= GF SUBFAM_hybrid_R2R place_explicit m m - - 0 1 0 0 0 - 90
#= GF SUBFAM_hybrid_R2R place_explicit o o - - 0 2 . 5 0 0 0 90
#= GF SU BF A M_ h y br i d_R 2R tick_label i target
#= GF SUBFAM_hybrid_R2R tick_label a guide
#= GF S U B F A M _ h y b r i d _ R 2 R s h a d e _ a l o n g _ b a c k b o n e A rgb :193 ,255 ,193
#= GF S U B F A M _ h y b r i d _ R 2 R s h a d e _ a l o n g _ b a c k b o n e a rgb :193 ,255 ,193
#= GF S U B F A M _ h y b r i d _ R 2 R s h a d e _ a l o n g _ b a c k b o n e E rgb :193 ,255 ,193
#= GF S U B F A M _ h y b r i d _ R 2 R s h a d e _ a l o n g _ b a c k b o n e n rgb :193 ,255 ,193
#= GF S U B F A M _ h y b r i d _ R 2 R s h a d e _ a l o n g _ b a c k b o n e D rgb :255 ,228 ,196
#= GF S U B F A M _ h y b r i d _ R 2 R s h a d e _ a l o n g _ b a c k b o n e d rgb :255 ,228 ,196
#= GF S U B F A M _ h y b r i d _ R 2 R s h a d e _ a l o n g _ b a c k b o n e h rgb :255 ,228 ,196
#= GF S U B F A M _ h y b r i d _ R 2 R s h a d e _ a l o n g _ b a c k b o n e H rgb :255 ,228 ,196
#= GF R2R_oneseq Pa21 - S892 shade_along_backbone KLOOP : K rgb :0 ,129 ,255

#= GF R2R_oneseq Pa21 - S892 shade_along_backbone ILOOP : P rgb :193 ,255 ,193
#= GF R2R_oneseq Pa21 - S892 shade_along_backbone ILOOP : Q rgb :193 ,255 ,193
#= GF R2R_oneseq Pa21 - S892 shade_along_backbone ILOOP : X rgb :193 ,255 ,193
#= GF R2R_oneseq Pa21 - S892 shade_along_backbone ILOOP : Y rgb :255 ,228 ,196
#= GF R2R_oneseq Pa21 - S892 shade_along_backbone ANA : C rgb :200 ,200 ,200
//
7. Stockholm seed alignment of the snoR9 family (RFAM ID:

RF00065) decorated with R2R labels to generate a series of
consensus and secondary structures (see Fig. 17 for the graphi-
cal rendering).
# STOCKHOLM 1 . 0
#= GF ID snoR9
#= GF AC RF00065
#= GF DE Small nucleolar RNA snoR9
#= GF AU Bateman A , Daub J
#= GF GA 50 . 0
#= GF NC 49 . 8
#= GF TC 68 . 7
#= GF SE Bateman A
#= GF SS Published ; PMID :12032319
#= GF TP Gene ; snRNA ; snoRNA ; CD - box ;
#= GF BM cmbuild -F CM SEED ; cmcalibrate - - mpi -s 1 CM
#= GF BM cmsearch -Z 274931 -E 1000000 - - toponly CM SEQDB
#= GF DR SO :0000593 SO : C_D_box_snoRNA
#= GF DR GO :0006396 GO : RNA processing
#= GF DR GO :0005730 GO : nucleolus
#= GF RN [1]
#= GF RM 12032319
#= GF RT Noncoding RNA genes identified in AT - rich h yp er the rm o phi les .
#= GF RA Klein RJ , Misulovin Z , Eddy SR ;
#= GF RL Proc Natl Acad Sci U S A 2002;99:7542 - 7547 .
#= GF CC snoRNA R9 is a member of the C / D class of snoRNA which contain
#= GF CC the C ( UGAUGA ) and D ( CUGA ) box motifs . R9 was identified in a
#= GF CC computational screen in AT - rich h y p e r th er m o ph iles [1] . R9 was
#= GF CC found to overlap with the smaller snoRNA R19 which is currently a
#= GF CC member of Pyrococcus C / D box snoRNA family Rfam : RF00095 .
#= GF WK http :// en . wikipedia . org / wiki / S m a l l _ n u c l e o l a r _ R N A _ s n o R 9
#= GF SQ 5
#= GS Pyrococcus_furiosus AC AE009950 . 1/163991 - 163864

#= GS Pyrococcus_abyssi_GE AC AJ248283 . 1/230575 - 230449
#= GS Pyrococcus_horikoshi AC BA000001 . 2/215834 - 215709
#= GS P . furiosus AC AF468960 . 1/1 - 128
#= GS Thermococcus_kodakar AC AP006878 . 1/47908 - 47779
Pyrococcus_furiosus GGGCCCGGUU .
CCCGCCCUCUCCGGGGAAUCGUGAACCGGGGGUUCCGACCGGGCCCACA ..
A U G G G A U G A U G A C C U U U U G C U U U A C U G A A C A C A U G A U G A C C A C G C C C U U C G C U G A C . CUAA AUAU UU GAC
Pyrococcus_abyssi_GE GGGCCCGGCU .
CCCGCCCUCUCCGGGGAAUCGUGAACCGGGGGUUCCGGCCGGGCCUACA ..G..
UUAUGAUGAACUUUUGCUUUGCUGAUGUGGUGAUGAGCACGCCCUUCGCUGAUACUCUCUCGUCCAU
Pyrococcus_horikoshi CGGCCCGGUU .
C C C G C C C U C U C C G G G G A A U C G U G A A C C G G G G G U U C C G A C C G G G C C G A C A . . GG .
G G A U G A A G A G C U U U U G C U U U G C U G A G C A G A U G A U G A C C A C G C C C U U C G C U G A C . CU . GCUAUUUGAC
P. furiosus GGGCCCGGUU .
CCCGCCCUCUCCGGGGAAUCGUGAACCGGGGGUUCCGACCGGGCCCACA ..
A U G G G A U G A U G A C C U U U U G C U U U A C U G A A C A C A U G A U G A C C A C G C C C U U C G C U G A C . CUAA AUAU UU GAC
Thermococcus_kodakar
GGGCCUGGCGUCCCGCCCUCCCCGGGGAAACGUGAACCGGGGCUUCCUGCCAGGCCUACACCGGGGGAUGAAGAGCUUUUGCUUUGCUGAC
.. UGUGAUGAGCACGCCCUUCACUGACCCCGUAUCAGCUCU
#= GC S S_ c o n s < <<<<<<<<........ <. <<<<<<<<<..... > >>>>>>>>>..... >>>>>>>>>..
.........................................................................
#= GC RF gGGCCCGGcu .
CCCgCCCUCUCCGGGGAAUCGUGAACCGGGGGuUCCggCCGGGCCcACA ..
a u g u u A U G A U G A a C U U U U G C U U U a C U G A a g a g a U G A U G A g C A C G C C C U U C G C U G A u . CU aaaUa uUugAu
#= GR Py r o c o c c u s _f u r i o s u s DEL_COLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
...----------------------------------------------------------------------
#= GR P y r o c o c c u s _ a b y s s i _ G E DEL_COLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
...----------------------------------------------------------------------
#= GR P y r o c o c c u s _ h o r i k o s h i DEL_COLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
...----------------------------------------------------------------------
#= GR P . furiosus DEL_COLS ...........................................................
...----------------------------------------------------------------------
#= GR T h e r m o c o c c u s _ k o d a k a r DEL_COLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.------------------------------------------------------------------------
#= GC R 2 R _ L A B E L p........1......2...........3..4... QQQQqQQQ5 ...6 RRRRrRRRs ..
.7......................................................................8
#= GC R 2 R _ X L A B E L _ K L O O P . . . . . . . . . . . . . . . . . . . . . . . . . . KKKKKKKKK . . . . . . . . . . . . . . . . . . . . . . . .
.........................................................................
#= GC R 2 R _ X L A B E L _ K L O O P L ................................k..........................
.........................................................................
#= GC R 2 R _ X L A B E L _ I L O O P . . . . . . . . . IIIIIIII . . . . . . . . . . . . . . . . . . . . . . . . . . JJJJJ . . . . . . . . . . .
.........................................................................
#= GC R 2 R _ X L A B E L _ I L O O P L .............i...............................j.............
.........................................................................
#= GC R 2 R _ X L A B E L _ A N A .........................................................
CCC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
#= GC R2R _XLABEL_ANAL ...........................................................
c........................................................................
#= GC S U B F A M _ L A B E L _ T E R M ..........x................................................
.........................................................................
#= GC S U B F A M _ I L O O P 7 5 _ R 2 R _ L A B E L - - - - - - - <. . . . . . . . . . B > - - - - - - - - - - - - - - - - - - - - - <. KLMNO . > - - - - - - - - -
-------------------------------------------------------------------------
#= GC S U B F A M _ I L O O P 8 5 _ R 2 R _ L A B E L - - - - - - - <. . . . . . . . . . B > - - - - - - - - - - - - - - - - - - - - - <. KLMNO . > - - - - - - - - -
-------------------------------------------------------------------------
#= GF R2R keep p s
#= GF R2R s h a d e _ a l o n g _ b a c k b o n e ILOOP : I rgb :0 ,255 ,0
#= GF R2R s h a d e _ a l o n g _ b a c k b o n e ILOOP : J rgb :0 ,255 ,0
#= GF R2R s h a d e _ a l o n g _ b a c k b o n e KLOOP : K rgb :0 ,129 ,255
#= GF R2R s h a d e _ a l o n g _ b a c k b o n e ANA : C rgb :200 ,200 ,200
#= GF R2R tick_label ANAL : c ANA box
#= GF R2R tick_label KLOOPL : k K - Loop
#= GF S U B F A M _ R E G E X _ P R E D ILOOP75 TERM [ . ]
#= GF S U B F A M _ R E G E X _ P R E D ILOOP85 TERM [ A -Z ]
#= GF SUBFAM_ILOOP75_R2R no5
#= GF SUBFAM_ILOOP75_R2R s h a d e _ a l o n g _ b a c k b o n e ILOOP : I rgb :0 ,255 ,0
#= GF S U BF AM _ I L OO P 7 5 _ R 2 R tick_label ILOOPL : i 7 nt
#= GF SUBFAM_ILOOP75_R2R outline_nuc ILOOP :I
#= GF SUBFAM_ILOOP75_R2R v a r _ b a c k b o n e _ r a n g e _ s i z e _ f a k e _ n u c s 1 M M 5 nt
#= GF SUBFAM_ILOOP75_R2R var_backbone_range_size_fake_nucs 1 B B
#= GF SUBFAM_ILOOP75_R2R var_backbone_range_size_fake_nucs 1 K K
#= GF SUBFAM_ILOOP75_R2R var_backbone_range_size_fake_nucs 1 L L
#= GF SUBFAM_ILOOP75_R2R var_backbone_range_size_fake_nucs 1 N N
#= GF SUBFAM_ILOOP75_R2R var_backbone_range_size_fake_nucs 1 O O
#= GF SUBFAM_ILOOP85_R2R no5
#= GF SUBFAM_ILOOP85_R2R s h a d e _ a l o n g _ b a c k b o n e ILOOP : I rgb :0 ,255 ,0
#= GF S U BF AM _ I L OO P 8 5 _ R 2 R tick_label ILOOPL : i 8 nt
#= GF SUBFAM_ILOOP85_R2R outline_nuc ILOOP :I
#= GF SUBFAM_ILOOP85_R2R v a r _ b a c k b o n e _ r a n g e _ s i z e _ f a k e _ n u c s 1 M M 5 nt
#= GF SUBFAM_ILOOP85_R2R var_backbone_range_size_fake_nucs 1 B B
#= GF SUBFAM_ILOOP85_R2R var_backbone_range_size_fake_nucs 1 K K
#= GF SUBFAM_ILOOP85_R2R var_backbone_range_size_fake_nucs 1 L L
#= GF SUBFAM_ILOOP85_R2R var_backbone_range_size_fake_nucs 1 N N
#= GF SUBFAM_ILOOP85_R2R var_backbone_range_size_fake_nucs 1 O O
#= GF R2R_oneseq Pyrococcus_furiosus s h a d e _ a l o n g _ b a c k b o n e ILOOP : I rgb :0 ,255 ,0

#= GF R2R_oneseq P yrococcus_furiosus tic k_label ILOOPL : i 5 ’ guide sequence
#= GF R2R_oneseq Pyrococcus_furiosus s h a d e _ a l o n g _ b a c k b o n e ILOOP : J rgb :0 ,255 ,0
#= GF R2R_oneseq P yrococcus_furiosus tic k_label ILOOPL : j 3 ’ guide sequence
#= GF R2R_oneseq Pyrococcus_furiosus s h a d e _ a l o n g _ b a c k b o n e KLOOP : K rgb :0 ,129 ,255
#= GF R 2R _o n es eq Pyrococcus_furiosus v a r _ b a c k b o n e _ r a n g e _ s i z e _ f a k e _ n u c s 10 7 8
#= GF R2R_oneseq P yrococcus _furiosus inline_nuc ILOOP :I
#= GF R2R_oneseq P yrococcus _furiosus inline_nuc ILOOP :J
#= GF R2R_oneseq P yrococcus _furiosus inline_nuc KLOOP :K
#= GF R2R_oneseq P yrococcus_furiosus outline_nuc Q
#= GF R2R_oneseq P yrococcus_furiosus outline_nuc q
#= GF R2R_oneseq P yrococcus_furiosus outline_nuc R
#= GF R2R_oneseq P yrococcus_furiosus outline_nuc r
#= GF R2R_oneseq P yrococcus_furiosus outline_nuc s
#= GF R2R_oneseq P yrococcus _furiosus tick_label q 8 bp
#= GF R2R_oneseq P yrococcus _furiosus tick_label r 9 bp
#= GF R2R_oneseq Pyrococcus_abyssi_GE s h a d e _ a l o n g _ b a c k b o n e ILOOP : I rgb :0 ,255 ,0

#= GF R2R_oneseq P yrococcus_abyssi_GE tic k_label ILOOPL :i 5 ’ guide sequence
#= GF R2R_oneseq Pyrococcus_abyssi_GE s h a d e _ a l o n g _ b a c k b o n e ILOOP : J rgb :0 ,255 ,0
#= GF R2R_oneseq P yrococcus_abyssi_GE tic k_label ILOOPL :j 3 ’ guide sequence
#= GF R2R_oneseq Pyrococcus_abyssi_GE s h a d e _ a l o n g _ b a c k b o n e KLOOP : K rgb :0 ,129 ,255
#= GF R 2R _o n es eq Pyrococcus_abyssi_GE v a r _ b a c k b o n e _ r a n g e _ s i z e _ f a k e _ n u c s 10 7 8
#= GF R2R_oneseq P yrococcus_aby ssi_GE inline_nuc ILOOP : I
#= GF R2R_oneseq P yrococcus_aby ssi_GE inline_nuc ILOOP : J
#= GF R2R_oneseq P yrococcus_aby ssi_GE inline_nuc KLOOP : K
#= GF R2R_oneseq P yrococcus_abyssi_GE outline_nuc Q
#= GF R2R_oneseq P yrococcus_abyssi_GE outline_nuc q
#= GF R2R_oneseq P yrococcus_abyssi_GE outline_nuc R
#= GF R2R_oneseq P yrococcus_abyssi_GE outline_nuc r
#= GF R2R_oneseq P yrococcus_abyssi_GE outline_nuc s
#= GF R2R_oneseq P yrococcus_ab yssi_GE tick_label q 8 bp
#= GF R2R_oneseq P yrococcus_ab yssi_GE tick_label r 9 bp
#= GF R2R_oneseq Pyrococcus_horikoshi s h a d e _ a l o n g _ b a c k b o n e ILOOP : I rgb :0 ,255 ,0

#= GF R2R_oneseq Pyrococcus_horikoshi ti ck_label ILOOPL :i 5’ guide sequence
#= GF R2R_oneseq Pyrococcus_horikoshi s h a d e _ a l o n g _ b a c k b o n e ILOOP : J rgb :0 ,255 ,0
#= GF R2R_oneseq Pyrococcus_horikoshi ti ck_label ILOOPL :j 3’ guide sequence
#= GF R2R_oneseq Pyrococcus_horikoshi s h a d e _ a l o n g _ b a c k b o n e KLOOP : K rgb :0 ,129 ,255
#= GF R 2R _o n eseq Pyrococcus_horikoshi v a r _ b a c k b o n e _ r a n g e _ s i z e _ f a k e _ n u c s 10 7 8
#= GF R2R_oneseq Pyrococcus_ horikoshi inline_nuc ILOOP : I
#= GF R2R_oneseq Pyrococcus_ horikoshi inline_nuc ILOOP : J
#= GF R2R_oneseq Pyrococcus_ horikoshi inline_nuc KLOOP : K
#= GF R2R_oneseq Pyrococcus_horikoshi outline_nuc Q
#= GF R2R_oneseq Pyrococcus_horikoshi outline_nuc q
#= GF R2R_oneseq Pyrococcus_horikoshi outline_nuc R
#= GF R2R_oneseq Pyrococcus_horikoshi outline_nuc r
#= GF R2R_oneseq Pyrococcus_horikoshi outline_nuc s
#= GF R2R_oneseq Pyrococcu s_horikoshi tick_label q 8 bp
#= GF R2R_oneseq Pyrococcu s_horikoshi tick_label r 9 bp
#= GF R2R_oneseq P . furiosus shade_along_backbone ILOOP : I rgb :0 ,255 ,0

#= GF R2R_oneseq P . furiosus tick_label ILOOPL : i 5 ’ guide sequence
#= GF R2R_oneseq P . furiosus shade_along_backbone ILOOP : J rgb :0 ,255 ,0
#= GF R2R_oneseq P . furiosus tick_label ILOOPL : j 3 ’ guide sequence
#= GF R2R_oneseq P . furiosus shade_along_backbone KLOOP : K rgb :0 ,129 ,255
#= GF R2R_oneseq P . furiosus v a r _ b a c k b o n e _ r a n g e _ s i z e _ f a k e _ n u c s 10 7 8
#= GF R2R_oneseq P . furiosus inline_nuc ILOOP : I
#= GF R2R_oneseq P . furiosus inline_nuc ILOOP : J
#= GF R2R_oneseq P . furiosus inline_nuc KLOOP : K
#= GF R2R_oneseq P . furiosus outline_nuc Q
#= GF R2R_oneseq P . furiosus outline_nuc q
#= GF R2R_oneseq P . furiosus outline_nuc R
#= GF R2R_oneseq P . furiosus outline_nuc r
#= GF R2R_oneseq P . furiosus outline_nuc s
#= GF R2R_oneseq P . furiosus tick_label q 8 bp
#= GF R2R_oneseq P . furiosus tick_label r 9 bp
#= GF R2R_oneseq Thermococcus_kodakar s h a d e _ a l o n g _ b a c k b o n e ILOOP : I rgb :0 ,255 ,0

#= GF R2R_oneseq Thermococcus_kodakar ti ck_label ILOOPL :i 5’ guide sequence
#= GF R2R_oneseq Thermococcus_kodakar s h a d e _ a l o n g _ b a c k b o n e ILOOP : J rgb :0 ,255 ,0
#= GF R2R_oneseq Thermococcus_kodakar ti ck_label ILOOPL :j 3’ guide sequence
#= GF R2R_oneseq Thermococcus_kodakar s h a d e _ a l o n g _ b a c k b o n e KLOOP : K rgb :0 ,129 ,255
#= GF R 2R _o n eseq Thermococcus_kodakar v a r _ b a c k b o n e _ r a n g e _ s i z e _ f a k e _ n u c s 10 7 8
#= GF R2R_oneseq Thermococcu s_kodakar inline_nuc ILOOP : I
#= GF R2R_oneseq Thermococcu s_kodakar inline_nuc ILOOP : J
#= GF R2R_oneseq Thermococcu s_kodakar inline_nuc KLOOP : K
#= GF R2R_oneseq Thermococcus_kodakar outline_nuc Q
#= GF R2R_oneseq Thermococcus_kodakar outline_nuc q
#= GF R2R_oneseq Thermococcus_kodakar outline_nuc R
#= GF R2R_oneseq Thermococcus_kodakar outline_nuc r
#= GF R2R_oneseq Thermococcus_kodakar outline_nuc s
#= GF R2R_oneseq Thermococ cus_kodakar tick_label q 8 bp
#= GF R2R_oneseq Thermococ cus_kodakar tick_label r 9 bp
#= GF Makefile skeleton - with - pairbonds

//
8. R2R meta files including the instructions to process the input

files associated with the different representations generated in
Figs. 16 and 17.
R F 0 0 0 6 5 _ g u i d e _ t a r g e t . cons . sto
R F 0 0 0 6 5 _ g u i d e _ t a r g e t . sto oneseq Pa21 - S892
RF00065_guide_target - hybrid . cons . sto
RF00065_seed . cons . sto

RF00065_seed . sto oneseq P y r o c o c c u s _ f u r i o s u s
RF00065_seed . sto oneseq P y r o c o c c u s _ a b y s s i _ G E
RF00065_seed . sto oneseq P y r o c o c c u s _ h o r i k o s h i
RF00065_seed . sto oneseq T h e r m o c o c c u s _ k o d a k a r
RF00065_seed . cons . sto skeleton - with - pairbonds
RF00065_seed - ILOOP75 . cons . sto
RF00065_seed - ILOOP85 . cons . sto
Acknowledgements
The author wishes to thank the French Centre National de la

Recherche Scientifique (CNRS) for its continued support, the NSF-
funded RNA Ontology Consortium, and Jim Procter for stimulat-
ing discussions.
References
1. Andronescu M, Bereg V, Hoos H, Condon A 8. Fourmann J-B, Tillault A-S, Blaud M, Leclerc
(2008) RNA STRAND: the RNA secondary F, Branlant C, Charpentier B (2013)
structure and statistical analysis database. Comparative study of RNA conformational
BMC Bioinform 9(1):340. doi: 10.1186/1471- dynamics during assembly of two box H/ACA
2105-9-340 ribonucleoprotein pseudouridine-synthases
2. Bateman A, Agrawal S, Birney E, Bruford EA, revealing uncorrelated efficiencies in assembly
Bujnicki JM, Cochrane G, Cole JR, Dinger and activity. PLoS ONE 8(7):e70313.
ME, Enright AJ, Gardner PP, Gautheret D, doi:10.1371/journal.pone. 0070313
Griffiths-Jones S, Harrow J, Herrero J, Holmes 9. Fourmy D, Yoshizawa S (2012) Protein-RNA
IH, Huang H-D, Kelly KA, Kersey P, Kozomara footprinting: an evolving tool. Wiley Interdiscip
A, Lowe TM, Marz M, Moxon S, Pruitt KD, Rev RNA 3(4):557–566. doi:10.1002/wrna. 1119
Samuelsson T, Stadler PF, Vilella AJ, Vogel 10. Gansner ER, North SC (2000) An open graph
J-H, Williams KP, Wright MW, Zwieb C (2011) visualization system and its applications to
RNAcentral: a vision for an international data- software engineering. Softw – Pract Exp
base of RNA sequences. RNA 17(11):1941– 30(11):1203–1233. doi:10.1002/ 1097-024X
1946. doi:10.1261/rna. 2750811 (200009)30:11<1203::AID- SPE338>3.3.
3. Berman HM, Olson WK, Beveridge DL, CO; 2-E
Westbrook J, Gelbin A, Demeny T, Hsieh SH, 11. Gansner ER, Koutsofios E, North SC, Vo KP
Srinivasan AR, Schneider B (1992) The nucleic (1993) A technique for drawing directed
acid database. A comprehensive relational data- graphs. IEEE Trans Softw Eng 19(3):214–
base of three-dimensional structures of nucleic 230. doi:10.1109/32.221135
acids. Biophys J 63(3):751–759. doi:10.1016/ 12. Griffiths-Jones S, Bateman A, Marshall M,
S0006-3495(92) 81649-1 Khanna A, Eddy SR (2003) Rfam: an RNA
4. Bruccoleri RE, Heinrich G (1988) An improved family database. Nucl Acids Res 31(1):439–
algorithm for nucleic acid secondary structure 441. doi: 10.1093/nar/gkg006
display. Comput Appl Biosci 4(1):167–173. 13. Gruber AR, Lorenz R, Bernhart SH, Neuböck
doi:10.1093/bioinformatics/4.1. 167 R, Hofacker IL (2008) The Vienna RNA
5. Byun Y, Han K (2009) PseudoViewer3: generat- websuite. Nucl Acids Res 36(Web Server
ing planar drawings of large-scale RNA structures issue):W70–W74. doi:10.1093/nar/gkn188
with pseudoknots. Bioinformatics 25(11):1435– 14. Hofacker IL, Fontana W, Stadler PF, Bonhoeffer
1437. doi:10.1093/bioinformatics/ btp252 LS, Tacker M, Schuster P (1994) Fast folding
6. Cannone JJ, Subramanian S, Schnare MN, and comparison of RNA secondary structures.
Collett JR, D’Souza LM, Du Y, Feng B, Lin N, Monatshefte für Chemie/Chem Mon
Madabusi LV, Muller KM, Pande N, Shang Z, 125(2):167–188. doi:10.1007/ BF00818163
Yu N, Gutell RR (2002) The comparative 15. Jossinet F, Westhof E (2005) Sequence to
RNA web (CRW) site: an online database of structure (S2S): display, manipulate and
comparative sequence and structure infor- interconnect RNA data from sequence to
mation for ribosomal, intron, and other structure. Bioinformatics 21(15):3320–3321.
RNAs. BMC Bioinform 3(2). doi: 10.1186/ doi:10.1093/bioinformatics/ bti504
1471-2105-3-15 16. Jossinet F, Ludwig TE, Westhof E (2010)
7. Darty K, Denise A, Ponty Y (2009) VARNA: Assemble: an interactive graphical tool to ana-
interactive drawing and editing of the RNA sec- lyze and build RNA architectures at the 2D
ondary structure. Bioinformatics 25(15): and 3D levels. Bioinformatics 26(16):2057–
1974–1975. doi:10.1093/bioinformatics/ btp250 2059. doi:10.1093/ bioinformatics/ btq321
17. Kaiser A, Krüger J, Evers DJ (2007) RNA mov- ary structure. Bioinformatics 19(2):299–300.
ies 2: sequential animation of RNA secondary doi:10.1093/bioinformatics/19.2. 299
structures. Nucl Acids Res 35(Web Server 28. Schlatterer JC, Brenowitz M (2009)
issue):W330–W334. doi:10.1093/nar/gkm309 Complementing global measures of RNA fold-
18. Lai D, Proctor JR, Zhu JYA, Meyer IM (2012) ing with local reports of backbone solvent
R-CHIE: a web server and R package for visu- accessibility by time resolved hydroxyl radical
alizing RNA secondary structures. Nucl Acids footprinting. Methods 49(2):142–147.
Res 40(12):e95. doi:10.1093/nar/gks241 doi:10.1016/j. ymeth.2009.04.019
19. Lemay J-F, Lafontaine DA (2007) Core 29. Shabash B, Wiese KC, Glen E (2012)
requirements of the adenine riboswitch Improving the portability and performance
aptamer for ligand binding. RNA 13(3):339– of jViz.RNA – a dynamic RNA visualization
350. doi:10. 1261/rna.142007 software. In: PRIB, pp 82–93. doi:10.1007/
20. Leonard CW, Hajdin CE, Karabiber F, 978-3-642-34123-6\_8
Mathews DH, Favorov OV, Dokholyan NV, 30. Waugh A, Gendron P, Altman R, Brown JW,
Weeks KM (2013) Principles for understand- Case D, Gautheret D, Harvey SC, Leontis N,
ing the accuracy of SHAPE-directed RNA Westbrook J, Westhof E, Zuker M, Major F
structure modeling. Biochemistry 52(4):588– (2002) RNAML: a standard syntax for
595. doi:10.1021/bi300755u exchanging RNA information. RNA 8(6):
21. Leontis NB, Altman RB, Berman HM, Brenner 707–717. doi:10.1017/ S1355838202028017
SE, Brown JW, Engelke DR, Harvey SC, 31. Weinberg Z, Breaker RR (2011) R2R–software
Holbrook SR, Jossinet F, Lewis SE, Major F, to speed the depiction of aesthetic consensus
Mathews DH, Richardson JS, Williamson JR, RNA secondary structures. BMC Bioinform
Westhof E (2006) The RNA ontology consor- 12:3. doi: 10.1186/1471-2105-12-3
tium: an open invitation to the RNA community. 32. Wiese KC, Glen E, Vasudevan A (2005) JViz.
RNA 12(4):533–541. doi:10.1261/rna. 2343206 Rna–a Java tool for RNA secondary structure
22. Mathews DH, Sabina J, Zuker M, Turner DH visualization. IEEE Trans Nanobiosci 4(3):212–
(1999) Expanded sequence dependence of 218. doi:10.1109/TNB.2005. 853646
thermodynamic parameters improvespredic- 33. Xu Z, Culver GM (2009) Chemical probing of
tion of RNA secondary structure. J Mol Biol RNA and RNA/protein complexes. Methods
288:911–940. doi:10.1006/jmbi.1999. 2700 Enzymol 468:147–165. doi:10.1016/ S0076-
23. Menzel P, Seemann SE, Gorodkin J (2012) 6879(09) 68008-3
RILogo: visualizing RNA–RNA interactions. 34. Yang H, Jossinet F, Leontis N, Chen L,
Bioinformatics 28(19):2523–2526. doi:10. Westbrook J, Berman HM, Westhof E (2003)
1093/bioinformatics/ bts461 Tools for the automatic identification and
24. Mortimer SA, Trapnell C, Aviran S, Pachter L, classification of RNA base pairs. Nucl Acids
Lucks JB (2012) SHAPE-Seq: high-through- Res 31(13):3450–3460. doi:10.1093/nar/
put RNA structure analysis. Curr Protoc gkg529
Chem Biol 4(4):275–297. doi:10.1002/ 35. Zarringhalam K, Meyer MM, Dotu I, Chuang
9780470559277. ch120019 JH, Clote P (2012) Integrating chemical foot-
25. Muller G, Gaspin C, Etienne A, Westhof E printing data into RNA secondary structure pre-
(1993) Automatic display of RNA secondary diction. PLoS ONE 7(10):e45160. doi:10.1371/
structures. Comput Appl Biosci 9(5):551–561. journal.pone. 0045160
doi:10.1093/bioinformatics/9.5. 551 36. Zhang C, Darnell RB (2011) Mapping in vivo
26. Ouyang Z, Snyder MP, Chang HY (2012) protein-RNA interactions at single-nucleotide
SeqFold: genome-scale reconstruction of RNA resolution from HITS-CLIP data. Nat Biotechnol
secondary structure integrating high-through- 29(7):607–614. doi:10.1038/nbt.1873
put sequencing data. Genome Res. 37. Zuker M, Stiegler P (1981) Optimal computer
doi:10.1101/gr.138545.112 folding of large RNA sequencesusing thermo-
27. Rijk PD, Wuyts J, Wachter RD (2003) RnaViz dynamics and auxiliary information. Nucl Acids
2: an improved representation of RNA second- Res 9:133–148. doi:10.1093/nar/9.1.133
Chapter 6
Modeling and Predicting RNA Three-Dimensional Structures

Jérôme Waldispühl and Vladimir Reinharz
Abstract
Modeling the three-dimensional structure of RNAs is a milestone toward better understanding and
prediction of nucleic acids molecular functions. Physics-based approaches and molecular dynamics simula-
tions are not tractable on large molecules with all-atom models. To address this issue, coarse-grained
models of RNA three-dimensional structures have been developed. In this chapter, we describe a graphical
modeling based on the Leontis–Westhof extended base-pair classification. This representation of RNA
structures enables us to identify highly conserved structural motifs with complex nucleotide interactions in
structure databases. Further, we show how to take advantage of this knowledge to quickly and simply
predict three-dimensional structures of large RNA molecules.
Key words Tertiary structure, RNA motifs, Extended secondary structure, Base-pair classification,
Modeling, Prediction
1 Introduction
RNAs perform a broad range of catalytic and regulatory functions

in cells, which often use the folding properties of these molecules.
RNA structures are typically described at two levels of organiza-
tion: the secondary structure, which enumerates the Watson–Crick
and Wobble base pairs that create the backbone of the molecular
structure; and the tertiary structure, which indicates the positions
of each atom in the molecule.
The study of secondary structures and base-pairing properties
can reveal fundamental insights into the functional mechanisms of
RNAs such as frameshift elements [1] and riboswitches [2].
However, the information provided by this representation is some-
times not sufficient to describe the complexity of intermolecular
and intramolecular interactions that govern RNA functions.
A seminal example is the sarcin–ricin factor-binding loop, which
possesses a sophisticated tertiary structure [3].
Classical approaches to predict and analyze three-dimensional
molecular structures make use of molecular dynamic simulations
[4]. This methodology has the potential to perform calculations on
101
102 Jérôme Waldispühl and Vladimir Reinharz
all-atom models, but suffers from high computational complexity.

Straightforward applications are thus limited to small molecules
(approximately 50 nt.) on a short period of time with very small
body motion. Coarse-grained modeling of RNA structures enables
us to overcome some of these limitations [5–8], but the practical
impact and theoretical horizon of these technologies remain
unchanged. In addition, molecular dynamic simulations often
require fine-tuning and can be challenging to run and interpret for
nonexperts.
Recently, several research groups developed alternate strategies
to model and predict RNA three-dimensional structures. One of
the most successful approach, applied in the MC-Fold|MC-Sym
pipeline [9], RNA2D3D [10] or 3dRNA [11], consists in predict-
ing a secondary structure first, and use it to build a three-dimensional
model. Other programs have employed fragment-assembly [12, 13]
or conditional random fields techniques [14]. It is worth noting
that, as we see in a previous chapter, comparative modeling tech-
niques can also be applied to RNA molecules if one structural
homolog has been already identified [15].
In this chapter, we introduce a versatile model for RNA struc-
tures, which enables us to describe essential features and fine-
grained details of RNA three-dimensional structures while
preserving the complexity of the representation to a minimum.
This methodological advance is key to large-scale analysis and to
better understanding of critical features of RNA molecular struc-
tures. In turn, this knowledge enables us to define new computa-
tional techniques to quickly and easily predict three-dimensional
structures of large RNA molecules.
First, we describe the base-pair classification at the base of this
model [16], and present rnaview [17]—a program that annotates
automatically RNA three-dimensional structures. Then, we intro-
duce the concept of RNA structural motifs and present recent
databases and online resources based on this definition [18, 19].
Finally, show how to use this knowledge to predict RNA three-
dimensional structures. The pipeline described in this chapter
works in two steps. First, we predict secondary structures with
RNAsubopt [20] and expand the prediction to a full base-pairing
interaction network with RNA-MoIP [21]. Next, we use this
graphical representation to build the three-dimensional structure
of the RNA with MC-Sym [9].
2 Material
2.1 Sequence The Nucleic Acids Database [22] is a repository of experimentally

and Structure Data determined three-dimensional RNA structures maintained at
Rutgers University, which is available at http://ndbserver.rutgers.
edu. This is a specialized version of the Protein Data Bank [23],
Modeling and Predicting RNA Three-Dimensional Structures 103
which is available at http://www.rcsb.org/pdb. The structures are

typically stored in files with the extension “.pdb”. In particular, the
PDB files store the spatial coordinates of each atom in the mole-
cule. Plain sequences can also be downloaded separately under the
FASTA format.
In this chapter, we illustrate our methods on the sequence and
experimentally determined tertiary structures of tRNA(Cys) of
Archaeoglobus fulgidus [24]. The PDB ID of this molecule is 2DU3.
2.2 Automatic The rnaview software is used to identify all base-pairing interac-
Annotation of 3D tions that represent in a RNA tertiary structure [17]. This program
Structures can be downloaded at: http://ndbserver.rutgers.edu/ndbmodule/
services/download/rnaview.html and is currently available for
Linux, UNIX, SUN, and MAC systems.
2.3 RNA Secondary The prediction of RNA secondary structures is performed with the
Structure Prediction RNAfold program available in the Vienna RNA package [20, 25].
Tools The software suite is available at http://www.tbi.univie.ac.at/
RNA/ and can run under LINUX, UNIX, MAC, and WINDOWS
systems. In this chapter, we used the version 2.1.2 of the Vienna
RNA package. Web services are also available at http://rna.tbi.uni-
vie.ac.at/.
2.4 Insertion of RNA We perform the insertion of RNA motifs inside predicted RNA
Motifs in Secondary secondary structures with the RNA-MoIP software [21] available
Structure at http://csb.cs.mcgill.ca/RNAMoIP/. Noticeably, this software
requires installing the Gurobi solver (http://www.gurobi.com/),
which is free for academic users.
2.5 Building RNA 3D We use the MC-Sym software to build tertiary structure from a
Structures from a RNA base-pairing interaction network [9]. MC-Sym is part of the MC-
Interaction Graphs tools package available at: http://www.major.iric.ca/MajorLabEn/
MC-Tools.html.
3 Methods
In this section, we describe how to extract from a three-dimensional

structure, the list of all base-pairing interactions. Then, we describe
how to use databases of recurrent motifs to predict the tertiary
structure of large RNA molecules.
3.1 Classification Watson–Crick (C–G and A–U) and Wobble (G–U) base pairs are
of Base-Pairing the most common type of interactions. They create a scaffold for
Interactions the tertiary structure. Nonetheless, the analysis of RNA crystal
structures revealed a diversity of base-pairing interactions that goes
far beyond these canonical base pairs. In order to facilitate the
description of RNA structures, Leontis and Westhof proposed a
complete nomenclature of base-pairing interactions [16] that aims

to provide a better description of the complexity of the tertiary
structure. Key to this model is to introduce a detailed representa-
tion of the base of nucleotides, which in turn enables us to define
a more sophisticated classification of base-pairing interactions.
Thereby, a base is abstracted with a right triangle modeling all
edges of the molecules that can be potentially involved in an inter-
action with other nucleotides (see Fig. 1). The three interacting
edges are: the Watson–Crick edge, the Hoogsteen edge, and the
Sugar edge. The hypotenuse is associated with the Hoogsteen
edge, and the vertex created by the Hoogsteen and Sugar edge
represents the root of the base.
Base-pairing interactions between nucleotides can now occur
between any of these edges. In addition, a complete description of
these interactions must also specify the relative orientation of the
bases (i.e., the glycosidic bond orientation). Hence, we indicate if
the bases are oriented in the same or in opposite directions. These
configurations are respectively named trans and cis configurations.
These parameters enable us to define a complete nomenclature
of all possible base-pairing interactions between nucleotides. This
catalog is shown in Fig. 1 and includes all 12 possible base pairs
that are found in RNA structures.
In secondary structure diagrams, base pairs are often repre-
sented with specific links. C–G base pairs are represented with two
parallel line, A–U base pairs with a single line and G–U base pairs
with a circle. Leontis and Westhof also assigned a symbol to each
base-pairing interaction of their classification. Watson–Crick edges
are represented with a circle, Hoogsteen edges with a square and
Sugar edges with a triangle. In addition, black symbols represent a
cis orientation, and white symbols a trans orientation. This nota-
tion is illustrated at the bottom of each base pair in Fig. 1.
Using this nomenclature, RNA tertiary structures can be
decomposed into a network of base-pairing interactions. Unlike
classical secondary structures, nucleotides are now no longer
restricted to interact with at most one other nucleotide. Instead,
they can be involved in multiple interactions and create, for instance,
base triples. Crossing interactions are also permitted. While this
modeling does not obviously encompass all details of three-dimen-
sional atomistic structures, it appears to store most of the informa-
tion that one needs to reconstruct full tertiary structures [9].
3.2 Annotation The rnaview program annotates automatically of all base-pairing

of Base-Pairing interactions found in RNA tertiary structure using the Leontis–
Interactions from RNA Westhof classification [16]. It can also produce an image of the
3D Structures interaction graph. The command line to run the program is:
rnaview –p input.pdb
Fig. 1 Base modeling and Leontis–Westhof base-pair classification. The base of a nucleotide is represented
with a right triangle. The hypotenuse represents the Hoogsteen edge (noted “H”), and the other sides are
associated with the Watson–Crick edge (noted “W”) and Sugar edge (noted “S”). This figure represents all 12
base-pair interactions with cis or trans orientation
where “input.pdb” is your input three-dimensional structure

and the “-p” flag is an optional parameter that indicates to the
program to create a visualization of the annotated structure.
Figure 2 shows the graphical output of rnaview on the tRNA(Cys)
from Archaeoglobus fulgidus.
The output files are stored in the same directory as the input
file. The list of base-pairing interactions is stored in a new text file
named after the input file and appended with the suffix “.out”.
Similarly, if you used the “-p” option, rnaview also creates a post-
script file with a drawing of the base-pairing interaction network.
10
U C
U G
C G
C G
G C 15U A G
G
5
A G
C C
50 U A G
G
G C 20 55 C
G
C G U
G
C C 5’
3’
C G
45 A U
C G
C G 25
U G
G A
40 G G
A
U G 30
C G
G C
G U
35
C U
Fig. 2 rnaview annotation of a crystal structure of tRNA(Cys) from Archaeoglobus

fulgidus (Fukunaga and Yokoyama [24])
The following code, calculated from the tRNA(Cys) from

Archaeoglobus fulgidus (2DU3), illustrates a typical output of
rnaview:
9_12, B: 9 U-G 12 B: S/W tran syn n/a
16_52, B: 16 U-A 52 B: W/H tran XXIV
18_51, B: 18 G-G 51 B: H/S tran n/a
18_53, B: 18 G-C 53 B: +/+cis syn XIX
The first column gives the indices, separated by an underscore,
of the two nucleotides involved in the base pair. These indices are
calculated by rnaview and correspond to their positions in the
sequence(s) entered in the program. They may eventually differ
from the indices stored in the PDB file, which are given in the third
and fifth columns.
The letters in the second and sixth columns indicate the chain
of the two nucleotides. In our case, the RNA molecule has a single
chain named “B” in the PDB file. The types of the nucleotides
forming the base pair are indicated in the fourth column.
Finally, the seventh and eighth columns describe the edges
involved in the base pair (Watson–Crick, Hoogsteen, or Sugar
edge), followed by the orientation of their interaction (cis or trans).
For instance, in the example above the first row says “the nucleo-
tide U at index 9, and the nucleotide G at position 12, form a trans
base pair between the Sugar edge (S) of nucleotide U and Watson–
Crick edge (W) of the nucleotide G.” It is worth noting that
classical Watson–Crick base pairs are annotated with +/+ (C–G
base pair) or −/− (A–U base pair). More details on the base-pair
notation, including “marginal” interactions, can be found in refs.
[17] or [26].
As expected, most base pairs represented in Fig. 2 are Watson–
Crick base pairs. Nonetheless, we note that noncanonical base pairs
tend to get concentrated in specific regions of the secondary struc-
ture backbone and create sophisticated local motifs. In fact, this
observation is recurrent in most annotated structures. It will give
rise to the notion of RNA motifs introduced in the next section.
Finally, it is worth noting that we can alternatively use the pro-
gram MC-Annotate [27] to perform similar annotations. However,
the classification used by MC-Annotate differs slightly from the
one used in the motif databases described further.
3.3 RNA Motifs The modeling of RNA tertiary structures into networks of base-
pairing interactions revealed recurrent motifs conserved across
families of structures. These structural patterns form a base toward
better understanding of complex organizations of nucleotides
inside hairpins, internal loops and beyond (See previous section for
a definition of the secondary structure elements). Several method-
ologies and databases have been developed to mine and store these
data. Among the most popular repositories, we count the RNA 3D
Hub maintained by the Bowling Green State University RNA

group at http://rna.bgsu.edu/rna3dhub/ [19], and the RNA 3D
Motif database developed at the university of Paris-Sud and avail-
able at http://rna3dmotif.lri.fr/ [18]. More recently, international
institute for molecular and cell biology in Warsaw introduced a
novel motif database RNAbricks [28], available at http://iimcb.
genesilico.pl/rnabricks/.
Beside databases of annotated structures (i.e., the RNA
Structure Atlas) and recurrent RNA 3D motifs (RNA 3D Motif
Atlas), the RNA 3D Hub offers large suite of online tools and Web
services. In particular, users can here retrieve local 3D motifs with
WebFR3D (http://rna.bgsu.edu/main/webapps/webfr3d/) or
align RNA tertiary structures with R3D Align (http://rna.bgsu.
edu/main/webapps/webr3dalign/).
Each database developed its own local 3D motif format
description, based on the Leontis–Westhof base-pair classification.
RNA-MoIP use the format introduced with RNA3Dmotif [18].
An example of an internal loop motif is provided below:
Bases: 26_G 27_A 28_G 29_A 30_G 39_U 40_G 41_G 42_U
(39_U)--- C/C - ---(40_G)
(30_G)--- 5/5 s ---(40_G)
(30_G)--- +/+c ---(39_U)
(40_G)--- C/C - ---(41_G)
(29_A)--- C/C - ---(30_G)
(28_G)--- 5/5 s ---(41_G)
(28_G)--- C/C - ---(29_A)
(41_G)--- C/C - ---(42_U)
(26_G)--- +/+c ---(42_U)
(27_A)--- C/C - ---(28_G)
(26_G)--- C/C - ---(27_A)
The first line starting with the key word “Bases” indicates the
nucleotides involved in this motif. For each nucleotide, we display
its index and its type separated by an underscore. Each following
line describes a base-pairing interaction in this motif. The syntax of
a base pair is structured as follows. On both ends of the row, the two
nucleotides (index and type) are shown between parentheses. Then,
the type of the interaction is shown in the middle, surrounded by
three dashes on each side. The first field indicates the edges involved
and the second the base-pair orientation (“c” for cis, “t” for trans,
and “s” for a stacking). An interaction annotated “C/C” stands for
a backbone connection (i.e., consecutive nucleotides) and thus may
not be considered as a base pair in our discussion. More information
can be found at http://rna3dmotif.lri.fr/help.html.
RNA-MoIP needs to decompose these motifs into compo-
nents to permit their insertion. Components are largest sequences
of contiguous nucleotides that belong to a motif. For instance, in

the example above, we have two components ([26_G,27_A,28_G,
29_A,30_G] and [39_U,40_G,41_G,42_U]). In fact, the defini-
tion of components follows the classical secondary structure loop
classification. Hairpins have one single component, bulges and
internal loops have two components, and k-way junctions have k
components. This definition has been introduced with RNA-MoIP
to facilitate the description of integer programming equations.
Figure 3 illustrates the definition of motifs. In this example, we
identify two motifs (a hairpin and a internal loop) inserted into a
single stem. The figure shows the 3D structures and the base-
pairing interaction graphs of each motif. We observe that the hair-
pin has one single component (i.e., one single stranded region),
while the internal loop has two.
3.4 Prediction Modeling RNA tertiary structure is the first step toward predicting
of RNA Tertiary RNA 3D structures. We now describe how the knowledge accu-
Structures from mulated in motif databases can be used together with RNA sec-
Sequence Data ondary structure prediction methods to predict the structure of
large RNA molecules from sequence data only. The strategy pre-
sented here works in two steps. First, we predict a secondary struc-
ture using classical software such as RNAfold [20] and refine this
Fig. 3 Example of 3D RNA motif insertions in a secondary structure. In green, we show the position of the
hairpin motif “1F7Y.B.6”, and in blue we indicate the position of the internal loop motif “1FKA.A.51”. On the left
side of the motif IDs, we display a 3D structure of the motif together with its base-pairing interaction graph
prediction by inserting RNA 3D motifs and adding noncanonical

base pairs inside the predicted secondary structure with RNA-
MoIP [21]. Next, we use this extended secondary structure to
reconstruct three-dimensional models of the molecule using the
MC-Sym software [9].
3.5 Prediction Our first objective is to create a base-pairing network from sequence
of a Secondary data. Since the majority of base pairs in RNA structures are cis
Structure Scaffold Watson–Crick base pairs, we first predict a secondary structure
(without pseudo-knots) that will be used to build a scaffold of the
interaction graph. Secondary structures (without pseudo-knot) can
be deterministically predicted with RNAfold, or stochastically gen-
erated with RNAsubopt. The command line used to predict the
minimum free energy (MFE) secondary structure with RNAfold is:
RNAfold --noPS < input.fasta
Where input.fasta is a text file (FASTA format recommended)
that stores your input sequence. The--noPS flag is not mandatory,
but it prevents the program to generate a postscript file drawing
the predicted secondary structure. The program returns the input
sequence with its MFE secondary structure in bracket format on
the line below.
However, as discussed in previous chapters, single energy min-
imized structures do not always provide the best secondary struc-
ture prediction. Instead, it is recommended to perform a deeper
exploration of the conformational space and to consider subopti-
mal structures [29, 30]. This approach to secondary structure pre-
diction, originally implemented in mfold [30, 31] and Sfold [32],
is available with the RNAsubopt program in the Vienna RNA
package. The command line for running RNAsubopt is:
RNAsubopt -e 3 < input.fasta
Where the “-e” option specifies the depth of the suboptimal
search. More specifically, this argument indicates the range
(in kCal/mol) from the MFE, within which all suboptimal struc-
tures must be returned. Obviously, the values of that range depen-
dent of the MFE of the sequence, and should be adjusted on a
case-by-case basis. In our experiments, a value of 3 kCal/mol gen-
erated 25 secondary structures; which appears to provide a good
representation of the suboptimal conformational landscape.
The list of suboptimal secondary structures generated by
RNAsubopt (available at http://csb.cs.mcgill.ca/RNAMoIP)
provides us a description of the set of potential secondary structure
backbones. It will be used as it in the next step.
It is worth noting that other software could have been used to
generate the initial secondary structures. RNAstructure [33, 34],
mfold [30, 35], or Sfold [32, 36] make similar prediction than
RNAsubopt. Further, recent software such as MC-Fold [9] and
RNAwolf [37] have been developed to predict extended secondary

structures (including all noncanonical base pairs) directly from
sequence data.
3.6 Prediction We describe how to use RNA-MoIP [21] to insert local motifs into
of a Complete secondary structures generated with RNAsubopt. By default,
Base-Pairing RNA-MoIP aims to inserts motifs from a repository build with
Interaction Graph RNA3Dmotif [18]. This repository of nonredundant motifs is
with Motif Insertion included in the distribution of RNA-MoIP and named “No_
Redondance_DESC” (see Note 1). Nonetheless, advanced users
can also either build themselves an up-to-date motif repository
using RNA3Dmotif, or use databases available on the RNA 3D
Hub [19].
RNA-MoIP is a flexible tool that allows modifications of the
secondary structure to permit the insertion of motifs. In particular,
the program has the capacity to remove a fixed amount of base
pairs from the input secondary structure. This feature is particu-
larly helpful if the predicted secondary structure has incorrectly
predicted base pairs that prevent motifs to be inserted.
Lets assume that you work in a directory that contains the
RNA-MoIP program (i.e., the python script named “RNAMoIP.
py”) and that Gurobi has been properly installed. The command
line to insert motifs in the sequence and secondary structure with
RNA-MoIP is:
gurobi.sh RNAMoIP.py -s <sequence> -ss <structure_list>
–d <path_to_repository>
Where the argument < sequence > is a string representing the
primary structure of the RNA sequence, <structure_list > is the
name of the file that stores the secondary structures in bracket for-
mat generated by RNAsubopt, and < path_to_repository > is the
location of the motif repository. It indicates the path to the motif
repository stored in the directory named “CATALOGUE,” which
is distributed with the RNA-MoIP package at http://csb.cs.mcgill.
ca/RNAMoIP/. Assuming that the directory “CATALOGUE” is
in your current directory, the value of < path_to_repository > is the
string “./CATALOGUE/No_Redondance_DESC/”.
RNA-MoIP accepts an additional parameter to control the
maximum number of base pairs that can be removed. This param-
eter can be adjusted with a float number (between 0 and 1) through
the option −r. By default, RNA-MoIP allows up to 30 % (i.e., −r
0.3) of base pairs to be removed. This is a reasonable choice as the
base-pair prediction positive predictive value (PPV) is roughly
60 % for classical secondary structure predictors such as RNAfold
and mfold [38]. Nonetheless, users can decrease this value if they
are confident in their predicted secondary structure.
Lets now run a concrete example on the tRNA(Cys) sequence,

and lets call “RNAsubopt.out” the file storing the output of
RNAsubopt. Then, the RNA-MoIP command line is:
gurobi.sh RNAMoIP.py -s
“GCCAGGGUGGCAGAGGGGCUUUGCGGCGGACUGCAGAUCCGCUUUACCCCGGUUCGAA
UCCGGGCCCUGGC”-ss RNAsubopt.out -d
“./CATALOGUE/No_Redondance_DESC/”
The program outputs first the usual output of Gurobi (the
mathematical solver used by RNA-MoIP). We are interested by the
output produced after the line starting with “Best objective.” It
contains the secondary structure used to insert the motifs, the IDs
of the inserted motifs, the positions of the deleted base pairs, and
the score of the solution. The RNA-MoIP command above will
output the following results:
Solution for the secondary structure:
(((((((..((((((…)))))).(((((.......))))).....(((((.......))))))))))))
Optimal solution nb: 1
Corrected secondary structure: ((((((…(((.........)))..

((((.......))))......((((.........)))).))))))
C-2DU6.D.2-12-22-1
C-3CUL.D.6-51-61-1
C-2DU3.D.3-30-38-1
C-2DU5.D.1-6-10-1
C-2DU5.D.1-24-27-2
C-2DU5.D.1-41-48-3
C-2DU5.D.1-64-66-4
D-15-19
D-14-20
D-13-21
D-26-42
D-52-60
D-7-65
The optimal solutions has as value:

-663.0
The first line starts with “Solution for the secondary structure”
and returns the best secondary structure extracted from the list of
suboptimal structure, which has been used to insert to motifs. The
next line (starting with “Optimal solution nb”) indicates how
many suboptimal structures have been used. In our case, only one
structure can be used to build the optimal solution.
The line starting with “Corrected secondary structure” shows
the structure used by RNA-MoIP. Since RNA-MoIP may remove
some base pairs to insert motifs, this structure can be different to

the one returned on the first line.
Then, the program outputs information about the inserted
motifs and deleted base pairs. The rows starting with a “D” are the
positions of the deleted base pairs. Hence, here the output tells us
that the base pairs (15,19), (14,20), (13,21), (26,42), (52,60),
and (7,65) have been removed to insert the motifs. We highlight
in red the deleted base pairs.
(((((((..((((((…)))))).(((((.......))))).....(((((.......))))))))))))
As indicated in the third line, the secondary predicted (or
updated) by RNA-MoIP is:
((((((…(((.........)))..((((.......))))......((((.........)))).))))))
The rows starting with a “C” display information about the
inserted motifs. Each row indicates where a component of a motif
(i.e., contiguous sequence of a local motif) has been used. The
syntax of these lines is:
C-<ID>-<first_index>-<last_index>-<component_number>
Where < ID > is the identifier (or name) of the inserted motif,
<first_index > is the first position and < last_index > is the last posi-
tion of the insertion. The last value < component_number > indi-
cates which component of the motif has been used.
In our example, four motifs have been inserted: three hairpins
and one 4-way junction. For each motif, Table 1 shows the ID,
type, and positions at which each component has been inserted.
Figure 4 illustrates these results and shows where the motifs and
components have been inserted by RNA-MoIP in the secondary
structure.
The last line in the output returns the value of the objective
function that has been optimized by RNA-MoIP. In our example,
the value returned is -663.0. Details about this function can be
found in ref. [21].
Table 1
Description of motifs inserted by RNA-MoIP in the suboptimal
secondary structures generated by RNAsubopt for the tRNA(Cys)
from Archaeoglobus fulgidus
Motif ID Type Indices of insertion sites

2DU6.D.2 Hairpin [12, 22]
3CUL.D.6 Hairpin [51,61]
2DU3.D.3 Hairpin [30, 38]
2DU5.D.1 4-Way junction [6,10], [24, 27], [41,48], [64,66]
Fig. 4 Location of RNA motifs inserted by RNA-MoIP in the secondary structure

of tRNA(Cys) from Archaeoglobus fulgidus. Three hairpins (2DU6.D.2 in green,
2CUL.D.6 in magenta, and 2DU3.D.3 in blue) and one 4-way junction (2DU5.D.1 in
red) have been inserted. We note the 4 components of the 4-way junction, which
correspond to the 4 single strands of the multi-loop
3.7 Reconstruction Now, we describe how to use a base-pairing network to build

of Tertiary Structures three-dimensional structures with the MC-Sym software [9]. This
from Base-Pairing operation requires writing scripts that will be used to run MC-Sym
Interaction Networks at http://www.major.iric.ca/MC-Sym/.
An MC-Sym script is composed of 6 parts. The first one pro-
vides a description of the sequence that is going to be modeled,
and the second indicates the set of fragments to be used. The third
part is the order in which the fragments will be merged. The fourth
and fifth parts describe geometrical constraints to satisfy during the
construction of the three-dimensional model. The last segment
defines a set of rules for the space exploration. Every line starting
by ‘//’ is a comment.
In this example, we show how to model the tRNA(Cys) mole-
cule of Archaeoglobus fulgidus, a 71-nucleotides long RNA into
which four motifs are inserted by RNAMoIP (See previous section).
Each part of the script is described and explained separately.
Part 1: Sequence Definition

The first step defines the sequence we are modeling. The syntax is:
// ==================== Sequence ====================
sequence(r A1 GCCAGGGUGGCAGAGGGGCUUUGCGGCGGACUGCAGA
UCCGCUUUACCCCGGUUCGAAUCCGGGCCCUGGC)
// ((((((…((((((…))))))..((((.......))))......(((((((…))))))).))))))
// 1234567890123456789012345678901234567890123456789
0123456789012345678901
// 1 2 3 4 5 6 7
The keyword “sequence” states that we are defining a sequence.

The program requires three arguments that are indicated between
the parentheses. First, “r” specifies that we are working with an
RNA sequence. Next, “A1” sets as identifier for the sequence “A”
and defines the first position as “1.” Finally, we end by the sequence
and close the parenthesis. The two lines of comment that follows
are just here to improve readability.
Part 2: Fragment Definition

This is probably the most important and delicate step of the script.
In MC-Sym, the basic units are called cycles. The second step aims
to define which cycles and motifs will be used to build the struc-
ture. Cycles, motifs, and other structural elementary blocks used
by MC-Sym are called fragments. The order in which they entered
is not important, but we must ensure that every position is included
in a fragment and that all fragments overlap.
In the following example, we illustrate how to insert a basic
cycle (i.e., not a motif). Here, the cycle is a stack between two
consecutive base pairs. This is the most common usage of cycles as
most Watson–Crick and Wobble base pairs are predicted by sec-
ondary structure predictors and thus not covered by RNA motifs.
The code below indicates how and where to map this cycle. In this
case, a stacked base pairs at positions 1, 2, 70, and 71.
ncm_01 = library(
pdb("MCSYM-DB/2_2/GCGC/*R20*.pdb.gz")
#1:#2, #3:#4<- A1:A2, A70:A71
rmsd(0.1 sidechain && !(pse || lp || hydrogen)))
We start with a unique identifier (i.e., “ncm_01”), followed by
the keyword “library.” It indicates to MC-Sym where to retrieve
the basic fragments.
On the second line, we provide specifications of the files
within that library. Here, our library is composed of PDB files, fol-
lowed by the path “MCSYM-DB/2_2/GCGC/*_1.pdb.gz”.
MCSYM-DB is the location of all standard library fragments.
MC-Sym fragments are sorted according to the number of con-
secutive nucleotides in each component. Here, we have two times
two consecutive nucleotides (1–2 and 70–71). Therefore, we must

look into the subfolder “2_2”. Next, we must explicitly tell which
sequence of nucleotides are involved, in the order 5′ to 3′. Since
the nucleotides at positions 1, 2, 70, and 71 are respectively “G,”
“C,” “G,” “C,” the next path is “GCGC.” We end by “*.pdb.gz”
to consider all possible structures for this specific fragment. Specific
subsets of those fragments can be chosen as specified in the MC-
Sym documentation.
On the third line, we provide detailed information on the
positions onto which the fragment will be mapped. The syntax is
composed of two parts, separated by the symbol “<-”. The left-
hand side indicates the segments associated with each consecutive
component. Segments are represented with an interval, and nucle-
otides are numbered sequentially according to their position in the
motif. The right-hand side defines the specific positions (still as
intervals or segments) onto which these components/segments
must match. In our sample code, we have two stretches of two
nucleotides. We write it “#1:#2, #3:#4”, which also implies that
the first part is composed of the two first nucleotides and the sec-
ond one of the next two. On the right hand side, “A1:A2,
A70:A71” describes which sequence must be mapped. The nucle-
otides at from position 1 to 2 for the first component, and the
nucleotides at from position 70 to 71 for the second component.
“A” is a unique identifier of the sequence to be used, as defined in
the first part. The last line is a set of constraints for this specific
fragment. Since this information is nonessential here, we kindly
redirect the reader to the MC-Sym documentation for more
details. It is also important to not forget to close the parenthesis
opened after the keyword “library”
Now, we must specify how to declare the fragments inserted by
RNAMoIP. The simplest one is a hairpin. Lets start with the hair-
pin “1DU6.D.2”, between positions “12” and “22”. The code is:
pdbs_hairpin_1 = library(
pdb ("/u/reinharz/RNAMOIP-DB-VIEW3D/2DU6.D.2/*.pdb")
#1:#11<- A12:A22
)
The syntax is very similar to those of the previous example.
The first line is the same as for the stacked base pair: a unique iden-
tifier followed by the keyword “library.” However, the path to the
library is different. All RNA-MoIP motifs are contained in a repos-
itory that is accessible with the path “/u/reinharz/RNAMOIP-
DB-VIEW3D/<Motif Identifier>”. The path must always end by
“*.pdb” to indicate that we want to consider all possible structures
in the folder. On the third line, we define the positions involved.
In this case, the motif contains 11 consecutive nucleotides, thus
the left hand side is “#1:#11”. Those nucleotides are at positions
“12” to “22” in our sequence, defining the right hand side.

We end by closing the parenthesis of “library.”
Finally, motifs with multiple stretches of consecutive nucleo-
tides are a slightly more complicated to specify. We illustrate that
case with the 4-way junction “2DU5.D.1”. Using the same syntax
as above, we write the code as follows.
pdbs_4_way = library(
pdb ("/u/reinharz/RNAMOIP-DB-VIEW3D/2DU5.D.1/*.
pdb")
#1:#5, #6:#9, #10:#17, #18:#20 <- A6:A10, A24:A27,
A41:A48, A64:A66
)
In this case, our motif is composed of 4 components, between
positions “6” and “10,” “24” and “27,” “41” and “48,” and “64”
and “66” in the input sequence. Each component of consecutive
nucleotides needs to be described individually.
Part 3: Merging Fragments

Once all fragments are specified, we must define in which order
they are going to be assembled together. One fragment must be
selected to start the construction. All others will be sequentially
added; with the constraint that each new fragment must overlaps
with any of the previous one. The syntax to assemble the first two
stacked base pairs, uses the unique names used in the fragments
definitions (i.e., ncm_01 and ncm_02), and is written as follows.
structure = backtrack
(
ncm_01
merge(ncm_02 1.5)
)
It is important to always start with the keywords “struc-
ture = backtrack”. The assembly is indicated between two paren-
theses. The starting point of our model is “ncm_01” and must be
written on the first line. On the second line we use the keyword
“merge” followed, between parentheses, by an argument that indi-
cates the next fragment, and another argument that specifies an
“error” threshold that can be tolerated by the program to extend
the structure. By default we use a value of “1.5”. More details on
this parameter can be found in the MC-Sym documentation.
Part 4 and 5: Constraints

MC-Sym requires several general parameters and constraints to
perform the three-dimensional reconstruction. A standard set of
values is provided in the following code and can be used as it is.
// ================ Backtrack Restraints ================

clash
(
structure
1.5 !(pse || lp || hydrogen)
)
backtrack_rst
(
structure
width_limit = 25 %,
height_limit = 33 %,
method = probabilistic
)
//===================RiboseRestraints===================
ribose_rst
(
structure
method = ccm,
pucker = C3p_endo,
threshold = 2.0
)
Part 6: Space Exploration Parameters

Finally, the last step is to determine the boundaries of the space to
explore. We also specify the time limit and the maximal number of
structures to output. MC-Sym will stop as soon as one of those
conditions is fulfilled. The other parameter options are standard
and can be found in the documentation of MC-Sym.
// ================ Exploration Initialization ==============

explore
(
structure
option(
model_limit = 1000,
time_limit = 30 m,
seed = 3210)
rmsd(3.0 sidechain && !(pse || lp || hydrogen))
pdb("structure" zipped)
)
Where, the key word “model_limit” specifies the maximal

number of structure to generate, and “time_limit” the maximal
amount of time allowed to run MC-Sym. Here, we use an upper
bound of 30 min with a maximum of 1,000 structures. The time
needed to generate structures can vary. In our benchmark [21], we
have seen that about half an hour of computation was often enough
to produce solutions with molecules with sizes up to one hundred
nucleotides. Within this time frame, the maximal number of struc-
tures created will rarely reach one thousand in that time.
Importantly, motifs with large number of components will highly
restrict the exploration of the space, thus the time needed to
explore the conformational space. In absence of any motif with
three or more components, increasing the time limit to 2 days is a
reasonable step. In that case, the number of structures could
increase to five or even ten thousands, and we can expect much
more diversity in the sample set.
It is difficult to determine in advance how many structure sam-
ples are necessary to provide a good representation of the confor-
mational landscape. Numbers between 100 and 1,000 are good
starts, but more might be occasionally needed. We acknowledge
that large numbers of structures are difficult to analyze. Fortunately,
as we will see in the next section, the MC-Sym server offers several
tools to facilitate these tasks.
The key word “seed” is a random number used to initialize
MC-Sym. Its value has little importance in the context of this dis-
cussion, and users can modify it if they wish.
The complete source code of this MC-Sym script is available at
http://csb.cs.mcgill.ca/RNAMoIP/.
3.8 Retrieving, Once submitted, the script will run on MC-Sym servers and the
Optimizing, Analyzing URL of a result page will be returned to the user. This page offers
and Visualizing multiple options to optimize, analyze, and retrieve the results. To
Structures access these options, the user must access the web page “com-
mands.html” located in the results directory.
The energy minimization of the MC-Sym structure is probably
one of the most important options. We recommend any user to
run it before analyzing or visualizing the results. The “steepest
descent” option returns satisfactory results in a short time, but
more sophisticated and slower techniques (e.g., simulated anneal-
ing) are also available.
Clustering of structures using the k-means algorithm is another
useful option that enables the user to automatically identify the most
significant structures in the set or structures returned by MC-Sym.
A PDB model of all predicted structures is available at the root
of the directory. Each model can be visualized with molecular
viewer such as PyMOL or Jmol. Figure 5 show an example of a
structure predicted with our RNA-MoIP and MC-Sym pipeline,
aligned with the experimental structure [24].
Fig. 5 3D structure of the tRNA(Cys) from Archaeoglobus fulgidus predicted with

the RNA-MoIP + MC-Sym pipeline (in blue), aligned with its crystallographic
model (Fukunaga and Yokoyama [24])
4 Notes
1. The motifs database used by RNA-MoIP includes motifs from

the PDB model 2DU3 (tRNA(Cys) from Archaeoglobus fulgi-
dus) database. Thus, the quality of the prediction achieved in
our example illustrates the performance of RNA-MoIP in the
best-case scenario. In applications on other (new) molecules, it
will be very unlikely to have motifs from the same molecule in
the motif repository. Fortunately, even in the absence of motifs,
the benchmark performed in ref. [21] shows that the quality
of the structure prediction remains high.
References
1. Bekaert M et al (2003) Towards a computa- potentials for RNA structure evaluation. RNA
tional model for -1 eukaryotic frameshifting 17(6):1066–1075
sites. Bioinformatics 19(3):327–335 6. Ding F et al (2008) Ab initio RNA folding by
2. Vitreschak AG et al (2004) Riboswitches: the discrete molecular dynamics: from structure
oldest mechanism for the regulation of gene prediction to folding mechanisms. RNA 14(6):
expression? Trends Genet 20(1):44–50 1164–1173
3. Szewczak AA et al (1993) The conformation 7. Jonikas MA et al (2009) Coarse-grained mod-
of the sarcin/ricin loop from 28S ribosomal eling of large RNA molecules with knowledge-
RNA. Proc Natl Acad Sci U S A 90(20): based potentials and structural filters. RNA
9581–9585 15(2):189–199
4. Sponer J et al (2012) Chapter 6 molecular 8. Poursina M et al (2011) Strategies for articulated
dynamics simulations of RNA molecules, in multibody-based adaptive coarse grain simula-
innovations in biomolecular modeling and tion of RNA. Methods Enzymol 487:73–98
simulations. R Soc Chem 2:129–155 9. Parisien M, Major F (2008) The MC-fold and
5. Bernauer J et al (2011) Fully differentiable MC-Sym pipeline infers RNA structure from
coarse-grained and all-atom knowledge-based sequence data. Nature 452(7183):51–55
10. Martinez HM, Maizel JV Jr, Shapiro BA cysteine biosynthesis in archaea. Nat Struct
(2008) RNA2D3D: a program for generating, Mol Biol 14(4):272–279
viewing, and comparing 3-dimensional models 25. Lorenz R et al (2011) ViennaRNA Package
of RNA. J Biomol Struct Dyn 25(6):669–683 2.0. Algorithms Mol Biol 6:26
11. Zhao Y et al (2012) Automated and fast build- 26. Waugh A et al (2002) RNAML: a standard
ing of three-dimensional RNA structures. Sci syntax for exchanging RNA information. RNA
Rep 2:734 8(6):707–717
12. Das R, Baker D (2007) Automated de novo 27. Lemieux S, Major F (2002) RNA canonical
prediction of native-like RNA tertiary struc- and non-canonical base pairing types: a recog-
tures. Proc Natl Acad Sci U S A 104(37): nition method and complete repertoire.
14664–14669 Nucleic Acids Res 30(19):4250–4263
13. Das R, Karanicolas J, Baker D (2010) Atomic 28. Chojnowski G, Walen T, Bujnicki JM (2014)
accuracy in predicting and designing noncanoni- RNA Bricks–a database of RNA 3D motifs and
cal RNA structure. Nat Methods 7(4):291–294 their interactions. Nucleic Acids Res
14. Wang Z, Xu J (2011) A conditional random 42(1):D123–D131
fields method for RNA sequence-structure 29. Ding Y, Chan CY, Lawrence CE (2005) RNA
relationship modeling and conformation sam- secondary structure prediction by centroids in
pling. Bioinformatics 27(13):i102–i110 a Boltzmann weighted ensemble. RNA 11(8):
15. Rother M et al (2011) ModeRNA: a tool for 1157–1166
comparative modeling of RNA 3D structure. 30. Zuker M (1989) On finding all suboptimal
Nucleic Acids Res 39(10):4007–4022 foldings of an RNA molecule. Science
16. Leontis NB, Westhof E (2001) Geometric 244(4900):48–52
nomenclature and classification of RNA base 31. Zuker M, Mathews DH, Turner DH (1999)
pairs. RNA 7(4):499–512 Algorithms and Thermodynamics for RNA
17. Yang H et al (2003) Tools for the automatic Secondary Structure Prediction: A Practical
identification and classification of RNA base Guide. In: Barciszewski J, Clark BFC (eds)
pairs. Nucleic Acids Res 31(13):3450–3460 RNA Biochemistry and Biotechnology.
18. Djelloul M, Denise A (2008) Automated motif Springer, Netherlands, pp 11–43
extraction and classification in RNA tertiary 32. Ding Y, Lawrence CE (2003) A statistical
structures. RNA 14(12):2489–2497 sampling algorithm for RNA secondary struc-
19. Leontis N, Zirbel CL (2012) Nonredundant ture prediction. Nucleic Acids Res 31(24):
3D structure datasets for RNA knowledge 7280–7301
extraction and benchmarking. In: Leontis N, 33. Reuter JS, Mathews DH (2010) RNAstructure:
Westhof E (eds) RNA 3D structure analysis software for RNA secondary structure prediction
and prediction. Springer, Berlin, pp 281–298 and analysis. BMC Bioinformatics 11:129
20. Hofacker IL et al (1994) Fast folding and 34. Bellaousov S et al (2013) RNAstructure: Web
comparison of RNA secondary structures. servers for RNA secondary structure predic-
Monatsh Chem 125(2):167–188 tion and analysis. Nucleic Acids Res 41(Web
21. Reinharz V, Major F, Waldispuhl J (2012) Server issue):W471–W474
Towards 3D structure prediction of large RNA 35. Zuker M (2003) Mfold web server for nucleic
molecules: an integer programming framework acid folding and hybridization prediction.
to insert local 3D motifs in RNA secondary Nucleic Acids Res 31(13):3406–3415
structure. Bioinformatics 28(12):i207–i214 36. Ding Y, Chan CY, Lawrence CE (2004) Sfold
22. Berman HM et al (1992) The nucleic acid web server for statistical folding and rational
database. A comprehensive relational database design of nucleic acids. Nucleic Acids Res
of three-dimensional structures of nucleic 32(Web Server issue):W135–W141
acids. Biophys J 63(3):751–759 37. Honer zu Siederdissen C et al (2011) A fold-
23. Bernstein FC et al (1977) The Protein Data ing algorithm for extended RNA secondary
Bank: a computer-based archival file for mac- structures. Bioinformatics 27(13):i129–i136
romolecular structures. J Mol Biol 112(3): 38. Do CB, Woods DA, Batzoglou S (2006)
535–542 CONTRAfold: RNA secondary structure
24. Fukunaga R, Yokoyama S (2007) Structural prediction without physics-based models.
insights into the first step of RNA-dependent Bioinformatics 22(14):e90–e98
Chapter 7
Fast Prediction of RNA–RNA Interaction

Using Heuristic Algorithm
Soheila Montaseri
Abstract
Interaction between two RNA molecules plays a crucial role in many medical and biological processes such
as gene expression regulation. In this process, an RNA molecule prohibits the translation of another RNA
molecule by establishing stable interactions with it. Some algorithms have been formed to predict the
structure of the RNA–RNA interaction. High computational time is a common challenge in most of the
presented algorithms. In this context, a heuristic method is introduced to accurately predict the interaction
between two RNAs based on minimum free energy (MFE). This algorithm uses a few dot matrices for
finding the secondary structure of each RNA and binding sites between two RNAs. Furthermore, a parallel
version of this method is presented. We describe the algorithm’s concurrency and parallelism for a multi-
core chip. The proposed algorithm has been performed on some datasets including CopA-CopT, R1inv-
R2inv, Tar-Tar*, DIS-DIS, and IncRNA54-RepZ in Escherichia coli bacteria. The method has high validity
and efficiency, and it is run in low computational time in comparison to other approaches.
Key words RNA secondary structure, RNA–RNA interaction, Heuristic, Parallel, Minimum free energy
1 Introduction
RNA–RNA interaction is applied in the gene regulation [1]. The

interaction between two RNAs is used in the treatment of many
cancers, including colon, breast, and pancreatic by silencing the gene
expression or genes involved in the development of malignancy.
In most living organisms, noncoding RNAs (ncRNAs) bind to their
target mRNAs to posttranscriptional regulation [2]. Furthermore,
small interfering RNAs (siRNAs) and microRNAs (miRNAs) play an
important role in the gene suppressing process [3].
RNA–RNA interaction prediction is one of the most challeng-
ing problems in bioinformatics that involves an interplay between
the secondary structure of each molecule on the one hand, and the
binding of the two molecules on the other hand [4].
The first method to predict RNA–RNA interaction is carried
out by concatenating two RNA sequences as a single sequence and
123
124 Soheila Montaseri
using dynamic programming techniques [5, 6]. A similar approach

has been presented based on a rigorous extension of secondary
structure models to multistranded cases. It calculates the partition
function of a secondary structure complex of multiple interacting
RNAs [7].
Reduction of computational time might be achieved by ignor-
ing all internal base pairings in both RNAs. This idea was first
introduced by RNAhybrid [8] and UNAFold [9]. It was later
implemented by RNAduplex from Vienna package [6].
Minimizing the free energy between two RNA molecules in a
number of energy models with growing complexity was introduced
in [10]. In both RNAup [11, 12] and intaRNA [13] the single
regions on each RNA are used to find the interaction between two
RNAs. In RNAplex [3], the possible hybridization sites for an
RNA in large RNA databases are found based on a slight simplifica-
tion of the energy model. In this model, the loop energy is a func-
tion of the loop size. An approach based on the stochastic context
free grammars was introduced [14]. In this algorithm, two real
values called transition probability and emission probability are
specified for each rule of the grammar.
InteRNA [15] computes the optimal set of interactions among
all loops as well as single-stranded sequences in the original sec-
ondary structures of interacting RNA molecules. This method uses
a heuristic approach for solving this problem. In another heuristic
algorithm, highly accessible regions are predicted in each sequence.
These regions include the loop regions in the native structure of
RNA strand. Then the optimal non-conflicting interaction regions
are found. In this algorithm, each accessible region can interact
with at most two accessible regions from the other sequence [2].
Approximation algorithms were introduced in ref. [4]. An RNA–
RNA interaction graph is created in which every edge represents a
possible bond in or between two RNAs. A set of node disjoint
edges is found to maximize the number of bonds. A statistical sam-
pling algorithm [1] calculates the interaction probabilities for any
given interval on the target RNA by grammars.
In this context, a heuristic method is presented for RNA–RNA
interaction prediction problem. This algorithm has two main steps.
Firstly, an RNA secondary structure (RSS) is constructed for each
RNA sequence. Secondly, the interaction between two RNA sec-
ondary structures is computed. In order to make an RSS, a dot
matrix is constructed for each RNA sequence. Each sub-diagonal
in the dot matrix can be considered as a stem. Using some of these
sub-diagonals, an RSS is generated. Then another dot matrix is
created for finding the interaction. Each sub-diagonal can be seen
as a part of a binding site between two RNAs. Finally, some of
these sub-diagonals are chosen as the interaction regions between
two RNAs. In these dot matrices, all possible base pairs in RNA
secondary structure and RNA–RNA interaction are considered,
RNA-RNA Interaction Prediction 125
and thus each set of consecutive unpaired bases can interact with
other complementary unpaired sets. Here, we also have imple-
mented a modeling for RNA–RNA interaction prediction based on
parallel heuristic approach. All diagonals in each dot matrix can be
considered independently of each other. Therefore, the sub-
diagonals on different diagonals and their minimum free energy
(MFE) are calculated concurrently. The sub-diagonals are sorted in
parallel based on their length and MFE.
2 Materials
2.1 Dataset The proposed algorithm has been performed on some standard
datasets such as CopA-CopT, R1inv-R2inv, Tar-Tar*, DIS-DIS,
and IncRNA54-RepZ in Escherichia coli bacteria [14]. These pair
RNAs have kissing hairpin structures.
Before presenting our algorithm, some definitions and nota-
tions about RNA are presented. An RNA molecule consists of a
sequence of four nucleotides: adenine (A), cytosine (C), guanine
(G), and uracil (U). Thus, an RNA sequence is defined by
R = r1r2 … rn in 5′ − 3′ direction, where |R| = n and
ri ∈ {A, C, G, U} (1 ≤ i ≤ n). On the other hand, the reverse R is
indicated as rnrn − 1 … r1 in 3′ − 5′ direction. A subsequence
rij = riri + 1 … rj from the sequence R is started from the position i
and ended at the position j in 5′ − 3′ direction. The reverse rij is
defined as rjrj − 1 … ri in 3′ − 5′ direction. The RNA structure is
formed by the creation of hydrogen bonds between complemen-
tary bases (A − U, C − G).
An RNA secondary structure is composed of stems and single
regions. Each stem contains some consecutive base pairs such as
ri. rj and ri + 1. rj − 1. Let rij and rkl be two subsequences from the
sequence R that form a stem. So the subsequence rij is bonded to
the reverse rkl base to base. For the stem, each base in rij is repre-
sented by ' (' and each base in rkl is displayed with ') '. A single
region is a loop or single strand. Each end of a loop is linked to a
stem, while one end of a single strand is connected to a stem. For
the single region, each base is represented by '. '. Thus the second-
ary structure of the RNA R is indicated with a sequence of charac-
ters ′(', ′) ' and ′. ′.
An RNA–RNA interaction structure is composed of stems in
each RNA and hybrids between the two RNAs. Each hybrid con-
tains some consecutive base pairs between the two RNAs without
any base pairs in each RNA. Let rij′ and rkl″ be two single regions of
RNA secondary structures R′ and reverse R″, respectively, which
are involved in a hybrid. Thus the subsequence rij′ is bonded to
the reverse rkl″ base to base. For the hybrid, each base in rij′ is
shown by ' [' and each base in the rkl″ is declared by '] '.
In RNA–RNA interaction prediction problem, two RNA

sequences R′ = r1′r2′ … rn′(|R ′ | = n) and R″ = r1″r2″ … rm″(|R″| = m)
are given as inputs. The goal is to predict the interaction structure
between R′ and R″ that maximizes the sum of the length (number
of 1’s) of stems and hybrids. Moreover, the free energy of the
interaction structure should be minimized.
2.2 Computational The proposed method is run on Intel(R) Core(TM)2 Duo proces-
Hardware sor T6670 2.20 GHz with 4 GB RAM to predict the interaction
structures. Parallel version of the algorithm implements parallelism
on a multicore system. Experiments are performed on an Intel(R)
core(TM) i7 2600k CPU 3.40 GHz with 8G RAM.
2.3 Computational The heuristic method for RNA–RNA interaction prediction is per-
Software formed in language C++. The parallel heuristic approach is imple-
mented in language Matlab.
3 Methods
3.1 RNA Secondary In this section, a heuristic algorithm is presented for RNA–RNA
Structure Prediction interaction prediction. This method has two main steps. In the first
step, an RNA secondary structure is constructed for each RNA
sequence as follows:
1. A stem dot matrix M Rn × n is constructed for the RNA sequence
R = r1r2 … rn (|R| = n) as follows:
ìï1 if ( ri , rj ) Î {( A, U ) , ( U, A ) , ( C, G ) , ( G, C )} ,
M R [i, j ] = í
îï0 else
where ri and rj (1 ≤ i, j ≤ n) are the i-th nucleotide in the
sequence R and the j-th nucleotide in the reverse R,
respectively.
2. In the stem dot matrix Mn × nR, each right-skewed consecutive
value of 1’s which is parallel to the main diagonal shows a stem
sub-diagonal. A set of stem sub-diagonals for the RNA R is
formed as follows:
D R = {< i, j , k , l >| 1 £ i £ k £ n &1 £ j £ l £ n} ,

where (i, j) and (k, l) are the start and end positions of a stem
sub-diagonal, respectively. Let dR ∈ DR and dR = 〈i, j, k, l〉. Then
dR shows how the subsequence rik is bonded to the reverse rjl in
R. If there are i′ and j′ where i < i′ < k, j < j′ < l and i′ + j′ = n + 1,
then dR has to be removed and two stem sub-diagonals
〈i, j, i′, j′〉 and 〈i′ + 1, j′ + 1, k, l〉 have to be inserted to the set DR.
3. Length and MFE are computed for each stem sub-diagonal.

The length of stem sub-diagonal is the number of 1’s in it.
MFE of each sub-diagonal is determined as the sum of the free
energies of two adjacent base pairs in it, as follows:
MFE ( s ) = å e ( ri ,i +1 , rj -1, j ) , (1)

( ri , rj ), ( ri+1 , rj -1 )Îs
where s represents a stem or hybrid, and e(ri,i + 1, rj − 1,j) shows
the free energy of two adjacent base pairs (ri, rj) and (ri + 1, rj − 1).
The MFEs of all two adjacent base pairs are depicted in
Table 1.
4. Stem sub-diagonals in the array DR are decreasingly sorted
based on the length, then each set of the sub-diagonals with
equal length is increasingly sorted based on MFE.
5. The sub-diagonals are selected to form RNA secondary struc-
ture based on order of setting in the set DR. Notice that the
selected sub-diagonals should not have overlapping. Let D be
a type of sub-diagonal sets and d1, d2 ∈ D, where d1 = 〈i1, j1, k1, l1〉
and d2 = 〈i2, j2, k2, l2〉. Overlapping between two sub-diagonals
d1 and d2 is defined as follows:
ì1 if$p : i1 £ p £ k1 & i2 £ p £ k2 ,
ï
Overlap ( d1 , d 2 ) = í1 if$p : j1 £ p £ l1 & j2 £ p £ l2 ,
ï0 else.
î
After performing the algorithm, S represents the secondary
structure of the RNA R. Accordingly, S′ and S″ are the structures
of RNAs R′ and R″, respectively. Each base pair in the stem regions
is represented by ' (' and ') '.
Table 1
MFE of all two adjacent base pairs
3.2 RNA–RNA In the second step, interaction structure between the two struc-
Interaction Prediction tures S′ and S″ is calculated as follows:
¢ ²
1. A hybrid dot matrix M nR´m, R is made up of two secondary struc-
tures of RNA sequences R′ = r1′, …, rn′ and R″ = r1″, …, rm″(|R′
| = n, |R″| = m), as follows:
ìï1 if ( ri¢ ,rj² ) Î {( A, U ) , ( U, A ) , ( C, G ) , ( G, C )} & ( si¢ , s ²j ) = ( ¢ .¢ ,¢ .¢ ) ,
M R , R [i, j ] = í
¢ ²
îï0 else,
where (ri′, rj″) are the i-th and j-th nucleotides and (si′, sj″) are
the i-th and j-th structures in the sequence R′ and reverse R″,
respectively (1 ≤ i ≤ n, 1 ≤ j ≤ m).
¢ ²
2. In the hybrid dot matrix M nR´m, R , each right-skewed consecu-
tive value of 1’s which is parallel to the main diagonal shows a
hybrid sub-diagonal. A set of hybrid sub-diagonals for the two
RNAs R′ and R″ is formed as follows:
D R , R = {< i, j , k , l >| 1 £ i £ k £ n 1 £ j £ l £ m} ,
¢ ²

where (i, j) and (k, l) are the start and the end positions of the
R¢ , R ² ¢ ²
hybrid sub-diagonal, respectively. Each d =< i, j , k , l >Î D R , R
indicates that the subsequence rik′ in R′ is bonded to the rjl″ in
reverse R″.
The remaining steps are similar to steps 3, 4, and 5 in
Subheading 3.1.
The time and space complexity of the algorithm are 0(k2 log k2)
and 0(k2), respectively, where k indicates the sum of the length of
two RNAs.
Example 1
Let R′ = AGUACCGAAAACU and R″ = CCGUUUGAGGUCGG
be two RNA sequences. Interaction between the two RNAs can be
shown as follows:
5¢ - AGUACCGAAAACU - 3¢ 5¢ - CCGUUUGAGGUCGG - 3¢

( ( ( [ [ [ .[ [ [ ) ) ) ( ( ( ] ] ] . . ] ] ] ) ) ),

Here, each RNA has one stem. In the left hand RNA, one stem is
found by the formation of bonding between AGU and the reverse
ACU. Consecutive open brackets and their corresponding closing
brackets are denoted as an interaction region between two RNAs.
Here, there are two interaction regions between the sequences R′
and R″. The first one is formed by binding between ACC and the
reverse GGU. The second one is generated by binding between
AAA and the reverse UUU.
3.3 Parallel Here, parallel version of the heuristic algorithm is described. Let N
Algorithm be the number of created threads. Each dot matrix is constructed
in parallel. Notice that an element (x, y) in the dot matrix is depen-
dent on a set of elements (x, j) and (i, y) for each 1 ≤ i, j ≤ n as shown
in Fig. 1. This dependency pattern allows that the main diagonal
and other diagonals parallel to it can be considered independent of
each other. So divide the diagonals between threads. Figure 2
Fig. 1 Dependency of element (x, y) to a set of elements (x, j) and (i, y) for each
1 ≤ i, j ≤ n
Fig. 2 Pattern of implemented computation

shows the sequence of computations of sub-diagonals. In this figure,

some samples of sub-diagonals are shown in different colors on
distinct diagonals.
Procedure Merge-Split in ref. [16] is called to sort the sub-
diagonals decreasingly based on the length. The Merge-Split is
recalled to sort the sub-diagonals with equal length increasingly
based on the MFE.
Creating the dot matrices and extracting the sub-diagonals
require time complexity 0(k2/N), where k and N indicate the sum
of the length of two RNAs and the number of threads, respectively.
Sorting the arrays of sub-diagonals is run in order of 0(k2 log k2/N).
Therefore, the time complexity of the algorithm is 0(k2 log k2/N).
Because of using dot matrices, computational space is 0(k2).
3.4 Algorithm One of the limitations of the proposed algorithm is that we have
Limitations first obtained the length of sub-diagonals and then computed their
MFEs for RNA–RNA interaction prediction, while the interaction
between two RNAs is formed based on MFE.
To evaluate the prediction accuracy of the method, sensitivity,
specificity, and F-measure have been employed. Let TP, FP, and
FN be the number of correctly predicted base pairs, the number of
false predicted base pairs, and the number of unpredicted base
pairs, respectively. So the sensitivity (Sn), specificity (Sp), and
F-measure (F) are defined as follows:
Sn = TP / (TP + FN ) , Sp = TP / (TP + FP ) , F = ( 2 * Sn * Sp ) / ( Sn + Sp ) .
We have compared the prediction accuracy of our parallel method,
TPIRNA, in these criteria with some similar approaches such as
InRNAs, RNAup [12] and EBM in Table 2. As it is shown, the
Table 2
Comparison of sensitivity, specificity, and F-measure
Sensitivity (%) Specificity (%) F-measure (%)

RNA–RNA
pairs TPIRNA inRNAs RNAup EBM TPIRNA inRNAs RNAup EBM TPIRNA inRNAs RNAup EBM
Tar-Tar* 100 100 100 100 100 87.5 83.3 93.3 100 93.3 90.9 96.5
R1inv- 100 90 100 90 100 90 77.8 94 100 90 87.5 92
R2inv
DIS-DIS 100 100 100 78.6 100 100 100 78.6 100 100 100 78.6
CopA- 96.74 100 55.6 90.09 78.18 84.6 65.2 80 86.86 91.7 60 85.1
CopT
IncRNA54- 86 87.5 75 91.7 75.51 79.2 85.7 83 80.41 83.1 80 87.1
RepZ
Average 96.55 95.5 86.12 90.23 90.73 95.5 82.4 85.9 93.45 91.6 83.69 87.9
Table 3
Comparison of running times of TPIRNA with 1, 2, 4, and 8 threads and also EBM
RNA–RNA pairs 1-TPIRNA 2-TPIRNA 4-TPIRNA 8-TPIRNA EBM

CopA-CopT 0.176 0.089 0.033 0.021 1206.47
DIS-DIS 0.162 0.075 0.016 0.011 382.13
Tar-Tar* 0.152 0.064 0.008 0.006 21.25
R1inv-R2inv 0.161 0.060 0.009 0.007 49.21
IncRNA54-RepZ 0.202 0.089 0.030 0.023 1328.62
Average 0.17 0.037 0.019 0.013 597.39
Fig. 3 Comparison of the average running times of TPIRNA using 1, 2, 4, and 8

threads and sequential EBM on our datasets
average accuracy of TPIRNA in the sensitivity, specificity, and

F-measure is 96.55 %, 90.73 %, and 93.45 %, respectively. According
to the table, the average accuracy of the proposed approach is
higher than these algorithms.
Table 3 represents the running times achieved using 1, 2, 4,
and 8 threads for TPIRNA in comparison to the sequential EBM
on our datasets. TPIRNA computes interaction of CopA-CopT in
0.021 s with 8 threads, while EBM takes approximately 20 min. In
the table, i -TPIRNA represents that TPIRNA uses i thread(s).
Comparison of the average running times of TPIRNA using 1, 2,
4, and 8 threads and also EBM on our datasets is shown in Fig. 3.
Table 4
Comparison of time and space complexity of some methods
Algorithm Time complexity Space complexity

TPIRNA O(k2 log k2/N) O(k2)
SPM O(n3m3) O(n2m2)
LM O(n3m3) O(n2m2)
inRNAs O(k4w) O(k2)
RNAup O(n3m) O(n2)
EBM O(n3m3) O(n2m2)
App O(n3m3) O(n2m2)
Pairfold O(k3) O(k2)
IntaRNA O(nm + nl3) O(nm)
Ripalign O(N )
6
O(N4)
PETcofold O(MIl3)
The dashed horizontal line represents the sequential running time

of EBM. It is clear from the figure that TPIRNA with one thread
performs much faster than EMB. The running time of TPIRNA is
decreased when the number of threads is increased.
The time and space complexity of some algorithms such as
Petcofold [17], IntaRNA [18], ripalign [19], Pairfold [5], App
[4], EBM, RNAup, LM [10], and SPM [10] is represented in
Table 4.
According to the table, minimum time complexity of these
algorithms except TPIRNA is O(k3), where k indicates the sum of
the lengths of two RNAs. As it is shown, TPIRNA is run in a time
complexity of O(k2 log k2/N). Thus our algorithm is performed in
lower computational time in comparison to other methods.
Acknowledgements
Special thanks to Professor Fatemeh Zare-Mirakabad and Nasrollah

Moghadam-charkari for their valuable consultancy in improving
this work.
References
1. Huang FWD, Qin J, Reidys CM et al (2010) 11. Mückstein U, Tafer H, Bernhart S et al (2009)
Target prediction and a statistical sampling Translational control by RNA-RNA interac-
algorithm for RNA-RNA interaction. tion: Improved computation of RNA-RNA
Bioinformatics 26:175–181 binding thermodynamics. Bioinform Res Dev
2. Salari R, Backofen R, Sahinalp SC (2010) Fast 13:114–127
prediction of RNA-RNA interaction. 12. Mückstein U, Tafer H, Hackermuller J et al
Algorithms Mol Biol 5:5–15 (2006) Thermodynamics of RNA-RNA bind-
3. Tafer H, Hofacker IL (2008) RNAplex: a fast ing. Bioinformatics 22:1177–1182
tool for RNA-RNA interaction search. 13. Busch A, Richter AS, Backofen R (2008)
Bioinformatics 24:2657–2663 IntaRNA: efficient prediction of bacterial
4. Mneimneh S (2009) On the approximation of sRNA targets incorporating target site accessi-
optimal structures for RNA-RNA interaction. bility and seed regions. Bioinformatics 24:
Trans Comput Biol Bioinform 6:682–688 2849–2856
5. Andronescu M, Zhang ZC, Condon A (2005) 14. Kato Y, Akutsu T, Seki H (2009) A grammati-
Secondary structure prediction of interacting cal approach to RNA-RNA interaction predic-
RNA molecules. J Mol Biol 345:987–1001 tion. Pattern Recogn 42:531–538
6. Bernhart S, Tafer H, Mückstein U et al (2006) 15. Aksay C, Karakoc E, Kin Ho C et al. (2007)
Partition function and base pairing probabilities ncRNA discovery and functional identification
of RNA heterodimers. Algorithms Mol Biol 1–3 via sequence motifs. Technical Report TR 1–9
7. Dirks R, Bios J, Schaeffer JM et al (2007) 16. Selim GA (1989) The design and analysis of
Thermodynamic analysis of interacting nucleic parallel algorithms. ISBN 0-33-23005b-3
acid strands. Soc Ind Appl Math 49:65–88 17. Seemann SE, Richter AS, Gesell T et al (2011)
8. Rehmsmeier M, Steffen P, Hochsmann M et al PETcofold: predicting conserved interactions
(2004) Fast and effective prediction of and structures of two multiple alignments of
microRNA/target duplexes. RNA 10: RNA sequences. Bioinformatics 2:211–219
1507–1517 18. Chitsaz H, Salari R, Sahinalp SC et al (2009) A
9. Markham NR, Zuker M (2008) UNAFold: partition function algorithm for interacting
software for nucleic acid folding and hybridiza- nucleic acid strands. Bioinformatics 25:i365–i373
tion. Methods Mol Biol 453:3–31 19. Li AX, Marz M, Qin J et al (2011) RNA–
10. Alkan C, Karakoc E, Nadeau JH et al (2006) RNA interaction prediction based on multiple
RNA-RNA interaction prediction and antisense sequence alignments. Bioinformatics 4:
RNA target search. J. Comput Biol 13:267–282 456–463
Part II
Analysis of High-Throughput RNA Sequencing Data

Chapter 8
Quality Control of RNA-Seq Experiments

Xing Li, Asha Nair, Shengqin Wang, and Liguo Wang
Abstract
Direct sequencing of the complementary DNA (cDNA) using high-throughput sequencing technologies
(RNA-seq) is widely used and allows for more comprehensive understanding of the transcriptome than
microarray. In theory, RNA-seq should be able to precisely identify and quantify all RNA species, small or
large, at low or high abundance. However, RNA-seq is a complicated, multistep process involving reverse
transcription, amplification, fragmentation, purification, adaptor ligation, and sequencing. Improper opera-
tions at any of these steps could make biased or even unusable data. Additionally, RNA-seq intrinsic biases
(such as GC bias and nucleotide composition bias) and transcriptome complexity can also make data imper-
fect. Therefore, comprehensive quality assessment is the first and most critical step for all downstream analy-
ses and results interpretation. This chapter discusses the most widely used quality control metrics including
sequence quality, sequencing depth, reads duplication rates (clonal reads), alignment quality, nucleotide
composition bias, PCR bias, GC bias, rRNA and mitochondria contamination, coverage uniformity, etc.
Key words Quality control, RNA-seq, High-throughput sequencing, Next-generation sequencing
1 Introduction
RNA-seq has led to a better understanding of the RNA universe by

providing unprecedented opportunities to interrogate the tran-
scriptome from different perspectives. These include gene expres-
sion profiling [1–4], new isoforms or alternative splicing
identification and quantification [5–7], novel transcripts such as
lincRNAs discovery [8–10], aberrant transcripts such as gene
fusion identification [11–13], and variant calling [14–17].
However, current library preparation protocols of RNA-seq are
still developing and possess several intrinsic biases and limitations,
such as nucleotide composition bias, GC bias, and PCR bias, which
could directly detriment many RNA-seq applications [18, 19].
In general, quality of RNA-seq experiments can be assessed at
two different levels. Raw sequence based metrics, which check
RNA-seq experiments at a “low level” because they do not require
sequence alignments. These assessments include read (i.e., a con-
secutive sequence of nucleotides) quality, read duplication rate
137
138 Xing Li et al.
(clonal reads), GC content, nucleotide composition bias, etc.

However, raw sequence-based metrics largely focuses on evaluat-
ing the success of sequencing technologies and themselves alone
cannot ensure the usability (biologic accuracy) of RNA-seq data.
Therefore, checking RNA-seq data at a “higher level” is also
imperative. These metrics include mapping statistics, coverage uni-
formity, saturation of sequencing depth, reads distribution over
gene structure, ribosomal RNA contamination, reproducibility
between biological replicates, etc.
1.1 Raw Sequence Phred quality score (Q) was originally developed by the program
Quality Phred to measure base-calling reliability from Sanger sequencing
chromatograms [20, 21]. It is defined as Q = −10 × log10(P) where
P is the probability of erroneous base calling. For example, a Phred
quality score of 30 means the chance that this base is called incor-
rectly is 1 in 1,000. Although the Phred program is rarely used in
next-generation sequencing field, Phred or Phred-like quality score
has become widely accepted to characterize the quality of DNA
sequences (see Note 1). Most often, Phred scores are reported as
their corresponding ASCII characters (33–126 or “!” to “V”) (see
Note 2 for FASTQ format), but SOLiD still uses numbers to rep-
resent quality scores.
There is no gold standard to tell if the quality of a particular
sequence is good or bad, as this is really depending on the purpose
of the study. For example, compared to expression profiling, vari-
ants calling tasks require much higher sequence quality. In general,
scores over 30 indicate very good quality, 20–30 indicate reason-
able good and <20 indicate poor quality. Parallel boxplots visualize
“per nucleotide quality score” by summarizing Phred qualities for
all reads at each position (Fig. 1a) [22, 23]. In addition, one can
also calculate the average quality score per read (“per sequence
quality score”) and check the quality score distribution of all
sequences (Fig. 1b).
1.2 Nucleotide GC content (or guanine-cytosine content) is the percentage of

Composition and GC bases in a DNA sequence that are either guanine or cytosine. It is
Content a simple way to measure nucleotide composition of DNA. The rea-
son to use GC rather than AT (or AU in RNA) is that GC content
carries more direct biologic meaning. GC pairs are more stable
than AT (3 vs. 2 hydrogen bonds) which has implications in PCR
experiments where the GC content of primers predicts their anneal-
ing temperature. Further, exons have much higher GC content
than introns and intergenic regions and cytosine is the target of
DNA methylation. People have found the dependence between
read coverage and the GC content of reference genome in high-
throughput sequence data. Therefore, evaluating GC content bias
in RNA-seq data is of great importance to both transcript detec-
tion and abundance quantification [18].
Assessing RNA-seq data quality 139
a b
0.25
65
0.20
60
Phred quality score
55
0.15
Density
50
0.10
45
0.05
40
35
0.00
Nucleotide postion on read 35 40 45 50 55 60 65
Phred quality score
Fig. 1 (a) Parallel boxplot showing “per nucleotide quality score.” All reads are overlaid together, and then
summarize Phred quality score (Y-axis) for each position of read from 5′ to 3′ end (X-axis). (b) “per sequence
quality score” distribution. For each read, “per sequence quality score” is calculated as the average Phred
quality score (X-axis) across all nucleotides
Assume RNA-seq reads were randomly sampled from expressed

transcripts, when we pileup reads together and calculate the nucle-
otide composition (percentage of A, C, G, and T) at each positions
or column, we expect little differences between columns. Random
fluctuations will be cancelled out because the large sample size
(i.e., hundreds of millions of reads). If we visualize nucleotide
composition versus nucleotide positions in a diagram [22] the lines
should be roughly flat at a value of 0.25 (Fig. 2a). In practical, the
first 12 bases starting from 5′ end of reads exhibit large deviation
from 0.25, this is due to the random hexamer priming during PCR
amplification [19]. A serious bias indicates the existence of over-
represented sequences, and such bias will influence coverage uni-
formity as well as transcripts abundance estimation. Per sequence
GC content can be roughly used to measure the randomness of
sequencing library as GC content of reads from random sequence
library follows normal distribution with the mean equals to the
overall GC content of the transcriptome (Fig. 2b). A poorly pre-
pared or contaminated library will exhibit a skewed distribution.
1.3 Duplicate Read duplication rate is affected by read length, sequencing depth,
Sequences (PCR transcript abundance and PCR amplification. Supposing the sequenc-
Duplication) ing library is purely random and read length is 36 bp, the chance
to get a duplicated read is 1/472 (or 4.5 × 10−44), this chance is still
slim even if the sequencing depth reaches hundreds of millions.
140 Xing Li et al.
a b
0.40
0.030
A
T
0.35 G
C
Nucleotide composition
0.020
0.30
Density
0.25
0.010
0.20
0.15
0.000
0 5 10 15 20 25 30 35 0 20 40 60 80 100
Nucleotide position on read GC content
Fig. 2 (a) Diagram showing nucleotide composition bias at the beginning of reads. All reads are overlaid
together, and then calculate nucleotide frequency (Y-axis) for each position of read (X-axis). Four nucleotides
were indicated using different colors. (b) “per sequence GC content” distribution
Therefore, the majority of duplicated reads were artificially gener-

ated from PCR amplification [22]. And because of this, duplication
rate is one way to check PCR amplification bias. To circumvent the
huge memory requirement, tools such as FastQC will only track the
first 200,000 short reads in each file for duplication level and creates
a graph plotting the count of sequences with different degrees of
duplication. By default, FastQC will raise a warning if there are more
than 20 % of duplicated sequences in total and a failure if this num-
ber reaches over 50 % as the sequencing library is seriously biased
and may not randomly sampling the target sequence.
1.4 Descriptive Mapping statistics are the simplest and most intuitive way to assess
Statistics if RNA sequencing was successful. These include mappability
(number of reads aligned to reference genome), number of reads
aligned to unique locations in the genome, and the number of
splice mapped reads and number of reads mapped to mitochon-
dria. It is difficult to derive reasonable or even empirical thresholds
to determine if a particular RNA sequencing was successful or not,
because these metrics really depend on read length, sequencing
depth, bioinformatic analysis parameters, sample preparation pro-
tocol, and tissue type. For example, compared with shorter reads,
longer reads will have better mappability, lower duplication rate,
higher proportion that aligned to unique genome location, more
spliced reads given the same sequencing depth. For the same
sequencing depth and same read length, number of splice reads
may be dramatically different between two RNA-seq datasets simply

because tissue origins are different. Muscle and heart tissues usu-
ally have much more mitochondria than other tissue types, and
therefore mitochondria reads could be a problem if they account
for a large proportion (i.e., >30 %), as this makes actual sequencing
depth much lower than expected.
1.5 rRNA/tRNA The goal of most RNA-seq studies is to interrogate functional mes-
Contamination sage RNA (mRNA). However, structure RNAs such as Ribosomal
RNA (rRNA) and transfer RNA (tRNA) are the most abundant
RNA species and constitute 60–90 % of total RNA in a cell. To
avoid having these RNAs dominate the sequencing data, it is nec-
essary to remove these RNA species before preparing libraries for
deep sequencing. Two approaches have been used to enrich
mRNA. The first approach starts with total RNA that has been
depleted of rRNA by using a set of oligos that binds to rRNA (such
as RiboMinus™), and the second method selects for transcripts by
isolating poly-A RNA as the staring materials for the construction
of sequencing libraries.
Even with ribosome depletion, a fair amount of ribosomal
sequences may still remain in the raw data. Small amounts of rRNA
contamination will not be a detriment to downstream analyses.
However, a larger amount of ribosomal reads usually suggests
rRNA depletion was inefficient or failed and additional sequencing
may be necessary. Assessing rRNA contamination is straightfor-
ward; aligning reads to reference genome and then counts how
many reads mapped to ribosome genes, or aligning reads directly
to ribosomal RNA sequences.
1.6 Saturation Test RNA-seq experiments are diverse in their aims and design goals.
of Sequencing Depth The amount of sequencing needed for a given sample is determined
by the goals of the experiment. For gene expression profiling,
where we are interested to find quantitative differences of known
genes between groups, modest sequencing depth is good enough
(e.g., 30 million pair-end reads with length >30 bp for mammalian
genomes). But for studies that involve investigation of alternative
splicing, gene fusion detection and novel transcript identification,
deeper read depths is required to be able to adequately cover not
just the exons but also exon–exon junctions. It is recommended by
ENCODE consortium that a minimum of 100–200 million
2 × 76 bp or longer reads is needed for mammalian genomes.
The saturation test is an approach to determine if current
sequencing depth is deep enough to satisfy a particular purpose.
It is fundamentally important because if sequencing was unsatu-
rated, estimated gene expression metrics such as RPKM (Reads Per
Kilobase exon per Million mapped reads) will be unstable and low
abundant isoforms will be undiscovered. In practical, we resample
5, 10, …, 100 % of the total mapped reads and RPKMs are
142 Xing Li et al.
a b
40
400
38
300
Number of junctions
36
RPKM
200
34
100
32 Annotated Junctions
Novel Junctions
30
0
5 25 45 65 85 5 25 45 65 85
Resampling percentage Resampling percentage
Fig. 3 Saturation test of sequencing depth. (a) Saturation test using RPKM (gene expression measurement).
RPKMs were recalculated for each resampled subset (blue dots) to test if RPKM values enter a steady state
(or saturated). (b) Saturation test using detected splice junctions (blue: annotated junction, red: novel junction).
Horizontal dashed line indicates all annotated junctions encoded in reference gene models
subsequently recalculated using each subset. For a particular tran-

script, RPKM values may vary at the beginning with very small
sample sizes, but finally could reach a plateau. The stable RPKM
values indicate a saturated sequencing depth; otherwise, sequenc-
ing depth should be increased until RPKM values enter a station-
ary stage (Fig. 3a). Saturation test for splice junction is similar;
splice junctions are detected for each resampled subset, and num-
ber of detected splice junctions will increase as the resample per-
centage increases, but finally will reach a fixed value. The junction
saturation test is very important for alternative splicing analysis, as
unsaturated sequencing depth would miss many low-abundance
yet bona fide splicing junctions. Due to the sensitivity of RNA-seq,
the number of identified novel splice junctions will increase as
sequencing depth goes deeper, and therefore saturation rarely
occurs for novel splice junctions even with billons of reads in mam-
malian genome (Fig. 3b).
1.7 Reproducibility For RNA sequencing technology, depending on the goal of

Between Replicates the experiment, replicates can be of two kinds, technical and
biological. Technical replicates are replicates obtained from the
same sample for purposes of studying batch effects and evaluating
the technology, including background noise, differences in
sequencing chemistry, instrument-to-instrument differences, etc.
Biological replicates are the most desired form of replicates, as
these provide us with the true variation among biological samples.
The use of such replicates comes into play for experiments that
involve comparison of two or more groups for differential expres-
sion analysis. It is recommended to have at least two biological
replicates per group in order to statistically determine the signifi-
cantly differentially expressed genes.
Evaluating the reproducibility between replicates is straight-
forward. Most often scatter plots are used to visualize the repro-
ducibility between expression measurements such as RPKM or
FPKM (Fragment Per Kilobase exon per Million reads). Logarithm
transformation of RPKM is necessary because of the large dynamic
range of RPKM values. After logarithm transformation, expression
values roughly follow a normal distribution and have a high
Pearson’s correlation coefficient (Fig. 4).
1.8 Coverage Gene body coverage describes the overall reads density over the
Uniformity mRNA regions (both UTR exon and CDS exon). Ideally, each
base has the same chance to be sequenced, and each site within
gene body has similar coverage. However, read density profiles can
be affected by library preparation protocol, PCR amplification,
RNA degradation, genome complexity and the underlying gene
3.0
r = 0.9904
2.5
Replicate-2 (RPKM,log10)
2.0
1.5
1.0
0.5
0.0
−0.5
−0.5 0.0 0.5 1.0 1.5 2.0 2.5 3.0
Replicate-1 (RPKM, log10)
Fig. 4 Scatter plot showing reproducibility between two RNA-seq datasets (technical replicates). Each blue dot
represents a gene, and the red dashed line is linear regression line
144 Xing Li et al.
model used. For example, RNA-seq data using poly-A selection

usually have higher coverage at 3′ end. PCR amplification effi-
ciency could be different for different DNA fragments site, this
also introduce uneven coverage. Low DNA complexity (or repeti-
tive) regions usually have higher coverage but this will depend on
how multi-hit reads were processed. For RNA-seq data that are
not strand specific, coverage profile is also affected by underlying
reference gene model; for example, an uneven coverage occurs
when two genes are overlapped in the genome and express differ-
ently. Coverage profile is the most intuitive way to check unifor-
mity, by normalizing all annotated genes into the same scale, and
then calculating coverage for each position (Fig. 5).
1.9 Reads After mapping reads to a reference genome, we can calculate the
Distribution (Intron, fraction of reads assigned to exons (including both UTR and CDS
Exon, UTR, etc.) exons), introns and intergenic regions based on the provided gene
model. In ideal conditions and for well-annotated organisms, most
of reads in RNA-seq data should be mapped to exonic regions.
However, in practice, a considerable amount of reads are mapped
to intron or intergenic regions. Except for mapping artifacts, inter-
genic/intronic reads are mainly from DNA contamination, pre-
mRNAs, new isoforms, or novel transcripts. Some UTR regions
are overrepresented (i.e., higher reads density) because of DNA
15000
read number
10000
5000
0 20 40 60 80 100
percentile of gene body (5'−>3')
Fig. 5 Coverage uniformity over gene body. All transcripts were scaled into the
same length (100 nucleotides) and then reads coverage (Y-axis) was calculated
for each position (X-axis) from 5′ to 3′ end
repeats or PCR bias, but most often they have lower reads density
due to RNA degradation. Specially, when poly-A RNA-seq proto-
col was used, reads are biased (i.e., overrepresented) in 3′UTR.
1.10 Strand In mammalian genomes, many genes encoded on different strands

Specificity are partially or completely overlapped. Therefore, conventional
RNA-seq cannot decode the complex transcriptome due the lack
of RNA polarity information. Strand specific RNA-seq experiments
can overcome these limitations and is better suited for genome
annotation (such as identify anti-sense noncoding RNA, demar-
cate the exact boundaries of adjacent genes transcribed on differ-
ent strand), de novo transcriptome assembly, accurate digital gene
expression and variant calling. Strandness is achieved by degrade
one cDNA strand or by ligating distinct RNA adapters to the 5′
and 3′ ends of each RNA molecule prior to cDNA synthesis [24].
The specificity of a strand-specific protocol can be calculated by
comparing the strand of reads to the strand of a reference gene.
For example, if there is only one gene located in the forward strand
in a particular region, one would expect the majority (>90 %) of
reads mapped to this region will be forward reads (determined
from protocol). Further, one can use spliced reads to measure
strand specificity, because the strandness of spliced reads can be
easily inferred from splicing motif (such as GT/AG).
2 Notes
1. Two different equations, standard Sanger format and Solexa/

Illumina format (prior to v1.3) have been used to calculate Q
[25]. After Illumina Pipeline v1.4, the quality scheme has been
changed into standard Sanger scale.
P
QSanger = -10 ´ log10 P; QSolexa ( prior _ to _ v1.3) = -10 ´ log10
1- P
2. FASTQ file stores nucleotide sequences and Phred qualities in

the same file. For example:
@read_1 ← (read identifier)
CCGGCCCCAGGCTCCTGTCTCCCCCCAGGTGTGTGGTGATGCCAGGCATG←sequence
+read_1 ← (read identifier, can be omitted)
@@;D?DADF=FHFBEBH+A2ADFGFH:;09?*/?BB=F)==FH3=@CH@?←
quality
@read_2 ← (another read identifier)
CTGCTCTCTCTTGCTGATGGACAAGGGGGCATCAAACAGCTTCTCCTCTG
+read_2 ← (another read identifier, can be omitted)
CCCFFFFFHHHGHJJIJJJJJJJIJJJJJJIIJIJJFGIIIJJIIIIIGJ
146 Xing Li et al.
References
1. Mortazavi A, Williams BA, McCue K et al prostate cancer using next-generation RNA
(2008) Mapping and quantifying mammalian sequencing. Gene Dev 21:56–67. doi:10.1101/
transcriptomes by RNA-Seq. Nat Methods gr.110684.110
5:621–628. doi:10.1038/nmeth.1226 13. Edgren H, Murumagi A, Kangaspeska S et al
2. Marioni JCJ, Mason CEC, Mane SMS et al (2011) Identification of fusion genes in breast
(2008) RNA-seq: an assessment of technical cancer by paired-end RNA-sequencing. Genome
reproducibility and comparison with gene Biol 12:R6. doi:10.1186/gb-2011-12-1-r6
expression arrays. Gene Dev 18:1509–1517. 14. Peng ZZ, Cheng YY, Tan BC-MB et al (2012)
doi:10.1101/gr.079558.108 Comprehensive analysis of RNA-Seq data
3. Wang Z, Gerstein M, Snyder M (2009) RNA- reveals extensive RNA editing in a human tran-
Seq: a revolutionary tool for transcriptomics. scriptome. Nat Biotechnol 30:253–260.
Nat Rev Genet 10:57–63. doi:10.1038/ doi:10.1038/nbt.2122
nrg2484 15. Bahn JHJ, Lee J-HJ, Li GG et al (2012)
4. Wilhelm BT, Landry J-R (2009) RNA-Seq— Accurate identification of A-to-I RNA editing
quantitative measurement of expression in human by transcriptome sequencing. Gene
through massively parallel RNA-sequencing. Dev 22:142–150. doi:10.1101/gr.124107.111
Methods 48:249–257. doi:10.1016/j.ymeth. 16. Park EE, Williams BB, Wold BJB, Mortazavi
2009.03.016 AA (2012) RNA editing in the human
5. Wang ET, Sandberg R, Luo S et al (2008) ENCODE RNA-seq data. Gene Dev 22:1626–
Alternative isoform regulation in human tissue 1633. doi:10.1101/gr.134957.111
transcriptomes. Nature 456:470–476. 17. Ramaswami G, Zhang R, Piskol R et al (2013)
doi:10.1038/nature07509 Identifying RNA editing sites using RNA
6. Katz Y, Wang ET, Airoldi EM, Burge sequencing data alone. Nat Methods.
CB (2010) Analysis and design of RNA doi:10.1038/nmeth.2330
sequencing experiments for identifying iso- 18. Benjamini Y, Speed TP (2012) Summarizing
form regulation. Nat Methods 7:1009–1015. and correcting the GC content bias in high-
doi:10.1038/nmeth.1528 throughput sequencing. Nucleic Acids Res
7. Trapnell C, Williams BA, Pertea G et al (2010) 40:e72. doi:10.1093/nar/gks001
Transcript assembly and quantification by 19. Hansen KD, Brenner SE, Dudoit S (2010)
RNA-Seq reveals unannotated transcripts and Biases in Illumina transcriptome sequencing
isoform switching during cell differentiation. caused by random hexamer priming. Nucleic
Nat Biotechnol 28:511–515. doi:10.1038/ Acids Res 38:e131. doi:10.1093/nar/gkq224
nbt.1621 20. Ewing B, Hillier L, Wendl MC, Green P
8. Cabili MN, Trapnell C, Goff L et al (2011) (1998) Base-calling of automated sequencer
Integrative annotation of human large inter- traces using phred. I. Accuracy assessment.
genic noncoding RNAs reveals global proper- Genome Res 8(3):175–85
ties and specific subclasses. Gene Dev 21. Ewing B, Green P (1998) Base-calling of auto-
25:1915–1927. doi:10.1101/gad.17446611 mated sequencer traces using phred. II. Error
9. Guttman M, Garber M, Levin JZ et al (2010) probabilities. Genome Res 8(3):186–94
Ab initio reconstruction of cell type-specific tran- 22. Babraham Bioinformatics – FastQC a quality
scriptomes in mouse reveals the conserved multi- control tool for high throughput sequence data.
exonic structure of lincRNAs. Nat Biotechnol http://www.bioinformatics.babraham.ac.uk/
28:503–510. doi:10.1038/nbt.1633 projects/fastqc/
10. Prensner JRJ, Iyer MKM, Balbin OAO et al 23. Wang L, Wang S, Li W (2012) RSeQC: quality
(2011) Transcriptome sequencing across a control of RNA-seq experiments. Bioin
prostate cancer cohort identifies PCAT-1, an formatics. Oxford, England. doi:10.1093/
unannotated lincRNA implicated in disease bioinformatics/bts356
progression. Nat Biotechnol 29:742–749. 24. Levin JZ, Yassour M, Adiconis X et al (2010)
doi:10.1038/nbt.1914 Comprehensive comparative analysis of strand-
11. Kannan K, Wang L, Wang J et al (2011) specific RNA sequencing methods. Nat
Recurrent chimeric RNAs enriched in human Methods. doi:10.1038/nmeth.1491
prostate cancer identified by deep sequencing. 25. Cock PJ, Fields CJ, Goto N, Heuer ML, Rice
Proc Natl Acad Sci U S A 108:9172–9177. PM (2010) The Sanger FASTQ file format for
doi:10.1073/pnas.1100489108 sequences with quality scores, and the Solexa/
12. Pflueger D, Terry S, Sboner A et al (2011) Illumina FASTQ variants. Nucleic Acids Res.
Discovery of non-ETS gene fusions in human 38(6):1767-71. doi:10.1093/nar/gkp1137
Chapter 9
Accurate Mapping of RNA-Seq Data

Kin Fai Au
Abstract
The mapping of RNA-Seq data on genome is not the same as DNA-Seq data, because the junction reads
span two exons and have no identical matches at reference genome. In this chapter, we describe a junction
read aligner SpliceMap that is based on an algorithm of “half-read seeding” and “seeding extension.” Four
analysis steps are integrated in SpliceMap (half-read mapping, seeding selection, seeding extension and
junction search, and paired-end filtering), and all toning parameters of these steps can be editable in a
single configuration file. Thus, SpliceMap can be executed by a single command. While we describe the
analysis steps of SpliceMap, we illustrate how to choose the parameters according to the research interest
and RNA-Seq data quality by an example of human brain RNA-Seq data.
Key words Junction reads, Junction, SpliceMap, Half-read seeding, Seeding extension
1 Introduction
In most studies, mapping is the first step of RNA-Seq data analysis.

RNA-Seq short reads can be mapped to reference genome or
annotated transcriptome. The annotated transcriptomes such as
RefSeq [1] and Ensembl [2] are the gene and gene isoform identi-
fications accumulated from many previous research, but they are
far from comprehensive even in human and mouse. Therefore,
when RNA-Seq data is mapped to annotated transcriptome, some
reads from undiscovered genes could be missed in the mapping
process. Some others could be mapped to incorrect positions
because the sequence similarity between undiscovered genes and
annotated genes. The incorrect mapping would cause bias or even
errors in the downstream analyses, such as abundance estimation.
Because abundance estimation from sequencing data is determined
by short read coverage. In many species, reference genome is bet-
ter sequenced and more complete than transcriptome. Therefore,
an accurate and comprehensive mapping of RNA-Seq data should
be performed against reference genome.
There are two types of reads in RNA-Seq data: exon reads and
junction reads. They should be considered separately when mapped
147
148 Kin Fai Au
to reference genome. An exon read is generated from a single exon

and thus it has identical match in genome. Therefore, exon reads can
be mapped directly to genome by regular short read aligners, which
is mostly used in DNA-Seq. A junction read spans two exons and
thus its sequence is the combination of two exonic sequences sepa-
rated by introns. Therefore, there is no identical match of junction
read in the reference genome. Instead of regular short read aligners,
we need a special aligner to map junction reads. Herein, we describe
a detailed procedure of a aligner SpliceMap [3] to map junction
reads to genome. Integrating with a regular short read aligner, such
as Bowtie [4], ELAND (Cox, unpublished software), and SeqMap
[5], SpliceMap maps both exon reads and junction reads to genome,
without bias on the existing transcriptome annotations.
2 Material
2.1 RNA-Seq Dataset In order to illustrate the procedure of SpliceMap, we use a human
brain RNA-Seq data from Illumina’s Human BodyMap 2.0 proj-
ect. It was released with Ensembl (release 62). The RNA was from
a 77-year-old Caucasian female. This RNA-Seq dataset contains
80,946,860 50 bp paired-end reads in FASTQ format. Each read
in FASTQ format contains four lines:
@HWI-BRUNOP16X_0001:3:1:9465:1024#0/1
TTAACAGTCGAGAGTGTGCTGAGAACTTAGACGGGATTTGGTAGGCCAAG
+HWI-BRUNOP16X_0001:3:1:9465:1024#0/1
TQFLTSSTTTggggfJKTSUcccccgggggggeggggggggfgffgggge
The first line is the read name and begins with a symbol “@.”
The second line is the raw sequence. The third line begins with a
symbol “+” and it is the extra info that was added during sequenc-
ing or preliminary analysis. In this example, the third line is just a
copy of the read name. The fourth line contains quality score of
each base. The quality score encodes the probability that the cor-
responding base call is incorrect. There are three types of quality
scores: phred-33, phred-64, and Solexa formats. Phred-33 format
encodes Phred quality score from 0 to 93 using the ASCII charac-
ters from 33 to 126. Phred-64 format encodes Phred quality score
from 0 to 62 using the ASCII characters from 64 to 126. Solexa
format encodes Phred quality score from −5 to 62 using the ASCII
characters from 59 to 126. Each character in quality score line
presents an integer as a sequencing quality which can be converted
to probability of false base call. Thus, please be aware that the same
character in three different quality score format can encode differ-
ent probabilities of false base call. Therefore, SpliceMap requires a
proper setting of quality score format in the configuration file. Our
human brain data set uses phred-64 format (see Subheading 3).
2.2 Reference In this chapter, we use hg19 as the reference human genome,
Genome which is in FASTA format, with each chromosome in a separate
RNA-Seq Mapping 149
file. These can be obtained from UCSC website http://hgdown-

load.soe.ucsc.edu/downloads.html (for hg19, the download link
is http://hgdownload.soe.ucsc.edu/goldenPath/hg19/bigZips/
chromFa.tar.gz). After unzipping, each chromosome is in its own
.fa file. However, the “*_random.fa" and "*_hap*.fa” files can be
deleted, as they are random chromosome segregations and haplo-
type sequences, respectively. Make sure all the genome files of dif-
ferent organisms are placed in separate directories.
In order to run SpliceMap with Bowtie, you also need to obtain
the other kind of genome files: Bowtie index, which can be down-
loaded from http://bowtie-bio.sourceforge.net/index.shtml (down-
load the same version as your genome files, probably UCSC). If the
genome you are interested in is not listed, you may need to build your
own index from the FASTA-formatted genome files by following the
instructions at http://bowtie-bio.sourceforge.net/tutorial.shtml.
2.3 Computational Mapping of high-throughput sequencing data requires high CPU

Hardware and high memory usage. Linux or Unix-based computer is more
efficient for this study. SpliceMap supports multi-threading, so
multi-core computer cluster is recommended to use to decrease
running time. The running speed is approximately proportional the
number of threads used. The size of the raw data, intermediate files
and final results (including coverage, junction detection and map-
ping results) is about 29, 32, and 33 GB for this example. Therefore,
100 GB of hard disk space is required for the mapping demonstra-
tion of this RNA-Seq data set. In general, the required disk space is
proportional to the size of RNA-Seq, so the required storage space
should be calculated accordingly. The memory usage varies at dif-
ferent steps of SpliceMap, but the peak usage is approximately pro-
portional to the size of the reference genome. In this example, it
requires at least 5 GB memory, but >10 GB is recommended.
2.4 Computational SpliceMap uses a “half-read seeding” and “seeding-extension” algo-

Software rithm to map junction reads. The first step is to find the alignment
of the half read of junction read on genome. Therefore, a regular
short read aligner is required. Bowtie, ELAND, and SeqMap are
compatible with SpliceMap. In this chapter, Bowtie is used (version
0.12.7, complied by gcc version 4.1.2 20080704). The precompiled
executable of Bowtie and prebuilt index of human hg19 genome
can be downloaded from http://bowtie-bio.sourceforge.net.
SpliceMap is an integrated pipeline to map both exon reads
and junction reads. The mapping of each read is independent, so
more available threads can allow parallel mapping processes and
reduce running time. SpliceMap 3.3.5.2 is used to in this chapter.
You can download the source code and the precompiled execut-
able files from SpliceMap homepage (www.stanford.edu/group/
wonglab/SpliceMap/download.html). If the executive files are
not compatible with your computer system, you can compile them
from the source code on your own computer (see Note 1).
150 Kin Fai Au
3 Methods
A junction read spans two exons and thus its sequence can be split
into two exonic fragments. The mapping of junction read is the
process to find where to split junction read and how to map two
exonic fragments to genome. It is intuitive that either exonic frag-
ment is longer or equal to the half length of original junction reads.
Therefore, either half of a junction read comes from a single exon,
and thus can be directly mappable to genome. This mapping hit
becomes a seeding to narrow the search region of the remaining
sequence. The second part is to extend the seeding alignment to
complete the mapping of the corresponding exonic fragment and
then find the mapping of the other exonic fragment in a small
neighbor region on the same chromosome (Fig. 1a). SpliceMap
contains four steps: half-read mapping, seeding selection, seeding
extension and junction search, and paired-end filtering (Fig. 1b).
SpliceMap integrates all steps in an easy one-button tool. The
parametric settings of each step are simply editable in the configu-
ration file “run.cfg.” In order to use SpliceMap on your own data,
you should follow these protocol:
1. Obtaining the genome files in the format “chr1.fa, chr2.fa, …”
and also the corresponding Bowtie index (see Subheading 2.2).
2. Create an empty directory, this will be the working directory.
3. Copy “run.cfg” from the SpliceMap package to the working
directory.
4. Edit run.cfg to include paths to your data files and genome
directories.
5. Edit the parametric settings for each steps, according to the
data quality and research interest
6. Execute “runSpliceMap run.cfg” while in your working directory.
7. After a certain time execution will conclude. You can find
results in the “output” directory.
At first, the information of SpliceMap input (data and genome)
should be edited correctly in run.cfg. The location of the reference
genome is required:
# Directory of the chromosome files in FASTA format
# Each chromosome should be in a separate file (can be concatenated)
# ie. chr1.fa, chr2.fa, …
# (single value)
genome_dir = /home/kinfai/genome/hg19/
Also the names of chromosome files can be defined optionally:

# Name of the chromosome file with wildcards
# If none is given the default of "chr*.fa" will be
used [this is compatible with the UCSC genome files]
# If you genome file is concatenated, just type the file name
RNA-Seq Mapping 151
Fig. 1 (a) SpliceMap flowchart and (b) SpliceMap algorithm

152 Kin Fai Au
# (optional)
# (single value)
chromosome_wildcard = chr*.fa
The other input of SpliceMap is the RNA-Seq data. The loca-
tion, data format and the quality format (see Subheading 2.1)
should be defined (see Note 2):
# These are the two lists of sequencer reads files.
# "reads_list2" can be commented out if reads are not
paired-end.
# Make sure the order of both lists are the same!
# Also, "reads_list1" must be the first mate of the
pair of reads.
# Note: pair-reads should be in the "forward-reverse" format.
# (multiple values)
> reads_list1
/scratch1/0-2-kinfai/50bp_mRNA_Seq/FCA/s_3_1_sequence.txt
<
> reads_list2
/scratch1/0-2-kinfai/50bp_mRNA_Seq/FCA/s_3_2_sequence.txt
<
# Format of the sequencer reads, also make sure reads are
# not split over multiple lines.
# Choices are: FASTA, FASTQ, RAW
# (single value)
read_format = FASTQ
# Format of the quality string if FASTQ is used
# Choices are:
# phred-33--Phred base 33 (same as Sanger format) [default]
# phred-64--Phred base 64 (same as Illumina 1.3+)
# solexa --Format used by solexa machines
# (single value)
quality_format = phred-64
The other parametric settings in run.cfg are to tune the perfor-
mance of four steps of SpliceMap. It is illustrated in the following
subsections by the example of human brain RNA-Seq data. Some
parameters are required while some others are optional. If optional
parameters start with the symbol “#,” they are commented out.
3.1 Half-Read Many Second Generation Sequencing platforms provide reads that
Mapping are not shorter than 50 bp. Thus, the half reads are not shorter
than 25 bp, which are long enough for reliable alignment on
genome by regular short read aligners. SpliceMap is compatible
with Bowtie, ELAND, and SeqMap for this mapping. In this
example, we select Bowtie:
# The short reads mapper used
# choices are "bowtie", "eland" or "seqmap"
# Bowtie is recommended
# this can be commented out if the mapping has already
been done and the
RNA-Seq Mapping 153
# appropriate ".t" files exist in the temp directory

[advanced].
# (single value)
mapper = bowtie
The tolerance of mismatches should be set based on the data
quality (please see Chap. 8). Consider SNP and the other sequence
variants between target sample and reference genome, the toler-
ance of mismatch on half–read seedings set to 1. However, if the
qualities of some base positions are very low, this value (seed_mis-
match) could be higher. As most current Second Generation
Sequencing platform provides high-quality data, this number is
not bigger than two generally.
# Maximum number of mismatches allowed in seeding for
junction search
# Choices are 0,1(default) or 2
# (optional)
# (single value)
seed_mismatch = 1
Low-quality bases or specific sequence (e.g., barcode or adap-
tor) may exists at two ends of some datasets, they should be clipped
off by the following setting:
# Full read length
# SpliceMap will only use the first "full_read_length"
bp for mapping.
# If the read is shorter than "full_read_length", the
full read will be used before head clip.
# If you comment this out SpliceMap will use as many
as possible.
# This is for the case where the reads might have N's
at the end.
# It is always desirable to cut off the N's
# (optional)
# (single value)
full_read_length = 50
# Number of bases to clip off the head of the read
# This clipping is applied after "full_read_length"
# (optional)
# (single value)
head_clip_length = 0
In this example, there is no clipping of base, because the data
qualities are high at all positions.
For instance, if we want to cut the last two bases off, we
should set:
full_read_length = 48
If we want to cut the first three bases off, we should set:
head_clip_length = 3
154 Kin Fai Au
Bowtie always requires a reference genome index for mapping.

You can either download a prebuilt index or build the own index
by (see Subheading 2.2). The path of the Bowtie index is required:
# Base of bowtie index, this should be the same genome
as the
# chromosome files
# eg. if you bowtie files are "genome/hg18/genome.1.ewbt", ?
# then your base dir is "genome/hg18/genome"
# (single value)
bowtie_base_dir = /home/kinfai/program/tophat-1.3.2.
Linux_x86_64/index/Homo_sapiens/UCSC/hg19/Sequence/
BowtieIndex/genome
Bowtie can utilize multiple threads to speed up the mapping
process. This can be set in the run.cfg as well:
# Number of threads to use for mapping
# Default value is 2
# (optional)
# (single value)
num_threads = 6
Bowtie also has an option to “try hard” to find as many hits as
possible but it is slightly slower than regular mode. Thus, this is
optional according to the research interest.
# Try hard?
# choices are "yes" or "no"
# Default value is yes. (it is not much slower, about 15%)
# I'm unsure if this is required, feel free to try with
# this option off and let me know your results.
# (optional)
# (single value)
try_hard = yes
3.2 Seeding The mapping hits of the half reads are used to narrow the search
Selection region of junction read mapping. But not all mapping hits of half
reads are used. Because there exist many repeat regions in genome,
especially mammalian genome, a half read may be mapped to mul-
tiple places. If all mapping hits are passed to the downstream map-
ping process, some half reads with extremely many repeats could
increase the computational intensity greatly. To balance the run-
ning time and sensitivity, we need to set a tolerance limit of multiple
mappability, which is determined by the complexity of the target
genome. According to the experience of human and mouse RNA-
seq analyses, we set the number of multi-hits not higher than 10:
# Maximum number of multi-hits
# If a 25-mer seed has more than this many multi-
hits, it will be discarded.
# (optional) Default is 10
# (single value)
max_multi_hit = 10
RNA-Seq Mapping 155
If two repeat regions are very closed, then the half-read map-
ping hits from two regions could be likely crossed over each other
in the downstream mapping process. This could introduce many
incorrect mapping hits of junction reads. We need to set a con-
straint of the distance between mapping hits from the same half
reads. Since the downstream mapping process is only performed
within a region of 99.9th percentile of intron size (400,000 bp in
human, see Fig. 2 and Subheading 3.3) around the half-read map-
ping hits, we exclude those hits that are within this region:
# Maximum intron size, this is absolute 99th-percen-
tile maximum.
# Introns beyond this size will be ignored.
# (optional) If you don't set this, we will assume a
mammalian genome (400,000)
# (single value)
max_intron = 400000
3.3 Seeding Half-read mapping hit is extended base by base until reach the
Extension canonical splice signal GT-AG. The extended hit is considered as a
and Junction Search possible hit of one exonic fragment of junction read. Then, the
other exonic fragment is searched in the neighbor regions. Two
exonic fragments of a junction read are separated by an intron, so
the search region is narrowed down to the 99.9th percentile of
intron size (400,000 bp in human, see Fig. 2 and Subheading 3.2).
Based on the data quality, we set a mismatch tolerance of the
entire read mapping. This value should be determined by the error
rate (please see the chapter “Quality controls of massive RNA
sequencing data”) but should not be smaller than seed_mismatch
(see Subheading 3.1). If seed_mismatch is bigger than read_mis-
match, then SpliceMap would output null result.
# Maximum number of mismatches allowed in entire read
# No limit on value, however SpliceMap can only identify
reads with
# a maximum of 2 mismatches per 25bp.
# Default is 2.
# (optional)
# (single value)
read_mismatch = 2
Some very long read may have a few low-quality bases at two
ends, while the remaining part are long and can be mapped reli-
ably. In order to rescue these reads from the constraints of read_
mismatch, an extra tolerance of unmappable termini is set. This
parameter is different from the clipping parameters in
Subheading 3.1. The base clipping should be used if some terminal
bases are of very low qualities for majority of reads. If only small
part of reads has quality problem at their terminus (especially for
very long reads), then max_clip_allowed is used instead. In this
chapter we use 50 bp reads as an example, so this optional param-
eter is commented out by starting with a symbol “#”:
156 Kin Fai Au
Fig. 2 The cumulative distribution function of intron lengths in human genome (RefSeq annotation)
# Maximum number of bases allowed to be soft clipped

from the ends of reads during
# alignment. This is required as mismatches near
junctions could cause parts of a
# a read to not map.
# Default is 40.
# (optional)
# (single value)
# max_clip_allowed = 40
The downstream processes after “Half-read Mapping” are
performed chromosome by chromosome. Thus, multiple cores can
be utilized to parallelize the mapping. This could require more
memory but reduce the running time:
# Number of chromosomes to process at once, to take
advantage of multi-core systems.
# This is not threading, so it will take extra mem-
ory. However, running 2 at a time should be fine.
# (optional) Default = 1
# (single value)
num_chromosome_together = 3
3.4 Paired-End Paired-end information can help to exclude some false mapping
Filtering hits. If a paired-end data is input, the paired-end filter is applied by
default. At first, three types of “good hits” are defined as: if two
halves of a full read are mapped to two locations that differ by
RNA-Seq Mapping 157
exactly the half-read length, this mapping hit of the full read is
called “exonic hit.” If a half-read mapping can be extended to rea-
sonable long (not shorter than full length minus 10), this mapping
hit is defined as “extension hit.” If a read can be mapped as a junc-
tion read by the processes above, then it is a “junction hit.” Three
constraints are used for paired-end filtering:
1. Both hits from a pair of reads are good hits.
2. The mapping hits from a pair of reads are at the same chromo-
some and not separated by the 99.9th percentile of intron size
(400,000 bp in human).
3. The directions and the positional order of the mapping hits from
a pair of reads are consistent to the sequencing platform setting.
If both mapping hits of two mates of a pair of reads pass three
constraints, then they are reported in a SAM file.
3.5 Mapping Results: The mapping results are written in a SAM (The Sequence
SAM File Alignment/Map) file “good_hits.sam” by SpliceMap (see Note 3).
If a read is multiply mapped, then there will be multiple entries in
the SAM file. Each line of the file good_hits.sam contains the fol-
lowing columns, separated by tab:
QNAME | FLAG | RNAME | POS | MAPQ | CIGAR | MRNM |
MPOS | ISIZE | SEQ | QUAL | OPTIONAL
where QNAME is the name of each query copied from the FASTA
or FASTQ file. A unique index is attached to the name;
FLAG is an integer that represents the information of read. The
explanation of each flag can be found at http://picard.sourceforge.
net/explain-flags.html;
RNAME is the name of reference sequence that the mapping hit
locates. For example, the reference sequences of hg19 has chromo-
some names as “chr1, chr2,…”;
POS is the mapping location;
MAPQ is 255 if uniquely mapped and 0 if multiply mapped;
CIGAR is the mapping pattern (details can be found in http://
samtools.sourceforge.net/SAM1.pdf);
MRNM and MPOS are the name of reference sequence and posi-
tion that the mapping hit of the mate read (the other read of the
pair) locates.
ISIZE is the distance between the mapping hits of two mates of the
paired-end reads.
SEQ and QUAL are the raw sequence and the mapping quality.
OPTIONAL contains the following items:
XS—Strand of junction
XC—Number of bases clipped
158 Kin Fai Au
NM—Number of mismatches in the read

MD—Describes location of mismatches
NH—Number of multi-hits
Here are the mapping hits of two mates of paired-end reads:
HWI-BRUNOP16X_0001:3:42:4915:16807#0[165352140] 65
chr1 14630 0 50M = 15006 327
TGGCTGTGTCCATGTCAGAGCAATGGCCCAAGTCTGGGTCTGGGGGGGAA
DFEEBADCD@<=<:=C@;>>45555FDDFFGGFGG@>?7?DFDFD@####
MD:Z:23C26 NM:i:1 XC:i:0 NH:i:5
HWI-BRUNOP16X_0001:3:42:4915:16807#0[165352140] 129
chr1 15006 0 33M757N17M = 14630 327
ACATCAAGGCCCCACCTTGGCTCGTGGCTCTCACTTGCTCCTGCTCCTTC
@>>>>AAAAA=;=;@AADA<@;@=8@A?A<25555DCFB?##########
XS:A:-MD:Z:8TG40 NM:i:2 XC:i:0 NH:i:8
The FLAG’s are 65 and 129, which mean that they are paired-
end reads and they are the first read and the second read respec-
tively. The CIGAR string of the first read is “50M,” so all 50 bases
are mapped together. Thus it is an exonic read. The CIGAR string
of the second read is “33M757N17M,” so the first 33 bases are
mapped together while the remaining 17 bases are mapped
together at a place 757 bases downstream. Thus, this is a junction
read. Both MAPQ’s are 0, so they are mapped to multiple places
and the other hits are reported in additional lines. In the SAM file,
some lines that begin with the symbol “@” are not the mapping
hits but just additional information.
3.6 Junction The mapping hits of junction reads can be converted to junction
Detection from detection and reported in a bed file “junction_color.bed.” The bed
Junction Read format is described in http://genome.ucsc.edu/FAQ/
Alignment FAQformat.html#format1. SpliceMap tags each junction with a
snippet of text describing the reliability of the junction (the fourth
column in bed file). The format of this snippet is
(nR)[width_nNR](nUR/nMR)
nR—Number of reads supporting this junction.
width—Range of different right lengths supporting this junction,
the larger the better.
nNR—Number of nonredundant reads supporting this junction.
nUR—Number of uniquely mappable reads supporting this
junction.
nMR—Number of multiply mappable reads supporting this
junction.
Using the parameters above, a number of junction filters pack-
aged with SpliceMap can control the specificity of the outputted
junctions. Different combinations of filters can be applied to tone
the balance of specificity and sensitivity. These filters only work on
the junctions (.bed files).
RNA-Seq Mapping 159
NnrFilter is the strictest filter. It filters junctions based on the

number of nonredundant supporting reads. Nonredundant is
defined as a read that is not mapped to exactly the same chromo-
some position. Empirical results show that using nNR > = 2, will
provide a specificity of about 90 %. The usage is:
nnrFilter infile.bed outfile.bed limit
where “limit” is the number of nonredundant supporting reads.
Suggested value is 2.
UniqueJunctionFilter removes junctions which are solely sup-
ported by multiply mapped reads. That is, junction can remain if it
has at least one uniquely mapped supporting read. Only junctions
with nNR = 1 are targeted by this filter. The usage is:
uniqueJunctionFilter infile.bed outfile.bed
NeighborFilter removes “lonely junction,” which is defined as
a junction with no exonic reads nearby. This filter seems to be able
to remove the few artifact junctions that appear in the middle of
nowhere. However, you may not want to use this filter if you are
looking for low coverage junctions. Only junctions with nNR = 1
are targeted by this filter. The usage is:
neighborFilter good_hits.sam infile.bed outfile.bed [limit]
where “limit” defines the number of nucleotides upstream and
downstream to check exonic reads nearby. The default value is
80 bp.
3.7 Results Using the settings above (6 threads for Bowtie and 3 cores for
from Human Brain SpliceMap), it took 69,580.6 s (19 h) to finish all SpliceMap steps.
RNA-Seq Data In total, 134,961,718 mapping hits from 127,349,786 reads are
reported in good_hits.sam. The mappable rate is 78.66 %
(127,349,786/(2 × 80,946,860)). Then, 193,127 nonredundant
junctions (junction_color.bed) are found from these mapping hits.
190,487 junctions pass the filter uniqueJunctionFilter:
uniqueJunctionFilter junction_color.bed junction_
nUM_color.bed
Or we can apply nnrFilter:
nnrFilter junction_color.bed junction_nNR_color.bed 2
139,190 junctions remain after the nonredundant read filter.
4 Notes
1. After extracting SpliceMap3352_linux-64.zip, the folder

should contain the following files and folders:
INSTALL src LICENSE bin run.cfg
160 Kin Fai Au
Try running “./bin/runSpliceMap.” If you see the following

output, the precompiled SpliceMap executive files are compat-
ible with your computer and you may skip this step.
SpliceMap3352_linux-64 moo$ ./bin/runSpliceMap
---== Welcome to SpliceMap 3.3.5.2 (55) ==---
Developed by Kin Fai Au and John C. Mu http://
www.stanford.edu/group/wonglab/SpliceMap/
__________
usage: ./runSpliceMap run.cfg
run.cfg -- Configuration options for this run, see
comments in file for details
See website for further details
However, if you see something like
SpliceMap3352_linux-64 $ ./bin/runSpliceMap
./bin/runSpliceMap: /usr/lib/libstdc++.so.6:
version `GLIBCXX_3.4.9' not found (required by ./
bin/runSpliceMap)
./bin/runSpliceMap: /usr/lib/libstdc++.so.6:
version `GLIBCXX_3.4.11' not found (required by
./bin/runSpliceMap)
Then the C++ standard libraries in your Linux distribu-
tion are not compatible and you need to build SpliceMap
from source code by following these instructions (for 64-bit
systems):
(a) Navigate to the “src” directory in the SpliceMap folder.
(b) Type “./install.sh bin,” this will install SpliceMap into the
example bin directory “bin.” Of course, you can install it
anywhere you like in future.
(c) Type “./install-bowtie.sh bin,” this will install Bowtie
into the example bin directory
or these instructions (for 32-bit systems):
(a) Navigate to the “src” directory in the example folder.
(b) Type “./install-32.sh bin,” this will install SpliceMap into
the example bin directory.
(c) Type “./install-bowtie-32.sh bin,” this will install Bowtie
into the example bin directory.
You may copy the SpliceMap executive files to any loca-
tion as long as all the executive files (including Bowtie) are in
the same directory or path.
2. A single set of paired-end data always contains two files which
are indexed by “1” and “2.” In the configuration file run.cfg,
two files should be listed under “> reads_list1” and “> reads_
list2,” respectively. If multiple sets of paired-end data are
RNA-Seq Mapping 161
input, the files listed under “> reads_list1” and “> reads_list2”
should be in the consistent order. Here is an example:
# These are the two lists of sequencer reads files.
# "reads_list2" can be commented out if reads are
not paired-end.
# Make sure the order of both lists are the same!
# Also, "reads_list1" must be the first pair.
# Note: pair-reads should be in the "forward-
reverse" format.
# (multiple values)
> reads_list1
/scratch1/0-2-kinfai/50bp_mRNA_Seq/FCA/s_3_1_
sequence.txt
sequence.txt
sequence.txt
<
> reads_list2
sequence.txt
sequence.txt
sequence.txt
<
3. In the output folder, there is a subfolder “debug_logs” which

contains a few log files that record the working status of each
step. You can learn the performances of each step and each
chromosome from them. Thus, they are helpful for debugging
if there are bugs and errors while SpliceMap is running.
References
1. Pruitt KD, Tatusova T, Maglott DR (2005) paired-end RNA-seq data by SpliceMap.
NCBI reference sequences (RefSeq): a curated Nucleic Acids Res 38:4570–4578
non-redundant sequence database of genomes, 4. Langmead B, Trapnell C, Pop M, Salzberg SL
transcripts and proteins. Nucleic Acids Res 33: (2009) Ultrafast and memory-efficient align-
501–504 ment of short DNA sequences to the human
2. Curwen V, Eyras E, Andrews TD, Clarke L, genome. Genome Biol 10:R25
Mongin E, Searle SM, Clamp M (2004) The 5. Jiang H, Wong WH (2008) SeqMap: mapping
Ensembl automatic gene annotation system. massive amount of oligonucleotides to the
Genome Res 14:942–950 genome. Bioinformatics 24:2395–2396
3. Au KF, Jiang H, Lin L, Xing Y, Wong WH
(2010) Detection of splice junctions from
Chapter 10
Quantifying Entire Transcriptomes by Aligned

RNA-Seq Data
Raffaele A. Calogero and Francesca Zolezzi
Abstract
Massive Parallel Sequencing methods (MPS) can extend and improve the knowledge obtained by conven-
tional microarray technology, both for mRNAs and noncoding RNAs. Although RNA quality and library
preparation protocols are the main source of variability, the bioinformatics pipelines for RNA-seq data
analysis are very complex and the choice of different tools at each stage of the analysis can significantly
affect the overall results. In this chapter we describe the pipelines we use to detect miRNA and mRNA
differential expression.
Key words Coding genes, Differential expression, Annotated genes, RNAseq workflow
1 Introduction
In less than a decade RNA-seq changed the way of doing transcrip-

tome analysis extending remarkably the knowledge derived by this
kind of studies [1–6]. However, RNA-seq technology is still under
evolution for both the wet and the dry sides. Now commercially
available RNA-seq kits allow the analysis of the coding transcrip-
tome of a single cell, directional sequencing that allows the quan-
tification of transcripts located on the opposite strand of the same
chromosomal locus, etc. Nevertheless, there are various criticalities
in a RNA-seq experiment: (1) the biological question, (2) the
experimental design, (3) the sequencing coverage, (4) the bioin-
formatics pipeline. Each of these points needs to be carefully
addressed to obtain biologically meaningful results.
1.1 RNA-Seq The biological question is the first and probably the most impor-
Criticalities tant point to be considered. Although RNA-seq provides a massive
amount of information if the experimental setting is not driven by
a clear biological question, the extraction of valuable biological
knowledge could be very challenging. The biological question is
very tightly connected with the experimental design, since only in
163
164 Raffaele A. Calogero and Francesca Zolezzi
the case of a clear biological question it is possible to correctly

design the experiment. On the contrary, if the aim of the experi-
ment is very vague (the so-called “fishing expedition experiment”),
it will be very difficult to design the RNA-seq experiment correctly.
While designing an experiment we need to keep in mind the familiar
expression; “garbage in garbage out.” Sequencing coverage and
replications are other critical issues. It is tricky to define which
would be the ideal number of reads for any experiment as well as
the number of biological replicates needed. The number of reads
required for each experiment is highly dependent on the aim of the
experiment itself. Estimating differential expression at gene level
will need a much lower number of reads compared the number of
reads needed to search for alternative splicing events or for the
detection of aberrant genomic rearrangement leading to chimeric
transcripts. However, if the sequencing depth is not sufficient to
achieve the aim of the experiment, one of the advantages of RNA-
seq is the possibility to increase, in a second time, the number of
reads associated to one or more specific sample simply by running
again the same library. This is feasible since the amount of cDNA
library produced is not a limiting factor and the technical bias
introduced during the sequencing procedure is negligible with
respect to the biological variance.
It is evident that, as in gene expression microarray experiments,
also in RNA-seq experiments the biological variability has to be
taken into consideration [7]. Thus biological replicates are indeed
needed. Looking at published RNA-seq experiments frequently the
number of replications is very limited, e.g., two biological replicates
for each experimental condition. This small number of replicates
limits the statistical power of the experiments and, as a conse-
quence, the statistics designed for RNA-seq are strongly biased by
the lack of a good estimation of the biological variance [8].
A well known bottom neck is given by the complexity of the
bioinformatics pipelines used for the RNA-seq analysis. We have
highlighted that, in case of miRNA-seq, it is extremely important
to have a clear insight of the tools used at each step of the pipeline
[4]. This issue can be extended to any other RNA-seq application.
Normalization and statistical tools are, for instance, a critical issue
for differential expression analysis at gene level [8, 9]. A critical
issue is also the detection of alternative splicing events. This field is
still in an early phase because, in spite of the growing number of
published methods on the detection of alternative splicing events,
we know very little about the strengths and weaknesses of
the different approaches. What at the moment is missing is a
benchmark experiments that allow an efficient comparison among
different pipelines.
In this chapter we present the approaches we use in our
laboratory for miRNA-seq and gene/transcript-level RNA-seq.
Transcriptome Quantification 165
2 Materials
2.1 Deep All the cDNA libraries where subjected to indexed sequencing run
Sequencing Data on an Illumina HiSeq 2000. We used 51 nts single end reads for
miRNA-seq because the use of 50 nts long reads makes easier the
detection and removal of linkers. In the case of mRNA-seq we use
pair end sequencing runs of 2 × 51 cycles as the best compromise
between sequencing cost and sequence specificity. Sequence direc-
tionality can provide valuable information specifically in the case
whole transcriptome analysis is performed, i.e., rRNA depletion
followed by sequencing of coding and noncoding RNAs.
2.2 Computational Being a relatively small lab we preferred a multicore solution with
Hardware respect to a cluster solution for RNA-seq analysis since manage-
ment of multiple cores on a single machine is easier with respect to
the management of a cluster solution. We run our analyses on two
AMD machines (64/48 cores), 512 Gb RAM each running linux
SUSE Enterprise 11/10. We decided also to invest significantly in
storage. Our storage is made of 30 × 2 Tb SATA disks and 6 × 2 Tb
hot spare disks configured as RAID 1 + 0, which guarantee a
reasonable security on hardware failure for the data under analysis.
We also have a 6 × 2 Tb SATA disks configured as RAID 5 for long
storage of raw data, i.e., fastq files.
2.3 Computational As aligner for RNA-seq projects we use STAR due to its elevated
Software performances [10]. We use SHRiMP [11] for miRNA-seq proj-
ects, because it has specific alignment parameters for miRNAs and
it was one of the aligner showing the best sensitivity in miRNA-seq
benchmark experiments [4].
As reference for miRNA alignment we use the latest version of
miRNA hairpins from miRbase (mirbase.org). The fasta file encom-
passing the hairpins of interest can be easily generated with the R
script shown below:
download.file("ftp://mirbase.org/pub/mirbase/CURRENT/
hairpin.fa.gz", "hairpin.fa.gz", mode="wb")
system("gunzip hairpin.fa.gz")
library(Rsamtools)
library(GenomicRanges)
library(Biostrings)
hairpins<- readRNAStringSet("hairpin.fa")
hsa<- hairpins[grep("hsa", names(hairpins))]
hsa<-DNAStringSet(hsa)
writeXStringSet(hsa, "hsa.fa", format="fastq")
Fig. 1 Table browser at USCS genome browser. The table browser allows the exportation in various formats of
annotation data. Specifically here it is shown the table selection to generate a GTF file for the human
transcriptome
The above example is given for human hairpins. Changing

“hsa” with another organism identifier of miRbase will produce
the specific subset of hairpins. As reference for mRNA-seq we use
the concatenated set of chromosomes downloadable at NCBI
(ftp://ftp.ncbi.nlm.nih.gov/genomes/) and the UCSC annota-
tion as GTF file. UCSC GTF annotation file can be easily gener-
ated using the table browser of the UCSC genome browser
(http://genome.ucsc.edu/cgi-bin/hgTables?command=start) as
shown in Fig. 1.
The majority of post alignment processing is done using R
(http://cran.r-project.org/) and Bioconductor (http://www.
bioconductor.org/) [12] frameworks (see Note 1).
3 Methods
3.1 Preprocessing Fastq files are the standard output of high-throughput sequencers,
including Illumina sequencers. Before aligning them to the refer-
ence genome it is important to check the overall quality of the
data. This can be done using FASTQC (http://www.bioinformatics.
babraham.ac.uk/projects/fastqc/), which is a stand-alone java
tool that allows the quality check of fastq as well as bam files.
For miRNA data analysis it is essential to remove linkers from
the fastq data. There are multiple tools to do it. We normally use the
perl script provided by mirTools suite (http://centre.bioinformatics.
zj.cn/mirtools/adaptortrim.php).
3.2 Sequence The first step in a RNA-seq pipeline is the alignment of the fastq
Alignment data to a reference dataset. miRNA data alignment is done with
SHRiMP [11]. The reference database for the alignment is created
with the following code:
$SHRIMP_FOLDER/utils/project-db.py --seed 0011111100111

1111100, 00111111110011111100, 00111111111100111
100, 00111111111111001100, 00111111111111110000 --h-
flag --shrimp-mode ls hsa.fa
The above code creates a set of files named: hsa-ls.seed.0, hsa-
ls.seed.1, hsa-ls.seed.2, hsa-ls.seed.3, hsa-ls.seed.4.
An example of the code needed to run SHRiMP is given below:
nohup $SHRIMP_FOLDER/bin/gmapper-ls --strata -o 1 --qv-
offset 33 -Q -L hsa-ls myfile.fastq -N number_of_cores --mode
mirna -w 170 % -E>myfile.sam 2>myfile.log &
In the case of RNA-seq pipeline the alignment is made with
STAR. As first step it is necessary to generate the reference data-
base. In this example hg19.fa is the file containing the concate-
nated set of all human chromosomes in fasta format and hg19.star
is the folder in which the database is created:
nohup STARstatic --runMode genomeGenerate –genomeDir
hg19.star --genomeFastaFiles hg19.fa --runThreadN number_
of_cores &
The example below refers to the alignment of pair end fastq
files against a reference database build in hg19.star folder.
nohup STARstatic --genomeDir hg19.star --readFilesIn my1R.
fastq my2R.fastq --runThreadN number_of_cores --outFile
NamePrefix ./mydata &
3.3 Postprocessing In the case of miRNA-seq the sam file generated by SHRiMP is
filtered to keep only the alignments with at least 16 contiguous
perfect matches. The script below associates the number of detected
counts to each miRNA present in the human hairpins dataset:
library(Rsamtools)
library(Biostrings)
asBam("my.sam", "my")
mybam<- scanBam("my.bam",param=ScanBamParam(what=c("r
name","cigar")))
cigar<- strsplit(mybam[[1]]$cigar, "S")
cigar1<- sapply(cigar, function(x){
good<- sub("M","",x[grep("M",x)])
})
cigar1<- as.numeric(cigar1)
reads<- as.character(mybam[[1]]$rname[which(cigar1>= 16)])
hairpins<- readRNAStringSet("hairpin.fa")
hsa<- hairpins[grep("hsa", names(hairpins))]
hsa<-DNAStringSet(hsa)
hsa.names<- strsplit(names(hsa), " ")

hsa.names<- sapply(hsa.names, function(x)x[1])
counts.mybam<- rep(NA, length(hsa.names))
names(counts.mybam)<- hsa.names
counts.mybam<- sapply(names(counts.mybam), function(x,y){
length(which(y==x))
}, y=reads)
Running the above script for all the bam files under analysis it
will be possible to generate the count table necessary for differen-
tial expression analysis. Since miRNA quantification relay on align-
ment over a very limited region of the miRNA hairpin no specific
normalization approach is required but only library size normaliza-
tion, which is embedded in the statistical model for differential
expression analysis (see Subheading 3.4). Before differential expres-
sion we filter out low or not expressed miRNAs, e.g., in a two
group experiment made in triplicate we remove all miRNAs being
characterized by having less than 10 counts in at least 50 % of the
analyzed samples.
In the case of mRNA-seq it is possible to quantify expression at
various levels:
● Exon
● Transcript
● Gene
Depending on the choice of the quantification approach
selected, a specific statistic is required for differential expression
(see Subheading 3.4). Since exons, transcripts and genes are char-
acterized by different size the sampling of the reads will be affected,
in the sense that it is more likely that reads will map more fre-
quently to a long exon compared to a short one. Furthermore, the
size of the sequence sampling, i.e., the total amount of the
sequenced reads for each sample, will also affect features coverage,
where feature refers to exon, transcript, or gene. To compensate
these biases Mortazavi [13] proposed RPKM (Reads Per Kilobase
per Million mapped reads) as a method of quantifying gene/tran-
script expression from RNA sequencing data by normalizing for
total read length and the number of sequencing reads. In case of
pair-end reads it is instead used FPKM (Fragments Per Kilobase
per Million mapped fragments, where a fragment). Recently it has
been shown that RPKM/FPKM do not represent the best approach
for data normalization [9]. In case of differential expression analy-
sis, other normalization approaches, embedded in the differential
expression model, are more indicated (see below Subheading 3.4).
The number of reads for a given transcript is not only deter-
mined by the transcript expression level since certain fragments are
preferentially detected in the mRNA-seq data acquisition process,
leading to nonuniform detection of expression between transcripts

[14–16]. Transcripts deconvolution, based on short reads data, is
not computationally trivial and requires the use of sophisticated
statistical models such as Cufflinks [17] or RSEM [18] in order to
estimate, rather than count, expression levels of these transcripts.
Since transcript expression estimation methods are still quite lim-
ited in performances we prefer to work, at least at the present time,
with a more conservative approach thus counting reads associated
to genes or to exons. It has to be noted that the mentioned
approaches work for differential expression analysis but are not
suitable for co-expression studies [19].
We use HTSeq (http://www-huber.embl.de/users/anders/
HTSeq/doc/count.html) to generate gene-level counts for each
of the sam file under analysis. A counts table can be imported in
DESeq [20, 21] for differential expression analysis using the fol-
lowing function:
newCountDataSetFromHTSeqCount(sampleTable, directory=".")
For exon-level analysis we use two python scripts provided as a
part of the DEXSeq [22] Bioconductor package: dexseq_prepare_
annotation (usage: python dexseq_prepare_annotation.py<in.
gtf><out.gff>) and dexseq_count.py (usage: python dexseq_count.
py--paired=PAIRED –stranded=no --minaqual=10 hg19.gff
my.sam my.counts), which use HTSeq to generate exon-level
counts. These counting procedures require some knowledge on
gene/exon annotation and results are, at a certain extent, depen-
dent on the type of annotation used. At the present time the most
used annotation databases are UCSC (genome.ucsc.edu) and
ENSEMBL (www.ensembl.org). Routinely in our lab we use
UCSC annotation, which is a transcript-oriented annotation. The
GTF annotation file can be easily generated as shown in Fig. 1.
Counts can be then imported in DEXSeq with the function:
read.HTSeqCounts(countfiles, design, flattenedfile)
where countfiles is a string vector containing the output files with the
paths from dexseq_count.py, design is vector of factors with infor-
mation corresponding to each of the countfiles and flattenedfile is
the annotation file generated by dexseq_prepare_annotation.py.
3.4 Differential Differential expression analysis for miRNAs and for gene level can
Expression Analysis be done applying the same statistical methods. We have observed
that the best tools for miRNA-seq differential expression are
DESeq [20] and baySeq [4, 23]. Their use is relatively simple and
a detailed description of the use of these tools is found by the fol-
lowing commands:
library(DESeq)
openVignette("DESeq")
library(baySeq)
openVignette("baySeq")
The counts for each experiment can be generated as indicated
above in post-processing paragraph.
DESeq estimates the variance in digital data and tests for dif-
ferential expression [20]. The method implemented in DESeq uses
the mean as a good predictor of the variance; that is, genes with a
similar expression level also have similar variance across replicates.
Hence, it predicts the variance from the mean. This estimation is
done by calculating for each gene, the sample mean and variance
within replicates and then fitting a curve to this data. The statistics
tests for differences between the base means of the two conditions.
baySeq is based on the NB (negative binomial) model.
Specifically, it estimates the empirical distribution on the parame-
ters of the NB distribution by bootstrapping from the data and the
subsequent acquisition of posterior likelihoods, thus estimating the
proportions of differentially expressed counts.
Both methods perform quite well on miRNA-seq data. Instead
in an analysis involving a much larger set of features, as in the case
of gene-level analysis, DESeq is often overly conservative [8].
Furthermore, the modification of the parameters in DESeq can
have significant effects on the results of the differential expression
analysis and the recommended parameters [20] are the best choice.
baySeq could produce highly variable results in case the differen-
tially expressed genes are up-regulated in one condition compared
to the other [8]. It is important to note that the behavior of DESeq
and baySeq in case of a very low number of replications, i.e., 2,
results in a poor FDR control [8].
As indicated above for the detection of alternative splicing
events we use an exon-level approach: DEXSeq [22]. The first
point to be highlighted is how counts are associated to exons.
Exons are fragmented on the basis of their association to a specific
isoform (Fig. 2) and the read counts are associated to them on the
Fig. 2 DEXSeq exon binning. Exons are fragmented on the basis of their associa-
tion to isoforms
basis of chromosomal location. Then DEXSeq uses the same

approach of DESeq to normalize the libraries on the basis of their
size and the variance of each exon is estimated using essentially the
same approach used in edgeR [24]. In calculating the exon vari-
ance it is possible to filter the data selecting only genes character-
ized by a maximum number of exons (maxExon parameter) and
exon’s total sum of counts over all samples is higher than a user
defined value (minCount). Specifically the minCount parameter
allows reducing significantly the number of evaluated genes and
thus reducing the number of multiple tests. Subsequently, a gener-
alized linear model is fitted:
sample + exon + condition * I ( exon == exonID ) (1)
and compared with the null model:

sample + exon + condition (2)
Then the deviances of both fits are compared using a X2-distribution
and p-values are corrected by the Benjamini–Hochberg adjustment
for multiple testing.
As in the case of DESeq a very detailed explanation of the pro-
cedure needed to run a differential expression analysis at exon level
can be seen using the following code:
library(DEXSeq)
openVignette("DEXSeq")
A particularly useful function in the package is the possibility
to generate an HTML report of the results using the following
code:
DEXSeqHTML(myExonSet, FDR=0.1)
4 Note
1. The procedure described for miRNA-seq is fully implemented

in oneChannelGUI [25] Bioconductor package. Since
oneChannelGUI is a graphical interface, user can run the
complete miRNA-seq analysis without the need to write any
line of R code.
Acknowledgement
This study was funded by grants from the Epigenomics Flagship

Project EPIGEN and FP7-Health-2012-Innovation-1 NGS-PTL
Grant no. 306242. SIgN, A*STAR, Singapore.
References
1. Maher CA, Kumar-Sinha C, Cao X, Kalyana- 13. Mortazavi A, Williams BA, McCue K, Schaeffer
Sundaram S, Han B, Jing X, Sam L, Barrette L, Wold B (2008) Mapping and quantifying
T, Palanisamy N, Chinnaiyan AM (2009) mammalian transcriptomes by RNA-Seq. Nat
Transcriptome sequencing to detect gene Methods 5(7):621–628
fusions in cancer. Nature 458(7234):97–101 14. Dohm JC, Lottaz C, Borodina T, Himmelbauer
2. McGettigan PA (2013) Transcriptomics in the H (2008) Substantial biases in ultra-short read
RNA-seq era. Curr Opin Chem Biol 17(1): data sets from high-throughput DNA sequenc-
4–11 ing. Nucleic Acids Res 36(16):e105
3. Arribere JA, Gilbert WV (2013) Roles for 15. Hansen KD, Brenner SE, Dudoit S (2010)
transcript leaders in translation and mRNA Biases in Illumina transcriptome sequencing
decay revealed by transcript leader sequencing. caused by random hexamer priming. Nucleic
Genome Res 23(6):977–987 Acids Res 38(12):e131
4. Cordero F, Beccuti M, Arigoni M, Donatelli S, 16. Wu Z, Wang X, Zhang X (2011) Using non-
Calogero RA (2012) Optimizing a massive uniform read distribution models to improve
parallel sequencing workflow for quantitative isoform expression inference in RNA-Seq.
miRNA expression analysis. PloS One 7(2): Bioinformatics 27(4):502–508
e31630 17. Trapnell C, Williams BA, Pertea G, Mortazavi
5. Carrara M, Beccuti M, Lazzarato F, Cavallo F, A, Kwan G, van Baren MJ, Salzberg SL, Wold
Cordero F, Donatelli S, Calogero RA (2013) BJ, Pachter L (2010) Transcript assembly and
State-of-the-art fusion-finder algorithms sensi- quantification by RNA-Seq reveals unanno-
tivity and specificity. Biomed Res Int 2013: tated transcripts and isoform switching during
340620 cell differentiation. Nat Biotechnol 28(5):
6. Carrara M, Beccuti M, Cavallo F, Donatelli S, 511–515
Lazzarato F, Cordero F, Calogero RA (2013) 18. Li B, Dewey CN (2011) RSEM: accurate
State of art fusion-finder algorithms are suit- transcript quantification from RNA-Seq data
able to detect transcription-induced chimeras with or without a reference genome. BMC
in normal tissues? BMC Bioinformatics Bioinformatics 12:323
14(Suppl 7):S2 19. Mostafavi S, Battle A, Zhu X, Urban AE,
7. Hansen KD, Wu Z, Irizarry RA, Leek JT Levinson D, Montgomery SB, Koller D
(2011) Sequencing technology does not elimi- (2013) Normalizing RNA-sequencing data by
nate biological variability. Nat Biotechnol modeling hidden covariates with prior knowl-
29(7):572–573 edge. PloS One 8(7):e68141
8. Soneson C, Delorenzi M (2013) A compari- 20. Anders S, Huber W (2010) Differential expres-
son of methods for differential expression anal- sion analysis for sequence count data. Genome
ysis of RNA-seq data. BMC Bioinformatics Biol 11(10):R106
14:91 21. Anders S, McCarthy DJ, Chen Y, Okoniewski
9. Dillies MA, Rau A, Aubert J, Hennequet- M, Smyth GK, Huber W, Robinson MD
Antier C, Jeanmougin M, Servant N, Keime (2013) Count-based differential expression
C, Marot G, Castel D, Estelle J et al (2012) A analysis of RNA sequencing data using R and
comprehensive evaluation of normalization Bioconductor. Nat Protoc 8(9):1765–1786
methods for Illumina high-throughput RNA 22. Anders S, Reyes A, Huber W (2012) Detecting
sequencing data analysis. Brief Bioinform differential usage of exons from RNA-seq data.
14(6):671–683 Genome Res 22(10):2008–2017
10. Dobin A, Davis CA, Schlesinger F, Drenkow J, 23. Hardcastle TJ, Kelly KA (2010) baySeq:
Zaleski C, Jha S, Batut P, Chaisson M, Gingeras empirical Bayesian methods for identifying dif-
TR (2013) STAR: ultrafast universal RNA-seq ferential expression in sequence count data.
aligner. Bioinformatics 29(1):15–21 BMC Bioinformatics 11:422
11. Rumble SM, Lacroute P, Dalca AV, Fiume M, 24. Robinson MD, McCarthy DJ, Smyth GK
Sidow A, Brudno M (2009) SHRiMP: accu- (2010) edgeR: a Bioconductor package
rate mapping of short color-space reads. PLoS for differential expression analysis of digital
Comput Biol 5(5):e1000386 gene expression data. Bioinformatics 26(1):
12. Gentleman RC, Carey VJ, Bates DM, Bolstad 139–140
B, Dettling M, Dudoit S, Ellis B, Gautier L, 25. Sanges R, Cordero F, Calogero RA (2007)
Ge Y, Gentry J et al (2004) Bioconductor: oneChannelGUI: a graphical interface to bio-
open software development for computa- conductor tools, designed for life scientists
tional biology and bioinformatics. Genome who are not familiar with R language.
Biol 5(10):R80 Bioinformatics 23(24):3406–3408
Chapter 11
Transcriptome Assembly and Alternative Splicing Analysis

Paola Bonizzoni, Gianluca Della Vedova, Graziano Pesole,
Ernesto Picardi, Yuri Pirola, and Raffaella Rizzi
Abstract
Alternative Splicing (AS) is the molecular phenomenon whereby multiple transcripts are produced from
the same gene locus. As a consequence, it is responsible for the expansion of eukaryotic transcriptomes.
Aberrant AS is involved in the onset and progression of several human diseases. Therefore, the character-
ization of exon–intron structure of a gene and the detection of corresponding transcript isoforms is an
extremely relevant biological task. Nonetheless, the computational prediction of AS events and the reper-
toire of alternative transcripts is yet a challenging issue.
Hereafter we introduce PIntron, a software package to predict the exon-intron structure and the full-
length isoforms of a gene given a genomic region and a set of transcripts (ESTs and/or mRNAs). The
software is open source and available at http://pintron.algolab.eu. PIntron has been designed for (and
extensively tested on) a standard workstation without requiring dedicated expensive hardware. It easily
manages large genomic regions and more than 20,000 ESTs, achieving good accuracy as shown in an
experimental evaluation performed on 112 well-annotated genes selected from the ENCODE human
regions used as training set in the EGASP competition.
Key words Alternative splicing, Spliced alignment, Gene structure, Expressed isoforms, Software
package
1 Introduction
Genes in eukaryotes may undergo to the Alternative Splicing (AS)

mechanism and, therefore, they may produce alternative mRNAs
(or alternative isoforms). AS is the major molecular mechanism
responsible for the expansion of eukaryotic transcriptomes and it is
involved in the onset and progression of several human diseases [1].
For this reason, two fundamental and challenging problems arise:
(1) characterizing the gene exon–intron structure and (2) finding
the set of alternative isoforms potentially expressed by a gene.
In the present work we introduce PIntron, a software package
designed for solving these problems through spliced alignments of
expressed transcripts to the genome. Its input data are the genomic
sequence containing a gene locus and the corresponding set of
173
174 Paola Bonizzoni et al.
expressed sequences such as ESTs and/or full-length mRNAs.

PIntron works as a pipeline composed of two main steps: the first
computes the spliced alignments of input transcripts to the genomic
sequence in order to reconstruct a consensus exon–intron gene
structure [2], while the second step assembles potential full-length
gene isoforms applying the graph-based method presented in [3].
Furthermore, PIntron tries to annotate the predicted isoforms
identifying some features such as the coding sequence (CDS), the
polyA tail, the polyadenylation signal (PAS), and the Nonsense-
Mediated Decay signal (NMD) [4].
PIntron produces an output in two file formats: GTF (Gene
Transfer Format) and JSON (JavaScript Object Notation). The
GTF format is a widely used standard format for representing a set
of full-length isoforms and can be easily visualized using genome
browsers as the UCSC (http://genome.ucsc.edu/). The JSON
format, instead, allows the representation of all details of a predic-
tion (such as the set of inferred introns and all full-length isoforms
along with all their characteristics) in a text-based format that is
both human-readable and machine-parsable in a variety of pro-
gramming languages.
PIntron is able to process complex gene structures or genes
associated with a large cluster of expressed sequences. Its efficiency
derives from the use of efficient data structures for exploring all
possible spliced alignments of input transcripts with respect to the
genome. Then, only those that support a parsimonious consensus
exon–intron gene structure are retained. Further ad hoc refinement
procedures are implemented to improve the prediction accuracy.
PIntron is customizable by a variety of input parameters (such
as the minimum length of exons and introns) and also nonexpert
users can execute the program without any problem.
The accuracy and efficiency of PIntron have been evaluated by
comparing results with those from ASPic [5] and Exogean [6],
which are two well-established tools for gene structure prediction.
Below we report technical details to install and run PIntron as
well as the characteristics of input and output files.
2 Materials
2.1 Installation PIntron has been developed and intensively tested on Linux and
should easily run also on OS X. It can be downloaded from http://
pintron.algolab.eu as a TAR or ZIP archive and the following addi-
tional software is required: Python v3.1 or newer, Perl v5 and a
recent version of the standard GNU toolchain (the C compiler
“gcc” and the build tool “make”, in particular). All these prereq-
uisites can be easily installed on popular Linux distributions
through their package manager.
Transcriptome Assembly and Alternative Splicing Analysis 175
PIntron is composed of several programs (each one performing

a step of the prediction pipeline) managed and executed by the
script “pintron”. The compilation step (for which detailed and
up-to-date instructions are available at the PIntron website) pre-
pares a directory containing the main script “pintron”, all needed
executables and sample input files (human gene TP53).
In order to successfully execute the pipeline, we suggest to
copy the directory containing the executables prepared at the com-
pilation step in a single (dedicated) directory that we will indicate
as the binary directory. Sometimes it may be convenient to create a
subdirectory of the user home folder, called “pintron”, in which
executables can be copied. In this case, no system administrative
privileges are required.
2.2 Input Data PIntron takes in input the genomic sequence of a gene locus and a
set of expressed sequences (mostly ESTs and mRNAs). In particu-
lar, it requires two input files: a FASTA file containing the genomic
sequence (genomic file) and a MultiFASTA file containing the
EST/mRNA sequences (transcript file). All nucleotide sequences
must be in uppercase letters.
In addition, the genomic file must contain the nucleotide
sequence on the same strand of the input gene and its FASTA
header must be in the format “>chrZZ:START:END:STRAND”,
where ZZ is the reference chromosome, START and END are the
start and end coordinates of the genomic region on the reference
chromosome, and STRAND is +1 (or, simply, 1) for a gene on the
plus strand of the chromosome or -1 for a gene on the minus
strand of the chromosome.
An example of genomic file is shown in Fig. 1 for human TP53
gene. Header characteristics are highlighted in boxes.
The transcript file must contain input sequences in a
MultiFASTA format (typically a UniGene file) and each header
should include the substring “/gb = XXXXXX”, where XXXXXX is
a unique identifier (like the GenBank identifier associated to
UniGene sequences). The substring “/clone_end = YY” is optional
and allows to specify the read strand (YY). In particular, YY = 3′
means plus strand (or 5′3′), while YY = 5′ means minus strand
(or 3′5′). In the latter case, the input transcript should be reversed
Fig. 1 Example of genomic file (genomic.txt) for human TP53 gene

and complemented before the alignment step. If the strand is not

specified, then the sequence is considered on plus strand by default.
RefSeq mRNAs are always considered on plus strand. PIntron rec-
ognizes RefSeq mRNAs if the corresponding GenBank (gb) iden-
tifier starts with “NM_”. In any case, if no valid alignments are
found using the suggested strand, the program will compute a
spliced alignment on the opposite strand.
Figure 2 represents an extract of a valid transcript file from the
UniGene cluster Hs.437460 associated to the human TP53 gene.
Two boxes highlight the GenBank identifier and its strand.
The user can optionally provide the coding sequence (CDS)
annotation for RefSeq mRNAs by means of an optional text file
that must be stored in the working directory and must be called cds.
Indeed, input RefSeq mRNAs are handled as full-length isoforms
(and thus not involved in the assembly process like EST sequences)
and the corresponding coding sequence is derived from the sup-
plied cds file (if any).
The format of cds file is quite simple and an example is pro-
vided in Fig. 3 in which the first RefSeq sequence (NM_000546)
is 2,591 bp long and composed of 11 exons; its coding sequence
starts at position 203 and ends at position 1,384 (stop codon
included). Its translation is composed of 393aa.
Fig. 2 Example of input transcript file (ests.txt) for human TP53 gene: EST sequence CN342738 from Unigene
cluster Hs.437460
Fig. 3 Example of CDS annotation file for human TP53 gene

In all cases in which the file cds is not present or the CDS
annotation is not provided or the isoform has been obtained by
assembling spliced ESTs, PIntron tries to annotate the coding
sequence by computing the left-most Open Reading Frame (ORF)
of at least 100 bp and having a non-weak context. If all predicted
ORFs of at least 100 bp have a weak context, then the left-most
ORF (if any) is reported. The ORF context [7] is given by a small
window “XNNATGY” of seven bases around the start codon ATG
(underlined). The context is strong if both X and Y are purines
(A or G), medium if only one of X and Y is a purine, and weak if
neither X nor Y is a purine. Symbol N stands for any base.
3 Methods
3.1 Description The PIntron pipeline (Fig. 4) is composed of two steps: (1) recon-
of the Pipeline struction of the exon–intron gene structure by computing
transcript-to-genome alignments (spliced alignments), and (2)
assembly of potentially expressed full-length isoforms.
Fig. 4 The PIntron pipeline

The first step takes in input the genomic sequence (genomic

file) and the expressed sequences (transcript file) and produces a set
of spliced alignments (one for each expressed sequence) that are
supposed to be compatible to an exon–intron gene structure.
During this step, a pool of alignments are computed [2] and, then,
only those biologically meaningful are retained and used to draft a
consensus gene structure using the methodology proposed in [8].
This method exploits the inherent redundancy of information in a
set of transcripts in order to select, among all possible alignments,
only those which allow to infer splice site junctions largely sup-
ported by input data. Finally, the predicted introns are classified in
U2 and U12, and putative splice sites are scored according to [9].
The alignment procedure of each input transcript is based on
the efficient construction of maximal sequences of perfect
transcript-genome matchings, called embeddings. An embedding
reveals the basic “building blocks” of the spliced alignment. The
embeddings are obtained from paths of a graph structure, called
embedding graph, whose vertices are the perfect matching
substrings of the genomic sequence and the input transcript.
An embedding graph gives a compact and implicit representation
of all the embeddings of a transcript with respect to the genome in
input. This procedure is efficient and runs in time linear in the
transcript length, in the genomic length, and in the size of the
resulting alignments.
In the second step, PIntron reconstructs all potentially
expressed isoforms by applying the graph-based method presented
in [3], and annotates inferred transcripts adding some specific
features such as the coding sequence (CDS), the polyA tail, and
the polyadenylation signal (PAS).
Accuracy of PIntron has been assessed by comparing its perfor-
mances with those [2] from ASPic [5] and Exogean [6]. In par-
ticular, the assessment has been conducted on 112 genes selected
from the ENCODE human regions and used as training set in the
EGASP competition [10]. Sensitivity and specificity of predictions
have been computed with respect to the GENCODE annotation.
Our results indicate that PIntron performs better than ASPic
and Exogean, except for the specificity at transcript level. In addition,
it is the only tool to successfully compute a gene structure for all
112 genes (ASPic and Exogean failed to compute a gene structure
for 19 and 8 genes, respectively).
Scalability of PIntron has been tested on 26 complex gene
structures that are computationally intensive to analyze. Although
Exogean is on average 30 % faster than PIntron, it is not able to
compute all gene structures.
3.2 Execution As explained in Subheading 2.1, PIntron is composed of several

executables coordinated by a script called “pintron”. In the follow-
ing, we assume that all the executables (together with the “pin-
tron” script) are placed in a directory indicated by the placeholder
“<BINDIR>”. Please replace the placeholder with the full path of

the binary directory (i.e., the directory where PIntron executables
have been copied to). If the binary directory is in the system PATH,
then it can be omitted.
3.2.1 Invocation The PIntron pipeline is executed by invoking the command

and Main Parameters “<BINDIR>/pintron”. Several options can be specified to cus-
tomize default parameter values as (in parentheses the equivalent
short form is reported):
● Options for input data:
– --genomic = FILE (-g FILE), the input file containing the
genomic sequence (default: “genomic.txt”).
– --est = FILE (-s FILE), the input file containing the
expressed sequences (default: “ests.txt”).
– --organism = NAME (-n NAME), the name of the organ-
ism (default: “unknown”).
– --gene = NAME (-e NAME), the gene name (default:
“unknown”).
● Options for output data:
– --output = FILE (-o FILE), the output file name (that will
be created in JSON format) (default: “pintron-full-output.
json”).
– --gtf = FILE (-t FILE), the output file name in GTF format
(default: “pintron-all-isoforms.gtf”).
– --strict-GTF-compliance, to force PIntron to report only
isoforms with an annotated CDS.
– --logfile = FILE (-l FILE), the log file name (default:
“pintron-pipeline-log.txt”).
– --general-logfile = FILE, the log file name for the main
script (default: “pintron-log.txt”).
● Additional parameters:
– --bin-dir = DIR (-b DIR), path to the binary directory
including PIntron executables. This option has to be speci-
fied if the binary directory is not listed in the PATH
variable.
Other advanced parameters (not discussed here) allow to adapt
the alignment step of the pipeline to the characteristics of the data
to analyze (see Note 1).
Two equivalent ways to launch PIntron on the provided sam-
ple files are:
<BINDIR>/pintron --bin-dir=<BINDIR> --genomic=
genomic.txt --est=ests.txt --organism=human
--gene=TP53 --output=pintron-full-output.json
--gtf=pintron-all-isoforms.gtf
<BINDIR>/pintron -b <BINDIR> -g genomic.txt -s

ests.txt -n human -e TP53 -o pintron-full-
output.json -t pintron-all-isoforms.gtf
3.2.2 Managing The pipeline could prematurely terminate if a runtime error occurs.
Runtime Errors There are two common causes for runtime errors: wrong input file
formats and exhaustion of computational resources. PIntron
attempts to check if the input files are strictly compliant with their
required format (as described in Subheading 2) and, in case it is
not able to correctly parse them, terminates with a runtime error
(or with unexpected/partial results). As a consequence, in case of
runtime errors or unexpected results, it is important to check the
formats of input files and rerun the pipeline. The other most com-
mon cause of runtime errors is the exhaustion of the computa-
tional resources. PIntron has been designed to run on standard
mid-range workstations and, as such, has strict default limits on
computational resources (running times and amount of memory).
Default settings are conservative and prevent to deplete all compu-
tational resources of the workstation.
The inspection of log files (see Note 2) or the customization of
default computational limits (see Note 3) could help in identifying
and removing the causes of runtime error. However, if the user is
not able to trace down the cause of a runtime error can report it
using the issue tracker at PIntron website (http://pintron.algolab.
eu/). The report should be as complete as possible and should
include all data needed to reproduce the error. Moreover, results
from all intermediate steps should be attached to the report. By
default, these (temporary) files are deleted at the end of the execu-
tion. To prevent this, it is possible to use the option --keep-
intermediate-files (−k, in short).
3.3 Output Data PIntron produces its output in two standard formats: GTF (Gene
Transfer Format) and JSON (JavaScript Object Notation). More
precisely, it outputs a file describing the predicted full-length iso-
forms in GTF format and a JSON file reporting all details of the
prediction (such as the exon–intron gene structure and the pre-
dicted full-length isoforms along with their annotations). GTF is a
well-known standard format in bioinformatics used to describe
gene features, while the JSON format allows to describe all results
into a unique file. The JSON format has been chosen since it is
human-readable and very easy to parse by means of libraries that
are available for almost all programming languages (Ruby, Python,
JavaScript, Perl and so on).
3.3.1 The GTF A GTF file is composed of records with nine tab-separated fields,
Output File and each record represents a feature. Here we omit the details of
the format and refer the reader to the official documentation
(http://mblab.wustl.edu/GTF22.html).
The GTF output file is specified by the option --gtf and con-
tains all predicted full-length isoforms if the execution option
--strict-GTF-compliance has not been specified, or only the subset
of the CDS-annotated isoforms, otherwise. A full-length isoform is
described in the GTF output as composed of features on the
genome. A feature can be an exon, a 5′ untranslated region, a 3′
untranslated region, a coding sequence, a start codon, or a stop codon.
Below, there is a description of GTF fields produced by
PIntron:
● seqname is the reference chromosome.
● source is always the string “PIntron”, since PIntron is the
program generating the feature.
● feature is the string describing the feature represented by the
GTF record, and its value belongs to the set {“exon”, “3UTR”,
“5UTR”, “CDS”, “start_codon”, “stop_codon”}.
● start is the (1-based) starting position of the feature on the
plus strand of the reference chromosome (that is specified by
the seqname field).
● end is the (1-based) ending position of the feature on the plus
strand of the reference chromosome (that is specified by the
seqname field).
● score is always a dot “.”, since PIntron does not produce this
value.
● strand is the strand of the input gene.
● frame can assume one of the following values:
– 0, if the value of the field feature is “start_codon” or
“stop_codon” (the feature is a start or a stop codon).
– 0, if the value of the field feature is “CDS” (the feature is
a coding sequence), and its first base is the first base of a
codon.
a coding sequence), and its first base is the second base of a
codon.
a coding sequence), and its first base is the third base of a
codon.
– A dot “.” (the frame is not defined), if the value of the field
feature is “exon” or “5UTR” or “3UTR”.
● attribute specifies the gene and the full-length isoform con-
taining the feature. More in detail:
– The value of the tag gene_id is the gene HUGO
name specified by the execution option --gene
(see Subheading 3.2).
Fig. 5 Example of GTF records (features) for human TP53 gene
– The value of the tag transcript_id has the format

“GENE.ID”, where GENE is the gene name (see the
above tag gene_id) and ID is a progressive index associated
to the full-length isoform containing the feature.
An example of GTF records produced by PIntron on the
human TP53 gene is shown in Fig. 5, where the eight records
(or features), related to one of the predicted isoforms, are reported.
3.3.2 The JSON A JSON file is a (key, value) dictionary, where value can be in turn
Output File a dictionary, a list, or simple value (like a string or a number). In
the following, we describe the JSON structure of the PIntron
output file, while we refer the interested reader to the official docu-
mentation (http://www.json.org/) for a complete description of
the format.
We have chosen key names that are self-explanatory, and we
have exploited the nesting nature of the JSON format to encode a
gene and its sets of introns and of full-length isoforms. The JSON file
produced by PIntron is specified by the execution option --output
(see Subheading 3.2). The value of a key ending with a question
mark (“?”) can be true or false, since it represents the answer to a
given question.
The whole PIntron prediction is represented by a dictionary
giving the predicted exon–intron gene structure and the full-length
isoforms potentially expressed by the gene. In particular, the top-
level dictionary gives the following data:
● The input genomic sequence (key = “genome”) that in turn is
represented by a dictionary specifying:
– The sequence identifier (key = “sequence_id”) through the
FASTA header of the genomic file (without the symbol “>”).
– The gene strand (key = “strand”), which is “+” if the gene is
on the plus strand of the chromosome, otherwise it is “-”.
– The sequence length (key = “length”).
● The number of input transcripts which have been successfully
aligned to the genomic sequence (key = “number_of_
processed_transcripts”).
● The pipeline version that produced that JSON file

(key = “program_version”).
● The version of the JSON schema (key = “file_format_version”).
● The set of predicted introns (key = “introns”).
● The number (key = “number_of_predicted_isoforms”) and the
set (key = “isoforms”) of predicted isoforms.
The values associated to the keys “introns” and “isoforms” are
two dictionaries describing the set of the predicted introns and the
set of the predicted full-length isoforms, respectively. In both these
dictionaries, keys are progressive identifiers, while each associated
value represents an intron and an isoform, respectively.
An intron is a dictionary giving the following data:
● The (1-based) start position of the intron both on the input
genomic sequence (key = “relative_start”) and on the reference
chromosome (key = “absolute_start”).
● The (1-based) end position of the intron both on the input
genomic sequence (key = “relative_end”) and on the reference
chromosome (key = “absolute_end”).
● The intron length (key = “length”).
● The 15 bp-long suffix of the donor exon (key = “donor_exon_
suffix”), and the 15 bp-long prefix of the acceptor exon
(key = “acceptor_exon_prefix”), flanking the intron.
● The 20 bp-long prefix (key = “prefix”) and the 20 bp-long suf-
fix (key = “suffix”) of the intron.
● The number of input transcripts supporting the intron
(key = “number_of_supporting_transcripts”).
● The set of the input transcripts supporting the intron
(key = “supporting_transcripts”) as a dictionary whose keys are
GenBank identifiers. The value associated to each GenBank
identifier is a dictionary representing an input transcript by
means of:
– The 15 bp-long suffix (key = “donor_factor_suffix”) of the
transcript factor aligned to the donor exon.
– The 15 bp-long prefix (key = “acceptor_factor_prefix”) of
the transcript factor aligned to the acceptor exon.
– The (1-based) start (key = “donor_factor_start”) and end
(key = “donor_factor_end”) positions of the transcript fac-
tor aligned to the donor exon; these endpoints are given
over the transcript itself.
– The (1-based) start (key = “acceptor_factor_start”) and end
(key = “acceptor_factor_end”) positions of the transcript
factor aligned to the acceptor exon; these endpoints are
given over the transcript itself.
● The average transcript-genome alignment error at the donor

(key = “donor_alignment_error”) and at the acceptor
(key = “acceptor_alignment_error”) splice sites (computed on a
15 bp-long suffix of the donor exon and on a 15 bp-long prefix
of the acceptor exon, respectively).
● The intron classification and the splice site scoring in donor
and acceptor according to [9]. Precisely:
– The intron type (key = “type”), where a value of 0 means a
U12 intron, a value of 1 means a U2 intron, and 2 denotes
an unclassified intron.
– The score (key = “BPS_score”) and the position (key = “BPS_
position”) of the Branch Point Sequence (BPS) from the
donor splice site; if the BPS does not exist, then the score
is 0, and the key “BPS_position” is not present.
– The splice site scores in donor (key = “donor_score”) and
acceptor (key = “acceptor_score”).
● The intron pattern (key = “pattern”), given by the first and the
last dinucleotides concatenated into a unique string (the string
“GTAG” denotes a canonical intron).
● When present, the repeat sequence (key = “repeat_sequence”),
that is the sequence giving ambiguities at the intron splice sites
as described in [11].
An isoform (the value related to a key in the dictionary
“isoforms”) is a dictionary reporting the following data:
● The sequence (key = “sequence”) of the isoform and its length
(key = “length”).
● If the isoform is a RefSeq mRNA (key = “from_RefSeq?”);
if that is true, then its GenBank identifier is also given
(key = “RefSeqID”).
● If the isoform is to be considered as a reference (key =
“reference?”).
● A string describing the splicing variants with respect to the
reference isoform (key = “variant_type”). When the isoform
itself is the reference, then this string is “[key = “RefSeqID”]
(Reference TR)”, where [key = “RefSeqID”] is the value (the
null string, possibly) of the key “RefSeqID”.
● If the coding sequence (CDS) has been annotated on the
isoform (key = “annotated_CDS?”); if that is true, then the fol-
lowing data are also given:
– The length of the coding sequence (key = “CDS_length”)
and its (1-based) start (key = “CDS_start”) and end
(key = “CDS_end”) positions on the isoform (stop codon
not included).
– If the CDS starts with the ATG codon (key = “start_

codon?”) (a false value denotes an incomplete CDS).
– If the CDS ends with a stop codon (key = “stop_codon?”)
(a false value denotes an incomplete CDS).
– If the start ATG codon, referred to the genome, is con-
served with respect to the start ATG codon of the reference
isoform (key = “reference_start_codon?”).
– If the stop codon, referred to the genome, is conserved
with respect to the stop codon of the reference isoform
(key = “reference_stop_codon?”).
– The length of the CDS translation (key = “protein_length”).
– If the CDS translation is incomplete (key = “protein_
incomplete?”).
– If the CDS is in frame (key = “reference_frame?”) with
respect to the CDS of the reference isoform.
● If a PolyAdenylation Signal (key = “PAS?”) and a polyA tail
(key = “polyA?”) have been detected on the isoform accord-
ing to [12].
● The introns composing the isoform (key = “introns”), as a list
of the keys of the dictionary “introns”.
● The Nonsense-Mediated Decay flag (key = “NMD_flag”), where
1 means that the CDS has been annotated, the exon contain-
ing the stop codon is not the last one, and its 3′UTR suffix is
longer than 50 bp. Otherwise (that is, the stop codon is not on
the last exon or the 3′UTR exon suffix is at most 50 bp), the
value is 0. When the CDS has not been annotated or the stop
codon is missing, then the NMD flag is set to −1.
● The number of exons composing the isoform
(key = “number_of_exons”).
● The set of exons composing the isoform (key = “exons”), as a
list of exons represented each one by a dictionary giving the
following data:
– The (1-based) start (key = “relative_start”) and end
(key = “relative_end”) positions on the input genomic
sequence.
– The (1-based) start (key = “absolute_start”) and end
(key = “absolute_end”) positions on the reference
chromosome.
– The length of the exon on the genome (key = “length”).
– The length of the exon on the transcript (key = “length_
on_transcript”).
– The exon sequence (key = “sequence”), which is considered
on the genome, when the isoform is not a RefSeq mRNA,
otherwise the sequence is considered on the transcript. In

the latter case, the length on the transcript (key = “length_
on_transcript”) is the actual exon length.
– The length (key = “5UTR_length”) of the 5′UTR exon
prefix (that is, the exon prefix belonging to the 5′UTR
region). A value greater than 0 means that the exon has a
prefix that belongs to the 5′UTR region. In that case, it
is also given the start (key = “absolute_5UTR_start”) and
the end (key = “absolute_5UTR_end”) positions on the
reference chromosome of the 5′UTR exon prefix.
– The length (key = “3UTR_length”) of the 3′UTR exon suf-
fix (that is, the exon suffix belonging to the 3′UTR region).
A value greater than 0 means that the exon has a suffix that
belongs to the 3′UTR region. In that case, it is also given
the start (key = “absolute_3UTR_start”) and the end
(key = “absolute_3UTR_end”) positions on the reference
chromosome of the 3′UTR exon suffix.
– The cumulative length (key = “cumulative_length”)
considered on the genome.
– The cumulative length (key = “cumulative_length_on_tran-
script”) considered on the transcript.
– The start (key = “start_codon_absolute_start”) and the end
(key = “start_codon_absolute_end”) positions, on the ref-
erence chromosome, of the portion of the start codon
(possibly the whole feature) belonging to the exon. This
value exists only if the exon contains the start codon or a
portion of it. In that case, the frame of the start codon
(key = “start_codon_frame”) is also given, and is the num-
ber of the leading start codon bases that are located in the
previous exon(s) of the isoform.
– The start (key = “stop_codon_absolute_start”) and the end
(key = “stop_codon_absolute_end”) positions, on the ref-
erence chromosome, of the portion of the stop codon
(possibly the whole feature) belonging to the exon. This
value exists only if the exon contains the stop codon or a
portion of it. In that case, the frame of the stop codon
(key = “stop_codon_frame”) is also given, and is the num-
ber of leading stop codon bases that are located in the pre-
vious exon(s) of the isoform.
– The start (key = “CDS_absolute_start”) and the end
(key = “CDS_absolute_end”) positions, on the reference
chromosome, of the portion of the coding sequence (pos-
sibly the whole feature) belonging to the exon. This value
exists only if the exon contains the coding sequence or a
portion of it. In that case, the frame of the coding sequence
(key = “CDS_frame”) is also given, and is the number of

the leading CDS bases that are located in the previous
exon(s) of the isoform.
4 Notes
1. The spliced alignment step of the PIntron pipeline has several

parameters that affect the resulting alignments. Their default val-
ues have been chosen to process most of the human genes and it
could be necessary to change them in order to process genes
with peculiar characteristics. The values of these parameters must
be provided in an optional configuration file “config.ini” (placed
in the working directory) with the following format:
<parameter>="<value>"
where <parameter> is the name of a parameter and < value> is
the provided value.
Some significant parameters are:
● min-factor-length, that is the minimum length of the com-
mon transcript-genome substrings used to compute the
alignment (default: 15). In other words, it is the minimum
length of exons expected for the input gene. Notice that
PIntron is able to find exons shorter than this length (down
to 3 nt) but they must be preceded and/or followed by
exons longer than this length. This parameter affects the
computational resources used during the alignment: lower
values for min-factor-length corresponds to increased time
and memory usage. For genes in highly repetitive regions or
with long input transcripts (more than 100 k nucleotides),
it might be necessary to set this parameter to 20 or more.
● min-intron-length, that is the minimum length allowed for
an intron (default: 60).
● max-intron-length, that is the maximum length allowed for
an intron, where 0 means that this parameter is ignored
(default: 0).
An example of row is:
min-factor-length="20"
If PIntron is executed with the option --keep-intermediate-files
(-k), then, after the execution, the file “config-dump.ini” will
contain the values of all the parameters actually used in the
spliced alignment step. The file can be then customized and,
after renaming it to “config.ini”, can be used in subsequent
invocations.
2. PIntron reports a log of its activities in two different files: the
general logfile and the pipeline logfile. The first one, named
“pintron-log.txt” by default, records all the events and

messages produced by the main script, while the second
one, named “pintron-pipeline-log.txt” by default, records all
the messages produced by the programs which compose the
PIntron pipeline. The name (and the path, possibly) of these
files can be changed using the two options --general-logfile and
--logfile at the invocation of the main script “pintron”. The
two logfiles are human-readable and their content (especially
that of the pipeline logfile) could help in identifying the cause
of a runtime error.
3. In case of runtime errors due to the limits imposed on the
computational resources, and in case the system has more
resources than those allowed by default, the user can change
the default limits imposed on the most computationally expen-
sive step—the alignment of the transcripts against the genomic
sequence—using the following execution options when invok-
ing the main script “pintron”:
● --set-max-factorization-time = INT, sets a time limit
(in minutes) for computing the spliced alignments (which is
often the most computational expensive step). The default
value is 60 (i.e., an hour).
● --set-max-factorization-memory = INT, sets a memory
limit (in MiB) for computing the spliced alignments
(which is often the most computational expensive step).
The default value is 3,000 MiB (approximately 3 GB).
References
1. Caceres J, Kornblihtt A (2002) Alternative protein-coding genes in eukaryotic genomic

splicing: multiple control mechanisms and DNA. Genome Biol 7(Suppl 1):S7
involvement in human disease. Trends Genet 7. Kozak M (1986) Point mutations define a
18:186–193 sequence flanking the AUG initiator codon
2. Pirola Y et al (2012) PIntron: a fast method that modulates translation by eukaryotic ribo-
for gene structure prediction via maximal pair- somes. Cell 44:283–292
ings of a pattern and a text. BMC Bioinformatics 8. Bonizzoni P et al (2009) Minimum factoriza-
13:S2 tion agreement of spliced ESTs. Lectures
3. Bonizzoni P et al (2009) Detecting alterna- Notes in Bioinformatics (LNCS). Proceedings
tive gene structures from spliced ESTs: a of 9th workshop on algorithms in bioinfor-
computational approach. J Comput Biol 16: matics WABI, vol 5724. pp 1–12
43–66 9. Sheth N et al (2006) Comprehensive splice-
4. Green RE et al (2003) Widespread predicted site analysis using comparative genomics.
nonsense-mediated mRNA decay of Nucleic Acids Res 34:3955–3967
alternatively-spliced transcripts of human nor- 10. Guigó R et al (2006) EGASP: the human
mal and disease genes. Bioinformatics ENCODE Genome Annotation Assessment
19(S1):i118–i121 Project. Genome Biol 7(Suppl 1):S2
5. Bonizzoni P, Rizzi R, Pesole G (2005) ASPIC: 11. Burset M, Seledtsov I, Solovyev V (2000)
a novel method to predict the exon-intron Analysis of canonical and non-canonical splice
structure of a gene that is optimally compatible sites in mammalian genomes. Nucleic Acids
to a set of transcript sequences. BMC Res 28:4364–4375
Bioinformatics 6:244 12. Zhang H et al (2005) PolyA_DB: a database
6. Djebali S, Delaplace F, Crollius HR (2006) for mammalian mRNA polyadenylation.
Exogean: a framework for annotating Nucleic Acids Res 33(Suppl 1):D116–D120
Chapter 12
Detection of Post-Transcriptional RNA Editing Events

Ernesto Picardi, Anna Maria D’Erchia, Angela Gallo, and Graziano Pesole
Abstract
The advent of deep sequencing technologies has greatly improved the study of complex eukaryotic
genomes and transcriptomes, providing the unique opportunity to investigate posttranscriptional molecu-
lar mechanisms as alternative splicing and RNA editing at single base-pair resolution. RNA editing by
adenosine deamination (A-to-I) is widespread in humans and can lead to a variety of biological effects
depending on the RNA type or the RNA region involved in the editing modification.
Hereafter, we describe an easy and reproducible computational protocol for the identification of
candidate RNA editing sites in human using deep transcriptome (RNA-Seq) and genome (DNA-Seq)
sequencing data.
Key words RNA editing, A-to-I editing, Deep sequencing, Bioinformatics, Genomics, RNA-Seq,
DNA-Seq
1 Introduction
RNA editing is a posttranscriptional mechanism occurring in a

wide range of organisms, affecting primary RNAs through the
insertion/deletion or the modification of specific nucleotide [1].
In humans, the most common RNA editing event is the deamina-
tion of adenosine to inosine (A-to-I) by members of adenosine
deaminase [2] family of enzymes acting on double stranded RNA
(dsRNA) [1]. Editing events in both coding and noncoding tran-
scripts can lead to a variety of biological effects depending on the
RNA type (mRNA or ncRNA) or RNA portion undergo editing
(5′UTR, 3′UTR, CDS, intron) [3]. RNA editing within the 5′-3′-
UTRs are associated to gene expression modulation (preventing
the efficient ribosome binding or the recognition by small regula-
tory RNAs) or perturbed RNA stability, whereas modifications in
coding protein regions may lead to amino acid replacements with
functional consequences.
RNA editing deregulation in human has been associated to
diverse neurological disorders and cancer, attesting its important
189
190 Ernesto Picardi et al.
biological role in preserving the cellular homeostasis [4, 5]. In

principle, the identification of RNA editing sites is straightforward
and based on the direct comparison between transcripts and their
corresponding genomic loci of origin. However, this naïve solu-
tion is not suitable for genome-wide investigations. Alternatively,
massive transcriptome sequencing (RNA-Seq) in combination with
whole genome sequencing (DNA-Seq) from the same individual
(to exclude variations due to SNPs) provide now a novel exciting
opportunity to explore at genome scale the RNA editing landscape
of complex organisms as human [6, 7].
Hereafter, we describe a computational method to detect can-
didate RNA editing sites in human by using matched RNA-Seq
and DNA-Seq data. The methodology is based on specific bioin-
formatic tools implemented in the package REDItools [8], freely
available under the MIT license at http://code.google.com/p/
reditools/.
2 Materials
REDItools are implemented in Python and tested on Linux

(CentOS 6.0) and Mac OS X (10.8).
2.1 Prerequisites Computer (64-bit architecture) running Linux, Mac OS X or other

UNIX-based operating systems with at least 4 GB of RAM and
2.1.1 Hardware
several GB of free disk space.
2.1.2 Short Read Mapper Although a plethora of mapping programs has been released up to
now, we obtained accurate results using GSNAP [9], freely avail-
able for Unix/Linux or Mac OS X systems from http://research-
pub.gene.com/gmap/ (see Note 1 for basic installation on Linux/
Unix systems). Nonetheless, other tools to perform the fast align-
ment of RNA reads onto the reference genome could be used,
even though the selected mapping software may affect the detec-
tion of RNA editing [7]. Note that the mapping program should
print out alignments in the standard SAM format [10].
2.1.3 Mandatory REDItools requires the Python interpreter (at least version 2.6)
Software (see Note 2). In Linux/Unix or Mac OS X it should be prein-
stalled. Alternatively, a copy can be downloaded from http://
www.python.org/download/ and installed according to instruc-
tions reported in the web page. To correctly read alignments in
BAM files, the external python module pysam (at least version
0.6) needs to be installed (available from https://github.com/
pysam-developers/pysam). In addition, SAMtools should be
downloaded and installed from http://samtools.sourceforge.net/
(version ≥0.1.18) [10].
Detection of RNA Editing Sites 191
2.1.4 Optional Software Sometimes optional software could be required to improve final
results. Indeed, mismapping errors, that are quite frequent and
represent the first source of false RNA editing calls, can be miti-
gated aligning reads (only ones carrying mismatches) using Blat
[7]. The complete Blat suite, including gfServer and gfClient exe-
cutables as well as the related documentation, can be obtained
from the following UCSC web page http://hgdownload.cse.ucsc.
edu/admin/exe/.
Duplicated reads due to PCR could be marked in BAM files
before running REDItools, even though this practice is still an
open question (especially at RNA level).
Duplicated reads can be marked using the Picard MarkDuplicate
tool.
$ java -Xmx2g -jar MarkDuplicate.jar INPUT=
myINPUT.bam OUTPUT=myOUTPUT.bam METRICS_FILE=
myMetrics.txt REMOVE_DUPLICATES=false ASSUME_
SORTED=true
The complete Picard package can be downloaded from http://
picard.sourceforge.net (version ≥1.57).
2.2 Installation Once the mandatory software described in Subheading 2.1.3 has

of REDItools been installed, the last version of REDItools package can be down-
loaded from http://code.google.com/p/reditools/.
REDItools include three main scripts:
–– REDItoolDnaRna.py: compares RNA-Seq and DNA-Seq data
from the same sample;
–– REDItoolKnown.py: explores known editing events in RNA-
Seq experiments;
–– REDItoolDenovo.py: predicts de novo RNA editing events
using only RNA-Seq data;
Several accessory scripts to filter, annotate, sort, and convert
REDItool tables are also included in the release.
In Unix/Linux or Mac OS X environment, the package can be
installed in the following way:
$ gunzip reditools-1.0.x.tar.gz
$ tar –xvf reditools-1.0.x.tar
$ cd reditools-1.0.x
$ python setup.py install
To check the installation, download the testing dataset from
http://150.145.82.212/ernesto/reditools/data/testREDItools.
tar.gz and execute the following commands:
$ gunzip testREDItools.tar.gz
$ tar –xvf testREDItools.tar
$ cd testREDItools
$ REDItoolDnaRna.py -i rnaseq.bam -j dnaseq.
bam -f reference.fa -o outputfolder
Alternatively, each REDItool script can be executed independently,

for example:
$ python REDItoolDnaRna.py -i testREDItools/
rnaseq.bam -j testREDItools/dnaseq.bam -f test-
REDItools/reference.fa -o testREDItools/
outputfolder
3 Methods
3.1 Mapping The mapping of RNA-Seq reads can be carried out using GSNAP
of RNA-Seq Reads and providing known splice junctions (from RefSeq, Gencode or
other specialized databases) (see Note 3) to improve the global
alignment process.
The command line for paired-end reads against the human
genome assembly hg19 (see Note 4) is:
$ gsnap -d hg19 -s known-splicesites -E 1000
-n1 -Q -O--nofails -A sam--gunzip--split-output=
outputGsnap -a paired R1.fastq.gz R2.fastq.gz
The option “-a paired” enables a procedure to remove adapt-
ers in paired-end reads (useful in case of not preprocessed data).
The options “-t” and “-B” take into account multiple working
threads and available RAM for allocating genomic indexes, respec-
tively. They speed up the mapping process and should be tuned
according to the hardware running GSNAP (see Note 5). For
human genome, 4 GB of RAM should be sufficient to preload
indices in memory (option “-B 4”).
At the end of the run, GSNAP will create nine separate output
files in the standard SAM format (thanks to the option--plit-
output), one for each alignment type (using specific filename suf-
fixes) (see Note 6). For an accurate RNA editing detection, only
the file with suffix “.concordant_uniq” will be taken into account
for downstream analyses. Indeed, it includes unique and concor-
dant alignments in order to exclude reads mapping to multiple
genomic locations with the same number of mismatches.
3.2 SAM to BAM REDItools accept pre-aligned reads in the standard BAM format
Conversion (see Note 7). For this reason, after the mapping, SAM alignments
need to be converted in the BAM format by means of SAMtools:
$ samtools view -bS outputGsnap.concordant_
uniq > outputGsnap.concordant_uniq.bam
$ samtools sort outputGsnap.concordant_uniq.
bam outputGsnap.concordant_uniq.sorted
$ samtools index outputGsnap.concordant_
uniq.sorted.bam
Optionally, duplicated reads due to PCR can be marked before

the RNA editing calling using the Picard MarkDuplicate tool,
according to the command:
$ java -Xmx2g -jar MarkDuplicate.jar INPUT =
outputGsnap.concordant_uniq.sorted.bam
OUTPUT = outputGsnap.concordant_uniq.sorted.
nodup.bam METRICS_FILE = outputGsnap.concordant_
uniq.sorted.nodup.metrics.txt REMOVE_
DUPLICATES = false ASSUME_SORTED = true
The resulting BAM file should be indexed using SAMtools as
above:
$ samtools index outputGsnap.concordant_uniq.
sorted.nodup.bam
3.3 Blat Correction Although GSNAP is quite accurate in aligning paired-end reads
(Optional) onto the complete human genome [11], several reads could be
ambiguously mapped, leading to false mismatches and, thus,
erroneous RNA editing calls. Misalignment errors can be miti-
gated realigning reads carrying mismatches by the classical Blat
program. The list of problematic reads can be generated using
the REDItoolBlatCorrection.py script and the following com-
mand line:
$ REDItoolBlatCorrection.py -i outputGsnap.
concordant_uniq.sorted.nodup.bam -f hg19.fa -F
hg19.2bit -o BlatCorrection -V -T
At the end of the run, the script will generate a directory
named “BlatCorrection” and, inside, several “.bad” files including
the list of reads prone to mismapping.
3.4 Detection of RNA REDItoolDnaRna.py is the main script to identify RNA editing
Editing Candidates candidates using matched DNA-Seq and RNA-Seq data. In par-
Using Matched ticular, it inspects all genomic regions covered by RNA-Seq reads
DNA-Seq position-by-position looking for nucleotide changes between the
and RNA-Seq Data reference genome and RNA sequences. DNA-Seq data are
employed to support the presence of RNA editing events exclud-
ing potential SNPs or somatic mutations.
REDItoolDnaRna.py requires at least three input files: RNA-
Seq alignments in BAM format, DNA-Seq alignments in BAM for-
mat and the reference genome in FASTA format. All inputs need
to be indexed using SAMtools (see Note 8).
Once all input data are ready, a list of RNA editing candidates
can be obtained with the following command line:
$ REDItoolDnaRna.py –i RNAseq.bam -j DNAseq.
bam -f hg19.fa -o myOUTPUT -b BlatCorrection -c
10,10 -m 20,20 -q 25.25 -u -a6-0 -v 3 -N 0.0
-n 0.1
where RNAseq.bam is the file of aligned RNA-Seq reads using

the above described methodology (Subheadings 3.1 and 3.2),
DNAseq.bam is the file of aligned DNA-Seq reads (see Note 9),
hg19.fa is the reference human genome assembly hg19 in FASTA
format, BlatCorrection is the optional directory containing prob-
lematic reads (obtained using the REDItoolBlatCorrection.py
script as suggested in Subheading 3.3) and myOUTPUT is the
directory in which results will be stored.
According to the above instance, REDItoolDnaRna.py will
explore all genomic positions supported by at least 10 RNA-Seq
reads and 10 DNA-Seq reads, filtering out bases with quality score
lower than 25, reads with map quality lower than 20, sites in the
first six bases of each read and positions with less than 3 indepen-
dent reads supporting the RNA editing event.
REDItoolDnaRna.py can infer the strand of each position when
strand oriented RNA-Seq reads are provided (see Note 10) [7].
A detailed description of available parameters and options is in
Table 1.
3.5 Exploring Known REDItoolKnown.py script has been developed to explore the
RNA Editing Sites impact of RNA editing on a given RNA-Seq dataset using known
in RNA-Seq Data editing events annotated in available databases as DARNED [12]
and RADAR [13] or collected from supplementary materials of a
variety of publications. REDitoolKnown.py requires at least three
input files: RNA-Seq alignments in BAM format, the reference
genome in FASTA format and a list of known RNA editing sites
(see Note 11).
Once all input data are ready, RNA editing can be explored by
means of the following command line:
$ REDitoolKnown.py -i RNAseq.bam -f hg19.fa
-l known_editing_sites.txt.gz -o myOUTPUT
where RNAseq.bam is the file of aligned RNA-Seq reads gen-
erated by the methodology described in Subheadings 3.1 and 3.2,
hg19.fa is the reference human genome assembly hg19 in FASTA
format, known_editing_sites.txt.gz is the table file containing
known RNA editing positions and myOUTPUT is the directory
including all output files.
A detailed description of available parameters and options is in
Table 2.
3.6 Output Each execution of REDItoolDnaRna.py or REDItoolKnown.py

creates an output folder containing a second folder in which scripts
store results. The main output is a simple textual table called out-
Table_XXXX where XXXX is a random number generated at each
run. The resulting REDItool table comprises the following fields:
–– Region: is the genomic region according to the reference genome.
–– Position: is the exact genomic coordinate (1-based).
Table 1
Parameters required for REDItoolDnaRna.py
Parameter Description
-i RNA-Seq BAM file
-j DNA-Seq BAM files separated by comma or folder containing BAM files. Note that
each chromosome must be present in a single BAM file only
-I Sort input RNA-Seq BAM file
-J Sort input DNA-Seq BAM file
-f Reference genome in fasta format. Note that chromosome names must match
chromosome names in BAMs files
-C It indicates how many bases to load in RAM during the execution [100,000 by default]
-k List of chromosomes to skip separated by comma. A file in which each line contains a
chromosome name can also be provided
-t It indicates how many processes should be launched [1 by default]
-o Output folder in which all results will be stored [rediFolder_XXXX by default in which
XXXX is a random number generated at each run]
-F It indicates the name of the internal folder containing output tables
-M If selected, pileup-like files are generated
-c Minimum read coverage for DNA and RNA reads, respectively (dna,rna) [10,10 by
default]
-Q Fastq offset value (dna,rna) [33,33 by default]. For Illumina FASTQ 1.3+, 64 should
be used
-q Minimum quality score for DNA and RNA reads, respectively (dna,rna) [25,25 by
deafult]. This option can influence the sequencing depth but is very useful to
exclude sequencing errors
-m Minimum mapping quality score for DNA and RNA reads, respectively (dna,rna)
[25,25 by default]. This parameter can mitigate errors due to misplaced reads
-O Minimum homoplymeric length for DNA and RNA reads, respectively (dna,rna) [5,5
by default]. Homopolymeric regions may confound alignment tools leading to
misalignments. As a consequence, substitutions occurring in such regions of a
predefined length (generally equal to or higher than five bases) should be excluded
-s For strand oriented RNA-Seq reads, it indicates which read has the orientation of the
RNA. Available values are: 1 for read1 in line with RNA; 2 for read2 in line with
RNA. Option 1 is equivalent to fr-secondstrand library, whereas option 2
corresponds to fr-firststrand library. The strand information is essential to exclude
substitutions occurring in antisense RNAs
-g It specifies how strand should be deduced. Valid options are: 1 maxValue and 2
useConfidence (the strand is assigned if over a prefixed frequency set by -x option)
-x Strand confidence value [0.70 by default]. It is used if -g 2 option is selected
-S If selected, it performs the strand correction. Once the strand has been inferred, only
bases according to this strand will be printed out
(continued)
Table 1
(continued)
-G Infer strand by GFF annotation (must be in GFF and sorted, otherwise the -X option
should be used). Sorting requires grep and sort unix executables
-K GFF File with positions to exclude (must be in GFF and sorted, otherwise the -X
option should be used). Sorting requires grep and sort unix executables
-T Work only on given GFF positions (must be in GFF and sorted, otherwise the -X
option should be used). Sorting requires grep and sort unix executables
-X Sort annotation files. It requires grep and sort unix executables
-e Exclude multi hits in RNA-Seq
-E Exclude multi hits in DNA-Seq
-d Exclude duplicated reads in RNA-Seq. PCR duplicates need to be marked in the input
BAM
-D Exclude duplicated reads in DNA-Seq. PCR duplicates need to be marked in the input
BAM
-p Use paired concordant reads only in RNA-Seq. It should be activated in case of paired
end reads
-P Use paired concordant reads only in DNA-Seq. It should be activated in case of paired
end reads
-u Consider mapping quality in RNA-Seq
-U Consider mapping quality in DNA-Seq
-a Trim x bases up and y bases down per read [0–0] in RNA-Seq
-A Trim x bases up and y bases down per read [0–0] in DNA-Seq
-b Blat folder containing problematic reads in RNA-Seq
-B Blat folder containing problematic reads in DNA-Seq
-l Remove substitutions in homopolymeric regions in RNA-Seq
-L Remove substitutions in homopolymeric regions in DNA-Seq
-v Minimum number of reads supporting the variation for RNA-Seq [3 by default]
-n Minimum editing frequency for RNA-Seq [0.1 by default]
-N Minimum variation frequency for DNA-Seq [0.1 by default]
-z Exclude positions with multiple changes in RNA-Seq
-Z Exclude positions with multiple changes in DNA-Seq
-W Select RNA-Seq positions with defined changes (separated by comma, for example:
AG,TC) [all by default]
-R Exclude invariant RNA-Seq positions
-V Exclude sites not supported by DNA-Seq
(continued)
Table 1
(continued)
-w File containing splice sites annotations
-r Number of bases near splice sites to explore [four by deafult]
--gzip Gzip output files
-h, −-help Print out these options
Table 2
Parameters required for REDItoolKnown.py
-i RNA-Seq BAM file
-I Sort input BAM file
-f Reference genome in FASTA format. Note that chromosome names must match
chromosome names in the BAM file
-l List of known RNA editing events (see Note 11)
-C It indicates how many bases to load in RAM during the execution [100,000 by default]
-k List of chromosomes to skip separated by comma. A file in which each line contains a
chromosome name can also be provided
-t It indicates how many processes should be launched [1 by default]
-o Output folder in which all results will be stored [rediFolder_XXXX by default in which
XXXX is a random number generated at each run]
-F It indicates the name of the internal folder containing output tables
-c Minimum read coverage for RNA reads [10 by default]
-Q Fastq offset value [33 by default]. For Illumina FASTQ 1.3+, 64 should be used
-q Minimum quality score for RNA reads [25 by default]
-m Minimum mapping quality score for RNA reads [25 by default]
-O Minimum homoplymeric length [5 by default]
-s For strand oriented RNA-Seq reads, it indicates which read has the orientation of the
RNA. Available values are: 1 for read1 in line with RNA; 2 for read2 in line with
RNA. Option 1 is equivalent to fr-secondstrand library, whereas option 2 corresponds
to fr-firststrand library. The strand information is essential to exclude substitutions
occurring in antisense RNAs
-g It specifies how strand should be deduced. Valid options are: 1 maxValue and 2
useConfidence (the strand is assigned if over a prefixed frequency set by -x option)
-x Strand confidence value [0.70 by default]. It is used if -g 2 option is selected
(continued)
Table 2
(continued)
-S If selected, it performs the strand correction. Once the strand has been inferred, only
bases according to this strand will be printed out
-G Infer strand by GFF annotation (must be in GFF and sorted, otherwise the -X option
should be used). Sorting requires grep and sort unix executables
-X Sort annotation files. It requires grep and sort unix executables
-K File with positions to exclude in the format chromosome name[TAB or space]coordinate
-e Exclude multi hits
-d Exclude duplicated reads in RNA-Seq. PCR duplicates need to be marked in the
input BAM
-p Use paired concordant reads only in RNA-Seq. It should be activated in case
of paired end reads
-u Consider mapping quality
-T Trim x bases up and y bases down per RNA read [0–0]
-B Blat folder containing problematic reads in RNA-Seq
-U Remove substitutions in homopolymeric regions
-v Minimum number of reads supporting the variation for RNA-Seq [3 by default].
-n Minimum editing frequency [0.1 by default]
-E Exclude positions with multiple changes
-P File containing splice sites annotations
-r Number of bases near splice sites to explore [4 by default]
-h Print out these options
–– Reference: is the nucleotide base in the reference genome.

–– Strand: is the strand info with notation 1 for + strand, 0 for −
strand and 2 for unknown or not defined strand.
–– Coverage-qxx: is the depth per site at a given xx quality score.
–– MeanQ: is the mean quality score per site.
–– BaseCount[A,C,G,T]: is the base distribution per site in the
order A, C, G, and T.
–– AllSubs: is the list of observed substitutions at a given site,
separated by a space. A character “-” is included in case of
invariant sites.
–– Frequency: is the observed frequency of substitution. In case of
multiple substitutions, it refers to the first in the AllSubs field.
REDItoolDnaRna.py includes five additional columns to take

into account information from DNA-Seq reads. Such columns are
indicated as:
–– gCoverage-qxx: is the depth per site at a given xx quality score
in DNA-Seq.
–– gMeanQ: is the mean quality score per site in DNA-Seq.
–– gBaseCount[A,C,G,T]: is the base distribution per site in the
order A, C, G, and T.
–– gAllSubs: is the list of observed substitutions at a given site,
separated by a space. A character “-” is included in case of
invariant sites.
–– gFrequency: is the observed frequency of substitution. In case of
multiple substitutions, it refers to the first in the gAllSubs field.
An example of output table is shown in Table 3.
The output folder includes also a parameters.txt file that con-
tains all selected options as well as the main command line.
3.7 Downstream All output positions reported in REDItool tables can be sorted,
Analyses filtered, and annotated using auxiliary scripts provided in the pack-
age. The complete list of available accessory scripts can be found at
the web page http://code.google.com/p/reditools/.
An important issue in RNA editing detection is the exclusion
of SNP sites also in case of DNA-Seq dataset. Indeed, positions
well supported by RNA-Seq reads could not be equally sup-
ported by DNA-Seq reads and, thus, RNA editing candidates
could be genuine SNPs. REDItools include the accessory
FilterTable.py script that is able to filter out known SNPs as well
as other preferred positions according to specialized databases as
dbSNP (see Note 12). A typical command line is:
$ FilterTable.py -i myREDItoolTable -s dbsnp.
gff.gz -S snp -E -p -o myREDItoolTable.nosnp
where myREDItoolTable is the output table from a REDItool
script, dbsnp.gff.gz is the file containing genomic SNPs in GFF
format (bgzipped using bgzip from SAMtools) and myREDI-
toolTable.nosnp is the final table. Option -S indicates the feature
to filter out (according to GFF file), −E prevents the printing of
filtered positions and -p reports simple statistics in the standard
output.
FilterTable.py can also be used to filter out (or in) editing can-
didates in repetitive regions, as Alu rich regions, employing anno-
tations stored in the RepeatMask table from UCSC. In this case, a
sample command line is:
$ FilterTable.py -i myREDItoolTable -s dbsnp.
gff.gz -F alu -E -p -o myREDItoolTable.alu
Table 3
Example of a REDItool output table
BaseCount gCoverage- gBaseCount

Region Position Reference Strand Coverage-q25 MeanQ [A,C,G,T] AllSubs Frequency q25 gMeanQ [A,C,G, T] gAllSubs gFrequency
chr21 15412990 A 1 18 37.22 [3, 0, 15, 0] AG 0.83 18 37.22 [3, 0, 15, 0] AG 0.83
chr21 15415901 A 1 13 37.15 [2, 0, 11, 0] AG 0.85 13 37.15 [2, 0, 11, 0] AG 0.85
chr21 15423330 A 1 11 38.27 [4, 0, 7, 0] AG 0.64 11 38.27 [4, 0, 7, 0] AG 0.64
chr21 15425640 A 1 8 36.12 [0, 0, 8, 0] AG 1.00 8 36.12 [0, 0, 8, 0] AG 1.00
chr21 15456434 T 1 90 34.96 [0, 6, 1, 83] TC TG 0.07 90 34.96 [0, 6, 1, 83] TC TG 0.07
chr21 15461406 A 1 83 37.27 [73, 0, 10, 0] AG 0.12 83 37.27 [73, 0, 10, 0] AG 0.12
chr21 15461417 A 1 90 36.26 [72, 0, 18, 0] AG 0.20 90 36.26 [72, 0, 18, 0] AG 0.20
chr21 15461444 A 1 64 37.22 [26, 0, 38, 0] AG 0.59 64 37.22 [26, 0, 38, 0] AG 0.59
chr21 15461479 A 1 70 36.96 [66, 0, 4, 0] AG 0.06 70 36.96 [66, 0, 4, 0] AG 0.06
chr21 15461486 A 1 68 37.06 [61, 0, 7, 0] AG 0.10 68 37.06 [61, 0, 7, 0] AG 0.10
chr21 15461503 A 1 76 37.26 [69, 0, 7, 0] AG 0.09 76 37.26 [69, 0, 7, 0] AG 0.09
chr21 15461511 A 1 81 37.68 [55, 0, 26, 0] AG 0.32 81 37.68 [55, 0, 26, 0] AG 0.32
Individual positions of REDItool tables can be annotated

using the accessory AnnotateTable.py script and a favorite database
as RefSeq or UCSC Known genes or Gencode (see Note 12).
AnnotateTable.py script is quite useful to understand the gene and
genomic context of detected RNA editing events. It can be
launched with the following command line:
$ AnnotateTable.py -a refgene.gtf.gz -i
myREDItoolTable -u -n RefGene -o myREDItoolT-
able.refgene
where myREDItoolTable is the output table from a REDItool
script, refgene.gtf.gz is the file containing RefSeq annotations in
GTF format (bgzipped using bgzip from SAMtools) and myREDI-
toolTable.refgene is the final annotated table.
Additional REDItool scripts can be used to format annota-
tions, search in REDItool tables and select specific positions
according to different criteria as the RNA-Seq or DNA-Seq cover-
age and the number of bases supporting the RNA editing change
as well as the RNA editing frequency. Detailed documentation for
accessory REDItool scripts is available at http://code.google.
com/p/reditools/ and http://150.145.82.212/ernesto/redi-
tools/doc.
4 Notes
1. The GSNAP installation on Unix/Linux systems is quite sim-

ple. Download the latest version from http://research-pub.
gene.com/gmap/ and execute the following commands:
$ gunzip gmap-gsnap-2014-02-28.tar.gz
$ tar –xvf gmap-gsnap-2014-02-28.tar
$ cd gmap-gsnap-2014-02-28
$ ./configure
$ make
$ make check
$ make install
Executables will be installed in /usr/local/bin, whereas
genome databases in /usr/local/share. If you don’t have root
privileges, you can install GSNAP and genome databases in
your $HOME using the command:
$ ./configure--prefix=/your/path--with-gmapdb=
/path/to/gmapdb
If you plan to use reads longer than 250 bases as those gen-
erated by the MiSeq Illumina sequencer, you need to define
the MAX_READLENGTH variable at the compile time.
$ ./configure MAX_READLENGTH = 300
2. REDItools are designed to work with python versions 2.6 and

superior. Python 3.x is not yet supported. You can check the
python installation on your computer as well as the release ver-
sion using the command:
$ python--version
If correctly installed you will see something like:
Python 2.7.2
3. A list of known splice sites can be obtained from RefSeq or
UCSC gene annotations (also Gencode annotations are a good
source of splice sites, http://www.gencodegenes.org/). For
human RefSeq splice sites, first download the annotation table
from UCSC according to the right genome assembly. If you
are aligning to genome hg19, you can retrieve the table from
ftp://hgdownload.cse.ucsc.edu/goldenPath/hg19/database/
refGene.txt.gz. This track can be handled using the programs
psl_splicesites and iit_store, included in the GSNAP-GMAP
package.
$ gunzip -c refGene.txt.gz | psl_splicesites
-s 1 > my.splicesites
$ cat my.splicesites | iit_store -o
splicesites
In order to be used, this file (that will have the .iit exten-
sion) has to be moved into the maps subdirectory of your
gmapdb genome by doing:
$ cp splicesites.iit /path/to/gmapdb/
<genome>/<genome > .maps
Splice sites can be included in every GSNAP run through
the -s flag. For example:
$ gsnap -d < genome > −s splicesites
4. GSNAP indexes for human genome hg19 can be generated
using the utility gmap_build included in the GSNAP-GMAP
package. The human genome hg19 can be downloaded from
UCSC http://hgdownload.soe.ucsc.edu/goldenPath/hg19/
bigZips/chromFa.tar.gz.
Indices can be built by doing:
$ gunzip chromFa.tar.gz
$ tar –xvf chromFa.tar
$ cd chromFa
$ cat *.fa > hg19.fa
$ gmap_build -d hg19 -c chrM hg19.fa
A folder called hg19 will be created in your gmapdb direc-
tory defined at the compiling time.
5. If GSNAP is running on a computer with multicore architec-
ture, you can speed up the mapping process using the -t flag
followed by the number of required threads. In a computer
with a quadcore CPU, for instance, you can run GSNAP using
-t 2, in order to reserve two threads for the mapping.
6. In GSNAP, the--plit-output option creates separate output

files for each output type, adding specific suffixes. For example,
the suffix “.concordant_uniq” refers to paired-end reads align-
ing concordantly to a single position in the genome. The com-
plete list of suffixes is available in the README file of GSNAP
(http://research-pub.gene.com/gmap/src/README).
7. The BAM format is the binary version of the SAM format.
More details about SAM and BAM formats can be found at
http://samtools.sourceforge.net/.
8. BAM and FASTA files can be indexed by SAMtools:
$ samtools index myBAM.bam
$ samtools faidx myFASTA.fa
Note that BAM files need to be sorted before the
indexing:
$ samtools sort myBAM.bam myBAM.sorted
9. DNA-Seq reads can be mapped using again GSNAP but
excluding -s and -N flags that are specific for RNA-Seq data.
10. Current RNA-Seq libraries protocols can generate strand ori-
ented and non strand-oriented RNA-Seq reads. Strand ori-
ented reads can greatly improve the detection of RNA editing,
mitigating the effect of antisense RNA. According to the pro-
tocol used to generate the RNA-Seq library, there are two
main scenarios: (1) the left-most end of the fragment (in tran-
script coordinates) is the first sequenced (or only sequenced
for single-end reads); (2) the right-most end of the fragment
(in transcript coordinates) is the first sequenced (or only
sequenced for single-end reads).
REDItools can take into account the strand through -s flag. A
value of 1 corresponds to the first scenario. On the contrary, a
value of 2 is equivalent to the second scenario.
11. Known RNA editing sites have to be provided in simple textual
files containing three tabulated columns: (1) the genomic
region, (2) the coordinate (1-based) and the strand (+ or −).
Editing tables have to be sorted and indexed by tabix before
their use in REDItools. In Linux/Unix environments, the
sorting can be easily done by the command:
$ sort -k1,1 -k2,2n mytable.txt >
mytable.
sorted.txt
$ bgzip mytable.sorted.txt
$ tabix -b 2 -e 2 mytable.sorted.txt.gz
Alternatively, editing tables can be sorted and converted in
the right format for REDItools using the auxiliary script table-
ToTabix.py:
$ tableToTabix.py -i mytable.txt -c 2 -e 2
Known RNA editing sites can be downloaded for human

genome hg19 assembly from DARNED (http://beamish.ucc.
ie/data_hg19.txt) or RADAR (http://www.stanford.
edu/~gokulr/database/Human_AG_all_hg19.txt) databases.
12. REDItools can use annotations in GTF format that includes
nine tab-delimited fields:
–– Chromosome name.
–– Source (for example hg19_refseq).
–– Feature (for example CDS, exon, snp…).
–– Start coordinate (1-based).
–– End coordinate (1-based).
–– Score (use a dot if unknown).
–– Strand (+ or −, use + if unknown).
–– Frame for CDS only (a dot if unknown).
–– Attributes. There are two mandatory attributes: gene_id
and transcript_id separated by semicolons. For example:
gene_id “C7orf55”; transcript_id “NM_197964”.
Before their use in REDItools, annotations in GTF have to be
sorted and indexed using the accessory script GFFtoTabix.py:
$ GFFtoTabix.py -i myAnnotation.gtf
Further details are in http://150.145.82.212/ernesto/redi-
tools/doc/.
Acknowledgments
This work was supported by the Italian Ministero dell’Istruzione,

Università e Ricerca (MIUR): PRIN 2009 and 2010; Consiglio
Nazionale delle Ricerche: Flagship Project Epigen, Medicina
Personalizzata and Aging Program 2012–2014. It was also sup-
ported by the Italian Ministry for Foreign Affairs (Italy–Israel
actions) to E.P. and AIRC to A.G.
References
1. Gott JM, Emeson RB (2000) Functions and Brent M, Prange C, Schreiber K, Shapiro N,
mechanisms of RNA editing. Annu Rev Genet Bhat NK, Hopkins RF, Hsie F, Driscoll T,
34:499–531 Soares MB, Casavant TL, Scheetz TE, Brown-
2. Gerhard DS, Wagner L, Feingold EA, Shenmen stein MJ, Usdin TB, Toshiyuki S, Carninci P,
CM, Grouse LH, Schuler G, Klein SL, Old S, Piao Y, Dudekula DB, Ko MS, Kawakami K,
Rasooly R, Good P, Guyer M, Peck AM, Derge Suzuki Y, Sugano S, Gruber CE, Smith MR,
JG, Lipman D, Collins FS, Jang W, Sherry S, Simmons B, Moore T, Waterman R, Johnson
Feolo M, Misquitta L, Lee E, Rotmistrovsky K, SL, Ruan Y, Wei CL, Mathavan S, Gunaratne
Greenhut SF, Schaefer CF, Buetow K, Bonner PH, Wu J, Garcia AM, Hulyk SW, Fuh E, Yuan
TI, Haussler D, Kent J, Kiekhaus M, Furey T, Y, Sneed A, Kowis C, Hodgson A, Muzny DM,
McPherson J, Gibbs RA, Fahey J, Helton E, of human Alu and non-Alu RNA editing sites.
Ketteman M, Madan A, Rodrigues S, Sanchez Nat Methods 9(6):579–581. doi:10.1038/
A, Whiting M, Madari A, Young AC, Wetherby nmeth.1982
KD, Granite SJ, Kwong PN, Brinkley CP, 8. Picardi E, Pesole G (2013) REDItools: high-
Pearson RL, Bouffard GG, Blakesly RW, Green throughput RNA editing detection made easy.
ED, Dickson MC, Rodriguez AC, Grimwood Bioinformatics 29(14):1813–1814.
J, Schmutz J, Myers RM, Butterfield YS, doi:10.1093/bioinformatics/btt287
Griffith M, Griffith OL, Krzywinski MI, Liao
9. Wu TD, Nacu S (2011) Fast and SNP-tolerant
N, Morin R, Palmquist D, Petrescu AS, Skalska
detection of complex variants and splicing in
U, Smailus DE, Stott JM, Schnerch A, Schein
short reads. Bioinformatics 26(7):873–881
JE, Jones SJ, Holt RA, Baross A, Marra MA,
Clifton S, Makowski KA, Bosak S, Malek J 10. Li H, Handsaker B, Wysoker A, Fennell T,
(2004) The status, quality, and expansion of Ruan J, Homer N, Marth G, Abecasis G,
the NIH full-length cDNA project: the Durbin R (2009) The sequence alignment/
Mammalian Gene Collection (MGC). Genome map format and SAMtools. Bioinformatics
Res 14(10B):2121–2127 25(16):2078–2079
3. Keegan LP, Gallo A, O’Connell MA (2001) 11. Engstrom PG, Steijger T, Sipos B, Grant GR,
The many roles of an RNA editor. Nat Rev Kahles A, Alioto T, Behr J, Bertone P, Bohnert
Genet 2(11):869–878. doi:10.1038/35098584 R, Campagna D, Davis CA, Dobin A, Gingeras
4. Silberberg G, Lundin D, Navon R, Ohman M TR, Goldman N, Guigo R, Harrow J, Hubbard
(2012) Deregulation of the A-to-I RNA edit- TJ, Jean G, Kosarev P, Li S, Liu J, Mason CE,
ing mechanism in psychiatric disorders. Hum Molodtsov V, Ning Z, Ponstingl H, Prins JF,
Mol Genet 21(2):311–321. doi:10.1093/ Ratsch G, Ribeca P, Seledtsov I, Solovyev V,
hmg/ddr461 Valle G, Vitulo N, Wang K, Wu TD, Zeller G
(2013) Systematic evaluation of spliced align-
5. Gallo A, Locatelli F (2011) ADARs: allies or
ment programs for RNA-seq data. Nat Methods
enemies? The importance of A-to-I RNA edit-
10(12):1185–1191. doi:10.1038/nmeth.2722
ing in human disease—from cancer to HIV-1.
Biol Rev Camb Philos Soc 87(1):95–110 12. Kiran A, Baranov PV (2010) DARNED: a
6. Picardi E, Horner DS, Chiara M, Schiavon R, DAtabase of RNa EDiting in humans.
Valle G, Pesole G (2010) Large-scale detection Bioinformatics 26(14):1772–1776. doi:10.1093/
and analysis of RNA editing in grape mtDNA bioinformatics/btq285
by RNA deep-sequencing. Nucleic Acids Res 13. Ramaswami G, Li JB (2014) RADAR: a rigor-
38(14):4755–4767. doi:10.1093/nar/gkq202 ously annotated database of A-to-I RNA edit-
7. Ramaswami G, Lin W, Piskol R, Tan MH, ing. Nucleic Acids Res 42(Database
Davis C, Li JB (2012) Accurate identification issue):D109–D113. doi:10.1093/nar/gkt996
Chapter 13
Prediction of miRNA Targets

Anastasis Oulas, Nestoras Karathanasis, Annita Louloupi,
Georgios A. Pavlopoulos, Panayiota Poirazi, Kriton Kalantidis,
and Ioannis Iliopoulos
Abstract
Computational methods for miRNA target prediction are currently undergoing extensive review and
evaluation. There is still a great need for improvement of these tools and bioinformatics approaches are
looking towards high-throughput experiments in order to validate predictions. The combination of large-
scale techniques with computational tools will not only provide greater credence to computational predic-
tions but also lead to the better understanding of specific biological questions. Current miRNA target
prediction tools utilize probabilistic learning algorithms, machine learning methods and even empirical
biologically defined rules in order to build models based on experimentally verified miRNA targets. Large-
scale protein downregulation assays and next-generation sequencing (NGS) are now being used to validate
methodologies and compare the performance of existing tools. Tools that exhibit greater correlation
between computational predictions and protein downregulation or RNA downregulation are considered
the state of the art. Moreover, efficiency in prediction of miRNA targets that are concurrently verified
experimentally provides additional validity to computational predictions and further highlights the com-
petitive advantage of specific tools and their efficacy in extracting biologically significant results. In this
review paper, we discuss the computational methods for miRNA target prediction and provide a detailed
comparison of methodologies and features utilized by each specific tool. Moreover, we provide an over-
view of current state-of-the-art high-throughput methods used in miRNA target prediction.
Key words MiRNA target prediction, Computational tools, Databases, High-throughput methods,
Biological features of miRNA target recognition
1 Introduction
1.1 State of the Art MicroRNAs (miRNAs) belong to a recently identified group of the
large family of noncoding RNAs [1]. The mature miRNA is usually
19–27 nt long and is derived from a larger precursor that folds into
an imperfect stem-loop structure. The mode of action of the
mature miRNA in mammalian systems is dependent on com-
plementary base pairing primarily to the 3′-UTR region of the
target mRNA, thereafter causing the inhibition of translation and/
or the degradation of the mRNA.
207
208 Anastasis Oulas et al.
Searching through all human genes (~25,000) and/or other

species for novel miRNA gene targets is a complicated task for
which fast, flexible, and reliable identification methods are required.
Currently available experimental approaches working towards this
goal are complex and sub-optimal [2]. Inefficiencies result from
various sources, including difficulty in isolating certain miRNAs by
cloning due to low expression, stability, tissue specificity, and tech-
nical difficulties of the cloning and repression assay procedures,
while selecting the right 3′UTR to investigate is often a challeng-
ing task of its own. Computational prediction of miRNA gene
targets from 3′UTR genomic sequences is an alternative technique
which offers a much faster, cheaper and effective way of identifying
putative miRNA gene targets. Moreover, by predicting the loca-
tion of a miRNA gene target, these methods enable experimental
biologists to concentrate their efforts on genomic regions more
likely to contain novel genes that undergo miRNA regulation, thus
facilitating the discovery process.
Due to the lack of negative data in this specific biological prob-
lem, the performance of current miRNA target prediction tools is
largely dependent on the overall number of predicted targets.
Some tools are very efficient in predicting true target sites (high
sensitivity), but at the same time display an extremely large number
of overall predictions (low specificity) [3–6]. In contrast, other
tools display an overall high specificity and a relatively low sensitiv-
ity [7–9]. In order to provide an estimation of a false positive rate,
false or mock miRNAs are often generated by randomly shuffling
the nucleotide sequence of experimentally supported miRNAs [10].
Performing target prediction with these “mock” miRNAs can pro-
vide an estimation of the overall false positive rate of a miRNA
target prediction tool.
Accurate prediction of novel miRNA gene targets requires the
consideration of certain characteristic properties of the miRNA–
target mRNA interaction. These properties are based on either
experimental [11–14], or computational evidence [15–19] and
can be used to build a classification scheme or predictive model.
For example, the foremost nucleotides at the 5′region of a mature
miRNA sequence are considered crucial for recognizing and binding
to the target mRNA. Initial research performed by Kiriakidou et al.
[20] has shown that almost consecutive complementarity of the
first 9 miRNA nucleotides to the 3′UTR of protein coding genes
is a prerequisite for translational repression. Moreover, Lewis et al.
[7] showed that complementary motifs to nucleotides 2–7 of
miRNA (commonly referred to as the miRNA seed region) remain
preferentially conserved in several species in a statistically signifi-
cant manner [21, 22].
In general, it is believed that binding of at least seven consecu-
tive Watson–Crick (WC) base pairing nucleotides between the
miRNA Targets Prediction 209
foremost 5′region of the miRNA and the mRNA target is required

for sufficient repression of protein production [7, 20].
Based on the above mentioned evidence, miRNA target pre-
diction programs rely heavily on sequence complementarity of the
miRNA seed region (nucleotides 2–7) to the 3′UTR sequences of
candidate target genes for identifying putative miRNA binding
sites [8, 23]. Furthermore, most prediction tools make use of
thermodynamics and evolutionary conservation at the binding
site, in order to minimize false positives (increase specificity) [8,
23, 24]. Some tools utilize additional features such as binding site
structural accessibility [4, 25, 26], nucleotide composition flanking
the binding sites [27, 28] or proximity of one binding site to
another within the same 3′ UTR [27, 29].
In summary, the general features employed for miRNA target
prediction are: (1) sequence complementarily at the 5′region of
the mature miRNA, better known as the seed region and com-
monly characterized by nucleotides 2–7, (2) secondary structure of
the miRNA–target mRNA hybrid molecule and the overall ther-
modynamics of the interaction expressed in free energy (ΔG) and
(3) species conservation observed via the use of full genome
sequence alignments.
In addition to computational tools, large-scale, high-throughput
transcriptomic and proteomic methods, such as microarrays and
pSILAC, have recently been used, often in conjunction with
computational tools, for the identification of novel miRNA gene
targets [6, 30]. These methods are particularly useful as they can
provide accurate protein repression data or gene expression data
that may be correlated or anti-correlated to miRNA expression.
Moreover, if such data is coupled to computational tools, it can
facilitate rapid and precise detection of novel miRNA gene targets,
while at the same time giving greater credence to computational
predictions. One must keep in mind that such proteomic data
may only provide indirect evidence for target genes, as the signal
obtained may be due to downstream effects and not a consequence
of a direct interaction between miRNA and predicted target gene.
Next-Generation Sequencing (NGS) methods have also been
used for the prediction of miRNA genes, their mature sequences
[31] and their downstream targets. One drawback of these tech-
niques is that multiple small RNA sequences are often missed due
to technical difficulties of the sequencing methodology, such as
library construction. Moreover, not all small RNA and under
expressed mRNA sequences detected by NGS methods are true
miRNAs and targets respectively, unless, some physical interaction
between the two sequences can be attributed to them. Experimental
verification of miRNA targets can be achieved via the use of lucif-
erase assays whereby the miRNA is expressed in vitro, while simul-
taneously expressing and monitoring the target messenger RNA
linked to a luciferase reporter gene [32–34]. This assay provides an
experimental verification of a direct interaction between the mature

miRNA and the target gene and furthermore provides evidence
that regulation is mediated via the miRNA silencing pathway.
However, the extent to which this interaction takes place in the
intact system in vivo cannot be inferred from luciferase assays alone.
Currently, miRNA target prediction tools are freely available
and are commonly built on sophisticated algorithms (i.e., machine
learning) trained to recognize certain biological features of
miRNA–target mRNA interactions. Validation of computational
methodologies is achieved using protein repression information
from large-scale proteomic studies (i.e., pSILAC) [30] as well as
experimentally verified miRNA gene targets form online databases,
like Tarbase (v5) [35]. Results from these analyses are used to
compare prediction accuracy of existing target prediction tools. It
is rare that target prediction tools are assessed for their ability to
identify de novo biologically significant interactions. This can be
achieved by directing predictions and high-throughput experi-
ments towards answering specific biological questions [36]. Such
approaches can provide raw material for experimental biologists to
investigate interesting biological questions, such as the molecular
basis of a disease like cancer.
1.2 The Biology As mentioned earlier, the mode of action of the mature miRNA in
of miRNA Target mammalian systems is dependent on complementary base pairing
Prediction to the 3′UTR region of the target mRNA, thereafter causing
the inhibition of translation and/or the degradation of the
1.2.1 Background mRNA. Despite the body of evidence supporting a role of miR-
NAs in cancer, their exact mechanism of action remains to be
elucidated because the downstream targets of miRNAs implicated
in cancer have yet to be defined. Towards this goal, miRNA target
prediction tools can offer a first indication as to which target genes
are regulated by miRNAs, thus providing new insights regarding
their specific functions and guiding future experiments.
1.2.2 Biological Features Early work on the principles of microRNA target recognition was
of miRNA–mRNA conducted by Julius Brennecke et al. [14]. Based on experimental
Interactions data, they separate the target sites on two classes based on their
base pairing motifs, “5′ dominant” which are sites that depend
critically on the pairing to the miRNA 5′ end and “3′ compensa-
tory” which include seed matches of four to six base pairs and
seeds of seven to eight bases that contain G–U base pairs, single
nucleotide bulges, or mismatches and a extensive pairing to the
3′end of the miRNA.
Additionally, they distinguish two subgroups of 5′ dominant
sites, canonical and seed. Canonical sites have good pairing to both
5′ and 3′ ends of the miRNA. Their repression efficiency was not
affected by a mismatch at position 1, 9, 10 or at the 3′ region, but
any mismatch in positions 2–8 strongly reduce the magnitude in

target regulation. Seed matches can be as few as four central nucle-
otides. Seed sites, on the other hand, have good 5′pairing but little
or no 3′pairing. In that case, seed matches should be more than 6
nucleotides long and any G–U base pairs cause a clear reduction in
the efficiency of the regulation [14].
In more recent studies, “canonical,” “marginal,” and “atypical”
categories were defined [27, 37]. Canonical are 7–8 nucleotide
seed-match sites, which include two 7mers and one 8mer. The
7mer-m8 site contains the seed match augmented by a match to
miRNA nucleotide 8. The 7mer-A1 site contains the seed match
augmented by an A at target position 1. The 8mer site comprises
the seed match flanked by both the match at position 8 and the A
at position 1. Marginal are 6 nt sites matching the seed region.
There are two groups, 6mer which refer to sites with perfect 6 nt
match to the miRNA seed (miRNA nucleotides 2–7) and offset
6mer which refer to 6mer starting from the third nucleotide.
Atypical are sites with productive 3′pairing and are divided into
3′-supplementary and 3′compensatory. In the first case, Watson–
Crick pairing usually centering miRNA nucleotides 13–16 supple-
ments a 6–8 nt seed site. At least 3–4 well-positioned contiguous
pairs are typically required for increased efficacy, which explains
why 3′-supplementary sites are atypical. On the other hand, in
3′-compensatory sites, Watson–Crick pairing usually centering on
miRNA nucleotides 13–16 can compensate for a seed mismatch
and thereby create a functional site [37].
More recently, “centered sites” have been described. They are
a class of miRNA target sites that lack both perfect seed pairing and
3′-compensatory pairing, and instead have 11–12 contiguous
Watson–Crick pairs to miRNA nt 4–15 [38]. In addition, based on
reporter and microarray results, it was suggested that the centered
site is a miRNA target site capable of downregulation comparable
with that observed for single 7 nt seed sites. Although they are
much less abundant, than both seed-matched sites and sites with
3′-supplementary pairing, centered sites are present in numbers
similar to 3′-compensatory sites and could help explain the pref-
erential conservation observed in the central region of most
miRNAs [38].
In 2012, Sung Wook (Chi et al. [39]) identified an alternative
binding mode used by miRNAs, by analyzing Ago HITS-CLIP
data. G-bulge sites (positions 5–6) are often bound and regulated
by miR-124 in mice brain. It was shown that bulged sites comprise
more than 15 % of all Ago–miRNA interactions in mouse brain and
are evolutionarily conserved. Position 6 is called the “pivot” nucle-
otide and suggests a model in which a transitional “nucleation
bulge” leads to functional bulge mRNA–miRNA interactions,
expanding the number of potential miRNA regulatory sites [40].
Fig. 1 Categories of miRNA–mRNA interactions. (a) Canonical are 7–8 nt seed-matched sites. They are divided
into three groups, 7mer-A1, 7mer-m8, and 8mer. (b) Atypical are sites with productive 3´ pairing. They consist
of two groups 3´ supplementary and 3´compensatory. (c) Marginal are 6 nt sites matching the seed region.
(d) Recently characterized interactions, which do not belong to other categories, “centered” and “G-bulge”
sites or occur in the mRNA’s coding region (CDS), are shown here. Vertical dashes indicate contiguous Watson–
Crick pairing. Underlined numbers on the miRNA sequence represent the seed site
One-sided or two-sided bulge in the seed region or even seed-

less sites have also been identified in 2008 and 2009 [41, 42]. In
the first case, Nanog, Oct4, and Sox2 coding regions are targeted
by miRNAs and embryonic stem cell differentiation modulated by
these interactions [41]. In the second case, miR-24 regulates MYC
and other cell-cycle genes by base pairing to “seedless” 3′UTR
sites with extensive base pairing elsewhere in the sequence [42].
A visual representation of the different types of target site
categories is shown in Fig. 1.
1.2.3 Biological Features A summary of the features which are widely used by miRNA target
Used by miRNA Target prediction methods is described below. In general, these features
Prediction Algorithms can be grouped into three categories: (1) Sequence—this feature
allows for the detection of base pairing between the miRNA and
the target site and moreover pinpoints the exact location of the
matching nucleotides (e.g., perfect “seed” match (nucleotides 2–7
at the 5′ part of a miRNA) or compensatory binding at the 3′
region); (2) Thermodynamics—the minimum free energy (ΔG) of
the miRNA–mRNA hybrid structure is very important for predict-
ing target sites. Programs such as RNAcofold [43] or RNAhybrid
[44, 45] can successfully predict hybrid structures between two
RNA molecules based on matching base pairing and furthermore
yield a minimum free energy for the overall structure. Since
miRNAs bind to their targets in a very specific and stable manner,
it is anticipated that the ΔG will be low. In fact, minimum free
energy calculations from experimentally predicted targets sites
display a very low value for ΔG [46] analogous to structures, which
display stable binding and interaction; (3) Conservation—target
sites are functional sequences in the 3′UTR of the transcribed
mRNA. This fact makes target sites subject to evolutionary conser-
vation across various organisms. A conservation analysis using
multiz full genome alignment files provided by UCSC (http://
genome.ucsc.edu/) shows that ~70 % of experimentally predicted
human miRNA targets sites are in fact conserved across eight other
organisms studied (including non-mammalian species). When used
as a feature for predicting new target sites, conservation may pro-
vide additional support for predictions.
2 Materials
Standard PC equipped with linux or windows (depending on tool

requirements) can be used to run all tools locally and standard
browsers can also be employed to access Web applications and
databases. For more computationally intensive tasks (i.e., RNA-seq
data assembly) a PC cluster can be utilized. An example of a PC
cluster for such jobs submissions is shown below:
CPU: 12 CPUs (2 * 6-core Intel® Xeon® X5680 @3.33 GHz)
(2010 model).
RAM: 64 GB (for 1 node—for memory intensive jobs), 48 GB (for
any additional nodes).
Storage: each node has 500 GB to 1 TB storage.
Machine Type: x86_64.
Operating System Linux, 2.6.35-gentoo-r15.
Accessible via ssh from central master/login node.
Job submission via PBS the queuing system on the master node of
the cluster.
Torque and Maui (resource manager and job scheduler) can be
used on the master node of the cluster (more information:
https://www.gridpp.ac.uk/wiki/Torque_and_Maui).
3 Methods
3.1 miRNA Online One of the earliest online databases for miRNA is miRBase [47]
Databases (http://www.mirbase.org). The current version of miRBase (v19)
provides an elaborate repository which includes ~19,000 miRNA
entries, ~22,000 mature miRNA sequences from over 160 species.
Moreover, all entries are extensively annotated using information
such as hairpin and mature sequences, genomic location, and rel-
evant paper(s) supporting the entry. In addition to facilitating
functional information for each miRNA entry, miRBase provides
links to popular miRNA target prediction tools and databases, such
as microRNA.org [48], DIANA-microT [49] and TargetScan [23,
50], as well as miRecords [51], DIANA-TarBase [52]. MirBase is
also equipped with additional information like information about
miRNA families and clusters, external links to Entrez Gene [53],
Rfam [54], and HGNC [55] as well as links to deep sequencing
experiments describing miRNA entries.
Databases archiving expression data for miRNAs in various
tissues and cell lines are becoming an absolute necessity in the new
era of RNAi. Data of this sort are often used in combination to pre-
diction algorithms, and are ideal for obtaining initial clues and direc-
tions for biologically driven hypotheses that can be concurrently
investigated in greater depth. One of the first databases describing
such data is smiRNAdb (http://www.mirz. unibas.ch/cloningpro-
files) [56]. smiRNAdb provides an elaborate miRNA expression map
for humans and rodents [56] derived from sequencing ~250 small
RNA libraries from 26 different cell types and organs. This database
offers various analysis steps and tools in order to compare expression
levels and profiles of miRNAs under different experimental condi-
tions, such as clustering analysis and principal component analysis.
Although computational prediction of miRNA gene targets
is a valuable asset, predictions should be supported by experi-
mental verification. Techniques used to verify target prediction
results can be either small or large scale [57]. The most widely
used small-scale, wet lab experimentation techniques are reporter
gene assays, followed by small-scale miRNA and/or mRNA
expression assessment via northern blotting or qPCR. ELISA
immunoassays or Western blotting can also be used to detect
changes in protein concentration of gene targets. High-
throughput proteomics methods like stable isotope labeling with
amino acids in cell culture/SILAC and transcriptomics methods
like microarrays and sequencing-based methodologies, such as
HITS-CLIP, PAR-CLIP, Degradome-Seq and RNA-Seq, can
further be used for large-scale miRNA target validation (further
described in detail below).
Numerous databases that contain curated experimentally veri-
fied miRNA targets are currently readily available. The first
comprehensive repository of experimentally validated targets was
Tarbase [52] (http://www.microrna.gr/tarbase). Currently, the

version of Tarbase (v6.0) provides an elaborate repertoire of over
65,000 miRNA–gene interactions, includes 21 different species
and, although initially contained only results from small scale
experiments (i.e., luciferase assays), now includes an extensive
index of target interactions obtained from large-scale, high-
throughput experiments, such as microarrays and proteomics as
well as interactions derived from sequencing data from HITS-
CLIP and PAR-CLIP experiments.
Recently, tools that make use of text-mining techniques have
been put into effect to mine through the enormous amount of
biological literature available online. They provide a faster and
easier way to collect useful and reliable information; however, this
information has to be curated manually before depositing it online.
These techniques lead to creation of databases like miRTarBase [58]
(http://mirtarbase.mbc.nctu.edu.tw), which provides another
large manually curated repertoire of experimentally verified targets.
It contains ~4,200 miRNA gene interactions (release 2.5, 10/2011)
from 14 species. The targets are obtained by ~1,400 articles and
are usually validated by reporter gene, western blot and microarray
experiments. miRTarBase also provides direct links to miRBase,
Entrez and others databases as well as the UCSC genome browser.
miRecords [51] (http://mirecords.biolead.org), also makes
use of text-mining techniques in combination with manual cura-
tion and currently hosts over 2,200 miRNA–gene interactions
entries from nine different animal species. MirSel [59] (http://
services.bio.ifi.lmu.de/mirsel/) is a similar database that contains
thousands of entries and like mirRecords allows for search and
filtering capabilities based on miRNAs or genes of choice, species,
and validation method. One main difference between the two
databases is that MirSel allows for a more automated extraction of
associations between microRNAs and genes from the biomedical
literature. These databases also provide information derived from
prediction tools which can be integrated into queries together
with other search criteria. Links are also provided to external data-
bases such as RefSeq, miRBase, mir2Disease, and others. Another
cumulative database is StarBase [60] (http://starbase.sysu.edu.cn).
StarBase focusses on sequencing experiments (i.e. PAR-CLIP,
HITS-CLIP, and Degradome-Seq) derived from six different spe-
cies and currently includes tens of thousands of miRNA–target
interactions In addition, StarBase provides several Web-based soft-
ware like miRPathway, miRGO, ClipSearch, DegradomeSearch to
explore microRNA regulatory networks.
3.2 miRNA Target This section describes some widely used microRNA target predic-
Prediction Tools tion tools in detail with particular emphasis to the target site cate-
in Detail gories each tool is specialized to recognize. We aim to provide an
overview of the most recently published tools and also to describe
tools that employ markedly different methodologies and prediction

features. We trust that the selection of these tools is a sufficient
sample to obtain an elaborate overview of the current state-of-
the-art tools and algorithms for miRNA target prediction. It is also
important to be able to compare the different types of tools and
assess their individual performance (for details on tool comparison
methods see Notes 1–3). Moreover most of these tools support a
graphic user or Web interface that permits fast target prediction
querying and also user intervention at numerous stages in the
prediction pipeline (see Note 4).
3.2.1 PicTar (http:// PicTar is a target prediction tool suitable for predicting “5′-domi-
pictar.mdc-berlin.de/) nant” target sites. However, it allows for targets with imperfect
seed matches given that they pass a heuristically defined binding-
energy threshold. Additionally, PicTar implements a maximum
likelihood approach to incorporate the combinatorial nature of
miRNA targeting. The program also implements cross-species
conservation constraints. It requires conservation between at least
five species for the portion of the target site that binds to the
miRNA seed. Conservation amounts to a seed match occurring at
overlapping positions in a cross-species UTR alignment. A more
recent version of PicTar provides precompiled target predictions
on the mouse genome based on comparative analyses of 17 verte-
brate genomes [8].
3.2.2 TargetScan/ TargetScan/TargetScanS is also appropriate for predicting

TargetScanS (http://www. “5′-dominant” target sites. It requires perfect complementarity
targetscan.org/index.html) with a miRNA seed and five general features of site context that
help site efficacy:
● AU-rich nucleotide composition near the site,
● proximity to sites for co-expressed miRNAs,
● proximity to residues pairing to miRNA nucleotides 13–16,
● positioning within the 3′ UTR at least 15 nt from the stop
codon,
● positioning away from the center of long UTRs.
The model also identifies non-conserved sites because it pre-
dicts site efficacy without recourse to evolutionary conservation
[23, 50].
3.2.3 MicroRNA.org MicroRNA.org incorporates both miRNA expressions and pre-

(http://microrna.org) dicted miRNA targets identified by the miRanda algorithm [61]
based on sequence complementarity between the mature miRNA
and the target site, and binding energy of the miRNA/mRNA
duplex [61. Targets scored with mirSVR [48].
Evolutionary conservation of the target site sequence and site
position in aligned 3′UTRs of homologous sequences are the two
main factors to calculate conservation. mirSVR utilizes a support

vector regression scoring scheme trained on a set of nine miRNA
transfection experiments performed on HeLa cells [48]. Scoring is
calculated using a weighted sum of context and sequence features
such as A/U content flanking the target site, target site accessi-
bility and position, 3′UTR length, base pairing and a conservation
score.
3.2.4 PITA (http://genie. PITA uses standard settings to identify initial seeds for each miRNA
weizmann.ac.il/pubs/ in 3′UTRs, applies a target accessibility model to each putative site,
mir07/) and then combines sites for the same miRNA to find a total inter-
action score for the miRNA and the UTR. This model also adds a
new dimension to miRNA target prediction, namely, the “flank
sites.” “Flank sites” are sites around the seed, which are required
to be unpaired. It has been found that a flank of 3 upstream and 15
downstream nucleotides results in the optimal the model perfor-
mance which surpasses all other methods. The model’s perfor-
mance is actually better than other methods even without using
the “3-15 flank” option [4].
3.2.5 Diana-microT Diana-microT uses a dynamic programming algorithm and like

(http://diana.cslab.ece. other tools requires seed complementarity to determine predic-
ntua.gr/microT/) tions. However, it does allow for some imperfect seed complemen-
tarity given there is compensatory binding in the 3′ regions of the
miRNA. Diana-microT also utilizes evolutionary conservation
between various organisms to further provide supporting evidence
for the credence of a predicted miRNA gene target [20, 49].
3.2.6 TargetProfiler TargetProfiler uses a hidden Markov model trained on sequence,

(http://mirna.imbb.forth.gr/ structure, and conservation features derived from experimentally
Targetprofiler.html) verified miRNA–mRNA interaction from Tarbase. Its allows scan-
ning of user defined target sequences and novel miRNAs as well as
querying on pre-compiled data derived from scanning all human
3′UTR sequences against all miRNA genes [62].
3.3 miRNA The aspect of obtaining functional annotation for miRNAs of

Functional Annotation interest is very important and therefore numerous tools and data-
Tools bases have been developed in order to shed some light on the bio-
logical significance of predicted miRNA–mRNA interactions.
Some common methodologies adopted by these tools are the use
of network analysis to obtain functional modules of miRNAs and
target genes. Statistical analysis using gene set enrichment and
online databases like GO and KEGG is also commonly being
employed, often in combination to network analysis, to obtain
functional information for miRNA and target gene interactions
taking into consideration multiplicity and co-operativity of binding.
Some recently published tools, available as Web applications, that
perform functional annotation are described below:
DIANA-miRPath [63] and miTalos [64]. These tools are

available as standalone tools or as downstream analysis steps for
miRNA target prediction pipelines. They make use of predicted as
well as validated miRNA-mRNA target interactions from multiple
online sources, which are subsequently subjected to pathway/
network analysis. Another available tool is GeneTrail [65], which
acts as an extension to existing target prediction pipelines.
GeneTrail performs gene set enrichment using information derived
from multiple online databases such as TRANSPATH, TRANSFAC,
KEGG and GO. MiRGator [66, 67] (http://mirgator.kobic.re.kr)
provides a more composite and integrated system for functional
annotation of miRNAs. miRGator performs statistical enrichment
analysis of target genes using databases like KEGG pathways, GO,
BioCarta and GenMAPP and integrates this information with
expression data derived from large-scale experiments. Moreover,
some well-established network analysis tools like Cytoscape [68]
now support various plugins (i.e., MiRScape) for miRNA–target
gene network analysis and visualization, gene set enrichment analysis,
merging of networks from miRNA–mRNA and protein–protein
interactions and links to a plethora of online databases.
In addition to Web server applications, there are also several
online databases that store functional information and associations
for miRNAs and target genes. One of these is miR2Disease [69]
(http://www.mir2disease.org), which utilizes text-mining to
extract useful functional information from biological literature and
after manual curation it provides a database which connects miRNA
function with numerous pathological diseases. miR2Disease
currently contains over 3,000 miRNA–disease connections for
~350 miRNAs and ~160 human diseases.
Similar to mir2Disease is HMDD [70] (http://202.38.
126.151/hmdd/mirna/md). HMDD also retrieves functional
information related to miRNAs and human disease from the
available literature. Yet another databse that aims to provide a link
between miRNAs and disease is PhenomiR [71] (http://mips.
helmholtzmuenchen. de/phenomir). This database focuses more
on miRNA which suffer from some kind of deregulation possibly
associated with some genetic perturbation event like deletion of
amplification. Hence a direct association is made between altered
miRNA expression and specific pathological disorders. Another
interesting online database is Patrocles [72] (http://www.patro-
cles.org/). Patrocles is a database of polymorphic miRNA–target
interactions, which, as the name suggests, is a metaphoric associa-
tion to Patrocles who was killed by Hector because he went to
battle wearing the armor of Achilles. Likewise, several mRNAs may
suffer mutations or single site polymorphisms (SNPs) that render
them as targets for miRNAs that otherwise would not regulate the
specific mRNA. This mistaken identity phenomenon has been
observed for Texel MSTN mRNA which mistakenly became the
target of miRNAs because of its disguise using a target octamer
motif borrowed from genuine target genes. Recently miRenviron-

ment [73] (http://202.38.126.151/ hmdd/tools/miren.html)
has been developed aiming to provide an association between
miRNA expression and environmental factors such as irradiation,
viral infections, and drug administrations.
3.4 High-Throughput Currently, the use of cutting edge high-throughput technologies,

Tools and Approaches like transcriptomics and proteomics, has a high impact on the field
for Detecting miRNA– of miRNA target prediction. Previous studies have used microar-
mRNA Interactions rays and pSILAC combined with computational tools, for the
identification of novel miRNA gene targets [6, 30]. These meth-
odologies allow for expression of genes, either at the mRNA or
protein level, to be correlated to miRNA gene expression. By
downregulating or over-expressing miRNAs of interest it is possible
to observe the regulatory effect of potential target genes using
these large-scale studies. The use of these methods in parallel to
computational predictions allows for computational results to be
validated via high-throughput methods and vice versa.
Next-Generation Sequencing (NGS) methods are now been
used for the prediction of miRNA genes and/or their mature
sequences [31] as well as obtain an expression profile that hints to
downstream targets. However, these techniques pose certain prob-
lems mostly attributed to technical difficulties of the sequencing
methodology such as library construction. Thus, due to the limit
of sequencing depth the non-abundant miRNA as well as the ones
which were expressed in a specific developmental stage cannot be
detected [74]. Nowadays, a massive parallel sequencing technol-
ogy has been developed which, aids in resolving most of the above
mentioned drawback. Using cutting-edge technology after con-
structing 187 libraries from diverse developmental stages, specific
tissue and cell-types, mutant conditions, and/or Argonaute
immunoprecipitations, the modENDCODE project managed to
sequence mature miRNAs from Drosophila melanogaster in unprec-
edented depth. Remarkably, the results of modENCODE project
revealed that several novel miRNA loci can be found in antisense
strands as well as in coding and untranslated regions of mRNAs,
giving new insights in miRNA biogenesis [75, 76]. Another
example of a massive parallel sequencing was derived from ten
different tissues of adult mice. Li et al. [76] managed to charac-
terize, 87,369,704 and 252,003 sequencing reads derived from
887 mature and 281 precursor miRNAs respectively. Their analysis
revealed new aspects of miRNA/pre-miRNA processing and
modification. Furthermore, based on the sequences of both mature
and precursor miRNAs miRGrep was developed, a miRNA discov-
ery pipeline which uncover 239 known mouse pre-miRNAs and 41
novel potential pre-miRs [76].
In parallel, well-defined computational algorithms may analyze
large-scale, high-throughput sequencing datasets and validate
whether the small-RNA sequences “reads” are true miRNAs.
Algorithms can evaluate the “reads” by analyzing their compati-

bility with features of miRNA such as miRNA biogenesis and
miRNA secondary structure [31, 74]. MiRDeep is one of the most
reliable algorithms which utilize Bayesian statistics to score the fit
of the reads to the biological model of miRNA biogenesis. Recently,
Friedlander et al. [74] evaluated the updated algorithm miRDeep2
with NGS using a human cell line before and after knocking down
the miRNA biogenesis pathway. The vast majority of the novel
miRNAs predicted by miRDeep2 and expressed in this cell line
were indeed specifically downregulated. MiRDeep2 may predict
novel miRNAs with the accuracy of 98.6–99.9 % [74, 77].
MIReNA, another computational algorithm looks for miRNA
sequences by exploring a multidimensional space defined by only
five parameters characterizing acceptable pre-miRNAs and detects
new miRNAs by homology to known miRNAs or from deep
sequencing data [78].
Nevertheless, not all small RNA and under-expressed mRNA
sequences detected by NGS methods are true miRNAs and targets
respectively, unless, some physical interaction between the two
sequences can be attributed to them. Ultraviolet (UV) cross-linking
and immunoprecipitation (CLIP) was developed to identify specific
protein–RNA interaction. CLIP relies on the principle that in vivo
protein–RNA interactions can be preserved and then mapped,
using ultraviolet C (UVC) light which can induce the formation of
covalent bonds. After that, cells can be lysed and protein–RNA
complexes can be immunoprecipitated with a specific antibody for
the protein of interest. Next, several additional molecular steps are
undertaken to obtain the RNA sequence [79]. Functional miR-
NAs are loaded into Argonaute protein and then bound to
their target gene in order to silence it. Therefore, this functional
complex made up of Argonaute–mRNA–miRNA can be verified
through CLIP technology. High-throughput sequencing of RNAs
isolated from cross-linking immunoprecipitation (HITS-CLIP)
have identified functional protein–RNA interaction sites. Chi et al.
[39] used this method in order to discover miRNA–target interac-
tions and Argonaute proteins bound to mRNA molecules in mouse
brain [39, 74]. Furthermore, Hafner et al. [80] developed an
improved version of CLIP method, called Photoactivatable-
Ribonucleoside-Enhanced Cross-linking and Immunoprecipitation
(PAR-CLIP). PAR-CLIP uses photoactivatable nucleotide ana-
logues which can enhance the efficiency of cross-linking. Upon
exposure to UVA light (365 nm) these nucleotides lead to base
transition—from thymidine to cytidine—at the cross-linked sites
and mutation analysis can therefore reveal binding sites with great
accuracy and precision [77, 80]. At the moment, several databases
of large-scale miRNA–target interactions using the HITS-CLIP
and PAR-CLIP have been developed by different research groups
[39, 60, 81]. High-throughput PAR-CLIP sequences can be
analyzed by computational tools such as miRTarCLIP, a novel
computational approach which has been developed by Chou et al.

[77] to identify miRNA–target interactions [77].
With the advent of high-throughput technologies the field of
miRNA target prediction has also been extended to encompass
identification of target genes from next-generation sequencing
(NGS) datasets.
Mirtools [82] is a RNA-Seq data analysis tool that works as a
web application as well as a standalone command-line program.
It makes use of online genomes as a reference for mapping
RNA-Seq reads. Thereafter, it is capable of locating known as well
as novel miRNA genes within the mapped reads using information
form miRBase and also performs miRNA target prediction using
the prediction tool miRanda[61].
Tools like miRanalyzer [83] (http://bioinfo2.ugr.es/miRana-
lyzer/miRanalyzer. php) are equipped with various steps for the
analysis of small RNA deep sequencing experiments. These steps
include read assessment and alignments construction using a tool
called Bowtie [84]. miRanalyser supports the detection of miRNA
gene targets using pre-compiled data for miRNA gene targets
identified using miRanda [61] and TargetScan [23, 50].
wapRNA [85] (http://waprna.big.ac.cn) is suited for the
analysis of mRNA-Seq/miRNA-Seq data from SOLiD and Solexa
technologies. It supports 10 species and can perform data analysis
steps for raw RNA-Seq data like read filtering and mapping to
reference genomes using Corona_Lite (SOLiD data) and BWA
(Solexa data). Furtherormore, wapRNA is capable of predicting
miRNA targets using RNAhybrid [44, 45] and miRanda [61].
For additional bibliography and information regarding the
experimental verification of miRNA gene targets using both high-
throughput and smaller scale wet lab techniques see Notes 5–8.
A comprehensive guideline of the major steps in the miRNA target
prediction pipeline are outlined in Note 9.
4 Notes
1. State-of-the-art tools are assessed using large-scale datasets as

well as experimentally verified data from online databases, and
their performance is recorded. Specifically, data from PAR-CLIP
[77] or pSILAC [30] can provide additional credence to
computationally predicted target genes. Furthermore, pairwise
ROC curves are commonly used for comparison between target
prediction tools. Tools that show a higher true positive rate and
low levels of false positive rate (high sensitivity and high speci-
ficity) obtain a higher area under the curve and are considered
to be the state of the art.
2. Another way to compare state-of-the-art tools is to use experi-
mentally verified miRNA targets from online databases like
Tarbase and therefore obtain an indication as to which tool

achieves a significant improvement in prediction accuracy.
3. Moreover, assessment of the number of common predictions or
overlap between the tools can be performed to highlight the
uniqueness of each tool and stress that each tool provides a valu-
able source of computational miRNA targets.
4. State-of-the-art tools are publicly available and often include
precompiled predictions for all miRNAs and gene targets stored
in relational databases that can be queried by the user via a Web
interface. Furthermore, they commonly provide an optimized
prediction algorithm for ad hoc predictions that makes use of
several biologically meaningful features of miRNA–target
mRNA interactions, combined with a user friendly interface
which allows for user flexibility in threshold adjustments and
assists in the identification of interactions of interest. The URLs
for some of these tools [86] are shown in Table 1, and most
tools allow for user intervention at various steps of the predic-
tion pipeline and can take into account both conserved and not
conserved miRNA targets in accordance to user preferences.
5. The capacity of prediction tools to predict miRNA targets with
high accuracy is the first major part of the prediction analysis
pipeline on the road to discovering new knowledge that is
biologically meaningful and contributes in answering specific
biological questions. At this stage it is important to take research
one step further and perform experimental verification on com-
putational predictions of biological significance. In previous
work [87] it has been shown that the use of prediction algo-
rithms to identify potential targets miRNAs can directly lead to
the experimental validation of these targets and the generation
of new biologically significant knowledge [88]. Importantly,
supporting evidence from recent deep sequencing studies can
provide expression data for specific miRNA genes and their
targets and correlate these to specific expression signatures of
biological conditions of interest [89].
6. PAR-CLIP experiments are expected to revolutionize the
research area of high-throughput miRNA target discovery due
to their ability to assess the direct interaction of miRNAs and
target mRNAs at a large scale, something previously performed
by expression correlation. Computational identification of highly
significant and evolutionary conserved target binding sites for
miRNA in the specific genes of interest using target prediction
tools provides the initiative for performing high-throughput
validation experiments as well as reporter gene assays.
7. In line with this, luciferase reporter assay results can show that
specific genes of interest are targeted by specific miRNAs and
this is depicted by decreased activity of the reporter gene in
Table 1
Databases and tools utilized for miRNA target predictions, including function and attributed to each
application as well as link to the site where the application can be found
Name of system Type (Database, tool, etc) Website

miRDB Tool—target prediction http://www.mirdb.org
microRNA.org Tool—target prediction http://microrna.org
PicTar Tool—target prediction http://pictar.mdc-berlin.de
TargetScan Tool—target prediction http://www.targetscan.org
RNA22 Tool—target prediction http://cm.jefferson.edu/rna22v1.0-homo_sapiens
TargetProfiler Tool—target prediction http://mirna.imbb.forth.gr/Targetprofiler.html
DIANA-microT- Tool—target prediction http://www.microrna.gr/microT-CDS
CDS
TargetMiner Tool—target prediction http://www.isical.ac.in/~bioinfo_miu/
targetminer20.htm
PITA Tool—target prediction http://genie.weizmann.ac.il/pubs/mir07/
GeneTrail Tool—gene set analysis http://genetrail.bioinf.uni-sb.de
miRGator Tool—Integrated system http://mirgator.kobic.re.kr
for functional
investigation of
miRNAs
miTalos Tool—prediction and http://mips.helmholtz-muenchen.de/mitalos
visualization of miRNA
signaling pathways
DIANA- Tool—prediction and http://www.microrna.gr/miRPathv2
miRPath visualization of miRNA
pathways
wapRNA Tool—RNA-Seq data http://waprna.big.ac.cn
analysis
DSAP Tool—RNA-Seq data http://dsap.cgu.edu.tw
analysis
miRanalyzer Tool—RNA-Seq data http://bioinfo2.ugr.es/miRanalyzer/miRanalyzer.php
analysis
miRCat Tool—RNA-Seq data http://srna-tools.cmp.uea.ac.uk/animal/cgi-bin/
analysis srna-tools.cgi? rm=input_form&tool=mircat
mirTools Tool—RNA-Seq data http://122.228.158.106/mr2_dev/
analysis
miRBase Database—miRNA http://www.mirbase.org
repository
S-MED/ Database—miRNA http://www.oncomir.umn.edu/SMED and http://
CC-MED expression data www.oncomir.umn.edu/colon/basic_search.php
(continued)
Table 1
(continued)
Name of system Type (Database, tool, etc) Website

smirnaDB Database—miRNA http://www.mirz.unibas.ch/cloningprofiles
expression data
miRNEYE Database—miRNA http://mirneye.tigem.it
expression data
microRNA. Database—miRNA http://www.microrna.org/microrna/getExprForm.do
org — miRNA expression data
Expression
miRTarBase Database—of manually http://mirtarbase.mbc.nctu.edu.tw
curated mirna targets
DIANA-TarBase Database—of manually http://www.microrna.gr/tarbase
miRecords Database—of manually http://mirecords.biolead.org
miRSel Database/tool—text http://services.bio.ifi.lmu.de/mirsel/
mined and manually
curated targets
StarBase Database—of sequence http://starbase.sysu.edu.cn
related miRNA
experiments
PhenomiR Database—miRNA http://mips.helmholtz-muenchen.de/phenomir
associated phenotypes
miR2Disease Database—miRNA http://www.mir2disease.org
associated diseases
HMDD Database—miRNA http://202.38.126.151/hmdd/mirna/md
associated diseases
miRenvironment Database—validated http://202.38.126.151/hmdd/tools/miren.html
miRNA environmental
associations
Patrocles Database—of http://www.patrocles.org/
polymorphic miRNA-
target interactions
wild-type binding site conditions and the increased activity in

mutated binding site conditions. Moreover, addition of anti-
sense LNAs for specific miRNAs to wild-type conditions is yet
another experiment that can provide supportive evidence for a
direct interaction of miRNA and target, given that reduced reg-
ulation is observed by increased reporter gene activity. When
designing these experiments care must be taken to correctly
address all case scenarios. One important aspect to be consid-
ered is that more than one miRNA may compete for the same
target site. In fact, a target site under investigation should be

scanned for potential targeting by other known miRNAs.
8. Moreover, using publicly available RNA-Seq [74], full genome
tiling array [90], and next-generation sequencing data [31] it is
possible to obtain addition information regarding the expres-
sion of known miRNA in the tissue or cell line under investiga-
tion. This can further direct ones’ experimental design and
guide experimentalists in performing the correct series of valida-
tion experiments. Competition between two miRNAs for the
same target site can lead to deviations in luciferase activity
during LNA silencing of miRNAs.
9. At this stage we believe that a comprehensive ranking framework
may be valuable for the experimental biologist who has a specific
interest in miRNA gene target prediction and verification. We
provide an intuitive way of ranking miRNA gene targets accord-
ing to the features provided by target prediction tools.
(a) Initially, it is necessary to view the score assigned to the
prediction (i.e., trained HMM score [62]). This will pro-
vide an initial indication as to the likelihood of the
prediction.
(b) Next the free energy (ΔG) of the interaction may provide
additional support for the stability of the hybrid molecule
and further strengthen predictions (the lower the ΔG the
more stable the interaction). Even manual inspection of the
secondary structure, as provided by some tools [9, 23, 62]
may prove to be an asset to the experienced miRNA
specialists.
(c) As previously documented [9, 23, 62], conservation is a
crucial filtering step in the miRNA target and gene predic-
tion pipelines. The more evolutionary conserved a stretch
of DNA the higher the chances that it generates functional,
transcribed, RNA sequences. Therefore, conservation is yet
another parameter which should be taken in consideration
when selecting a potential miRNA target for experimental
verification. In this perspective, conservation score provided
by numerous target prediction tools, may prove to be
helpful.
(d) Another additional feature which is provided by most tools
is the number of target sites predicted on the same 3′UTR
by the same miRNA. If a miRNA targets a specific 3′UTR
more than once this increases the chances that the given
interactions are in fact functional, as in evolutionary terms
multiple target sites would not exist in the same 3′UTR
unless they were of some functional purpose. Similarly if a
given target site is a hotspot for multiple miRNAs then this
also increases chances that one of these is in fact regulatory
for similar reasoning as described above.
(e) The use of high-throughput (PAR-CLIP, RNA-Seq,

pSILAC) data can then provide addition validity to compu-
tational results. Functional annotation using online data-
bases and network analysis, combined with statistical tools,
like gene set enrichment analysis and hypothesis testing, can
allow the expert RNA scientist to obtain an indication of
functional modules present in the specific analysis being
conducted. Moreover meaningful biological question or
hypothesis can be investigated and used to direct additional
research and experiments.
(f) Finally small scale experiments can further validate and nar-
row down the biological question under hand and allow for
fine tuning of the analysis pipeline.
References
1. Lee RC, Feinbaum RL, Ambros V (1993) The 11. Khvorova A, Reynolds A, Jayasena SD (2003)
C. elegans heterochronic gene lin-4 encodes Functional siRNAs and miRNAs exhibit strand
small RNAs with antisense complementarity to bias. Cell 115(2):209–216
lin-14. Cell 75(5):843–854 12. Lee Y, Jeon K, Lee JT et al (2002) MicroRNA
2. Huttenhofer A, Vogel J (2006) Experimental maturation: stepwise processing and subcellu-
approaches to identify non-coding RNAs. lar localization. EMBO J 21(17):4663–4670
Nucleic Acids Res 34(2):635–646 13. Lim LP, Glasner ME, Yekta S et al (2003)
3. Miranda KC, Huynh T, Tay Y et al (2006) A Vertebrate microRNA genes. Science
pattern-based method for the identification of 299(5612):1540
MicroRNA binding sites and their correspond- 14. Brennecke J, Stark A, Russell RB et al (2005)
ing heteroduplexes. Cell 126(6):1203–1217 Principles of microRNA-target recognition.
4. Kertesz M, Iovino N, Unnerstall U et al (2007) PLoS Biol 3(3):e85
The role of site accessibility in microRNA tar- 15. Helvik SA, Snove O Jr, Saetrom P (2006)
get recognition. Nat Genet Reliable prediction of Drosha processing sites
39(10):1278–1284 improves microRNA gene prediction.
5. Griffiths-Jones S, Grocock RJ, van Dongen S Bioinformatics 23(2):142–149
et al (2006) miRBase: microRNA sequences, 16. Hertel J, Stadler PF (2006) Hairpins in a
targets and gene nomenclature. Nucleic Acids Haystack: recognizing microRNA precursors
Res 34(Database issue):D140–D144 in comparative genomics data. Bioinformatics
6. Betel D, Wilson M, Gabow A et al (2008) The 22(14):e197–e202
microRNA.org resource: targets and expres- 17. Lim LP, Lau NC, Weinstein EG et al (2003)
sion. Nucleic Acids Res 36(Database The microRNAs of Caenorhabditis elegans.
issue):D149–D153 Genes Dev 16(8):991–1008
7. Lewis BP, Burge CB, Bartel DP (2005) 18. Sewer A, Paul N, Landgraf P et al (2005)
Conserved seed pairing, often flanked by aden- Identification of clustered microRNAs using an
osines, indicates that thousands of human ab initio prediction method. BMC Bioinform
genes are microRNA targets. Cell 6:267–281
120(1):15–20 19. Yousef M, Nebozhyn M, Shatkay H et al
8. Krek A, Grun D, Poy MN et al (2005) (2006) Combining multi-species genomic data
Combinatorial microRNA target predictions. for microRNA identification using a Naive
Nat Genet 37(5):495–500 Bayes classifier. Bioinformatics
9. Maragkakis M, Reczko M, Simossis VA et al 22(11):1325–1334
(2009) DIANA-microT web server: elucidat- 20. Kiriakidou M, Nelson PT, Kouranov A et al
ing microRNA functions through target pre- (2004) A combined computational-experi-
diction. Nucleic Acids Res 37(Web Server mental approach predicts human microRNA
issue):W273–W276 targets. Genes Dev 18(10):1165–1178
10. Maragkakis M, Alexiou P, Papadopoulos GL 21. Friedman RC, Farh KK, Burge CB et al (2009)
et al (2009) Accurate microRNA target predic- Most mammalian mRNAs are conserved
tion correlates with protein repression levels. targets of microRNAs. Genome Res 19(1):
BMC Bioinform 10:295 92–105
22. Witkos TM, Koscianska E, Krzyzosiak WJ functional sites with centered pairing. Mol Cell
(2011) Practical aspects of microRNA target 38(6):789–802
prediction. Curr Mol Med 11(2):93–109 39. Chi SW, Zang JB, Mele A et al (2009)
23. Lewis BP, Shih IH, Jones-Rhoades MW et al Argonaute HITS-CLIP decodes microRNA-
(2003) Prediction of mammalian microRNA mRNA interaction maps. Nature 460(7254):
targets. Cell 115(7):787–798 479–486
24. Enright AJ, John B, Gaul U et al (2003) 40. Chi SW, Hannon GJ, Darnell RB (2012) An
MicroRNA targets in Drosophila. Genome alternative mode of microRNA target recogni-
Biol 5(1):R1 tion. Nat Struct Mol Biol 19(3):321–327
25. Long D, Lee R, Williams P et al (2007) Potent 41. Tay Y, Zhang J, Thomson AM et al (2008)
effect of target structure on microRNA func- MicroRNAs to Nanog, Oct4 and Sox2 coding
tion. Nat Struct Mol Biol 14(4):287–294 regions modulate embryonic stem cell differ-
26. Marin RM, Vanicek J (2011) Efficient use of entiation. Nature 455(7216):1124–1128
accessibility in microRNA target prediction. 42. Lal A, Navarro F, Maher CA et al (2009) miR-
Nucleic Acids Res 39(1):19–29 24 Inhibits cell proliferation by targeting E2F2,
27. Grimson A, Farh KK, Johnston WK et al MYC, and other cell-cycle genes via binding to
(2007) MicroRNA targeting specificity in “seedless” 3′UTR microRNA recognition ele-
mammals: determinants beyond seed pairing. ments. Mol Cell 35(5):610–625
Mol Cell 27(1):91–105 43. Hofacker IL (2003) Vienna RNA secondary
28. Schmidt T, Mewes HW, Stumpflen V (2009) A structure server. Nucleic Acids Res
novel putative miRNA target enhancer signal. 31(13):3429–3431
PLoS One 4(7):e6473 44. Rehmsmeier M, Steffen P, Hochsmann M et al
29. Baek D, Villen J, Shin C et al (2008) The (2004) Fast and effective prediction of
impact of microRNAs on protein output. microRNA/target duplexes. RNA
Nature 455(7209):64–71 10(10):1507–1517
30. Selbach M, Schwanhausser B, Thierfelder N 45. Kruger J, Rehmsmeier M (2006) RNAhybrid:
et al (2008) Widespread changes in protein microRNA target prediction easy, fast and flex-
synthesis induced by microRNAs. Nature ible. Nucleic Acids Res 34(Web Server
455(7209):58–63 issue):W451–W454
31. Friedlander MR, Chen W, Adamidi C et al 46. Rusinov V, Baev V, Minkov IN et al (2005)
(2008) Discovering microRNAs from deep MicroInspector: a web tool for detection of
sequencing data using miRDeep. Nat miRNA binding sites in an RNA sequence. Nucleic
Biotechnol 26(4):407–415 Acids Res 33(Web Server issue):W696–W700
32. Cimmino A, Calin GA, Fabbri M et al (2005) 47. Kozomara A, Griffiths-Jones S (2010) miR-
miR-15 and miR-16 induce apoptosis by tar- Base: integrating microRNA annotation and
geting BCL2. Proc Natl Acad Sci U S A deep-sequencing data. Nucleic Acids Res
102(39):13944–13949 39(Database issue):D152–D157
33. Mayr C, Hemann MT, Bartel DP (2007) 48. Betel D, Koppal A, Agius P et al (2010)
Disrupting the pairing between let-7 and Comprehensive modeling of microRNA tar-
Hmga2 enhances oncogenic transformation. gets predicts functional non-conserved and
Science 315(5818):1576–1579 non-canonical sites. Genome Biol 11(8):R90
34. Sylvestre Y, De Guire V, Querido E et al (2007) 49. Reczko M, Maragkakis M, Alexiou P et al
An E2F/miR-20a autoregulatory feedback (2012) Functional microRNA targets in pro-
loop. J Biol Chem 282(4):2135–2143 tein coding sequences. Bioinformatics
35. Papadopoulos GL, Reczko M, Simossis VA 28(6):771–776
et al (2009) The database of experimentally 50. Garcia DM, Baek D, Shin C et al (2011) Weak
supported targets: a functional update of seed-pairing stability and high target-site abun-
TarBase. Nucleic Acids Res 37(Database dance decrease the proficiency of lsy-6 and
issue):D155–D158 other microRNAs. Nat Struct Mol Biol
36. Lee Y, Yang X, Huang Y et al (2010) Network 18(10):1139–1146
modeling identifies molecular functions tar- 51. Xiao F, Zuo Z, Cai G et al (2009) miRecords:
geted by miR-204 to suppress head and neck an integrated resource for microRNA-target
tumor metastasis. PLoS Comput Biol interactions. Nucleic Acids Res 37(Database
6(4):e1000730 issue):D105–D110
37. Bartel DP (2009) MicroRNAs: target recogni- 52. Vergoulis T, Vlachos IS, Alexiou P et al (2012)
tion and regulatory functions. Cell TarBase 6.0: capturing the exponential growth
136(2):215–233 of miRNA targets with experimental support.
38. Shin C, Nam JW, Farh KK et al (2010) Nucleic Acids Res 40(Database issue):
Expanding the microRNA targeting code: D222–D229
53. Maglott D, Ostell J, Pruitt KD et al (2005) of microRNAs. Nucleic Acids Res 36(Database
Entrez Gene: gene-centered information at issue):D159–D164
NCBI. Nucleic Acids Res 39(Database 68. Shannon P, Markiel A, Ozier O et al (2003)
issue):D52–D57 Cytoscape: a software environment for inte-
54. Griffiths-Jones S, Moxon S, Marshall M et al grated models of biomolecular interaction net-
(2005) Rfam: annotating non-coding RNAs in works. Genome Res 13(11):2498–2504
complete genomes. Nucleic Acids Res 69. Jiang Q, Wang Y, Hao Y et al (2009) miR2Dis-
33(Database issue):D121–D124 ease: a manually curated database for
55. Seal RL, Gordon SM, Lush MJ et al (2011) microRNA deregulation in human disease.
genenames.org: the HGNC resources in 2011. Nucleic Acids Res 37(Database
Nucleic Acids Res 39(Database issue):D98–D104
issue):D514–D519 70. Lu M, Zhang Q, Deng M et al (2008) An anal-
56. Landgraf P, Rusu M, Sheridan R et al (2007) A ysis of human microRNA and disease associa-
mammalian microRNA expression atlas based tions. PLoS One 3(10):e3420
on small RNA library sequencing. Cell 71. Ruepp A, Kowarsch A, Schmidl D et al (2010)
129(7):1401–1414 PhenomiR: a knowledgebase for microRNA
57. Thomson DW, Bracken CP, Goodall GJ (2011) expression in diseases and biological processes.
Experimental strategies for microRNA target Genome Biol 11(1):R6
identification. Nucleic Acids Res 72. Hiard S, Charlier C, Coppieters W et al (2010)
39(16):6845–6853 Patrocles: a database of polymorphic miRNA-
58. Hsu SD, Lin FM, Wu WY et al (2010) miRTar- mediated gene regulation in vertebrates.
Base: a database curates experimentally vali- Nucleic Acids Res 38(Database
dated microRNA-target interactions. Nucleic issue):D640–D651
Acids Res 39(Database issue):D163–D169 73. Yang Q, Qiu C, Yang J et al (2011) miREnvi-
59. Naeem H, Kuffner R, Csaba G et al (2010) ronment database: providing a bridge for
miRSel: automated extraction of associations microRNAs, environmental factors and pheno-
between microRNAs and genes from the bio- types. Bioinformatics 27(23):3329–3330
medical literature. BMC Bioinform 11:135 74. Friedlander MR, Mackowiak SD, Li N et al
60. Yang JH, Li JH, Shao P et al (2010) starBase: (2012) miRDeep2 accurately identifies known
a database for exploring microRNA-mRNA and hundreds of novel microRNA genes in
interaction maps from Argonaute CLIP-Seq seven animal clades. Nucleic Acids Res
and Degradome-Seq data. Nucleic Acids Res 40(1):37–52
39(Database issue):D202–D209 75. Berezikov E, Robine N, Samsonova A et al
61. John B, Enright AJ, Aravin A et al (2004) (2011) Deep annotation of Drosophila mela-
Human MicroRNA targets. PLoS Biol nogaster microRNAs yields insights into their
2(11):e363 processing, modification, and emergence.
62. Oulas A, Karathanasis N, Louloupi A et al Genome Res 21(2):203–215
(2012) A new microRNA target prediction 76. Li N, You X, Chen T et al (2013) Global profil-
tool identifies a novel interaction of a putative ing of miRNAs and the hairpin precursors:
miRNA with CCND2. RNA Biol insights into miRNA processing and novel
9(9):1196–1207 miRNA discovery. Nucleic Acids Res
63. Vlachos IS, Kostoulas N, Vergoulis T et al 41(6):3619–3634
(2012) DIANA miRPath v. 2.0: investigating 77. Chou CH, Lin FM, Chou MT et al (2013) A
the combinatorial effect of microRNAs in computational approach for identifying
pathways. Nucleic Acids Res 40(Web Server microRNA-target interactions using high-
issue):W498–W504 throughput CLIP and PAR-CLIP sequencing.
64. Kowarsch A, Preusse M, Marr C et al (2011) BMC Genomics 14(Suppl 1):S2
miTALOS: analyzing the tissue-specific regula- 78. Mathelier A, Carbone A (2010) MIReNA:
tion of signaling pathways by human and finding microRNAs with high accuracy and no
mouse microRNAs. RNA 17(5):809–819 learning at genome scale and from deep
65. Backes C, Keller A, Kuentzer J et al (2007) sequencing data. Bioinformatics
GeneTrail–advanced gene set enrichment anal- 26(18):2226–2234
ysis. Nucleic Acids Res 35(Web Server 79. Konig J, Zarnack K, Luscombe NM et al
issue):W186–W192 (2011) Protein-RNA interactions: new
66. Cho S, Jun Y, Lee S et al (2010) miRGator genomic technologies and perspectives. Nat
v2.0: an integrated system for functional inves- Rev Genet 13(2):77–83
tigation of microRNAs. Nucleic Acids Res 80. Hafner M, Landthaler M, Burger L et al (2010)
39(Database issue):D158–D162 Transcriptome-wide identification of RNA-
67. Nam S, Kim B, Shin S et al (2008) miRGator: binding protein and microRNA target sites by
an integrated system for functional annotation PAR-CLIP. Cell 141(1):129–141
81. Khorshid M, Rodak C, Zavolan M (2011) 86. Vlachos IS, Hatzigeorgiou AG (2013) Online
CLIPZ: a database and analysis environment resources for miRNA analysis. Clin Biochem
for experimentally determined binding sites of 46(10–11):879–900. doi:10.1016/j.clinbio-
RNA-binding proteins. Nucleic Acids Res chem.2013.03.006, Epub 2013 Mar 18.
39(Database issue):D245–D252 Review. PMID: 23518312 [PubMed—indexed
82. Wu J, Liu Q, Wang X, Zheng J, Wang T, You for MEDLINE]
M, Sheng Sun Z, Shi Q (2013), mirTools 2.0 87. Oulas A, Boutla A, Gkirtzou K et al (2009)
for non-coding RNA discovery, profiling, and Prediction of novel microRNA genes in cancer-
functional annotation based on high-through- associated genomic regions – a combined com-
put sequencing. RNA Biol 10(7): 1087–1092 putational and experimental approach. Nucleic
83. Hackenberg M, Rodriguez-Ezpeleta N, Acids Res 37(10):3276–3287
Aransay AM (2011) miRanalyzer: an update on 88. Simoneau M, Aboulkassim TO, LaRue H et al
the detection and analysis of microRNAs in (1999) Four tumor suppressor loci on chro-
high-throughput sequencing experiments. mosome 9q in bladder cancer: evidence for two
Nucleic Acids Res 39(Web Server novel candidate regions at 9q22.3 and 9q31.
issue):W132–W138 Oncogene 18(1):157–163
84. Langmead B, Trapnell C, Pop M et al (2009) 89. Han Y, Chen J, Zhao X et al (2011) MicroRNA
Ultrafast and memory-efficient alignment of expression signatures of bladder cancer
short DNA sequences to the human genome. revealed by deep sequencing. PLoS One 6(3):
Genome Biol 10(3):R25 e18286
85. Zhao W, Liu W, Tian D et al (2011) wapRNA: 90. Kapranov P, Cheng J, Dike S et al (2007) RNA
a web-based application for the processing of maps reveal new RNA classes and a possible
RNA sequences. Bioinformatics 27(21):3 function for pervasive transcription. Science
076–3077 316(5830):1484–1488
Chapter 14
Using Deep Sequencing Data for Identification

of Editing Sites in Mature miRNAs
Shahar Alon and Eli Eisenberg
Abstract
Deep sequencing has many possible applications; one of them is the identification and quantification of
RNA editing sites. The most common type of RNA editing is adenosine to inosine (A-to-I) editing.
A prerequisite for this editing process is a double-stranded RNA (dsRNA) structure. Such dsRNAs are
formed as part of the microRNA (miRNA) maturation process, and it is therefore expected that miRNAs
are affected by A-to-I editing. Indeed, tens of editing sites were found in miRNAs, some of which change
the miRNA binding specificity. Here, we describe a protocol for the identification of RNA editing sites in
mature miRNAs using deep sequencing data.
Key words miRNA, A-to-I editing, Deep sequencing, Bioinformatics, Genomics
1 Introduction
Adenosine to inosine (A-to-I) editing is catalyzed by enzymes of

the adenosine deaminase that act on RNA (ADAR) family, which
posttranscriptionally changes adenosine to inosine, the latter being
treated by cell machinery similar to guanosine [1]. As ADARs bind
to double-stranded RNA (dsRNA), they may act on the double-
strand formation of the primary transcript of a microRNA (miRNA)
gene (pri-miRNA) [2]. A-to-I editing of pri-miRNA can affect the
processing of the pri-miRNA to precursor miRNA (pre-miRNA)
or the processing of the pre-miRNA to mature miRNA [2]. The
mature miRNA is a single-stranded RNA, ~21 bases long, that can
bind to partially complementary targets located in the 3′ untrans-
lated region of specific mRNAs. Its binding blocks the translation
or triggers degradation of the target mRNAs [3]. Pri-miRNA and
pre-miRNA editing events may lead to an edited version of the
mature miRNAs. Bases 2–8 at the 5′ end of the mature miRNA
were found to be critical for the target recognition [3]. Therefore,
if the editing occurs in this recognition site, known as the “seed”
region, a change in the set of target genes is expected. A striking
231
232 Shahar Alon and Eli Eisenberg
example is mouse miR-376 in which editing in the recognition site

alters the target specificity of the miRNA and profoundly affects
cellular processes [4]. Editing outside the recognition site can also
be of interest as it may affect the loading of the mature miRNAs to
the RNA-induced silencing complex (RISC), thereby altering the
effectiveness of the mature miRNAs [1]. It should be mentioned
that it was recently discovered that mature miRNAs undergo large-
scale RNA modifications in the form of adenylation and uridylation
in the 3′ end [5]. However, the functional significance of these
changes is not fully understood.
In addition to A-to-I editing, it is possible that other types of
RNA editing modify the mature miRNA sequence. Proper utiliza-
tion of deep sequencing data has the potential to unravel the full
extent of A-to-I editing in miRNAs as well as other types of RNA
editing [6–8]. However, technical problems in the analysis of deep
sequencing datasets may result in a false detection of miRNA mod-
ification events [9]. Indeed, recent deep sequencing studies have
raised a debate surrounding the extent of these other types of
modifications in the human transcriptome in general and in miR-
NAs in particular [10–13]. Here we describe a protocol which uti-
lizes deep sequencing data to characterize RNA editing in mature
miRNAs while avoiding technical data analysis problems [13].
2 Materials
1. Hardware: 64-bit computer running Linux (including other

UNIX-based operating systems and Cygwin inside Windows)
or Mac OS with at least 8 GB of RAM (preferably 16GB).
2. Short read alignment program: In this protocol we use Bowtie
[14] which can be downloaded for Linux (see Note 1) or Mac
OS from http://bowtie-bio.sourceforge.net/index.shtml.
The Bowtie Web site provides installation instructions.
3. Genome indexes: The alignment program Bowtie uses genome
indexes to speed up computation time. Pre-built genome
indexes for many organisms can be downloaded via the Bowtie
Web site (mentioned above). The Mouse genome indexes
(UCSC mm9) are needed for the example below. Therefore,
they should be downloaded and placed in the “indexes” folder
inside the Bowtie folder.
4. Software: The Perl programming language should be prein-
stalled. Perl for Linux or Mac OS can be downloaded from
ActiveState http://www.activestate.com/activeperl. The
ActiveState Web site provides installation instructions.
5. Mathematical package: The Perl package Math-CDF is
required for the Perl scripts. This package can be downloaded
from CPAN (http://www.cpan.org/). The CPAN Web site
provides installation instructions.
Identification of Editing Sites in Mature miRNAs 233
6. Scripts: Three Perl scripts are required for this protocol:

http://www.tau.ac.il/~elieis/miR_editing/Process_reads.pl
http://www.tau.ac.il/~elieis/miR_editing/Analyze_mutation.pl
http://www.tau.ac.il/~elieis/miR_editing/Binomial_analysis.pl
7. miRNA definitions files: For the example below, download
these three text files which contain data about mouse miR-
NAs: the mature/star sequences, the alignment of the pre-
miRNA sequence against the genome, and the secondary
structure of the pre-miRNA:
http://www.tau.ac.il/~elieis/miR_editing/Mature_file.txt
http://www.tau.ac.il/~elieis/miR_editing/Mouse_pre_
miRNA_aligned_against_the_genome.txt
http://www.tau.ac.il/~elieis/miR_editing/RNAfold_file_
mouse.txt
One may place these files in the same working folder as the
Perl scripts for the purpose of the example below.
8. Sequencing data: This protocol refers to the output of Illumina
sequencing machines. As the following analysis relies on the
quality scores of the sequencing reads, the Fastq formatted
output of the sequencing run is required. To run the example
given below, download the dataset SRR346417 from the SRA
NCBI http://www.ncbi.nlm.nih.gov/sra/ (see Note 2). Place
the downloaded file in the same folder as the Perl scripts.
3 Methods
3.1 Filtering The Fastq file of the sequencing reads could be the raw reads with-
Low-Quality Reads out any filter, or filtered reads using Illumina software tools. In the
and Trimming latter case, proceed directly to Subheading 3.2. Otherwise, continue
Sequence Adapters with this step. Raw sequencing reads likely contain parts of the
adapter sequence. Therefore, these sequences must be identified
and removed. Moreover, low-quality reads (as defined by the read
quality score) are unlikely to be informative and therefore should be
removed. There are several published tools for raw Fastq filtering
[15, 16], or one can use our in-house filtering script “Process_reads.
pl.” The script allows the user to define “low-quality reads” by
choosing (a) a quality score cutoff and (b) the maximum number of
locations allowed to have lower quality score compared to the cho-
sen cutoff. For example, one can use a cutoff of 20 for the Phred
quality score (which ranges from 0 to 40) and the maximum num-
ber of locations can be set to 3 [13]. The script also trims the adapt-
ers (see Note 3) and the resulting trimmed read can be filtered if it
is too long or too short (as defined by the user). For example, as the
expected length of mature miRNAs is ~21 bases, one can remove
reads with length longer than 28 bases or shorter than 15 bases.
As an example, we will use the publicly available dataset of

mature miRNAs from the mouse cerebellum (accession number
SRR346417).
1. Use the head command to view the Fastq file (see Note 4):
$ head −100 sra_data.fastq
it is clear that these are raw, 36 bases long, reads.
2. Run the Perl script “Process_reads.pl” (see Note 5):
$ perl Process_reads.pl sra_data.fastq sra_data_filtered.fastq
the output on the screen should indicate that from ~25 million
reads, ~19 million reads passed the filtering procedure.
RUN TIME: ~10 min (using Intel Core i7 processor).
3.2 Aligning The filtered and trimmed reads should be aligned against the
the Reads Against genome of interest and not against an miRNA sequence database
the Genome [13]. We require unique best alignment, that is, the reads cannot
be aligned to other locations in the genome with the same number
of mismatches (Fig. 1). Lastly, we recommend using only align-
ments with up to one mismatch (in our datasets, allowing up to
two mismatches did not aid in detecting more editing sites; instead,
it added unreliable alignments). These steps taken together solve,
by and large, the cross mapping problem that significantly hinders
identification of true editing sites in mature miRNAs [9, 13].
The last two bases (the 3′ end) of mature miRNA undergo
extensive adenylation and uridylation [5]. It is thus recommended
that these bases will not be considered in the alignment. Naturally,
doing so prevents detection of editing in these locations. However,
not taking this measure and still demanding low number of mis-
matches will severely reduce the number of alignments obtained.
Fig. 1 A schematic representation of the alignment procedure. The sequencing

reads should not be aligned to more than one location in the genome with the
same number of mismatches
The following command line runs the alignment program

Bowtie in a manner compatible with the above considerations
(see Notes 6–8):
$ bowtie -p 8 -n 1 -e 50 -a -m 1 --best --strata --trim3 2 The_bowtie_
folder/indexes/mm9 sra_data_filtered.fastq > sra_data_filtered.
output
The output on the screen should indicate that from ~19 million
reads, ~8 million reads were uniquely aligned against the
mouse genome.
RUN TIME: ~1 min (using Intel Core i7 processor with eight
threads)
3.3 Mapping The purpose of this step is to move from reads aligned against the
the Mismatches genome (the end-point of the previous step) to counts of each of
Against the Pre- the four possible nucleotides at each position along the pre-miRNA
miRNA Sequences sequence, for all the pre-miRNAs. Performing this transformation
will allow us to focus our analysis only on bona fide miRNA and to
use, in the following step, binomial statistics to detect significant
modifications inside them.
We use pre-built files for the transformation: the alignment of
the pre-miRNA sequences against the genome and the mature/
star sequences of miRNAs. Taken together, these files give the
location of the pre-miRNA and the mature/star miRNA inside the
genome (see Notes 9 and 10).
The pre-built files and the Bowtie output file from the previous
step are the input for the script “Analyze_mutation.pl.” A key user-
defined input for this script is the minimum quality score allowed
in the location of the mismatch in order for it to be counted.
Naturally, the higher this number is, the lower the probability for
sequencing error. However, taking very high quality score filter
will give small number of counted modifications. We suggest using
30 as the minimum quality score allowed (see also below and [13]).
Running this script on large sequencing datasets (large Bowtie
output files) requires extensive internal memory resources. A way
around this hurdle is to divide the sequencing dataset into sev-
eral smaller files; afterwards the output files can be merged
(see Subheading 3.4). We suggest dividing the Bowtie output file
if the number of reads is higher than 2.5 million when using 8 GB
of RAM.
1. As in our example the Bowtie output file has ~8 million reads,
dividing the file is required (see Note 11):
$ split -l 2500000 sra_data_filtered.output part_
This command will divide the Bowtie output file into 4 files:
part_aa, part_ab, part_ac, and part_ad.
2. The script “Analyze_mutation.pl” can now be run:

$ perl Analyze_mutation.pl part_aa main_output_a.txt
RUN TIME: ~20 min (using Intel Core i7 processor with
8 GB of RAM).
3. Run the other parts of the Bowtie output file:
$ perl Analyze_mutation.pl part_ab main_output_b.txt
$ perl Analyze_mutation.pl part_ac main_output_c.txt
$ perl Analyze_mutation.pl part_ad main_output_d.txt
RUN TIME: ~60 min (using Intel Core i7 processor with
8 GB of RAM).
The main output of this script is a text file containing the
counts of each of the four possible nucleotides at each position
along all the pre-miRNA sequences. This output file is later used
for the binomial test (see Note 12).
The binomial statistics requires the number of successes (the
number of mismatches of a given type in a given position), the
number of trials (the total number of counts in the given position),
and the sequencing error probability. The sequencing error prob-
ability can be estimated using the Phred score. For example, if only
mismatches with Phred score of 30 are allowed, the expected base
call error rate should be 0.1 % in each position. The aligned read
data can be used in order to examine the actual sequencing error in
the retained reads (see Note 13).
3.4 Using Binomial In this step binomial statistics is performed on the output file of the
Statistics to Remove previous step to separate sequencing errors from statistically sig-
Sequencing Errors nificant modifications (Fig. 2).
Importantly, binomial statistics do not require any arbitrary
expression level filter. It is well suited even for low-expressed miR-
NAs with low number of sequencing reads, and the P-values com-
puted reflect the absolute number of reads detected, small or large
as the case may be.
This analysis is performed for every position (except the last
two positions of the miRNA due to the extensive adenylation and
uridylation) in every mature/star miRNA separately. As multiple
tests are performed, the resulting P-value for each position must be
corrected accordingly. The script performing this analysis,
“Binomial_analysis.pl,” gives the user an option to use Bonferroni
or Benjamini–Hochberg corrections.
1. Run the script “Binomial_analysis.pl” using the files from the
previous step (see Notes 14 and 15):
$ perl Binomial_analysis.pl main_output_a.txt main_output_b.
txt main_output_c.txt main_output_d.txt > binomial_output.txt
RUN TIME: <1 min (using Intel Core i7 processor).
Fig. 2 A schematic representation of the mismatch filtering step. Sequencing

errors, 3′ adenylation and uridylation events, and SNPs are removed, leaving
only possible RNA editing events
2. View the output file “binomial_output.txt” using text editor.

The locations of the significant modification are given as well
as the statistical description of the modifications (see Note 16).
Overall, in this example 38 modifications should appear (see
Note 17). Most of the modifications (30 out of the 38) are
A-to-G, which can be a result of A-to-I editing. What can be
the nature of the other types of modifications? Among the
possible explanations are the following: (a) problems in the
definition of the miRNA (see Note 18), (b) 5′ adenylation
and uridylation (see Note 19), (c) SNPs or somatic mutations
(see Subheading 3.5), or (d) non-A-to-I RNA editing events.
3.5 Removing SNPs Known SNPs may be filtered from the statistically significant modi-
from the List fications detected in the previous step. This can be done using
of Statistically Galaxy (http://galaxy.psu.edu/).
Significant 1. In the Galaxy site choose the “Use Galaxy” option.
Modifications
2. Use the “Get Data” link, and then “UCSC Main” to down-
load data from the UCSC Table Browser.
3. To download the mouse SNP dataset, choose “Mouse” in the
“genome” field and “Variation and Repeats” in the “group”
field. Make sure that the “track” field is set to “SNPs” and the
“region” field is set to “genome.” Then press “get output”
and afterwards press “send query to Galaxy.”
4. Upload the output file of the previous step, “binomial_output.

txt,” using the “Get Data” link, and then “Upload File” link.
Choose “interval” in the “File Format” field and locate the file
using the “Browse” button. Choose “Mouse mm9” in the
“Genome” field and then press “Execute.”
5. Subtract the genomic locations of the SNPs from the locations
of the identified mismatches. Use “Operate on Genomic
Intervals” and then “Subtract.” Choose to subtract the SNP
dataset from the identified mismatch dataset, make sure that
the minimal overlap is set to 1 base, and then press “Execute.”
After a while the subtracted file should appear as a new tab in
the “History” panel (right side of the screen).
6. View the resulting subtracted file. The file should appear in the
upper right corner of the screen under the “History” tab. It
can be viewed using the eye icon. In this case, the resulting file
should be identical to the file “binomial_output.txt,” meaning
that none of the identified modifications are known SNPs.
RUN TIME: ~30 min (the execution time depends on the
speed of the Internet connection and the current load of
the Galaxy servers).
TOTAL RUN TIME: ~2 h (using Intel Core i7 processor
with 8 GB of RAM).
4 Notes
1. By Linux we also refer to UNIX-based operating systems and

Cygwin inside Windows.
2. Downloading the sequencing data from the NCBI SRA: (a)
Type the run accession number (starts with SRR) or the exper-
iment accession number (starts with SRX). Alternatively, type
the study accession number (starts with SRA) or search by
keywords and then choose the experiment of interest. (b) Use
the “Run” link to choose the run of interest. (c) Choose
“Filtered Download.” (d) Choose FASTQ as the “Download
Format” and press “Download.” (e) Unzip the file “sra_data.
fastq.gz” using gunzip command (in Linux and Mac). The
“Filtered Download” tool can also allow the user to filter the
obtained reads as explained in the NCBI SRA site. The down-
load process can be made faster by using the Aspera plug-in
using the link in the NCBI SRA site.
3. Although Illumina small RNA Adapters are universal, different
versions of the protocol may result in variations in the adapter
sequences. The user should check the adapter sequences given
in the Perl script “Process_reads.pl” and change them if
needed. Viewing and editing the Perl script can be done by
using a text editor.
4. Commands meant to be executed from the Linux command

line are prefixed with a “$” character. The commands are
meant to be run from the example working directory.
5. Edit the Perl script “Process_reads.pl” using a text editor if a
change in the default parameters is required.
6. Note that the allowed Phred score in the mismatch position is
not limited (-e 50). This means that even mismatches with
high quality score will be allowed to align.
7. The -p option of Bowtie allows the use of multiple threads in
parallel. The number of available threads is naturally depen-
dent on the hardware.
8. By default, the Fastq file is treated by Bowtie as written in
Phred + 33 format. Phred + 33 format means that the number
33 is added to the Phred score and the result is translated into
single ASCII character. Note that previous versions of the
Illumina software produced Fastq files in Phred + 64 format.
Therefore, it is recommended to view the Fastq file (using, for
example, the Linux “head” command) to determine the Phred
to ASCII conversion. If the Fastq file is in Phred + 64 format,
the flag --phred64-quals should be added to the Bowtie exe-
cution line.
9. Pre-built files for human are given below:
http://www.tau.ac.il/~elieis/miR_editing/Human_pre_
miRNA_aligned_against_the_genome.txt
http://www.tau.ac.il/~elieis/miR_editing/RNAfold_file_
human.txt
Open the Perl script “Analyze_mutation.pl” using a text edi-
tor and change the names of the pre-built miRNA files
accordingly.
10. The miRBase release 17 was used for the human and mouse
definition files. Human and mouse definition files for miRBase
release 18 are available in
http://www.tau.ac.il/~elieis/miR_editing/Release18.zip
The results of the example described in this protocol are not
affected by using the latter release.
It is simple to create the definition files for various other spe-
cies. The miRNA sequences can be downloaded from
miRBase (http://www.mirbase.org/), and the alignment
of the pre-miRNA sequence against the genome can be
achieved using Bowtie. Finally, RNAfold [17] can be used
to identify the secondary structure of the pre-miRNA.
11. The -l option in the Linux split command allows dividing a file
by the number of lines. Splitting by line (and not by size) is
required in order to prevent cutting the output lines in the
end of the file.
12. This output representation can also be used to detect significant

variation from the miRBAse mature/star definitions (in terms
of beginning and end locations). Only modifications inside the
mature/star sequence as defined in miRBAse are analyzed in
the script. Additional output file is a text file containing the
number of reads for each mature/star miRNA; this file (termed
“miRNA_expression.txt”) can be used for expression analysis.
13. Merging the data from all the mature and star miRNAs, the
rate of any type of mismatch from the 12 possible mismatches
can be calculated. In this calculation the location of the mis-
match should also be taken into account, as biological modifi-
cations appear in the 3′ of the miRNA and sequencing quality
is lower towards the end of the read. Therefore, the calculation
should be performed separately in each position along the
mature/star miRNA. The results are presented in the main
output file of “Analyze_mutation.pl.” The rate relevant for the
statistical analysis to follow is the maximal mismatch rate of any
single type of mismatch. If it is higher than the rate expected
from the Phred score, the higher estimate should be used [13].
14. The Perl script “Binomial_analysis.pl” can process any number
of input files. The counts of each of the four possible nucleo-
tides at each position along the pre-miRNA sequence, for all
the pre-miRNAs, are merged and analyzed together. Naturally,
the input files should all be created using the same miRNA
definition files (see Subheading 3.3).
15. Additional input file for the script “Binomial_analysis.pl” is
the alignment of the pre-miRNA against the genome. This file
is required to translate the location of the mismatch to genomic
coordinates. See Subheading 2, item 7, and Note 9 for pre-
built files for human and mouse. If change in organism is
required, open the Perl script using a text editor and change
the names of the pre-built miRNA files accordingly.
16. The genomic locations in the output of “Binomial_analysis.pl”
are in the format, Chromosome (tab) Start location (tab) End
location, for example chr16 (tab) 77599184 (tab) 77599185.
The mismatch is located in the End position, that is, in base
77599185 in this example. This format, which includes Start
location and End location with 1 base separating them, was
chosen to be compatible with the genomic representation of
SNPs in the UCSC Table Browser (see Subheading 3.5).
17. The identity of the detected modifications depends on the
specific parameters used in “Binomial_analysis.pl”: the esti-
mate of the sequencing error probability, the multiple correc-
tion type, and the P-value used as a cutoff after multiple
correction. Here, the parameters used were 0.001, Bonferroni
correction, and 0.05, respectively.
18. miRNAs with inconsistent definitions can be identified by

visually inspecting the distribution of the reads along the pre-
miRNAs in question. This data can be viewed in the output
files of “Analyze_mutation.pl,” for example, “main_output_a.
txt.” In addition, miRBase (http://www.mirbase.org/) can
be used to view the distribution of reads along each pre-
miRNA using publicly available deep sequencing datasets. For
example, miR-5105 was identified as modified in position 22,
but the distribution of the reads along the pre-miRNAs does
not have a single peak around the mature miRNA sequence.
Instead, the reads appear to come from multiple regions inside
the pre-miRNA, which may suggest that this is not a bona fide
miRNA. The distribution of reads along the pre-miRNA of
miR-5115 indicates that the actual location of the mature
miRNA is significantly different from the location mentioned
in miRBase. Therefore, the modifications in miR-5105 and
miR-5115 should be taken with caution.
19. We note that some low-abundance isomirs display 5′ sequence
modifications similar to the biological modifications reported
at the 3' of mature miRNAs. We previously identified several
isomers that start one or two bases upstream from a different
isomer and display sequence modifications in the 5′ end in the
form of adenylation and uridylation. In all these events the
abundance of the modified isomer was at least an order of mag-
nitude lower than the unmodified isomer [13]. The script
“Binomial_analysis.pl” automatically identifies these events
and discards them from the analysis. However, the file “bino-
mial_output.txt” contains three modifications which are likely
to be a result of 5′ adenylation and uridylation: in miR-126,
miR-337, and miR-23b. Viewing the distribution of reads
along the pre-miRNAs (see Note 18), it can be seen that the
modifications in miR-337 and miR-23b are in isomirs which
are more than an order of magnitude lower compared to the
expression levels of the main isomer. They were not filtered out
because they are located in positions 4 and 3 along the mature
miRNA sequence, respectively, whereas the maximal position
to discard is set to 2 in “Binomial_analysis.pl.” miR-126 was
not filtered out because the fold change between the 5′ isomir
and the main isomer was only ~5 whereas the default is set to
10 in “Binomial_analysis.pl.” These parameters can be easily
changed by editing “Binomial_analysis.pl” using a text editor.
Acknowledgments
This work was supported by a grant from the Israel Science

Foundation [379/12] and a grant from the United States–Israel
Binational Science Foundation (grant number 2009/290),
Jerusalem, Israel, to E.E. S.A. was supported by a Clore Fellowship.
References
1. Nishikura K (2010) Functions and regulation of editing sites in mature microRNAs in high-
of RNA editing by ADAR deaminases. Annu throughput sequencing libraries. Genome Res
Rev Biochem 79:321–349 20:257–264
2. Yang W, Chendrimada TP, Wang Q et al 10. Li M, Wang IX, Li Y et al (2011) Widespread
(2006) Modulation of microRNA processing RNA and DNA sequence differences in the
and expression through RNA editing by ADAR human transcriptome. Science 333:53–58
deaminases. Nat Struct Mol Biol 13:13–21 11. Lin W, Piskol R, Tan MH et al (2012)
3. Bartel DP (2004) MicroRNAs: genomics, bio- Comment on “Widespread RNA and DNA
genesis, mechanism, and function. Cell 116: sequence differences in the human transcrip-
281–297 tome.”. Science 335:1302
4. Kawahara Y, Zinshteyn B, Sethupathy P et al 12. Piskol R, Peng Z, Wang J et al (2013) Lack of
(2007) Redirection of silencing targets by evidence for existence of noncanonical RNA
adenosine- to-inosine editing of miRNAs. editing. Nat Biotechnol 31:19–20
Science 315:1137–1140 13. Alon S, Mor E, Vigneault F et al (2012)
5. Burroughs AM, Ando Y, de Hoon MJL et al Systematic identification of edited microRNAs
(2010) A comprehensive survey of 3′ animal in the human brain. Genome Res 22:
miRNA modification events and a possible role 1533–1540
for 3′ adenylation in modulating miRNA target- 14. Langmead B, Trapnell C, Pop M et al (2009)
ing effectiveness. Genome Res 20:1398–1410 Ultrafast and memory-efficient alignment of
6. Ekdahl Y, Farahani HS, Behm M et al (2012) short DNA sequences to the human genome.
A-to-I editing of microRNAs in the mamma- Genome Biol 10:R25
lian brain increases during development. 15. Stocks MB, Moxon S, Mapleson D et al (2012)
Genome Res 22:1477–1487 The UEA sRNA Workbench: a suite of tools
7. Warf MB, Shepherd BA, Johnson WE et al for analysing and visualising next generation
(2012) Effects of ADARs on small RNA pro- sequencing microRNA and small RNA datas-
cessing pathways in C. elegans. Genome Res ets. Bioinformatics 28:2059–2061
22:1488–1498 16. Zhu E, Zhao F, Xu G et al (2010) mirTools:
8. Vesely C, Tauber S, Sedlazeck FJ et al (2012) microRNA profiling and discovery based on
Adenosine deaminases that act on RNA induce high-throughput sequencing. Nucleic Acids
reproducible changes in abundance and Res 38:W392–W397
sequence of embryonic miRNAs. Genome Res 17. Zuker M, Stiegler P (1981) Optimal computer
22:1468–1476 folding of large RNA sequences using thermo-
9. de Hoon MJL, Taft RJ, Hashimoto T et al dynamics and auxiliary information. Nucleic
(2010) Cross-mapping and the identification Acids Res 9:133–148
Chapter 15
NGS-Trex: An Automatic Analysis Workflow

for RNA-Seq Data
Ilenia Boria,* Lara Boatti,* Igor Saggese, and Flavio Mignone
Abstract
RNA-Seq technology allows the rapid analysis of whole transcriptomes taking advantage of next-generation
sequencing platforms. Moreover with the constant decrease of the cost of NGS analysis RNA-Seq is
becoming very popular and widespread. Unfortunately data analysis is quite demanding in terms of bioin-
formatic skills and infrastructures required, thus limiting the potential users of this method.
Here we describe the complete analysis of sample data from raw sequences to data mining of results
by using NGS-Trex platform, a low user interaction, fully automatic analysis workflow. Used through a
web interface, NGS-Trex processes data and profiles the transcriptome of the samples identifying expressed
genes, transcripts, and new and known splice variants. It also detects differentially expressed genes and
transcripts across different experiments.
Key words Next-generation sequencing, RNA-Seq, Transcriptome profile, Differentially expressed

genes, Gene expression profiling, NGS-Trex
1 Introduction
RNA-Seq technology enables to fix a snapshot of RNA expressed

in a cell taking advantage of next-generation sequencing platforms.
RNA-Seq is emerging as the major quantitative transcriptome pro-
filing system [1, 2] but it also allows the detection of differentially
expressed genes and transcripts across different experiments [3].
Moreover—unlikely microarray technique—it is not biased by a
priori knowledge, thus making possible the identification of unan-
notated transcripts and splice variants.
Reads obtained after sequencing may be assembled into tran-
scripts or can be used to identify known genes and transcripts
available in databases. The first approach is mainly used when a
reference genome is not available for the organism under investigation.
*Author contributed equally with all other contributors.
243
244 Ilenia Boria et al.
This technique is known as de novo transcriptome assembly. On

the other hand whenever a reference genome (and its annotation)
is available this information can be used to drive the assembly and
the identification process. The latter case is known as genome-
guided transcriptome analysis and is the focus of this manuscript.
A conceptual RNA-Seq analysis pipeline is relatively easy to
be drawn, and four main steps can be identified: Filtering/
Demultiplexing, Mapping, Annotation, and Data Mining.
First of all, raw sequencing data must be filtered to remove
low-quality reads and cloning linkers. Moreover, since samples are
often multiplexed in a single sequence run, a demultiplexing pro-
cess is required to subdivide the reads in their respective sample
datasets. High-quality reads are then mapped onto the reference
genome and are compared to annotation to assign them to known
genes and transcripts. By comparing reads with gene models it is
possible to quantify known elements and to detect unannotated
features such as splice variants. Finally differential expression at
both gene and transcript level can be estimated by comparing reads
assigned to genes in different datasets.
While the conceptual analysis pipeline to deal with NGS data is
relatively easy to be drawn, the actual implementation can be chal-
lenging. Big efforts in terms of infrastructure and specific skills are
in fact required to handle the large amount of data produced by
high-throughput sequencing, thus making difficult for researchers
to fully exploit the potential of RNA-Seq/NGS technologies.
To date several software packages and bioinformatics tools
have been developed to address this issue; they can be categorized
into command-line-based packages, GUI-based software, and
web-based applications. Command-line- and GUI-based stand-
alone packages require dedicated hardware resources and software
installation, while web-based tools are ready to be used. Currently
available web-based systems include Galaxy [4], GeneProf [5], and
NGS-Trex [6] all of them with various advantages and drawbacks.
In the following section we will show the steps required to
analyze an RNA-Seq dataset using NGS-Trex system available at
www.ngs-trex.org.
2 Materials
2.1 RNA-Seq For our sample analysis we will process raw RNA-Seq data obtained
Datasets by Birzele et al. [7]. Briefly, the authors have used Illumina’s
Genome Analyzer platform to explore the differences in transcrip-
tome profile between resting and activated human CD4+ T-helper
and regulatory T-cells, including differential expression of genes
and splice variants.
Raw sequence files, corresponding to the four samples examined
by Birzele et al., are available at the NCBI Sequence Read Archive
(SRA) under accession number SRP006674. To download sequences
An Automatic Analysis Workflow for RNA-Seq Data 245
Table 1
List of datasets described by Birzele et al. Each dataset is the result of the forward and reverse
paired end sequences merged together
Dataset Accession number Description Number of reads

SRR192532 SRR192532_1 mRNAseq of resting Human 15,277,248
SRR192532_2 CD4 Th cells
SRR192536 SRR192536_1 mRNAseq of activated Human 10,335,712
SRR192536_2 CD4 Th cells
SRR192425 SRR192425_1 mRNAseq of resting Human Treg cells 13,655,820
SRR192425_2
SRR192534 SRR192534_1 mRNAseq of activated Human Treg cells 8,237,652
SRR192534_2
in fastq format it is more easy to use the European Nucleotide

Archive (ENA) hosted at the European Bioinformatic Institute
(www.ebi.ac.uk/ena) (see Note 1). By inserting the SRP code of the
dataset in the “text search” form the user is redirected to a table with
a link to the fastq files. For each dataset two distinct files—contain-
ing the forward and the reverse paired end sequences (file 1 and file
2)—are available. We need to download both files for each dataset.
Table 1 summarizes the downloaded files.
2.2 Computational As described in 6 NGS-Trex is an automatic system to store, ana-

Software lyze, and mine RNA-Seq data. It is available through a simple web
interface (www.ngs-trex.org) and allows the user to upload raw
sequences files in fastq/fasta format and easily obtain an accurate
characterization of the transcriptome profile after the setting of few
parameters. The system is also able to assess differential expression
at both gene and transcript level (i.e., splicing isoforms) by com-
paring the expression profile of different samples (e.g., normal and
disease status).
3 Methods
The overall procedure to analyze RNA-Seq data with NGS-Trex is

depicted in Fig. 1. It can be summarized into three major steps: (1)
upload of raw sequences; (2) analysis; (3) results exploration and
retrieval.
NGS-Trex organizes high-throughput sequencing data into
projects and each project includes one or more datasets or
samples.
To use the system we need to login using a demo account as
described in the homepage or we can obtain our personal credentials.
Fig. 1 Schema of NGS-Trex procedure for RNA-Seq data analysis (image modified from original in NGS-Trex
web page)
Demo account is fully functional but has two main limitation: (1) the
size of each sample is limited to 10 Gb (2) projects and files are
“world accessible” so anyone can read and modify our data.
3.1 Data Submission The first step is to upload files to NGS-Trex server. This is the only
procedure requiring an external tool. Indeed we need an FTP
3.1.1 File Upload
client (such as FireFTP, FileZilla, or any other client supporting
sftp transfer protocol) to upload sequences. The exact procedure
depends on the client used, but we have to provide the server name
(www.ngs-trex.org) and our credentials (username and password).
Only fasta and fastq formats (*.fa,*.tfa,*.fasta,*.fna,*.fq,*.fastq)
are supported by the system.
3.1.2 Create a New By clicking the Create New Project link we can fill a form with the
Project project name (we enter “SRP006674”), a brief description (“Cell
line datasets—Birzele et al. [7] NAR 39(18)”), and we can select
the reference genome to be used for the analysis (we select “Homo
sapiens (NCBI Build 36)” to use the same release used in Birzele
work). On form submission we are redirected to the project main
page. This page includes three sections: (1) the project summary:
with the description of the project and a summary of already ana-
lyzed dataset, (2) the list of datasets belonging to the project
(which is empty on newly created projects), and (3) the Add
Dataset form.
3.1.3 Create Datasets At this stage we need to create the required datasets by selecting
the Add Dataset panel.
The Select file to upload link opens a window that shows a list
of the previously uploaded files. It is also possible to upload files
through this interface but—given the size of the files involved—it
is much more reliable to use the ftp approach described before.
As previously noticed, each dataset has two files. NGS-Trex does
not explicitly deal with paired end sequences but treats forward
and reverse sequences as single stranded. For this reason we need
to assign the two distinct files to the same dataset. There is no need
to preprocess files as we can select both files (pressing “ctrl” key—
“cmd” key on mac OS—while selecting files). For example we can
select “SRR192536_1” and “SRR192536_2”, we then right-click
on one of the highlighted files and press on “Select” menu item.
We then assign “CD4_activated” as label and “CD4 Th activated—
Birzele et al. [7]” as short description and press “Add Dataset”
button.
We can repeat the same procedure for all the other datasets.
Now Dataset panel of SRP006674 project shows a table with
our four dataset: “CD4_activated”, “TREG_activated”, “CD4_
resting”, and “TREG_resting”. We are ready to analyze our data.
3.2 Analysis Although NGS-Trex analysis can be run with default options, it is
possible to tune many parameters through a simple form accessible
via the Set Analysis Params link provided—for each sample—in the
Datasets panel of the project main page. Through this form it is
possible to optimize the three main steps of NGS-Trex analysis
workflow: (1) sequences preprocessing, (2) mapping criteria to
align reads to the reference genome, and (3) annotation strategy to
assign mapping reads to annotated genes and transcripts.
3.2.1 Sequences With sequence preprocessing we can instruct the system to remove
Preprocessing cloning linkers and define the rules to deal with reverse stranded
reads.
“Which kind of data do you want to analyze?” Based on the plat-
form used to obtain the raw data, we can currently select “454” or
“Illumina”. This option controls the algorithm used for the map-
ping of the reads onto the reference genome.
“Cut-off value for percentage of read length that should be of given
quality” (default: 70 %). This threshold controls the number of bases
of the reads that satisfy the quality threshold defined in the “Cut-off
value for quality score for high-quality filtering” value (see Note 2).
“Do you have linker/cloning sequences to remove?” By answering Yes
a form where we can insert the sequence of the 5′ and 3′ linkers
appears. (Both linkers have to be written in 5′–3′ orientation.) It is
also possible to trim some extra nucleotides between the linker and
the main sequence. The system searches for exact matches between
the provided linker sequences and the reads.
“Look for linker in Reverse (if not found in forward)?” If sequences
are not oriented we can expect to find 5′ and 3′ linkers also with a
reverse order.
“Reverse complement sequence if linker found reversed?” If linkers
are found with reverse order the read can be simply trimmed
(maintaining unchanged its orientation) or it can be reverse-
complemented to reflect the correct order of linkers. This option
can be very critical for the following Annotation step (as described
below) as it allows proper orientation of the reads thus making
more reliable the Annotation procedure.
For the four datasets of the SRP006674 project, we used the
default options in trimming step (since we have no adapters to
remove). No quality filtering was applied as described in Birzele et al.
3.2.2 Mapping “Which trimmed sequences should be mapped?” (default: Map all
Reads (also those without linkers)). We can define which sequences
have to be compared to genome after the trimming step choosing
among the available options. Indeed we can proceed with the anal-
ysis using all reads or we can filter out the reads on the bases of the
number of linker identified. In our example this option is irrelevant
and we leave the default value.
“Only consider Reads mapping with similarity” (default: 95 %). We
can define the cut-off value for minimum sequence similarity
between the reads and the genome.
“Only consider Reads mapping with coverage” (default: 90 %). We
can set the minimum overlap of the read with the genome. The
value can be set as percentage or as number of nucleotides referred
to the length of trimmed read.
“Allow a mismatch of up to N nt at the 5′end of the read”. This
parameter filters out the reads with more than N unaligned bases
at the 5′ end. “0” is the most restrictive option while “-1” does not
apply any restriction.
“Allow a mismatch of up to N nt at the 3′end of the read”. Like the

previous filter but referred to the 3′ end of the read.
“Are indels allowed?” We can define whether taking into account
also indels within the alignment.
“Only consider Reads with N multiple mappings” (default: 5). This
parameter defines the maximum number of times the same read is
allowed to map to the genome (useful to remove repeated ele-
ments). If we do not want to apply any filter on multiple mapping
we can set 0.
For our datasets we used the default values proposed by the
NGS-Trex system meaning that among the reads mapped with
TopHat, we selected those aligning to the Homo sapiens (NCBI
Build 36) genome with a similarity of at least 95 % and an overlap
of at least 90 % and having a number of multiple matches less than
5 over the genome. Indels have been accepted.
3.2.3 Annotation The annotation process is a multistep procedure to assign reads to

gene-specific clusters.
NGS-Trex performs the assignment with three degrees of con-
fidence “P”, “G”, and “T” (Fig. 2) on the basis of few parameters
reported in the annotation section of the analysis setup page.
A read is tagged as “proximal”—P—if it maps to a region
flanking the gene (surrounding) and extending as much as a
bounding limit. It is tagged as “genic”—G—if it overlaps gene
coordinates by satisfying some threshold values (described
below). Finally genic reads are aligned to RefSeq transcripts and
tagged as “transcript read”—T—if they show contiguous map-
ping to transcripts.
Fig. 2 Read to gene assignment schema reporting proximal reads as “P”, genic reads as “G”, transcript reads
as “T”, and extragenic reads as “O”. The bounding limit for “P” reads is labeled as “a” and the minimum over-
lap with the gene for “G” assignment is labeled as “b” (image modified from original by Boria et al. (2013))
More in details:
“Gene upstream/downstream region classify Read as ‘Proximal’”.
Bounding limit upstream and downstream a gene to classify a read
as “P”. Default value is 0 nt which means that “P” reads are not
considered.
“Minimum overlap to classify Read as ‘Genic’”. Minimum overlap
between a read and a gene to classify the read as “G”. Default value
(1 nt) is very tolerant as it implies that a read overlapping a gene
with a single nucleotide is labeled as “G”.
“Minimum similarity between Read and Transcript” (default
95 %). Similarity cut-off value between read and RefSeq sequence
to label the read as “T”.
“Extension to assign Reads to Transcript” (default: 10 nt). A Read
overlapping the RefSeq transcript at its 5′ or 3′ end can extend it
up to N nt. If the read extends the transcripts exceeding this
threshold it is not assigned to the transcript but it is treated as a
putative (extended) variant of the annotated transcript.
“Trim to assign Reads to Transcript” (default: 3 nt). The maxi-
mum number of allowed un-aligned nucleotide at reads ends.
Unlike previous filter, this one is restricted to reads not aligning
at RefSeq ends.
All sequences mapping outside gene coordinates beyond the
defined surrounding region are tagged as “O”—out of gene.
“Are sequences oriented?” If reads are oriented (default: N) the rela-
tive orientation between the read and the gene is used to optimize
the assignment process. This can help the assignment of multiple
mapping reads or overlapping genes which are two main problems
of the annotation process [8].
By selecting Y (default value) in “Try to solve ambiguities?”
field, NGS-Trex attempts to solve ambiguities by using relative
orientation of genes and reads and the confidence level of read
assignment, being the order from higher to lower “T”, ”G”, ”P”
(Fig. 2). See 6 for detail.
3.2.4 Sequences Once filled the analysis setup form as described, let us click the
Post-processing “Upload parameters” button and the “submit job to analysis queue”.
The progresses of each dataset analysis will be shown in the corre-
sponding status column.
3.3 Data Mining Upon the completion of the analysis process (required time
depends on the size of the datasets and on the load of the server),
the status column of the datasets table is set to “Done” enabling us
to explore the results.
A first overview of the obtained results is displayed as a sum-
mary table in the project main page. For each analyzed dataset
the table shows the total number of reads in the sample, the reads
average length, the number of mapping reads, and the number of

genes identified in the sample. More information can be obtained
from “Dataset” panel.
3.3.1 Statistics By clicking the “Statistics” link we can access—for each dataset—
several statistical summaries:
“Analysis Parameters”. Displays a table summarizing all the param-
eters used for the analysis.
“Mapping”. Flowchart showing the flow of reads among the dif-
ferent steps of the mapping procedure. This diagram makes it easy
to monitor the quality of the sequences, as it shows how many
reads do not map (or map with low quality) with the reference
genome. Information about rRNA like sequences, about repetitive
sequences, and spliced reads is also reported.
“Annotation”. Flowchart which describes reads assignment at
gene and transcript level.
“Results Overview”. Table with a summary of some information
about the dataset like the number of identified genes, transcripts,
new introns, and differentially expressed genes. In Table 2 the
example of results overview for the dataset “CD4_activated”.
Table 2
Results overview for dataset “CD4_activated” as reported in Statistic
page of NGS-Trex
Represented genes 19,688

Represented transcripts 24,463
Differentially expressed genes 15,366
Differentially expressed genes 13,956
(min number of reads mapping gene region = 10;
pValue < 0.05)
Differentially expressed genes (min number of reads 11,061
mapping known exons = 10; pValue < 0.05)
mapping gene region = 10; pValue < 0.01)
mapping known exons = 10; pValue < 0.01)
Differentially expressed introns (pValue < 0.05) 10,379
Differentially expressed new introns (pValue < 0.05) 132
Differentially expressed introns (pValue < 0.01) 5,169
Differentially expressed new introns (pValue < 0.01) 54
New introns 574
Canonic D/A sites 574
Fig. 3 “Mapping” diagram for the dataset “CD4_activated” as reported in the Statistics page of NGS-Trex
Let us explore the “Mapping” diagram for the dataset “CD4_

activated” (Fig. 3). Starting from 10,335,712 “Total Reads”,
10,335,712 “Processed Reads” were filtered by the preprocessing
step and next mapped to the Homo sapiens (NCBI Build 36)
genome. We obtained 7,093,241 “Mapped” reads satisfying map-
ping criteria (Subheading 3.2.2), 955,404 “Mapped low sim” reads
aligning with similarity and overlap values below the defined
thresholds, and 2,287,067 “Unmapped” reads for which no
matches have been found in the reference genome.
Before submitting them to the annotation process, the
“Mapped” reads were filtered to remove those sequences mapping
more than five times to the genome, “Mapped high freq”, and those
aligning to ribosomal RNAs, “Mapped to rRNAs”. The resulting
6,863,096 “Mapped low freq” sequences represent the total num-
ber of reads submitted to the annotation step. Among these
sequences 6,062,247 map to one single genome position, “Unique
mapped”, while 800,849 map to more than one genome position,
“Multiple mapped”. Moreover 144,733 reads show a spliced align-
ment: “Spliced”, while 6,718,363 map with ungapped match:
“Unspliced”.
As shown in the “Annotation” panel of the “CD4_activated”

dataset statistics, we obtained 698,448 reads mapping out of gene
coordinates, “Intergenic”, and 6,164,648 reads “Genic” reads
including 5,088,923 reads labeled “T”, 1,075,725 labeled “G”,
and 0 tagged “P” as described in Subheading 3.2.3.
3.3.2 Query Results Through the “Query Results” link of the project navigation bar
(top left side of the page), it is possible to obtain the most useful
information by performing several queries on data. Again the
“Query Result” page offers several panels that we will explore in
details.
In “Query Gene” panel we can search a single gene using either
the Entrez Gene Identifier (EntrezGeneID) or the HUGO
accession.
The result is a short summary showing the count of reads
assigned to the gene and the count of both annotated and new
introns identified by the analysis. Furthermore the differential
expression pattern of the selected gene within the different datasets
is provided.
In “Advanced Search” panel we can filter the genes with three
indexes: depth, coverage, and focus index. Depth is the maximum
number of overlapping reads assigned to a gene. Coverage is the
total number of reads assigned to the gene. Focus index is the ratio
between depth and coverage: if all reads are distributed along the
gene the depth value is lower than the coverage and the focus index
is low, if all reads are focused on the same region the depth value is
similar to the coverage value and the focus index tends to 1.
“Reads mapping within the gene boundaries but not overlapping
with known exons should be considered?” To carry out the search we
must also decide whether we want to use all reads assigned to genes
(G,P,T) answering “Yes” or (with a more restrictive approach) to
limit the search to reads labeled as “T” according to the previously
described classification (by answering “No”). “All selected samples
should satisfy criteria on the same gene?” option searches genes
satisfying the filtering criteria in all selected samples (“Yes”) or in
at least one of the selected samples (“No”).
For each gene the result table shows: the eg_id, the HUGO
name, the sequence coverage, the sequencing depth, the focus
index, and the RPKM value.
“DE Genes” and “DE Introns” panels are very similar. They allow
the query differentially expressed genes between a reference and
the others datasets. We need to select the reference sample, pValue,
minimum number of reads, and relative expression level (i.e., we
can select genes o introns which are overexpressed in reference
dataset). As for the advanced search it is possible to limit the query
Fig. 4 Relative position of reads with respect to CDS and UTRs regions (Reads 5′ UTR = A, Reads 5′ UTR-
CDS = D, Reads CDS = B, Reads 3′ UTR-CDS = E, Reads 3′ UTR = C) in Transcripts panel (image modified from
original in NGS-Trex web page)
to reads assigned to annotated transcripts (T) or to use all reads

assigned to the gene (G,T,P). DE introns can also be queried on
specific donor acceptor sites. “Exclude if whole gene shows same
pattern?” allows to hide differentially expressed introns if the dif-
ferential expression is exhibited also at gene level (see Note 3).
Output table shows several useful information such as raw
number of reads, fold change value, pValue. If the output refers to
DE splice site also genomic coordinates and donor/acceptor sites
are shown.
“New introns”. With this query form we can obtain a list of unan-
notated splicing events. In addition to the basic parameters
described for the other panels, we can select whether to consider
only splice sites with canonical donor and acceptor dinucleotides.
The resulting table shows the chromosomal location (chromo-
some name, chromosome start, chromosome end), the strand, the
eg_id and the HUGO name of the gene it belongs to, the sequence
around the 5′ splice site (Donor) and the 3′ splice site (Acceptor),
and the donor/acceptor dinucleotides (D/A).
“Transcripts”. Transcripts query panel allows us to investigate the
relation between reads and known transcripts by identifying the
relative position of reads with respect to CDS and UTRs regions.
The most critical parameter is “Which reads should be considered?”
form. We can select among six different options as shown in Fig. 4.
Using this tool we get a table reporting for each transcript: the
eg_id and the HUGO name of the gene, the RefSeq accession
number identifying the transcript, the rna length, the rna strand,
the coding sequence location on the transcript (cds start and cds
end), the number of reads supporting the transcript, and the map-
ping position of reads related to the transcript.
We can also limit the result only considering reads aligning
with reference transcript with no indels.
For all query panels the results are shown as tables and can be
downloaded as tab-delimited text file clicking the “Download
Table” button at the top right of the “results” page. We can also
visualize our data through a simple genome browser accessible via
the “eg_id” link corresponding to the genes shown in tables and

simply download the reads assigned to a gene or mapped onto a
genomic region.
For example if we want to identify genes overexpressed in
“TREG_activated” dataset compared to the other datasets we can
use the “DE Genes” panel. Specifically we select all the samples in
the datasets table of the project main page, click the “Query
Results” link of the project navigation bar, and select the “DE
Genes” panel.
We set “TREG_activated” as “Reference Sample” and
“Significantly more frequent in Reference” as “Gene expression”
pattern. Moreover we choose “0.05” as “pValue” threshold and
“1,000” as “Minimum Reads number” and we select “Y” in “Reads
mapping within the gene boundaries but not overlapping with known
exons should be considered?”
By submitting the query we obtain a table reporting 1,148
genes differentially expressed in the reference sample. If we order
the table by descending fold change (fc) value and we consider the
top five listed genes, we find that three (CTLA4, IL2RA, IKZF2)
out of them have already been reported by Birzele et al. as Treg-
specific markers.
4 Notes
1. We downloaded RNA-Seq datasets in fastq format from ENA

at EBI since NGS-Trex does not deal with SRA files.
2. Together with nucleotide sequences, NGS platforms typically
produce a per-base quality score which provides the probability
that a given base is called incorrectly by the sequencer. The
quality scores range from 0 to 40, where higher numbers
report more confident base calls.
3. The probability of a gene X to be differentially expressed
within the reference dataset A compared to dataset B is com-
puted by the cumulative hypergeometric distribution as
follows:
pValue = PN , n; M , m = k = mnnkN - nM - kNM
Where N is the total number of reads supporting genes within

the two datasets, n the total number of reads supporting the
gene X within the two dataset, M the number of reads sup-
porting genes within the dataset B, and m the number of reads
supporting the gene X within the dataset B. Overrepresentation
is defined to be significant when pValue is less than 0.01 or
than 0.05. The Benjamini–Hochberg procedure is used to
compute adjusted pValues.
References
1. Wang RL, Biales A, Bencic D, Lattier D, Kostich computational research in the life sciences.
M, Villeneuve D, Ankley GT, Lazorchak J, Toth Genome Biol 11(8):R86
G (2008) DNA microarray application in eco- 5. Halbritter F, Vaidya HJ, Tomlinson SR (2011)
toxicology: experimental design, microarray GeneProf: analysis of highthroughput
scanning and factors affecting transcriptional sequencing experiments. Nat Methods 9:7–8
profiles in a small fish species. Environ Toxicol 6. Boria I, Boatti L, Pesole G, Mignone F (2013)
Chem 27:652–663 NGS-Trex: Next Generation Sequencing
2. Wang L, Feng Z, Wang X, Zhang X (2010) Transcriptome profile explorer. BMC
DEGseq: an R package for identifying differen- Bioinformatics 14(Suppl 7):S10
tially expressed genes from RNA-Seq data. 7. Birzele F, Fauti T, Stahl H, Lenter MC, Simon E,
Bioinformatics 26:136–138 Knebel D, Weith A, Hildebrandt T, Mennerich D
3. Mutz KO, Heilkenbrinker A, Lönne M, Walter (2011) Next-generation insights into regulatory T
JG, Stahl F (2013) Transcriptome analysis using cells: expression profiling and FoxP3 occupancy in
next-generation sequencing. Curr Opin Human. Nucleic Acids Res 39(18):7946–7960
Biotechnol 24(1):22–30 8. Costa V, Angelini C, De Feis I, Ciccodicola A
4. Goecks J, Nekrutenko A, Taylor J, Team G (2010) Uncovering the complexity of transcrip-
(2010) Galaxy: a comprehensive approach for tomes with RNA-Seq. J Biomed Biotechnol
supporting accessible, reproducible, and transparent 2010:1–20
Chapter 16
e-DNA Meta-Barcoding: From NGS Raw Data

to Taxonomic Profiling
Fosso Bruno, Marzano Marinella, and Monica Santamaria
Abstract
In recent years, thanks to the essential support provided by the Next-Generation Sequencing (NGS)
technologies, Metagenomics is enabling the direct access to the taxonomic and functional composition of
mixed microbial communities living in any environmental niche, without the prerequisite to isolate or
culture the single organisms. This approach has already been successfully applied for the analysis of many
habitats, such as water or soil natural environments, also characterized by extreme physical and chemical
conditions, food supply chains, and animal organisms, including humans. A shotgun sequencing approach
can lead to investigate both organisms and genes diversity. Anyway, if the purpose is limited to explore the
taxonomic complexity, an amplicon-based approach, based on PCR-targeted sequencing of selected
genetic species markers, commonly named “meta-barcodes”, is desirable. Among the genomic regions
most widely used for the discrimination of bacterial organisms, in some cases up to the species level, some
hypervariable domains of the gene coding for the 16S rRNA occupy a prominent place.
The amplification of a certain meta-barcode from a microbial community through the use of PCR prim-
ers able to work in the entire considered taxonomic group is the first task after the extraction of the total
DNA. Generally, this step is followed by the high-throughput sequencing of the resulting amplicons libraries
by means of a selected NGS platform. Finally, the interpretation of the huge amount of produced data
requires appropriate bioinformatics tools and know-how in addition to efficient computational resources.
Here a computational methodology suitable for the taxonomic characterization of 454 meta-barcode
sequences is described in detail. In particular, a dataset covering the V1–V3 region belonging to the bacte-
rial 16S rRNA coding gene and produced in the Human Microbiome Project (HMP) from a palatine
tonsils sample is analyzed. The proposed exercise includes the basic steps to manage raw sequencing data,
remove amplification and pyrosequencing errors, and finally map sequences on the taxonomy.
Key words Metagenomics, e-DNA, Meta-barcoding, Microbiome, Taxonomy, Pyrosequencing,

16S rRNA
1 Introduction
Metagenomics is today one of the most advanced and fast growing

research areas. It has opened an unprecedented comprehensive
window on the microbial world, which can now be revealed also
in all those parts which remained obscure due to the inability to
257
258 Fosso Bruno et al.
cultivate in the laboratory about 99 % of environmental species.

Metagenomics bypasses these limits via the direct extraction and
investigation of the total genetic material from any environmental
niche without the preliminary need to isolate the individual
organisms.
The recent amazing advances in next generation sequencing
(NGS) technologies have strongly supported the development of
this methodological approach by allowing the analysis of increas-
ingly large and complex microbial communities through the pro-
duction of massive amounts of sequence data in a short time and at
lower cost. To date the NGS-based metagenomics has been applied
for multiple purposes, such as inferring the health status of envi-
ronments through the analysis of their microbial biodiversity [1],
studying the human microbiome in physiological and pathological
conditions [2], analyzing the human nutrition, following the
food’s path, from production to digestion, and its impact on
human physiology [3], and finally discovering novel biomolecules
and microorganisms that could be fruitfully introduced in indus-
trial biotechnological processes [4, 5].
Even if in the last years Illumina and, to a lesser extent, SOLID
NGS platforms are gaining increasing popularity in metagenomic
projects due to their high level of throughput, Roche 454 is still
the most competitive platform thanks to the length of its reads [6]
which, in the latest version GS FLX Titanium XL +, grow up to
∼1,000 bp. In this regard it should be noted that the length of the
produced sequences has a significant influence on the accuracy of
their next annotation, in particular in situations where reliable ref-
erence sequences belonging to the same species residing in the
sample are not available. In this case 454 longer reads allow obtain-
ing better results mainly based on the comparison with ortholo-
gous genomes. On the other hand the higher output per run of
Illumina or SOLID could provide enough coverage to analyze
more complex samples and better detect the rare variants.
A common shortcoming in the protocols of all the above-
mentioned NGS technologies is represented by the errors, intro-
duced during the amplification step, which could lead to an
overestimation of biodiversity by creating false variant sequences.
Also a certain intrinsic degree of inaccuracy of the sequencing
chemistry contributes to this bias. Recently, several studies have
tried to assess the sequencing errors and artifacts produced by the
different NGS platforms in metagenomic studies and to compare
their overall ability to describe the original microbial community.
For example, the main bias associated with Roche 454 is due to
the generation of errors in homopolymers (three or more identi-
cal consecutive bases). The Illumina protocol seems not to have
the same limitation although even this platform is affected by
several systematic errors in base calling. Anyway, Luo et al. [7]
have shown that the Roche 454 FLX Titanium and the Illumina
Meta-Barcoding Taxonomic Profiling 259
Genome Analyzer (GA) II, despite their obvious differences in

read length and sequencing protocols, provide a comparable
overview of the biodiversity of the analyzed sample.
Organisms and gene functions of the microbiome living in a
specific environment can be both addressed by a shotgun sequenc-
ing of its entire metagenome. Conversely, if the purpose of the
analysis is limited to investigate the taxonomic complexity of the
microbial community, an environmental DNA (e-DNA) meta-
barcoding [8] approach could constitute an effective and less
expensive solution. It involves the culture-independent PCR-
targeted sequencing of a selected genetic taxonomic marker
(meta-barcode) that should be amplified by means of universal
primers and cycles conditions in virtually all the species belonging
to the explored taxonomic range.
An optimal meta-barcode should possess three major features:
(1) be ubiquitous in the group of organisms of interest (e.g.,
Bacteria, Fungi), (2) its sequence structure should be marked by
domains with very different degrees of variability (hypervariable
regions, able in discriminating at lower taxonomic ranks, flanked
by primers-annealing very conserved regions) and finally, (3) its
size should be compatible with the sequences produced by the
most recent versions of NGS platforms.
According to these requirements, selected regions of the inter-
nal transcribed spacer (ITS) of the ribosomal RNA (rRNA) gene
cluster and of the 16S rRNA gene are commonly used to identify
fungal and bacterial taxa, respectively [9–12]. Depending on the
selected region and on its length, the analysis will allow to reach a
certain depth in the taxonomic classification. For instance, the anal-
ysis of partial 16S rRNA gene sequences provides a direct and rela-
tively inexpensive access to bacterial communities biodiversity
achieving, in the case of the recent versions of 454 technology, a
resolution down to the genus level [13]. 16S rRNA surveys provide
information about relative taxa abundance too [14]. A limitation of
the 454 pyrosequencing is that the amount of obtained partial 16S
rRNA gene sequences might be inadequate to represent the entire
biodiversity of the environmental sample. In particular, the rare
species may be difficult to detect. Illumina overcomes this problem
thanks to its higher throughput but at the same time the shorter
length of its reads can adversely affect the accuracy of the assign-
ment to the appropriate taxonomic classes [13].
Taking into account the current progress of biomolecular
technologies, the potential power of e-DNA meta-barcoding is
limited basically by two aspects. The first one is its dependency on
PCR which, unfortunately, can intrinsically introduce errors as
substitutions and/or insertions/deletions and bias due to the dif-
ferent propensity of universal primers to anneal to the same
genomic region in different taxa. In addition, the rare species could
be overwhelmed by the most abundant ones, which are favored
during the amplification step. The second aspect concerns some

issues regarding the bioinformatic analysis of the large amount of
amplicons sequences produced through NGS and the computa-
tional challenges it entails.
Currently, the bioinformatic methods used to assign the
metagenomic sequences to their taxonomic classes [15] adopt
essentially three approaches:
●● In similarity-based methods, the taxonomic characterization of
a metagenomic dataset is commonly based on its comparison
to curated collection of sequences with known function and
origin [10]. These methods offer high accuracy, even for short
sequences, but they suffer three general shortcomings: (1) the
time needed for the similarity analysis grows with the size of
the chosen reference database, (2) the absence of reference
sequences for unknown species may affect the assignment
accuracy [16], and (3) the availability of comprehensive and
well annotated reference database requires considerable invest-
ments [8]. Among the largest reference resources currently
available for rRNA gene-based identification of microorgan-
isms there are Greengenes [17], Ribosomal Database Project
II (RDP II) [18], and SILVA [19]. ITSOneDB has been
recently developed as reference resource for meta-barcoding
analysis based on ITS1, a specific intergenic region of rRNA
gene cluster in Fungi [10].
●● In composition-based methods, several genome features, such as
GC content, codon-usage, or oligonucleotides frequencies,
are used to assign the metagenomic sequences to taxonomi-
cally annotated reference genomes. The reference data are
modelled using different methods, such as the interpolated
Markov models (IMMs), the naïve Bayesian classifiers, and the
k-means/k-nearest-neighbor algorithms. The principal short-
coming of these approaches is the high requirement of com-
putational power and time for the models building step.
Despite this phase, the metagenomic classification is generally
faster than that performed by means of similarity-based meth-
ods. Finally, composition-based methods offer a good classifi-
cation accuracy, even for short sequences, that can be further
improved using some statistical methods (e.g., bootstrap or
cross-validation).
●● In phylogenetic-based methods, the metagenomic sequences are
placed onto a reference phylogenetic tree according to a model
of evolution such as that implemented in maximum likelihood
(ML), Bayesian or neighbor-joining (NJ) methods. Nowadays,
although these methods promise a high accuracy, the compu-
tational demands for their application is often too high.
●● Once meta-barcodes have been sequenced by means of NGS
technologies, before applying one of the methods listed above
to assign the reads to their original taxonomic ranks, many

other different tools should be used for filtering and denoising
the raw data [13, 14, 16]. These preliminary bioinformatic
steps are essential in order to obtain a consistent taxonomic
inference.
Here, a simple and comprehensive protocol for the analysis of
NGS reads from 16S rRNA gene is described in all its steps from
the raw data to the taxonomic classifications of bacterial organisms.
The procedure refers, in particular, to the analysis of data produced
by the 454 platform but it can be applied, except for the denoising
step, to Illumina sequences too.
2 Materials
2.1 HMP (Human The proposed protocol is here executed by starting from a publicly
Microbiome Project) available dataset of bacterial 16S rRNA gene sequences stored in
Data Set the NCBI Sequence Read Archive (SRA) collection. In particular,
the considered dataset has been produced by the Human
Microbiome Project (HMP) in which several body sites, including
gastrointestinal and female urogenital tracts, oral cavity, nasal and
pharyngeal tracts, and skin, were studied by pyrosequencing the
V1–V3 and V3–V5 hypervariable regions of 16S rRNA gene. The
SRR331235 datarun covering the V1–V3 region obtained from
palatine tonsils sample of a male subject was chosen to be submitted
to the exercise described below. The forward primer ATTACCG
CGGCTGCTGG was used for the unidirectional pyrosequencing
as reported in SRA collection experiment page at the link http://
www.ncbi.nlm.nih.gov/sra?term=SRR331235. Moreover a nine-
nucleotide tag (TCGAGGAAC) was inserted upstream of the
V1–V3 region [20] in order to discriminate the reads belonging to
the different samples concurrently sequenced in the same run.
The SRR331235 datarun was downloaded as a binary sra file
of about 5.3 Mb. The dataset consists of 4,504 reads with an aver-
age length of 535.80 nt.
Since the data provided by the 454 technology are in the
binary Standard Flowgram Format (SFF), the downloaded sra file
was converted in the sff by using the sratoolkit (see Note 1). The
aim of this step was to reproduce in the exercise the case in which
the researcher has to analyze directly the data produced by this
type of platform.
The information included in the sff file can be visualized using
the Roche SFFtools by converting it in a textual flat file, as extensively
described in the next section. In particular, two principal sections
can be found in this file: a common header section including some
general information about the sequencing output (such as the
number of reads stored in the file and the number of nucleotide
flows performed during the sequencing) [21] and a read-specific
section for each of the generated sequences. In the read-specific section

the following information are stored:
●● The Flowgramm field provides information about the light sig-
nal measured for each nucleotide flow (e.g., a 0.0 signal is mea-
sured if there is not nucleotide incorporation while a 1.0 signal
indicates just a single nucleotide addition) [21];
●● The Flow Index field contains the progressive index associated
to each nucleotide flow providing base addition in the called
sequence;
●● The Bases field contains the nucleotide sequence;
●● The Quality Scores field contains information about the accu-
racy of the sequencing. The quality score is estimated for each
of the bases in the sequence and its reported values are in
Phred-score format and represent the −10log10 of the proba-
bility that there is a base-call error (e.g., a Phred-score of 30
means that there is a probability of 0.001 that the called base
is wrong) [22].
During the protocol computation two other common biologi-
cal formats are produced:
●● The fasta file format is a common standard in computational
biology consisting in a description line that starts with a “>”
character followed by one or more lines in which the sequence
is reported. An example of fasta format is shown below:
>GMF8DCB01AT97I
ATTACCGCGGCTGCTGGCACGTAGTTAGCCGTGGCTTTCTGGT
TAGATACCGTCAAGGCACACAGGGGATAGG
●● The fastq file format stores both the sequence and the base-call
quality information in a single file. Each sequence is described
by four lines: a description line starting with “@” digit report-
ing the sequence accession number, a single line containing the
sequence, a third line always starting with a “+” character and
the ASCII encoded quality-score line. Below an example is
shown:
@GMF8DCB01AT97I
ATTACCGCGGCTGCTGGCACGTAGTTAGCCGTGGCTTTCTGGT
TAGATACCGTCAAGGCACACAGGGGATAGG
+
IIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIHHHIIIII
IIIIIIIIIIIIIIIIIIIIC5555=GF@@
2.2 Computational The management and the processing of large metagenomic datasets
Hardware require powerful storage and computational capabilities therefore
Linux and Unix-based machines are recommended.
The storage of data from a typical 454 GS FLX Titanium run

usually takes up to 2 GB, and then it is necessary to consider addi-
tional free space for the analysis. Just think that by performing the
proposed protocol starting from a flat sff file of about 14 MB,
nearly 150 MB of data are produced just during the first step of the
analysis, the denoising phase.
With regard to the computational capabilities it is reasonable
to carry out all the analysis on a multi-processor/core machine
with at least 20 GB of RAM.
2.3 Computational The protocol described below includes the use of the following
Software software:
1. SFFtools, a part of Roche Data Analysis package, is not freely
available but 454 users can request it from http://454.com/
products/analysis-software/index.asp. In order to use
SFFtools to manage the sff file, the installation of the whole
Roche Data Analysis package is required on a Linux operating
system by using root privileges.
2.
AmpliconNoise, available at http://code.google.com/p/
ampliconnoise/, which is a collection of programs used to
reduce the base errors intrinsically introduced during the 454
amplicons sequencing procedure [23]. It is supported on
Linux and Mac OS X platforms but a particular attention is
required during the installation and the configuration of the
package because it is required the setting of some environment
variables (as fully described in the documentation file provided
with the installation package). Moreover, AmpliconNoise is
designed to work on multi-processors/cores system using the
Message Passing Interface (MPI) libraries that must be previ-
ously installed and configured. The MPI libraries allow paral-
lelising the process on multi-processor/core machines for the
High Performance Computing (HPC).
3. FastQC, a Java-based tool, which is available at http://www.
bioinformatics.babraham.ac.uk/projects/fastqc/ and used for
a simple and quick quality check of sequences outputted
directly by the sequencer (also referred as raw data).
4. RDP classifier is a naïve Bayesian classifier for fast taxonomic
assignment of sequences and confidence estimation
(Assignment procedures and confidence values estimation are
amply described in Subheading 3.3). It can be downloaded
from http://rdp-classifier.sourceforge.net.
5. Python, an object-oriented programming language available at
http://www.python.org. It is preinstalled in several Linux-based
Operating Systems (e.g., Ubuntu) and in Mac OS X but not in
Windows Operating Systems. During the proposed exercise
two Python modules are used: BioPython (http://biopython.
org/wiki/Main_Page) and Numpy (http://www.numpy.org).
In order to correctly follow the protocol, the previous listed

modules must be installed and operating properly. It is impor-
tant to note that the Python modules installation procedures
may change depending on the operating system and it is rec-
ommended to follow the installation instruction available in
the aforementioned module Web page.
3 Methods
Conceptually we can define three main steps, listed below, in the

taxonomic assignment of meta-barcodes, previously amplified from
the microbial community and sequenced by 454 technology.
●● Raw data management and evaluation: the 454 raw data are
converted in file format suitable for the subsequent analysis;
●● Denoising: the AmpliconNoise suite is applied to remove the
errors, introduced in the sequences by both amplification and
pyrosequencing processes, which could leads to incorrect
inference of microbiota diversity;
●● Assignment of cleaned sequences to their taxonomic group.
The entire protocol can be performed on a bash terminal. The
operations performed directly into the bash environment are pre-
ceded by the “$” prompt while those executed into the Python
shell are specified by the “>>>” prompt.
3.1 Raw Data The first step of the protocol consists in the conversion of the
Management binary sff file produced by the sequencing platform 454 into two
and Evaluation different file formats, a textual flat file and a fastq file, and in the
statistical evaluation of the raw data.
The conversion of the sff file into a flat textual file is performed
by the Roche SFFtools software. The produced file will be submit-
ted to the denoising step. Some alternative methods that are shown
in Note 2 are available to manage the sff file.
The command used in the bash terminal to achieve this con-
version is the following:
$ sffinfo SRR331235.sff > SRR331235.sff.txt
The sff is also converted into a fastq file suitable to perform the
raw data quality check by means of FastQC. This operation is car-
ried out in the Python shell by using the BioPython module and
consists of two steps:
(a) the invocation of the BioPython SeqIO command class, which
is useful to manage biological data files, by using the following
commands:
$ Python
And then:
>>>from Bio import SeqIO
(b) the sff file conversion in a fastq file by using the command
SeqIO.convert that requires four instructions: the file to be
converted, its format, the output file name and the final file
format. The command to perform this conversion is the
following:
>>>SeqIO.convert(“SRR331235.
sff”,”sff”,”SRR331235.fastq”,”fastq”)
Afterwards, the produced fastq file is processed by means of
FastQC, by using the following command:
$ fastqc SRR331235.fastq
FastQC performs several statistical evaluations on the sequenc-
ing data and builds an html report file divided in 12 subsections.
The descriptions of the latter are included in several different html
files (you can open them using any Internet browser) contained in
the “help” folder, available within the FastQC package.
For example, in one of the 12 subsections obtained in this
exercise, the “Basic Statistics,” the dataset under investigation
results to consist of 4,504 sequences ranging from 86 to 1,049
nucleotides and with a GC content of 52 %. Furthermore, in the
“Per base sequence quality” subsection it is shown a chart of the
quality values ranges across all bases at each sequence position in
the fastq file related to the analyzed dataset (Fig. 1). In Fig. 1 it is
possible to observe a reduction of the quality at the 3′ end. This
behavior is due to the intrinsic limits of the sequencing technology
which tends to lose sensitivity in the last steps of the process.
3.2 Denoising The second step of the entire protocol, called denoising, consists in
the application of AmpliconNoise. AmpliconNoise is a suite of
programs designed to reduce the errors introduced in the sequences
by both amplification and pyrosequencing before mapping them
on the taxonomy.
In order to exhaustively explain all the procedures of this
phase, the manual procedure is shown but it is possible to perform
them by means of bash scripts that are supplied in the installation
package. Moreover, AmpliconNoise can be alternatively applied
within the Mothur suite [24] (see Note 3).
The processing starts with a filtering step applied to remove
low quality reads using the Perl script FlowMinMax.
In the bash terminal, type:
$ FlowsMinMax.pl TCGAGGAACATTACCGCGGCTGCTGG
test_book < SRR331235.sff.txt
40
38
36
34
32
30
28
26
24
Quality score
22
20
18
16
14
12
10
8
6
4
2
0
1 2 3 4 5 6 7 8 9 15-19 30-34 45-49 70-79 100-149 250-299 400-449 600-699 900-999
Position in read (bp)
Fig. 1 The per base sequence quality chart produced by FastQC for the dataset analyzed in this exercise (datarun
SRR331235 produced by the Human Microbiome Project (HMP))
FlowMinMax requires two arguments, the tag (TCGAGGAAC),

eventually added in order to discriminate the sequences belonging
to different samples processed in the same sequencing run, fol-
lowed by the primer(s) sequence (s) (ATTACCGCGGCTGCTGG),
used for the sequencing of the selected meta-barcode, and an out-
put argument (test_book) to process the flat textual sff file derived
from the previous data conversion step. FlowMinMax works on
both sequences and flowgrams that generated them. As regards the
analysis of sequences, an exact match is searched between each of
them and both the primer and the specific tag. If the match is not
found the read is discarded. As regards the flowgrams, FlowMinMax
verifies if there are ambiguous flow signals or if no signals are reg-
istered for all the four nucleotide in the region corresponding to
the first 360 flows of pyrosequencing where reliability should be
higher [20]. If one of these two conditions is verified the corre-
sponding read is discarded. Moreover all flowgrams are truncated
at 720th flow (if the 454 GS FLX Titanium sequencing technology
has been adopted) due to an evident decrease in the quality of the
sequences after this point [25].
After the application of the filters described above,
FlowMinMax provides two output files: test_book.dat in which the
retained flowgrams are collected and test_book.fasta containing

the corresponding sequences.
In this exercise 3,229 flowgrams and the related sequences
pass the filtering step. In particular, in the test_book.dat file the
following information are reported: in the first line the number of
flowgrams which passed the filter followed by the number of base
flows retained after the cutoff procedure (in this case 3,229 flow-
grams and 720 flows) are reported. In the subsequent lines the
reads accession numbers, their corresponding flowgrams and num-
ber of flows are shown.
The second step of AmpliconNoise, called dereplication, con-
sists in removing the redundant sequences in order to get a more
slender dataset and reduce the computational power required for
the next bioinformatic steps.
The dereplication command to type on the bash terminal is the
following:
$ FastaUnique -in test_book.fa > test_
book_U. fa
FastaUnique joins all the identical sequences and returns a
file_test_book_U.fa in which the fasta description line of each
“representative” sequence is followed by a number indicating the
amount of sequences identical to it. An example of line obtained in
this exercise, is:
> GMF8DCB01BB75Z_2
where “2” represent the number of identical sequences including
the “representative” one.
In the case of the dataset used for the exercise, FastaUnique
identifies 1,636 unique sequences.
Moreover, the information about the dereplication is stored in
a map file (test_book.map) in which for each line is reported the
accession number of the “representative” sequence followed by the
accession number of all the reads identical to it.
Example: In the exercise case, the dereplication information
related to the unique sequence GMF8DCB01BB75Z_2 is the
following:
330,154,GMF8DCB01BB75Z,2:330,2965:GMF8DCB01B
B75Z,GMF8DCB01B09TA
In particular, several data are stored in the shown line:
●● 330 is the original ID of the representative sequences. Simply
it means that in the test_book.fa file GMF8DCB01BB75Z was
the 330th sequence;
●● 154 is the new ID associated to the representative sequence
and it means that in the test_book_U.fa file produced by
FastaUnique, GMF8DCB01BB75Z is the 154th sequence;
●● 2 is the number of sequences that are represented by
GMF8DCB01BB75Z;
●● 330 and 2965 are the original ID of the two sequences

represented by GMF8DCB01BB75Z;
●● GMF8DCB01BB75Z and GMF8DCB01B09TA are the
accession number of the joined sequences.
The unique sequences reported in the fasta file (test_book_U.
fa) produced by FastaUnique are then clustered according to their
mutual similarity typing three different commands in the bash ter-
minal. This process lead to reduce the computational weight in the
subsequent removal of pyrosequencing errors.
The three commands to be sequentially entered are:
$ mpirun -np 8 NDist -i -in test_book_U.fa >
test_book_U_I.ndist
$ FCluster -a -w -in test_book_U_I.ndist -out
test_book_U_I > test_book_U_I.fcout
$ SplitClusterEven -din test_book.dat -min test_
book.map -tin test_book_U_I.tree -s 2500 -m 250
> test_book_U_I.stats
Ndist requires two arguments, the fasta input file “-in” and the
“–i” flag to print the output in a format suitable for the FCluster
running. The exact Needleman–Wunsch algorithm is implemented
in NDist to perform pairwise alignments with a fixed gap penalty
score of 1.5. Then, the program produces a distance file (test_
book_U_I.ndist) by applying the QuikDist algorithm [26] which
assigns to both the mismatch and the gap a penalty score of 1. The
genetic distance calculated on the basis of these computed differ-
ences for each pair of sequences is normalized both for the align-
ment length of each pair of sequences and the total number of
comparisons performed in the dataset.
FCluster performs a hierarchical clustering starting by the dis-
tance file produced by NDist (-in). As default, it performs a com-
plete linkage clustering. This is a hierarchical clustering method
that sets the distance between two clusters as the maximum dis-
tance observed between the cluster members.
In the exercise case, sequences that are the result of the derep-
lication steps are considered and, for this reason, an average linkage
accounting for the amount of sequences represented by the unique
one is used, by typing the arguments, “–a –w”. In the average link-
age clustering procedure the distance between two clusters corre-
sponds to the average distance between the cluster members.
Three output are produced by FCluster:
●● a file (test_book_U_I.list) in which the genetic distances which
guide the clustering are listed. In particular, for each value of
distance the corresponding number N of formed clusters is
reported and for each cluster the number of included sequences
is indicated;
●● a file (test_book_U_I.otu) in which the clusters sizes, measured

at a distance separation of 0.01(e.g., 0.00, 0.01, 0.02 and so on)
by the QuickDist algorithm implemented in Ndist, are stored;
●● a file (test_book_U_I.tree) where clusters are mapped on a
hierarchical tree in newick format.
SplitClusterEven splits the flowgrams (-din), stored in the
test_book.dat file by using as reference the map file test_book.map
(-min), both produced by FlowMinMax, and by following the
hierarchical tree (-tin) produced by FCluster. The arguments “–s”
and “–m” control the cluster maximum and minimum dimension.
The data for each cluster are stored in different folder labelled as
C000,….C00N+. In the exercise case just one cluster is generated
(C000) using default “–s” and “–m” parameters.
In order to reduce the computational time, in this exercise the
MPI libraries (mpirun) are used and FCluster is applied on eight
nodes (-np argument) (see Note 4).
The clustered data obtained in the previous steps are then pro-
cessed in order to decrease the pyrosequencing errors noise.
The commands to type are:
$ mpirun -np 8 PyroDist -in C000/C000.dat -out
C000/C000 > C000/C000.fout
$ FCluster -in C000.fdist -out C000_X > C000.fout
$ mpirun -np 8 PyroNoiseM -din C000.dat -out
C000_s60_c01 -lin C000_X.list -s 60.0 -c 0.01 >
C000_s60_c01.pout
In order to reduce the pyrosequencing noise, AmpliconNoise
tries to estimate the “real sequences” and their relatives abundance
directly from the raw data. For this reason, it starts with an initial
clustering, performed by PyroDist and FCluster. In particular,
PyroDist computes a distance matrix, working directly on flow-
grams by taking in input the dat file (-din), in order to produce
data suitable for FCluster computation (C000.fout). FCluster per-
forms an average linkage clustering on the distance matrix pro-
duced by PyroDist (see before for the software details).
Then, the procedure continues by applying the Expectation
Maximization algorithm (EM) [27] implemented in
PyroNoiseM. The EM algorithm is based on the idea of maximiz-
ing the probability that a flowgram in a cluster was generated from
the estimated real sequence for that cluster. Accordingly, the algo-
rithm moves the flowgrams among the clusters and re-computes
the probability. The EM computation ends when no more shifts
are allowed because the estimated probability is already maxi-
mized. In the PyroNoiseM command-line the flowgrams to be
processed are indicated using the “-din” flag while the clustering
data produced by FCluster on PyroDist data are supplied with the
argument “-lin”. Two parameters guide the clustering procedure:
“-c” is the cutoff for the initial clustering step and “–s” rules the
cluster size (60.0 is a good optimization between noise removal
and computational requirements). At the end of the PyroNoiseM
processing, a number of files labelled with the “–out” parameter
(C000_s60_c01) are produced:
●● a fasta file where the denoised sequences associated with
their accession numbers, called as output_label_index_weight,
are reported. Weight indicates the number of raw sequences
associated to each denoised one and index is just a number
which identifies univocally the specific denoised sequences
(C000_s60_c01_cd.fa);
●● a quality file for the denoised sequences (C000_s60_c01_
cd.qual). It is a fasta-like format file where, for each denoised
sequence, the accession number and the estimated Phred-score
for each base is annotated;
●● a mapping file in which the accession numbers of the raw reads
are associated with their corresponding denoised sequences.
(C000_s60_c01_cd.mapping);
●● a directory (C000_s60_c01) containing a fasta file for each
denoised sequence. In each file the unique sequences, pro-
duced by the Perl script FastaUnique, and the corresponding
reads are listed.
The sequence and mapping data produced by PyroNoiseM are
then copied and renamed in the test_book_pyro_denoised.fa and
test_book_pyro_denoised.mapping files, respectively, using the
following commands:
$ cp C000/C000_s60_c01_cd.fa > test_book_pyro_
denoised.fa
$ cp C000/C000_s60_c01.mapping > test_book_
pyro_denoised.mapping
Subsequently, the denoised sequences stored in the test_book_
pyro_denoised.fa file are truncated at 400 nt (see the AmpliconNoise
documentation) by typing:
$ Truncate.pl 400 < test_book_pyro_denoised.fa
> test_book_pyro_denoised _T400.fa
The sequences contained in the produced test_book_pyro_
denoised _T400.fa file are processed in order to remove the barcode
and the primer sequences before the SeqNoise application by means
of sed, a stream editor which works using regular expressions.
In the exercise case, it is possible to substitute (s) the barcode
and the primer at the 5′ end (^) of each sequence with an empty
string by typing:
$ sed 's/^TCGAGGAACATTACCGCGGCTGCTGG//' test_
book_pyro_denoised _T400.fa > test_book_pyro_
denoised _T400_P.fa
The test_book_pyro_denoised _T400_P.fa file is

produced as output.
The last three steps of the denoising procedure are applied to
remove the amplification errors noise. The corresponding com-
mands to enter in the terminal are:
$ mpirun -np 8 SeqDist -in test_book_pyro_
denoised_T400_P.fa > test_book_pyro_denoised _
T400_P.seqdist
$ FCluster -in test_book_pyro_denoised_T400_
P.seqdist -out test_book_pyro_denoised_T400_P >
test_book_pyro_denoised _T400_P.fcout
$ mpirun -np 8 SeqNoise -in test_book_pyro_
denoised_T400_P.fa -din test_book_pyro_denoised_
T400_P.seqdist -lin test_book_pyro_denoised_
T400_P.list -min test_book_pyro_denoised.map-
ping -out test_book_denoised_T400_s25_c08 -s 25.0
-c 0.08 > test_book_pyro_denoised_T400_s30_c08.
snout
As discussed before for pyrosequencing noise reduction, the
amplification errors are removed by using an initial clustering, per-
formed by SeqDist and FCluster, to infer the “real sequences” and
then by applying an EM algorithm implemented in SeqNoise.
SeqDist works directly on the sequences, outputted by the
PyroNoiseM computation and then truncated using the Perl scripts
Truncate.pl, in order to generate a distance matrix (test_book_
pyro_denoised.seqdist) in a format suitable for FCluster. It works
using an implementation of Needleman–Wunsch algorithm to per-
form pairwise alignment starting from a fasta file containing the
pyro-denoised sequences (-in). For the distance matrix computa-
tion a different weight is given to homopolymer (4.0) and single
(15.0) gaps.
FCluster works as already described and, in particular, it per-
forms a hierarchical clustering based on the distance matrix pro-
duced by SeqDist.
SeqNoise removes the amplification noise using an EM algo-
rithm clustering procedure. It takes in input the distance matrix
produced by SeqDist (-din), a weighted fasta file (-in) with the
associated mapping file (-min) and the initial hierarchical clustering
data produced by FCluster (-lin). In particular, the weighted fasta
file indicated using the “-in” flag is the file produced by PyroNoiseM
and then processed by the Perl script Truncate.pl. In this file the
description line of each denoised sequence is formed as output_
label_index_weight where weight indicates the number of raw
sequences clustered in each denoised one. The raw reads included
in the denoised one are listed in the mapped file indicated by the
“-min” flag and produced by PyroNoiseM. Two parameters guide
the clustering procedure: “-c” (0.08) for the initial cutoff and “–s”
(30.0) for the clustering size (for more information refers to the
AmpliconNoise documentation). The “–out” flag (test_book_
denoised_T400_s25_c08) is used to label the output data pro-
duced by SeqNoise which are contained in the following files:
●● a fasta file where the denoised sequence associated with their
accession number formed as output_label_index_weight are
reported. Weight indicates the number of raw sequences
clustered in each denoised one and index is just a number
which identifies univocally the specific cluster (test_book_
denoised_T400_s25_c08_cd.fa);
●● a mapping file in which the accession numbers of the sequences
produced by PyroNoiseM computation are associated with
their corresponding denoised sequences obtained by SeqNoise
(test_book_denoised_T400_s25_c08.mapping);
●● a directory (test_book_denoised_T400_s25_c08) containing
a fasta file for each sequence derived from the SeqNoise
denoising procedure. In each file the unique sequences and
the corresponding reads, all used to generate the denoised
sequences, are listed;
●● if the –min parameter is added to the command-line, an
optional mapping file (test_book_denoised_T400_s25_c08_
cd.mapping), in which the accession numbers of the raw reads
are associated with the corresponding denoised sequence
derived from the full denoising procedure, is created.
In the exercise described here, after the complete denoising
procedure, AmpliconNoise identifies 399 real sequences, that are
theoretically unaffected by sequencing and amplification errors,
starting from the 3,229 reads that passes the initial filtering step.
3.3 Taxonomic The third step of the protocol consists in the taxonomic classifica-
Classification tion of the denoised sequences by using the RDP classifier.
RDP classifier [28] is a naïve Bayesian classifier which works
rapidly on short sequences and without the need to align them to
a taxonomically annotated reference database. It is developed by
the same team which maintains RPD II [29] and it is able to assign
sequences from kingdom to genus levels.
As other textual Bayesian classifiers, it works in a space of fea-
tures consisting in all the 8 nt long sub-string that could be found
in a sequence whatever their position. In this system W = {w1,…
,wd} is defined as the set of all possible 8 nt long sub-string that can
be retrieved in a given sequence. Given a collection of N sequences,
n(wi) represents the number of times that wi is observed in the set
N. The expected probability to observe wi in considering the entire
set N is:
n ( wi ) + 0.5
Pi =
N +1
If a classification at genus level is chosen, the conditional probability

of observing a specific genus given the single wi can be calculated.
Let us consider a training dataset (M) of sequences belonging
to the genus G. m(wi) is the number of sequences in M containing
the word wi. The Conditional probability that a sequence belong-
ing to G contain wi is:
é m ( wi ) + Pi ùû
P ( wi | G ) = ë
M +1
The joint probability P(S|G) that a sequence S (e.g., the 16S rDNA
variable region V1–V3), belonging to a specific genus G, contains
the entire set of possible words, V = {v1,…,vf}, in which it can be
splitted can be calculated as follows:
P ( S | G ) = ÕP ( vi | G )

If a sequence S with unknown taxonomic origin is considered, the
likelihood that S belongs to a genus G can be estimated by using
the Bayesian theorem as follows:
P ( S | G ) P (G )
P (G | S ) =
P(S )

Assuming that all genera are equally probable the terms P(G) and
P(S) can be ignored. A sequence is assigned to the genera with the
highest probability score but the actual numerical probability esti-
mate is ignored.
For each query sequence, the collection of all eight-character
sub-sequences (words) in the query is first calculated and then 100
bootstrap trials are performed. So, for each bootstrap trial, a subset
of one-eighth of the words is randomly chosen (with replacement)
and the words in this subset are used to calculate the joint proba-
bility. The number of times a genus is selected out of 100 boot-
strap trials is used as an estimate of confidence in the assignment to
that genus. By default a confidence threshold of 0.8 is used to
assign a sequence to a genus [28].
Some features of the dataset to be analyzed must be verified
prior to apply RDP classifier, in order to optimize the choice of the
confidence value threshold. Indeed, it has been demonstrated that
a threshold value of 0.5 can be used without loss of assignment
accuracy for short sequences (<250 nt) [30]. Therefore, it is con-
venient to extract the required information about the denoised
sequences by using Python shell.
Below the commands to be typed are reported. The sentences
starting with “#” are comments used to describe the operation and
they are not processed by Python.
$ Python
#First let us importing the class commands needed
to make the analysis
>>>from Bio import SeqIO
>>>import numpy
>>>import fpformat
#an empty list is created to store the infor-
mation about the sequence lengths
>>>size = []
# the length of each sequence is stored in the
sizes list by using the multifasta parser SeqIO.
parse
>>>for sequence in SeqIO.parse(“test_book_
denoised_T400_s25_c08_cd.fa”,”fasta”):
size.append(len(sequence))
#the average length is estimated using numpy.
mean command and the computed value is rounded
#to 2 decimal positions
>>>mean = fpformat.fix(numpy.mean(size),2)
#the standard deviation is estimated using
numpy.std command and the computed value is
rounded
#to 2 decimal positions
>>>sd = fpformat.fix(numpy.std(size),2)
#estimation of the minimum length
>>>min_length = min(size)
#estimation of the maximum length
>>>max_length = max(size)
#simply the estimated information are printed
>>>print “The denoised sequences are”,len(size)
The denoised sequences are 399
>>>print “the average length is”,mean,” and the
standard deviation is “,sd
the average length is 345.62 and the standard
deviation is 43.60
>>>print “the shortest sequence is long“,min_
length,”while the longest “,max_length
the shortest sequence is long 205 while the lon-
gest 374
The previous operations provide the information that the data-
set considered in the exercise contains sequences shorter than
250 nt. Accordingly, a confidence threshold values of 0.5 is fixed
for the RDP classifier computing by using the “--conf” parameter.
The flag “--hier_outfile” will allow to produce the output file con-
taining the assignment count for each annotated taxonomical rank
and the flag “--assign_outfile” will specify the output file contain-
ing the assignment details for each sequence.
The output file created with the “--hier_outfile” flag is tabular

and includes five fields:
●● Tax id: the taxonomy ID of the rank to which the sequences
are assigned;
●● Lineage: the full lineage path from root to the listed taxonomic
rank;
●● Name: the name of the considered taxonomic rank;
●● Rank: the taxonomic class to which the considered rank cor-
responds (e.g., class, order, family, genus);
●● The number of reads assigned to the considered rank.
In summary, the tabular file obtained using the flag “--hier_
outfile” provide a summary of the taxonomic assignment obtained
with RDP classifier and it can be easily managed by means the most
popular spread sheet programs.
The output file, obtained using the flag “--assign_outfile” is
tabular and includes the following fields:
●● accession number of the analyzed sequences;
●● the full lineage path resulting from the taxonomic assignment
and the confidence values estimated for each rank in the path.
In the exercise case, since the assigned sequences resulted from
several clustering and denoising steps, this file must be parsed in
order to associate to each rank the actual number of reads instead
their derived denoised sequence:
Following are the bioinformatics commands to perform the
latest described operations.
In particular, in order to perform the RDP-classifier, type:
-jar $ java -Xmx1g MultiClassifier.jar--ier_
outfile=hierchical_assignment.txt--ssign_outfile
= reads_classification.txt--onf = 0.5 test_book_
denoised_T400_s25_c08_cd.fa
Then the entire output assignment file (reads_classification.
txt) can be parsed and analyzed by means of a Python script, called
“extract_data.py” in order to create the submentioned tabular file
in which the number of actual reads is associated to each rank. The
obtained file is in csv format and it is easily manageable using com-
mon spread sheet programs.
First, the “extract_data.py” script must be written in the current
working directory. Below the commands included in it are listed:
l = open("read_classification.txt")
name2rank = {}
name2reads = {}
for line in l.readlines():
line = line.strip()
fields = line.split("\t")
acc = fields[0]
s = acc.split("_")
reads = int(s[-1])
stop = len(fields)
index = 2
while index < stop:
node = fields[index]
rank = fields[index+1]
conf = float(fields[index+2])
if conf >= 0.5:
name2rank.setdefault(rank,set())
name2rank[rank].add(node)
name2reads.setdefault(node,0)
name2reads[node] += reads
index += 3
else:
index = stop
#now we can print the data in a csv file
csv=open("data_summary_results.csv","w")
for rank in name2rank.keys():
csv.write(rank+"\n")
for node in name2rank[rank]:
c s v . w r i t e ( n o d e + " \ t " + s t r ( n a m e 2 r e a d s
[node])+"\n")
The script can be applied simply by typing “python extract_
data.py” and in few seconds it produces the “data_summary_
results.csv” file.
By observing the produced data, Firmicutes (1,099 reads,
34.03 %) and Proteobacteria (1,083 reads, 33.54 %) result the most
represented phyla, while Bacteroidetes, Fusobacteria and
Actinobacteria correspond only to 15.89 % (513 reads), 20.13 %
(650 reads) and 6.41 % (207 reads), respectively. Really, these data
are consistent with those obtained in previous human microbiota
surveys [31, 32].
4 Notes
1. The sratoolkit is a collection of tools provided for accessing

the information compressed in the sra files. The tools package
is available at http://trace.ncbi.nlm.nih.gov/Traces/sra/sra.
cgi?view=software . In order to extract the 454 data stored in
the downloaded .sra file, the sff-dump has been applied as
described below:
sff-dump SRR331235.sra -O sff_file_dir/
SRR331235.sff
the “-O” option is used to redirect the output in folder (in our
case “sff_file_dir”).
2. As reported in AmpliconNoise documentation, there are some
alternative methods to manage the sff file such as by using the
free software Flower [33] or the process_sff.py script included in
QIIME [34]. Moreover, the Mothur suite provides the com-
mand sffinfo.
3. In Mothur suite, PyroNoise and SeqNoise are re-implemented
in shhh.flows and shhh.seqs commands, respectively.
4. All the AmpliconNoise analysis is computationally very heavy.
For this region, the several steps are performed by using the
MPI libraries in order to parallelise the job on 8/10 nodes. By
means of these measures, the pyro- and seq-noise run times
can be reduced to 30 and 20 min respectively, if the tested
dataset is considered.
References
1. Hemme CL, Deng Y, Gentry TJ et al (2010) assignment in metagenomics. Brief Bioinform
Metagenomic insights into evolution of a heavy 13(6):682–695
metal-contaminated groundwater microbial 11. Tringe SG, Hugenholtz P (2008) A renais-
community. ISME J 4(5):660–672 sance for the pioneering 16S rRNA gene. Curr
2. Ottman N, Smidt H, de Vos WM et al (2012) Opin Microbiol 11(5):442–446
The function of our microbiota: who is out 12. Nilsson RH, Kristiansson E, Ryberg M et al
there and what do they do? Front Cell Infect (2008) Intraspecific ITS variability in the king-
Microbiol 2:104 dom fungi as expressed in the international
3. Dutton RJ, Turnbaugh PJ (2012) Taking a sequence databases and its implications for
metagenomic view of human nutrition. Curr molecular species identification. Evol Bioinform
Opin Clin Nutr Metab Care 15(5):448–454 Online 4:193–201
4. Knight R, Jansson J, Field D et al (2012) 13. Teeling H, Glöckner FO (2012) Current
Unlocking the potential of metagenomics opportunities and challenges in microbial
through replicated experimental design. Nat metagenome analysis—a bioinformatic per-
Biotechnol 30(6):513–520 spective. Brief Bioinform 13(6):728–742
5. Barnard D, Casanueva A, Tuffin M et al (2010) 14. Gilbert JA, Field D, Swift P et al (2010) The
Extremophiles in biofuel synthesis. Environ taxonomic and functional diversity of microbes
Technol 31(8–9):871–888 at a temperate coastal site: a 'multi-omic' study
6. Shokralla S, Spall JL, Gibson JF et al (2012) of seasonal and diel temporal variation. PLoS
Next-generation sequencing technologies for One 5(11):e15545
environmental DNA research. Mol Ecol 21: 15. Bazinet al, Cummings MP (2012) A compara-
1794–1805 tive evaluation of sequence classification
7. Luo C, Tsementzi D, Kyrpides N et al (2012) programs. http://drum.lib.umd.edu/handle/
Direct comparisons of Illumina vs. Roche 454 1903/13346
sequencing technologies on the same microbial 16. Simon C, Daniel R (2011) Metagenomic anal-
community DNA sample. PLoS One 7:e30087 yses: past and future trends. Appl Environ
8. Taberlet P, Coissac E, Pompanon F et al (2012) Microbiol 77(4):1153–1161
Towards next-generation biodiversity assess- 17. DeSantis TZ, Hugenholtz P, Larsen N et al
ment using DNA metabarcoding. Mol Ecol (2006) Greengenes, a chimera-checked 16S
21(8):2045–2050 rRNA gene database and workbench compati-
9. Blaalid R, Kumar S, Nilsson RH et al (2013) ble with ARB. Appl Environ Microbiol
ITS1 versus ITS2 as DNA metabarcodes for 72(7):5069–5072
fungi. Mol Ecol Resour 13(2):218–224 18. Cole JR, Chai B, Marsh TL et al (2003)

10. Santamaria M, Fosso B, Consiglio A et al Ribosomal Database Project. The ribosomal
(2012) Reference databases for taxonomic database project (RDP-II): previewing a new
autoaligner that allows regular updates and the 27. Chuong BD, Batzoglou S (2008) What is the
new prokaryotic taxonomy. Nucleic Acids Res expectation maximization algorithm? Nat
31(1):442–443 Biotechnol 26(8):897–899
19. Pruesse E, Quast C, Knittel K et al (2007) 28. Wang Q, Garrity GM, Tiedje JM et al (2007)
SILVA: a comprehensive online resource for Naïve bayesian classifier for rapid assignment of
quality checked and aligned ribosomal RNA rRNA sequences into the new bacterial taxonomy.
sequence data compatible with ARB. Nucleic Appl Environ Microbiol 73(16):5261–5267
Acids Res 35(21):7188–7196 29. Cole JR, Wang Q, Cardenas E et al (2009) The
20. Roche Applied Sciences (2008) Genome Ribosomal Database Project: improved align-
sequencer data analysis software manual. Roche ments and new tools for rRNA analysis. Nucleic
Diagnostics GmbH, Germany Acids Res 37(Database issue):D141–D145.
21. Metzker ML (2010) Sequencing Technologies - doi:10.1093/nar/gkn879
the Next Generation. Nat Rev Genet 11(1): 30. Claesson MJ, O'Sullivan O, Wang Q et al
31–46 (2009) Comparative analysis of pyrosequenc-
22. Ewing B, Green P (1998) Base-calling of auto- ing and a phylogenetic microarray for explor-
mated sequencer traces using phred II error ing microbial community structures in the
probabilities. Genome Res 8(3):186–194 human distal intestine. PLoS One 4(8):e6669
23. Quince C, Lanzen A, Davenport RJ et al 31. Gosalbes MJ, Abellan JJ, Durbán A et al (2012)
(2011) Removing noise from pyrosequenced Metagenomics of human microbiome: beyond
amplicons. BMC Bioinformatics 12:38 16s rDNA. Clin Microbiol Infect 18(4):47–49
24. Schloss PD (2009) A high-throughput DNA 32. Andersson AF, Lindberg M, Jakobsson H et al
sequence aligner for microbial ecology studies. (2008) Comparative analysis of human gut
PLoS One 4(12):e8230 microbiota by barcoded pyrosequencing. PLoS
25. Balzer S, Malde K, Lanzén A et al (2010) One 3(7):e2836
Characteristics of 454 pyrosequencing data- 33. Malde K (2011) Flower: extracting informa-
enabling realistic simulation with flowsim. tion from pyrosequencing data. Bioinformatics
Bioinformatics 26(18):i420–i425 27(7):1041–1042
26. Huse SM, Huber JA, Morrison HG et al 34. Caporaso JG, Kuczynski J, Stombaugh J et al
(2007) Accuracy and quality of massively paral- (2010) QIIME allows analysis of high-
lel DNA pyrosequencing. Genome Biol 8(7): throughput community sequencing data. Nat
R143 Methods 7(5):335–336
Chapter 17
Deciphering Metatranscriptomic Data

Evguenia Kopylova, Laurent Noé, Corinne Da Silva,
Jean-Frédéric Berthelot, Adriana Alberti, Jean-Marc Aury,
and Hélène Touzet
Abstract
Metatranscriptomic data contributes another piece of the puzzle to understanding the phylogenetic struc-
ture and function of a community of organisms. High-quality total RNA is a bountiful mixture of ribosomal,
transfer, messenger and other noncoding RNAs, where each family of RNA is vital to answering questions
concerning the hidden microbial world. Software tools designed for deciphering metatranscriptomic data
fall under two main categories: the first is to reassemble millions of short nucleotide fragments produced by
high-throughput sequencing technologies into the original full-length transcriptomes for all organisms
within a sample, and the second is to taxonomically classify the organisms and determine their individual
functional roles within a community. Species identification is mainly established using the ribosomal RNA
genes, whereas the behavior and functionality of a community is revealed by the messenger RNA of the
expressed genes. Numerous chemical and computational methods exist to separate families of RNA prior to
conducting further downstream analyses, primarily suitable for isolating mRNA or rRNA from a total RNA
sample. In this chapter, we demonstrate a computational technique for filtering rRNA from total RNA using
the software SortMeRNA. Additionally, we propose a post-processing pipeline using the latest software
tools to conduct further studies on the filtered data, including the reconstruction of mRNA transcripts for
functional analyses and phylogenetic classification of a community using the ribosomal RNA.
Key words Metatranscriptomics, High-throughput sequencing, 16S rRNA phylogenetic analysis
1 Introduction
Metatranscriptomics is the study of total RNA sequenced from a

group of microorganisms directly extracted from their native envi-
ronment. It successfully exploits the well-known RNA-Seq tech-
nique to transcriptome profiling based on leading high-throughput
sequencing technologies, such as Illumina Solexa (Illumina Inc.),
454 (Roche Diagnostics), and Ion semiconductor sequencing (Ion
Torrent Systems Inc.). RNA-Seq offers key advantages over
previous technologies, such as array-based (like tiling [1]) and tag-
based (like SAGE [2] or CAGE [3]). First, it is not limited to
279
280 Evguenia Kopylova et al.
detecting transcripts that correspond to existing genomic

sequences. It is then particularly attractive for non-model organ-
isms without an existing genome assembly. It can reveal the precise
location of splicing junctions, transcription start and stop, to a
single-base resolution. Furthermore, 30-bp short reads from RNA-
Seq give us information about how two exons are connected,
whereas longer reads or pair-end short reads should reveal connec-
tivity between multiple exons. These factors make RNA-Seq useful
for studying complex transcriptomes.
When applied to whole environments, RNA-seq can sequence
millions of uncultured organisms in parallel and provide an authen-
tic representation of species richness within a community at the
time of the sampling. Phylogenetic structure of a community is
primarily established using the 16S and 18S ribosomal RNA
(rRNA) genes, which remain highly conserved among different
species of bacteria, archaea and eukarya [4]. Other inquiries can be
made regarding the functionality of a community by studying the
messenger RNA (mRNA) of the actively transcribed genes. Thus, it
is of primary interest to sort out the rRNA from the mRNA in the
total RNA for answering the most basic questions regarding the
species composition, gene regulation, and protein information of a
metatranscriptome. Subject to prokaryotic or eukaryotic cells, the
rRNA content can represent up to 80–85 % of total RNA and the
protein coding mRNA as little as 1–6 % [5]. If one wants to focus
exclusively on mRNA, there exist prior-to-sequencing methods to
help isolate and enrich mRNA from the total RNA. These methods
focus on the depletion of rRNA with technologies such as subtrac-
tive hybridization (Life Technologies), exonuclease digestion [6]
and the duplex-specific nuclease treatment (DSN) [7]. However,
even selective removal of rRNA genes does not guarantee definite
depletion, and some rRNA genes can still remain in the metatran-
scriptomic sample. Alternatively, if the goal is sort apart RNA with-
out any loss of information, in the case of studying rRNA for species
identification, noninvasive methods for separating RNA are also
preferred. Such methods implement a bioinformatics pipeline to
catalog families of RNA using publicly available RNA databases.
To this end, we introduce SortMeRNA [8], which is the fastest
existing tool for filtering out rRNA from a total RNA dataset. The
underlying algorithm takes into account the high-conservation
property of rRNA sequences and achieves a high sensitivity.
This chapter is organized as follows. Following this introduc-
tion, in Subheading 2 we list the tools used for data cleaning and
sorting metatranscriptomic reads. In Subheading 3, we describe
the steps for filtering RNA from a total metatranscriptome using
the software SortMeRNA. Lastly, in Subheading 3.3 we summarize
the latest software tools developed to further process and analyze
sorted metatranscriptomic data, including methods for assembly,
mapping, functional analysis, and taxonomic analysis.
Metatranscriptomic Data Analysis 281
2 Materials
2.1 Software The software tool SortMeRNA is a fast and accurate filter for iden-
tifying and sorting rRNA in a set of metatranscriptomic reads.
SortMeRNA is written in C++ and freely distributed under the GPL
license as a stand-alone version or as a Galaxy wrapper. Both distri-
butions, including the user manual with installation instructions,
can be downloaded from http://bioinfo.lifl.fr/RNA/sortmerna/.
SortMeRNA supports multi-threading and has been tested on
Linux (Ubuntu, Fedora, CentOS, and Debian) and Mac OS 10.6.8
systems. For compilation, a g++ compiler version 4.3 or higher is
required. Our experiments were performed using version 1.8 of
SortMeRNA, on an Intel(R) Xeon(R) CPU W3520 2.67 GHz
machine with 8 GB of RAM and one thread.
2.2 Ribosomal RNA To filter out rRNA, SortMeRNA performs a search against a data-
Databases base of known rRNAs, such as the public rRNA databases SILVA
[9], Greengenes [10], Rfam [11], and RDP-II [12], that provides
access to millions of rRNA sequences. SortMeRNA can be used
with any of these databases, any fraction of them, or with any cus-
tom database. The only requirement is that sequences are in the
FASTA format. For easier use, the distribution package of
SortMeRNA also includes eight nonredundant representative
databases for 16S bacteria and archaea, 18S eukarya, 23S bacteria
and archaea, 28S eukarya, 5S bacteria/archaea, and 5.8S eukarya,
which were derived from the raw SILVA rRNA database sets using
the tools ARB [13] and UCLUST [14]. We first remove contami-
nated sequences, such as chimeric rRNA and partial mRNA, and
then use the tool UCLUST to construct a nonredundant database
according to the identity threshold of the curated sequences.
2.3 Metatran- In the following Subheading 3, we use SortMeRNA to filter rRNA

scriptomic Reads from a total metatranscriptomic dataset ERA236149 sequenced
from a marine environment. Total RNA, was chemically fragmented
and converted into single-stranded cDNA using random hexamer
priming. Then, the second strand was generated to create double-
stranded cDNA. The library was prepared following Illumina’s
protocol and sequenced using 101 base-length read chemistry on
the Illumina HiSeq2000. After quality and adapters cleaning, we
obtain 20,968,598 useful reads longer than 30 nucleotides.
Cleaning the reads is an important step to enhance software
performance and the quality of the results [15] (Table 1). Our
procedure is as follows:
1. Filter the raw data to remove any clusters that have “too much”
intensity corresponding to bases other than the called base.
Table 1
List of examples of relevant software which can be used for performing
read quality control
Filter Program
Vector/Adapter clipping TagCleaner [16], FASTX-Toolkit
Low quality trimming PRINSEQ [17], FASTX-Toolkit
Low complexity regions dust [18]
Read length PRINSEQ
Clone removal PRINSEQ
Error correction Coral [19]
2. Remove adapters and primers on the whole read and low qual-
ity nucleotides from both ends (while quality value is lower
than 20) and then continue next steps with the longest
sequence without adapters and low quality bases.
3. Remove sequences between the second unknown nucleotide
(N) and the end of the read. This step is done on trimmed
(adapters + quality) reads.
4. Discard reads shorter than 30 nucleotides after trimming.
5. Remove reads that mapped onto run quality control sequences
(PhiX genome).
3 Methods
3.1 Command-Line SortMeRNA takes two mandatory inputs:

3.1.1 Input 1. A FASTA or FASTQ file of reads.
2. A FASTA file for the database of full-length rRNA sequences.
The file of rRNA sequences must be first indexed using the
command “buildtrie” and then the index is automatically loaded
when “sortmerna” is launched. For example, we are interested in
filtering out all bacterial and archaeal 16S rRNA from the
metatranscriptomic dataset ERA236149 using SILVA database.
Firstly, we construct the index using the rRNA databases provided
in the distribution package:
buildtrie --db ./rRNA_databases/silva-bac-16s-database-id85.fasta
buildtrie --db ./rRNA_databases/silva-arc-16S-database-id95.fasta
This has to be done only once if one wants to process several
sets of reads.
3.1.2 How to Run We filter the 16S rRNA from our ERA236149 dataset using our
SortMeRNA two previously indexed databases. The command-line is as follows:
sortmerna -n 2 -- db ./rRNA_databases/silva-bac-16s-database-id85.
fasta ./rRNA_databases/silva-arc-16S-database-id95.fasta --I
ERA236149.fastq --accept rrna --bydbs --other nonrrna --log bilan
Where the options correspond to:
Input:
-n <N> is the number <N> of databases to search,
--db <string> is for a space separated list of <N> databases,
--I <filename> is for the input reads,
Output:
--accept <filename without extension> is for the output file(s)
of matching reads,
--bydbs is to output the matching reads to <N> different files
(one for each database),
--other <filename without extension> is the output file for
non-matching reads,
--log <filename> gives a statistics file for the percentages of
reads matching to each of the <N> databases.
Paired reads: paired reads are managed using the options--paired-
in and--paired-out. Although SortMeRNA does not use parity
information during filtering, when run in multi-threading mode,
the output of matching and non-matching reads can lose the parity
order of the original file. To maintain parity of reads, if one read
matches and the other does not, the option--paired-in will output
the paired reads into the--accept file, whereas the option--paired-
out will output the paired reads into the--other file. If neither of
the options are set, by default the output of SortMeRNA may not
maintain parity order.
3.1.3 Output From our example above, there will be four output files: two files
for rRNA classified reads for each 16s-bacteria, 16s-archaea
database, one file for the non-rRNA reads, and one file for the
overall statistics of the ERA236149 dataset tested above. The sort-
ing of ~21 million reads was performed 1.5 h using one thread.
From the statistics file, approximately 27 % of these reads were
classified as 16S rRNA to one of two reference databases.
3.1.4 Advanced Options Memory management: The size of ERA236149.fastq is 5.2 GB, by
default SortMeRNA works internally with 1 GB of memory allo-
cated for the reads, without any required user intervention to split
the file into multiple parts. If the user has sufficient memory to
load all 5.2 GB of reads into memory, the option -m <M> allows
to set the memory limit to M bytes. Likewise, the memory limit

can be reduced to under 1 GB for systems with RAM restrictions.
Multi-threading: The parameter -a <A> will use A number of
threads, it is set to A = 1 by default.
3.2 Galaxy SortMeRNA is also available through Galaxy [20], which is an

open, Web-based platform for computational research in the life
sciences. Galaxy provides users with a graphical work-flow manage-
ment system, and emphasizes accessibility, reproducibility and
transparency. Most of the same options are available in Galaxy as
the stand-alone version of SortMeRNA (excluding the accept by
database--bydbs option).
3.2.1 Installation The Galaxy wrapper for SortMeRNA is available on the Toolshed,
the place for sharing Galaxy tools with any Galaxy instance. Any
local administrator may install SortMeRNA by browsing the
Toolshed, and Galaxy takes care of downloading and installing
SortMeRNA, as well as indexing the ribosomal RNA databases
provided with SortMeRNA.
3.2.2 Input The Galaxy main graphical interface has three parts (see Fig. 1): the
left panel lists all the available tools; the right panel is the history of
Fig. 1 Galaxy interface for SortMeRNA

all the steps of the current analysis and their outputs. The central
panel displays all of the input, output and parameter criteria for
launching SortMeRNA.
The input FASTA or FASTQ file of reads must be first imported
to the Galaxy history. Small files may be uploaded from the local
hard drive, by selecting the “Upload file from your computer” in
the “Get data” section; bigger files can be transferred from an
accessible URL or by FTP. Once done, our file “ERA236149.fasta”
and its associated metadata are visible in the right panel history.
We bring up SortMeRNA by selecting “Filter with Sort-
MeRNA” in the left panel. We select the desired reads file from the
history—only FASTA/FASTQ files can be selected. Most of the
options selected by default are appropriate for our experiment;
we tick the 16S databases. We want to retrieve both the classified
reads and non-rRNA reads, as well as the statistics file, so we tick
all three options.
3.2.3 Output The SortMeRNA outputs are automatically imported to the his-
tory as Galaxy datasets, and are recognized as FASTA/FASTQ
files, which can be used for further analysis inside Galaxy.
3.2.4 Advanced Options It is possible to use personal databases to filter our reads against.
We must first import them to Galaxy as FASTA files, just like the
reads file. In the SortMeRNA interface, we then select “Databases
from your history” as “Databases to query,” and add as many files
as we want. The indexation of the rRNA sequences file will be
automatically invoked by Galaxy before running SortMeRNA—
note however that the index must be computed each time, which
is time-consuming. Alternatively, more databases can be added to
the default list by the Galaxy administrator.
3.3 Further Analyses Once SortMeRNA has been run, one might want to perform fur-
ther analyses of the filtered dataset. In this section we touch on
various stages of filtered data manipulation, such as the reconstruc-
tion of full-length RNA from short reads, taxonomic assignation
for ribosomal RNA, and functional analysis of messenger RNA
using well-maintained protein databases. Please refer to Fig. 2
for an illustration of the pipeline. Many of the tools shown in this
pipeline are also available in Galaxy.
3.3.1 Assembly De novo assembly of metatranscriptomic data is in the early stages

of development, and up to now there does not exist a tool specifi-
cally designed to assemble metatranscriptomic data. It is seldom
conducted on the total RNA sample, largely because assembly of
rRNA genes is a very resource intensive task and susceptible to
chimera formation from the variable and conserved regions on the
SSU rRNA molecule [21, 22]. However, if the user has a sorted
file of rRNA reads, they can perform full-length reconstruction
RAW DATA
Metatranscriptomic reads
(total RNA)
2.1 CLEAN
• PRINSEQ
• FASTX-Toolkit
• digital normalization
(see khmer)
2.2 SORT
SortMeRNA
mRNA rRNA
PROCESS
2.3 ASSEMBLE
2.4 MAP
• Oases
• TopHat2 ASSEMBLE/MAP
• Trinity
• BLAST • BLAST
• khmer
• CLC Genomics • EMIRGE
• CLC Genomics Workbench
Workbench (CLC Bio)
(CLC Bio)
ANALYZE
2.6 TAXONOMY
2.5 FUNCTION
• mothur
• FragGeneScan • MG-RAST • QIIME
• Glimmer • MEGAN 4 • RDP Classifier
• ARB
Fig. 2 Metatranscriptomic analysis flowchart
using the program EMIRGE [23], which can also provide relative
abundances of rRNA in the community sample. For reconstruct-
ing full-length mRNA transcripts from a metatranscriptome,
we describe two work-around methods applicable for raw or
normalized data. The first approach is to assemble short read

sequencing data into longer contigs using bioinformatic tools such
as the CLC Genomics Workbench (CLC Bio) or SOAPdenovo
[24]. Afterward, the longer contigs can be further assembled using
Newbler (454 Life Sciences) (designed for 454 pyrosequencing,
known to work with Illumina contigs [25]), or merged together
using AMOS Minimus2 [26] to reconstruct optimal-length tran-
scripts. The second approach is to use a transcriptome assembler
such as Oases [27] or Trinity [28], which can detect alternative
splicing events and duplicated genes, and predict distinct full-
length transcripts through assembly merging. If the reads will be
used for assembly, reducing high-coverage data and selecting reads
with low error rates can greatly improve the quality and computing
performance. The program khmer [29] includes a single-pass digi-
tal normalization prefilter to even out random sampling variation
of your metatranscriptome by removing redundant and errone-
ous sequences. Likewise, filtering out low-abundance k-mers
(-d option in SOAPdenovo) will reduce the computational require-
ments for all de novo assemblers using a de Bruijn graph approach.
3.3.2 Mapping Metatranscriptomic read mapping assists in recovering the arrange-

ment of short RNA reads with respect to long reference sequences,
such as genes, metagenomic contigs/ORFs, complete genomes, or
rRNA molecules. This analysis can be particularly useful for meta-
transcriptomes of low level coverage and high diversity, where no
significant overlap exists between the sequenced reads to facilitate
de novo assembly. The most mature and informative tool for local
sequence alignment is BLAST [30], although for large-scale high-
throughput metatranscriptomic projects, the execution time is
substantially slower in comparison with new-generation aligners
and therefore more practical to be integrated in downstream analy-
ses on smaller sized data. The most popular read-mapping software
applying is Bowtie [31], which is the core alignment tool behind
many cutting-edge read mapping methods, such as TopHat2 [32]
and EMIRGE, due to its small memory footprint and rapid
alignment. In addition to Bowtie1, Bowtie2 [33] supports both
local and gapped alignment, which makes it more practical for
reads prone to homopolymer errors such as those produced by 454
and Ion Torrent technologies.
3.3.3 Functional Analysis For assessing the potential functions of mRNA sequences, it is routine
of mRNA to search the reads or assembled transcripts (see Subheading 3.3.1)
for homologs in a public reference database such as the NCBI non-
redundant protein database (NCBI-nr) [34], the RDP-II database,
KEGG (for enzymes and pathway networks) [35], or SEED
(protein sequences with known functional roles) [36]. The
MG-RAST [37] server and MEGAN4 [38] are two services pro-
viding comprehensive tools for the annotation and functional
analysis of metagenomes and metatranscriptomes (prokaryotes

only for MG-RAST), including comparative analysis between mul-
tiple community datasets. Both services have been documented to
generate comparable results [39]; however, MEGAN4 can illus-
trate hierarchal representations of functional assignments and sup-
ports easy local installation for scientists not inclined to upload
their unpublished data. The tool MEGAN4 uses the KEGG classi-
fication to calculate molecular interaction and reaction pathways
including metabolism and cellular processes for both prokaryotes
and eukaryotes, and the SEED classification to assign each read to
the highest scoring gene with a known functional role. Alternatively,
MG-RAST provides access to computational power needed for
high-quality annotations through a CLOUD client.
Gene prediction programs such as FragGeneScan [40] and
Glimmer [41], provide an alternative ab initio approach to detect
novel genes in sufficiently long assembled contigs when no close
homologues of coding regions can be found in a reference data-
base. The underlying method of these tools utilizes Hidden Markov
models, which are trained on different species of prokaryotes,
eukaryotes, viruses, phages, and plasmids to recognize protein cod-
ing and noncoding regions of input DNA sequences. Given the
length of genes in prokaryotes can easily exceed 1,000 nucleotides
[42], the user should be aware of the restricted identification of
open reading frames (ORFs) when working with very short reads,
since the in-frame start and stop codons may be split between reads.
3.3.4 Taxonomic Due to the high conservation of rRNA genes between species and
Classification of rRNA their presence in all cells, the 16S and 18S small subunit rRNA
serve as signatures to infer whole organism phylogeny [43]. For a
reliable representation of species diversity, short reads covering
rRNA variable regions are more accurately classified than those
covering higher-conserved regions. Therefore, tools which are
more sensitive to capture sequence variation will render a higher
taxonomic resolution of a given metatranscriptome. In the follow-
ing two paragraphs, we describe two methods for rRNA classifica-
tion, either by using the raw unassembled short reads, or with an
assembled set of full-length rRNA genes.
Short rRNA fragments (reads): MEGAN4 implements visual and
interactive phylogenetic trees from a given set of rRNA reads, and
allows the user to carefully examine the alignment of each read
with its assigned species. To use MEGAN4, the user must first run
a BLASTN search to compare a set of reads to the SILVA database
and then import the BLAST results file to calculate taxonomic clas-
sification using the NCBI taxonomy. Although the searching step
can be computationally expensive for large set of reads and relaxed
parameter settings, the BLAST algorithm can be efficient at finding
the reads covering variable regions in low coverage samples.
The integrated software suite mothur [44] provides one of the

most complete packages to preprocess and assign 16S pyrose-
quencing or Illumina reads to operational taxonomic units (OTUs),
measure species diversity and display the results using Venn dia-
grams, heat maps, and dendrograms. The open source tool QIIME
[45] provides most of the similar community analysis tools as
mothur, however with more advanced visualization representa-
tions for network analyses, histograms, and phylogenetic trees.
Lastly, the RDP Classifier [46] uses a fast, alignment-free naïve
Bayesian algorithm to classify both short and full-length prokary-
otic rRNA sequences into higher-order taxonomic ranks and the
results can be input to QIIME for computing diversity estimates.
Full-length rRNA sequences: Following the reconstruction of full-
length rRNA genes from a pool of isolated short rRNA reads, we
can apply classical phylogenetic procedures to align them to exist-
ing rRNA databases. The ARB package provides a graphic environ-
ment for visualizing phylogenetic trees, aligning sequences, and
editing primary and secondary structures of both SSU and LSU
rRNA genes. The phylogenetic tree building tools include both
phenetic and cladistic methods, allowing for an exhaustive study
of relationships among a group of organisms and their pathways
of evolution.
Acknowledgments
This research was supported by the French National Agency for

Research (grant ANR-2010-COSI-004) and the French National
Sequencing Center (Genoscope).
References
1. Kapranov P et al (2007) RNA maps reveal 6. Boissinot K, Huletsky A, Peytavi R et al (2007)
new RNA classes and a possible function for Rapid exonuclease digestion of PCR-amplified
pervasive transcription. Science 316(5830): targets for improved microarray hybridization.
1484–1488 Clin Chem 53(11):2020–2023
2. Velculescu VE et al (1995) Serial analysis of 7. Yi H, Cho YJ, Won S et al (2011) Duplex-
gene expression. Science 270(5235):484–487 specific nuclease efficiently removes rRNA for
3. Shiraki T et al (2003) Cap analysis gene expres- prokaryotic RNA-seq. Nucleic Acids Res
sion for high-throughput analysis of transcrip- 39(20):e140
tional starting point and identification of 8. Kopylova E, Noe L, Touzet H (2012)
promoter usage. Proc Natl Acad Sci U S A SortMeRNA: fast and accurate filtering of
100(26):15776–15781 ribosomal RNAs in metatranscriptomic data.
4. Janda JM, Abbott SL (2007) 16S rRNA gene Bioinformatics 28(24):3211–3217
sequencing for bacterial identification in the 9. Quast C, Pruesse E, Yilmaz P et al (2013) The
diagnostic laboratory: pluses, perils, and pit- SILVA ribosomal RNA gene database project:
falls. J Clin Microbiol 45(9):2761–2764 improved data processing and web-based
5. Sorek R, Cossart P (2010) Prokaryotic tran- tools. Nucleic Acids Res 41(D1):D590–D596
scriptomics: a new view on regulation, physiol- 10. DeSantis TZ, Hugenholtz P, Larsen N et al
ogy and pathogenicity. Nat Rev Genet 11(1): (2006) Greengenes, a chimera-checked 16S
9–16 rRNA gene database and workbench compatible
with ARB. Appl Environ Microbiol 72(7): short-read de novo assembler. GigaScience.
5069–5072 doi:10.1186/2047-217X-1-18
11. Griffiths-Jones S, Bateman A, Marshall M et al 25. Mason OU, Hazen TC, Borglin S et al (2012)
(2003) Rfam: an RNA family database. Nucleic Metagenome, metatranscriptome and single-
Acids Res 31(1):439–441 cell sequencing reveal microbial response to
12. Cole JR, Wang Q, Cardenas E et al (2008) Deepwater Horizon oil spill. ISME J 6(9):
The Ribosomal Database Project: improved 1715–1727
alignments and new tools for rRNA analysis. 26. Sommer DD, Delcher AL, Salzberg SL et al
Nucleic Acids Res 37:D141–D145 (2007) Minimus: a fast, lightweight genome
13. Ludwig W, Strunk O, Westram R et al (2004) assembler. BMC Bioinformatics.
ARB: a software environment for sequence doi:10.1186/1471-2105-8-64
data. Nucleic Acids Res 32(4):1363–1371 27. Schulz MH, Zerbino DR, Vingron M et al
14. Edgar RC (2010) Search and clustering orders (2012) Oases: robust de novo RNA-seq assem-
of magnitude faster than BLAST. Bioinformatics bly across the dynamic range of expression lev-
26(19):2460–2461 els. Bioinformatics 28(8):1086–1092
15. Brown CT, Howe A, Zhang Q et al (2013) A 28. Grabherr MG, Haas BJ, Yassour M et al (2011)
reference-free algorithm for computational Full-length transcriptome assembly from
normalization of shotgun sequencing data. RNA-Seq data without a reference genome.
https://www.e-biogenouest.org/resources/46 Nat Biotechnol 29:644–652
16. Schmieder R, Lim YW, Rohwer F et al (2010) 29. Pell J, Hintze A, Canino-Koning R et al (2012)
TagCleaner: identification and removal of tag Scaling metagenome sequence assembly with
sequences from genomic and metagenomic probabilistic de Bruijn graphs. Proc Natl Acad
datasets. BMC Bioinformatics 11:341 Sci U S A. doi:10.1073/pnas.1121464109
17. Schmieder R, Edwards R (2011) Quality con- 30. Altschul SF, Gish W, Miller W et al (1990)
trol and preprocessing of metagenomic datas- Basic local alignment search tool. J Mol Biol
ets. Bioinformatics 27(6):863–864 215(3):403–410
18. Morgulis A, Gertz EM, Schäffer AA et al 31. Langmead B, Trapnell C, Pop M et al (2009)
(2006) A fast and symmetric DUST imple- Ultrafast and memory-efficient alignment of
mentation to mask low-complexity DNA short DNA sequences to the human genome.
sequences. J Comput Biol 13(5):1028–1040 Genome Biol 10:R25
19. Salmela L, Schroder J (2011) Correcting 32. Kim D, Pertea G, Trapnell C et al (2013)
errors in short reads by multiple alignments. TopHat2: accurate alignment of transcrip-
Bioinformatics 27(11):1455–1461 tomes in the presence of insertions, deletions
20. Goecks J, Nekrutenko A, Taylor J, The Galaxy and gene fusions. Genome Biol 14(4):R36
Team (2010) Galaxy: a comprehensive 33. Langmead B, Salzberg SL (2012) Fast gapped-
approach for supporting accessible, reproduc- read alignment with Bowtie 2. Nat Methods
ible, and transparent computational research in 9(4):357–359
the life sciences. Genome Biol 11(8):R86. 34. Pruitt KD, Tatusova T, Maglott DR (2005)
doi:10.1186/gb-2010-11-8-r86 NCBI Reference Sequence (RefSeq): a curated
21. Radax R, Rattei T, Lanzen A et al (2012) non-redundant sequence database of genomes,
Metatranscriptomics of the marine sponge transcripts and proteins. Nucleic Acids Res
Geodia barretti: tackling phylogeny and func- 33:D501–D504
tion of its microbial community. Environ 35. Kanehisa M, Goto S, Sato Y et al (2012)
Microbiol 14(5):1308–1324 KEGG for integration and interpretation of
22. Fan L, McElroy K, Thomas T (2012) large-scale molecular datasets. Nucleic Acids
Reconstruction of ribosomal RNA genes from Res 40:D109–D114
metagenomic data. PLoS One 7(6):e39948. 36. Overbeek R, Begley T, Butler RM et al (2005)
doi:10.1371/journal.pone.0039948 The subsystems approach to genome annota-
23. Miller CS, Baker BJ, Thomas BC et al (2011) tion and its use in the project to annotate
EMIRGE: reconstruction of full-length ribo- 1000 genomes. Nucleic Acids Res 33(17):
somal genes from microbial community short |5691–5702
read sequencing data. Genome Biol 12(5):R44 37. Meyer F, Paarmann D, D’Souza M et al (2008)
24. Luo R, Liu B, Xie Y et al (2012) SOAPdenovo2: The metagenomics RAST server—a public
an empirically improved memory-efficient resource for the automatic phylogenetic and
functional analysis of metagenomes. BMC between the two kingdoms. Mol Biol Evol
Bioinformatics 9:386. doi:10.1186/1471- 23(6):1107–1108
2105-9-386 43. Lane DJ, Pace B, Olsen GJ et al (1985) Rapid
38. Hudson DH, Mitra S, Ruscheweyh HJ et al determination of 16S ribosomal RNA
(2011) Integrative analysis of environmental sequences for phylogenetic analyses. Proc Natl
sequences using MEGAN4. Genome Res Acad Sci U S A 82(20):6955–6959
21(9):1552–1560 44. Schloss PD, Westcott SL, Ryabin T et al
39. Mitra S, Rupek P, Richter DC et al (2011) (2009) Introducing mothur: open-source,
Functional analysis of metagenomes and meta- platform-independent, community-supported
transcriptomes using SEED and KEGG. BMC software for describing and comparing micro-
Bioinformatics 12(Suppl 1):S21 bial communities. Appl Environ Microbiol
40. Rho M, Tang H, Ye Y (2010) FragGeneScan: 75(23):7537–7541
predicting genes in short and error-prone 45. Caporaso JG, Kuczynski J, Stombaugh J et al
reads. Nucleic Acids Res 38(20):e191 (2010) QIIME allows analysis of high-
41. Delcher AL, Bratke KA, Powers EC et al throughput community sequencing data. Nat
(2007) Identifying bacterial genes and endo- Methods 7(5):335–336
symbiont DNA with Glimmer. Bioinformatics 46. Wang Q, Garrity GM, Tiedje JM et al (2007)
23(6):673–679 Naive Bayesian classifier for rapid assignment
42. Lin X, Hong C, Xiaohua H et al (2006) of rRNA sequences into the new bacterial
Average gene length is highly conserved in taxonomy. Appl Environ Microbiol 73(16):
prokaryotes and eukaryotes and diverges only 5261–5267
Chapter 18
RIP-Seq Data Analysis to Determine RNA–Protein

Associations
Abstract
Next-generation sequencing (NGS) technologies have opened new avenues of unprecedented power for
research in molecular biology and genetics. In particular, their application to the study of RNA-binding
proteins (RBPs), extracted through immunoprecipitation (RIP), permits to sequence and characterize all
RNAs that were found to be bound in vivo by a given RBP (RIP-Seq). On the other hand, NGS-based
experiments, including RIP-Seq, produce millions of short sequence fragments that have to be processed
with suitable bioinformatic tools and methods to recover and/or quantify the original sequence sample. In
this chapter we provide a survey of different approaches that can be taken for the analysis of RIP-Seq data
and the identification of the RNAs bound by a given RBP.
Key words Immunoprecipitation, Next-generation sequencing, RNA binding protein, RIP-Seq,

CLIP-Seq, PAR-CLIP
1 Introduction
Next-generation sequencing (NGS) technologies have opened

new avenues of unprecedented power for research in molecular
biology and genetics. Other than their straightforward application
to the de novo sequencing or resequencing of genomes, genomic
loci, or transcriptomes, they can be used with excellent results to
the characterization and/or quantification of the sequences con-
tained in a given DNA or RNA sample [1]. Indeed, in the case of
RNAs, the ever decreasing cost of sequencing has made NGS the
method of choice not only for the characterization of transcrip-
tomes and identification of alternative splicing, but also for simpler
tasks like the quantification of transcript abundance, rapidly replac-
ing oligonucleotide microarrays as the de facto standard for these
applications [2]. In the latter case, as in the traditional Sanger
method, sequencing is usually performed after the isolated RNA
has been converted into a cDNA library.
293
RNAs in cells are associated with RNA-binding proteins

(RBPs) to form ribonucleoprotein (RNP) complexes [3]. RBPs
determine the structure and interactions of the RNAs, and their
biogenesis, stability, function, and cellular localization. Eukaryotic
genomes can encode hundreds of RBPs (thousands in vertebrates),
each of which has unique RNA-binding specificity. Indeed, another
broad field of application of NGS technologies is the characteriza-
tion of DNA or RNA samples, extracted for example through an
ImmunoPrecipitation (IP) experiment [4]. That is, given a DNA-
or RNA-binding protein of interest for which suitable antibodies
are available, the protein can be extracted from cells together with
DNA or RNA sequences that were bound to it in vivo. Since RBP
binding occurs simultaneously in the cells analyzed thousands of
times for the same RNA molecule, once separated from the protein
the DNA or RNA fragments bound should be found in many cop-
ies, and thus be enriched in the sample extracted. NGS technolo-
gies offer the most straightforward solution to the problem of
identifying enriched DNA/RNA fragments in the sample, that is,
to sequence the sample itself. The more times a given fragment is
sequenced, the more likely it is to be actually bound in vivo by the
protein investigated [5, 6].
The main issue that in general one has to face with NGS tech-
nologies applied to IP samples is that at the present time they do
not permit to sequence the entire DNA or RNA molecules obtained
from the IP, but only a short fragment at either or both 5′ ends.
A single sequencing lane of Illumina/Solexa platform, which has
become the method of choice for this kind of experiments, pro-
duces millions of sequence reads whose length does not exceed
75–100 bps. Thus, bioinformatic expertise is needed first to try
and reconstruct from the short sequence reads which were the
original DNA or RNA sequences bound by the protein, and also if
the respective abundance estimated from the short NGS sequence
reads can be considered to be due to a (statistically) significant
enrichment [7].
2 Materials
We list in the following bioinformatic tools and software that can

be used to process data from RIP-Seq and similar experiments.
Details on the purpose and application of each one can be found in
Subheading 3.
TopHat [8]: a fast splice junction mapper for RNA-Seq reads.
It aligns RNA-Seq reads to mammalian-sized genomes using the
ultrahigh-throughput short read aligner Bowtie [9], and then ana-
lyzes the mapping results to identify splice junctions between
exons. Source code and executables available at http://ccb.jhu.
edu/software/tophat/.
RIP-Seq Data Analysis 295
GSNAP [10]: another mapping software especially devised for

mapping short reads derived from RNA-Seq, taking into account
the presence of introns. Available at http://research-pub.gene.
com/gmap/.
RSEM [11]: A software package including tools for the map-
ping of RNA-Seq reads against a reference transcriptome annota-
tion, and the quantification of the expression level of each transcript.
It includes EBSeq [12], for the identification of differentially
enriched transcripts and/or genes. Available at http://deweylab.
biostat.wisc.edu/rsem/.
DESeq [13]—DEXSeq [14]: R packages to analyze count data
from high-throughput sequencing assays such as RNA-Seq and
test for differential expression and/or differential exon usage. Take
as input RNA-Seq reads aligned to the genome with tools like
TopHat. In order to be run, R and Bioconductor need to be
installed beforehand. Available at http://www.bioconductor.org.
Cufflinks [15]: a series of tools for RNA-Seq analysis, including
transcript assembly, transcript level estimation, and testing for dif-
ferential expression and regulation in RNA-Seq samples. It accepts
as input RNA-Seq reads aligned to the genome with tools like
TopHat. Available at http://cufflinks.cbcb.umd.edu/.
RIPSeeker [16]: Infer and discriminate RIP peaks from the
alignment of RIP-seq sequences to the genome. It also provides a
suite of bioinformatic tools addressing issues ranging from post-
alignment processing to visualization and annotation. In order to
be run needs R and Bioconductor to be installed beforehand.
Available at http://www.bioconductor.org.
PARalyzer [17]: Aligns sequencing reads generated by the
PAR-CLIP (Photoactivatable-Ribonucleoside-Enhanced Cross-
linking and Immunoprecipitation) protocol to the genome, and
identifies RBP binding sites by exploiting T=>C transitions induced
by cross-linking. Available at https://ohlerlab.mdc-berlin.de/
software/PARalyzer_85/.
3 Methods
3.1 RIP-Seq and Its In general, the steps that have to be followed for an IP experiment
Variants are similar regardless of the protein studied or the antibody
employed. From a bioinformatic point of view, the main differ-
ence is on whether and how the protein is cross-linked (fixed) to
the DNA/RNA before the IP [4]. For RBPs, this sometimes may
lead to different choices in the bioinformatic treatment of the
data. Sequencing of the RNA sample with cross-linking of RBPs
is often found in literature referred to as CLIP-Seq or HITS-
CLIP [18]. Also, PAR-CLIP [17] is based on the incorporation
of photoreactive ribonucleoside analogs, such as 4-thiouridine
(4-SU) and 6-thioguanosine (6-SG) into nascent RNA transcripts.
Cross-linking of the RBP to RNAs is then performed by irradiation

of the cells by UV light of 365 nm. The key feature of PAR-CLIP
is that the cross-linked sites are revealed by thymidine to cytidine
(T=>C) transitions in the cDNAs prepared from immunopurified
RNPs of 4-thiouridine-treated cells, and hence can be identified by
comparing the sequence reads obtained to the reference genome/
transcriptome sequences, and highlighting the transitions induced
by cross-linking.
On the other hand, even in the presence of highly specific anti-
bodies, IP experiments tend to be noisy (see Note 1). That is, in
theory nucleotide sequences isolated from the IP should be only
those that were bound by the DNA- or RNA-binding protein, and
thus, the only ones that were sequenced. But, in practice, IP’ed
samples contain a variable but sizable fraction of spurious sequences
that are due to a number of different factors, from poor specificity
of the antibody employed, to contamination from DNA of the lab
technician performing the experiment. It is thus common practice
to perform in parallel with the IP a control experiment, aimed
exactly at extracting only aspecific DNA or RNA reproducing
experimental noise. Different strategies can be employed in the
design of a control. Ideally, one should perform exactly the same
IP with the same antibody on the same cell line or tissue, but where
the protein studied is not expressed, that is, the corresponding
gene has been knocked out or down. Since in most of the cases this
approach is unfeasible, another idea is to perform an IP using an
antibody recognizing a protein not present in the cells studied.
Mouse or rabbit anti-IgG antibodies are often employed for this
task. Enriched DNA or RNA sequences will be then determined by
comparing enrichment in the IP’ed sample against enrichment in
the control sample (see Note 2) [7].
3.2 Bioinformatic Regarding the bioinformatic analysis of sequences derived from a

Analysis of RIP-Seq RIP-Seq experiment, we are still far from having a de facto stan-
Data dard, and just a handful of methods have been published so far.
On the other hand, as discussed in the previous sections, a RIP-Seq
(or CLIP-Seq or PAR-CLIP) experiment shares many similarities
to either full transcriptome sequencing (RNA-Seq)—since what is
sequenced is a subset of the transcriptome itself—or sequencing
applied to IP’ed chromatin (DNA bound by a protein, ChIP-Seq),
with the difference that IP’ed sequences in this case are whole
RNAs. Since for these two fields of NGS application literature for
tools and methods is booming, one possibility is to treat RIP-Seq
sequences as if they were derived from a RNA-Seq or ChIP-Seq
experiment, and apply to data produced the methods devised for
the latter (see Note 3).
Some general principles, however, remain the same, and
regardless of how they were isolated RNAs retrotranscribed into
cDNAs that have to be sequenced are fragmented at random
(either before or after retro-transcription). Then, fragments of

suitable size (usually 200 bp) are selected, sequenced at either
(single-end) or both (paired-end) strands producing sequence
reads of 50–100 bps. At this point, the rationale is simple: once
each sequence read has been assigned to its original transcript or
gene, then the abundance of the transcript in an RNA sample
should be proportional to the number of reads derived from it in
the sample, normalized by the RNA length. “Enrichment” for a
transcript or gene in an IP’ed sample can therefore be evaluated by
comparing the normalized count of sequence reads derived from it
in the IP with the read count obtained in the control experiment,
and by evaluating whether the difference in abundance can be
considered to be “significant” and not the effect of random fluc-
tuations of the data [7].
3.3 Sequencing As with most NGS applications for which the genome of the organ-
RNAs: Mapping ism studied is available, the first step of the analysis consists on the
Against the Genome mapping of sequence reads on the genome, that is, finding on the
genome the corresponding sequence region or, in case of RNAs,
the one that had been transcribed in the RNA itself. The sequence
read length currently available allows for unambiguous mapping of
nearly all the sequence reads not coming from repetitive DNA
regions, and sequence quality is good enough to allow at most two
or three substitutions per read.
Another aspect that has to be kept in mind is that if mature
RNAs are sequenced, a sizable fraction of the sequence reads
obtained will cover an exon–exon junction on the RNA. Hence,
mapping reads of this kind without allowing for insertions on the
genome will not find any match, since the alignment of the read
would have to be split by the intervening intron(s), as shown in
Fig. 1. However, fast and reliable sequence mapping tools for
Fig. 1 Result of the mapping of RNA-Seq reads derived from mature mRNAs on
the genome. Reads that covered exon–exon junctions will be split on the genome,
and their mapping will encompass the intervening intron
RNA-Seq experiments have been designed explicitly taking into

account this fact, such as TopHat and GSNAP. Tools of this kind
perform first a mapping without insertions on the genome, identi-
fying candidate “exons” corresponding to genomic regions cov-
ered by the mapped reads. Then, they align those reads that were
not mapped at the first round trying to split them on the candidate
exons identified at the first step. Further details can be found in
the RNA-Seq chapter of this volume. The mapping strategy to be
used (with or without insertions) depends on the RBP studied,
since some bind mature RNAs while others, like splicing factors,
pre-(m)RNAs. In case of doubt, we anyway advise the reader to
employ TopHat or GSNAP or a similar tool allowing for the map-
ping of reads spanning exon–exon junctions.
The reconstruction of exons and transcripts from the mapping
of the reads on the genome is not essential for organisms with a
reference gene annotation complete with splicing isoforms, or if no
genome but only a collection of transcript sequences is available.
In cases like these, the straightforward choice is to map the
sequence reads directly against the transcriptome, allowing only
substitutions, with tools like Bowtie. In this case reads mapping to
multiple positions (usually corresponding to different transcripts)
should not be discarded, since the different transcripts usually cor-
respond to isoforms of the same gene sharing some exons.
3.4 RIP-Seq: Using Starting from the considerations outlined in the previous sections,
Read Counts at the end of the mapping phase the mapping coordinates of each
sequence read can be crossed with the genomic coordinates of
exons, and of the respective transcripts or genes to which they
belong. This step, which can be performed by simple in-house
developed scripts, or by using a module contained in more general-
purpose analysis tools, yields the basic information needed to assess
enrichment, that is, a “read count” associated with each gene or
transcript annotated on the genome, in turn proportional to the
enrichment of the transcript in the samples sequenced. At this
point, if a control experiment is available, a simple but effective
strategy we implemented with good results [19], borrowing from
the analysis of SAGE data [20], evaluates differential enrichment
for a gene or transcript with respect to the control by using a 2 × 2
Chi-Square test. This kind of test is able to compare the relative
enrichment of the gene or transcript across the two experiments,
that is, it evaluates whether the relative enrichment of the gene dif-
fers, and if enough reads are associated to the gene to assume sig-
nificant difference and not a random fluctuation of the data deriving
from low read counts. Since a separate statistical test is performed
on each gene or transcript, suitable corrections for multiple test-
ing, like Bonferroni or Benjamini–Hochberg, should be then
applied to the resulting p-values, yielding family-wise error rates
and false discovery rates.
3.5 Treating RIP-Seq RNA-Seq, that is, sequencing of whole transcriptomes, has become
as RNA-Seq the de facto standard also for expression studies, aimed not only at
the reconstruction of the repertoire of alternative splicing and
RNAs expressed by a given cell line or tissue but more simply, given
an existing gene/transcript annotation, at the quantification of the
transcript level of genes, possibly at the single isoform level. The
application of RNA-Seq has boomed in the last few years, and sev-
eral different approaches have been introduced for its analysis. For
example, if a reliable gene annotation is available, DESeq identifies
differentially expressed genes starting from read counts associated
with them, with an approach similar to the one described in the
previous section. Instead of a Chi-Square test a negative binomial
distribution is employed, and mean and variance of read counts are
linked by local regression. DEXSeq performs similar computations
at the single exon level, in order to identify differentially enriched
alternative splicings. RSEM starts from the mapping of reads against
a reference transcriptome, and outputs for each gene or transcript
of the annotation abundance estimates (including estimation of the
relative abundance of alternative transcripts of the same gene),
95 % credibility intervals, and visualization files. Differentially
enriched genes or isoforms can be in turn identified through the
EBSeq package included in the RSEM software distribution.
Also, the Cufflinks pipeline is currently one of the most widely
used approaches in RNA-Seq analysis. Without delving into the
details, starting from the mapping of sequence reads against the
genomes, an estimate of the transcript level of each gene, and of
each annotated isoform of each gene can be computed, expressed
as Fragments Per Kilobase of exon per Million fragments mapped
(FPKM), a measure which is already normalized with respect to
the transcript length and the overall number of sequence reads
produced by an experiment. FKPM values associated with a given
gene or transcript can be compared across two or more samples,
and tested for significant differences denoting “under-” or “over-
expression” in the different conditions. Cufflinks returns plenty of
additional information, together with the FKPM values, like the
raw read counts associated with genes and transcripts, estimates of
the statistical significance of difference in gene (or isoform) expres-
sion, or alternative promoter usage.
Regardless of the approach, it is straightforward to see how
methods for RNA-Seq analysis can be applied to a sample extracted
by RIP-Seq. The reliability of the results depends on the distribu-
tion of the reads that were produced in the RIP-Seq sample, that
is, how the RBP-bound RNAs are enriched with respect to the
control, which in turn depends on the specificity of the antibody
employed. Also, RNA-Seq analysis tools need a suitable sequenc-
ing depth, and at least two replicates, to calculate the significance
of enrichment in a reliable way. An interesting feature of this kind
of analysis is that the FPKM (or analogous) values are computed at
the single transcript and isoform level, that is, return an abundance
estimate and differential expression for each of the alternative tran-
scripts of a gene. Thus, this might be of great interest if the RBP
studied is for example a splicing factor, or it is known to or sus-
pected to bind only some specific alternatively spliced transcripts.
3.6 Treating RIP-Seq Instead of RNAs, Chromatin Immunoprecipitation (ChIP) experi-

as ChIP-Seq ments permit to isolate DNA fragments bound by a protein of
interest, for example a transcription factor or a histone carrying a
given post-translational modification. The fragments that undergo
sequencing are usually of 200–300 base pairs, and each can be
sequenced at either 5′ end of the double strand. Thus, once
mapped onto the genome, sequence reads obtained in this way
tend to define the boundaries of the actual binding regions of the
protein. That is, the binding site is defined by a cluster of reads
mapping on the forward strand of the genome within a few base
pairs from one another, followed by another cluster mapping in
the same way on the antisense strand about 200–300 base pairs
downstream, that is, the two clusters mark the 5′ ends of the
original DNA fragments extracted from the IP (see Fig. 2).
Fig. 2 A typical ChIP-Seq enrichment profile corresponding to a DNA bound region. The actual protein-binding
site is usually located in correspondence to a local maximum (peak) of the sequencing coverage plot in the region
“Peak calling” algorithms, which identify DNA regions significantly

enriched, differ in the strategies used to assess enrichment (over a
control experiment), but nearly every single tool takes advantage
of the bimodality of sequence read enrichment on the genome in
correspondence of binding sites [7]. In RIP-Seq this feature is not
present, and instead of presenting a well-defined area of enrich-
ment as in ChIP-Seq corresponding to the protein binding site,
read enrichment on the genome is spread all over the exons of
bound RNAs, with all the transcribed region that should be
enriched with respect to control. Thus, an attempt can be made
nevertheless, but perhaps methods aimed at identifying broader
regions of enrichment [21] are more suitable than “peak-callers”
fine-tuned and optimized for the detection of isolate peaks corre-
sponding to transcription factor binding, like MACS [22].
3.7 RIPSeeker A recent tool that has been explicitly designed for RIP-Seq data
analysis is RIPSeeker, which however shares some similarities with
other ChIP-Seq analysis methods, especially those developed to
identify regions enriched for histone modifications. Once again,
the starting point of the analysis is the mapping of sequence reads
on the genome performed with TopHat, allowing for read split
mapping across different exons. The genome is then in turn divided
into “bins” of a given size, and the read count in each bin is com-
puted. The general idea is to first identify candidate RBP-bound
enriched bins by training a hidden Markov model (HMM) with an
Expectation Maximization procedure. Thus, the trained HMM
yields for each bin the probability of being a “RBP-bound region,”
or not. Once the HMM has been trained, the genome is scanned
with the HMM itself, yielding the Viterbi (highest likelihood) par-
tition of the genome into “RBP-bound” or “not bound” regions
given by merging of consecutive “bound” or “unbound” bins.
Comparison to other methods for RNA-Seq or ChIP-Seq seems to
demonstrate the advantages offered by this tool. As for other meth-
ods for NGS data processing, the usage is quite straightforward,
needing as input only the result of the read mapping, with most of
the needed parameters calculated automatically (like bin size) or set
to suitable default values.
3.8 PARalyzer As already described, cross-linking performed in PAR-CLIP exper-

iments induces a T=>C transition in the sequenced cDNAs. Thus,
bioinformatic analysis can exploit this feature in the identification
of the bound RNAs. PARalyzer is a tool devised for this task. Reads
are first aligned to the genome, with up to two mismatches
restricted to T to C conversions. Reads overlapping on the genome
by at least a single nucleotide are grouped together. Within each of
these read-groups, PARalyzer generates two smoothened kernel
density estimates, one for T to C transitions and one for non-
transition events. PARalyzer reports as interaction sites read groups
whose nucleotides maintain a minimum read depth (i.e., the

nucleotides have to appear in at least a minimum number of reads),
and where the likelihood of T to C conversion is higher than non-
conversion (that is, a significant number of T’s in a given base pair
of the genome are converted to C’s in sequenced reads). It should
be noticed that in this case it is possible to infer also the most likely
binding site positions on the RNA, which should correspond to
the T=>C conversions, rather than only determining which are the
RNAs bound by the protein studied.
4 Notes
1. The noisiness of IP experiments, either ChIP or RIP, makes

hard or often impossible to obtain final and clear results by
using just a single sequencing sample, even if compared to a
control. Indeed, the usual situation is that many “interesting”
targets for the protein studied fall into a sort of “gray area”, for
which the significance of the enrichment changes for example
according to the statistical framework employed or simply the
p-value and FDR thresholds chosen.
2. It is common practice nowadays to perform replicates of an
experiment, perform the bioinformatic analysis of the data on
each replicate, and merge the results at the end. Biological rep-
licates account for variability in the IP experiment, while techni-
cal replicates re-sequence the same starting material and account
for variability and error in sequencing. While the sequencing
steps can be nowadays considered to be reliable, at least for
RNA or IP sequencing, biological replicates are useful to iden-
tify cases in which “something went wrong” in the IP. Our
advice is, given also the decreasing cost of sequencing experi-
ments, and the possibility to multiplex different samples on the
same sequencing lane, to perform at least three biological repli-
cates. The final merging of the results can be then performed
according to different criteria, e.g., pooling together the read
counts and/or FKPM values of the replicates and performing
significance tests on these values, or process the replicates sepa-
rately with a “majority vote” at the end (a given gene, tran-
script, or region has to pass significance thresholds in two
experiments out of three). The latter strategy is more suitable in
cases where one replicate shows little correlation to the others.
3. The best results can be obtained by applying more than one
single analysis tool, like for example raw read count compari-
son, together with the FPKM analysis of Cufflinks, and a RIP-
Seq specific method like RIPseeker. A critical comparison of the
results, supported by experimental validation, will highlight the
key features of the experiment performed and of the RBP
investigated.
References
1. Horner DS, Pavesi G, Castrignanò T et al (2010) Stewart RM, Kendziorski C (2013) EBSeq: an
Bioinformatics approaches for genomics and empirical Bayes hierarchical model for infer-
post genomics applications of next-generation ence in RNA-seq experiments. Bioinformatics
sequencing. Brief Bioinform 11(2):181–197 29(8):1035–1043
2. Marioni JC, Mason CE, Mane SM et al (2008) 13. Anders S, Huber W (2010) Differential expres-
RNA-seq: an assessment of technical repro- sion analysis for sequence count data. Genome
ducibility and comparison with gene expres- Biol 11(10):R106
sion arrays. Genome Res 18(9):1509–1517 14. Anders S, Reyes A, Huber W (2012) Detecting
3. Glisovic T, Bachorik JL, Yong J et al (2008) differential usage of exons from RNA-seq data.
RNA-binding proteins and post-transcriptional Genome Res 22(10):2008–2017
gene regulation. FEBS Lett 582(14):1977–1986 15. Trapnell C, Roberts A, Goff L et al (2012)
4. Collas P (2010) The current state of chroma- Differential gene and transcript expression
tin immunoprecipitation. Mol Biotechnol analysis of RNA-seq experiments with TopHat
45(1):87–100 and Cufflinks. Nat Protoc 7(3):562–578
5. Furey TS (2012) ChIP-seq and beyond: new 16. Li Y, Zhao DY, Greenblatt JF (2013)
and improved methodologies to detect and RIPSeeker: a statistical package for identifying
characterize protein-DNA interactions. Nat protein-associated transcripts from RIP-seq
Rev Genet 13(12):840–852 experiments. Nucleic Acids Res 41(8):e94
6. Yuan J, Muljo SA (2013) Exploring the RNA 17. Corcoran DL, Georgiev S, Mukherjee N et al
world in hematopoietic cells through the lens (2011) PARalyzer: definition of RNA binding
of RNA-binding proteins. Immunol Rev sites from PAR-CLIP short-read sequence
253(1):290–303 data. Genome Biol 12(8):R79
7. Pepke S, Wold B, Mortazavi A (2009) 18. Licatalosi DD, Mele A, Fak JJ et al (2008)
Computation for ChIP-seq and RNA-seq HITS-CLIP yields genome-wide insights into
studies. Nat Methods 6(11 Suppl):S22–S32 brain alternative RNA processing. Nature
8. Trapnell C, Pachter L, Salzberg SL (2009) 456(7221):464–469
TopHat: discovering splice junctions with 19. Mihailovic M, Wurth L, Zambelli F et al
RNA-Seq. Bioinformatics 25(9):1105–1111 (2012) Widespread generation of alternative
9. Langmead B, Trapnell C, Pop M, Salzberg SL UTRs contributes to sex-specific RNA binding
(2009) Ultrafast and memory-efficient align- by UNR. RNA 18(1):53–64
ment of short DNA sequences to the human 20. Michiels EM, Oussoren E, Van Groenigen M
genome. Genome Biol 10(3):R25 et al (1999) Genes differentially expressed in
10. Wu TD, Nacu S (2010) Fast and SNP-tolerant medulloblastoma and fetal brain. Physiol
detection of complex variants and splicing in Genomics 1(2):83–91
short reads. Bioinformatics 26(7):873–881 21. Micsinai M, Parisi F, Strino F et al (2012)
11. Li B, Dewey CN (2011) RSEM: accurate tran- Picking ChIP-seq peak detectors for analyzing
script quantification from RNA-Seq data with chromatin modification experiments. Nucleic
or without a reference genome. BMC Acids Res 40(9):e70
Bioinform 12:323 22. Zhang Y, Liu T, Meyer CA et al (2008) Model-
12. Leng N, Dawson JA, Thomson JA, Ruotti V, based analysis of ChIP-Seq (MACS). Genome
Rissman AI, Smits BM, Haag JD, Gould MN, Biol 9(9):R137
Part III
Web Resources for RNA Data Analysis

Chapter 19
The ViennaRNA Web Services

Andreas R. Gruber, Stephan H. Bernhart, and Ronny Lorenz
Abstract
The ViennaRNA package is a widely used collection of programs for thermodynamic RNA secondary
structure prediction. Over the years, many additional tools have been developed building on the core
programs of the package to also address issues related to noncoding RNA detection, RNA folding kinetics,
or efficient sequence design considering RNA-RNA hybridizations. The ViennaRNA web services provide
easy and user-friendly web access to these tools. This chapter describes how to use this online platform to
perform tasks such as prediction of minimum free energy structures, prediction of RNA-RNA hybrids, or
noncoding RNA detection. The ViennaRNA web services can be used free of charge and can be accessed
via http://rna.tbi.univie.ac.at.
Key words RNA secondary structure prediction, RNA-RNA interaction prediction, Noncoding
RNA detection, Sequence design
1 Introduction
The ViennaRNA package started as a small collection of tools

aiming to provide fast implementations of algorithms for predic-
tion and comparison of RNA secondary structures [1]. Over the
years, the collection of algorithms and tools has substantially grown
[2], now providing a comprehensive set of programs for thermo-
dynamic RNA secondary structure prediction, including programs
for efficient prediction of RNA-RNA interactions. In order to open
this program package to a broad community, the ViennaRNA
group provided web interfaces to the most commonly used tools
early on [3]. In 2008, the original ViennaRNA secondary struc-
ture server was replaced by a revamped server attributing to recent
developments in web programming and offering web interfaces to
an extended set of tools [4].
In its current form, the ViennaRNA web services (http://rna.
tbi.univie.ac.at) offer easy and user-friendly access to a large vari-
ety of computational tools useful to the RNA community. Services
can basically be grouped into five categories as illustrated in Fig. 1.
307
308 Andreas R. Gruber et al.
Fig. 1 Outline of the ViennaRNA web services. Based on their functionalities, programs are grouped into five
categories
The core set of web services consists of tools for thermodynamic

RNA secondary structure prediction including the prediction of RNA-
RNA interactions. We briefly outline individual tools below and
give detailed descriptions of the programs in Subheading 3.
The most highly accessed service is the RNAfold server. It
allows users to determine the minimum free energy (MFE) RNA
secondary structure of a single DNA or RNA sequence, while the
RNAalifold server predicts the consensus structure of a set of
aligned sequences [5, 6]. Exploring folding kinetics of an RNA
molecule of interest can easily be accomplished with the barriers
server [7]. For prediction of RNA-RNA interactions, the
ViennaRNA web services offer access to several tools implementing
different algorithms to address this complex problem. RNAcofold
computes the secondary structure of an RNA-RNA dimer by a
simple algorithm that concatenates the two sequences [8]. On the
contrary, RNAup first computes the energy needed to open exist-
ing secondary structures in both molecules and then reports the
single most favorable interaction site [9]. RNApredator tackles a
problem of particular interest to the bacterial community, namely,
the genome-wide prediction of RNA molecules targeted by bacte-
rial small RNAs mediated by RNA-RNA hybridization [10]. On the
side of sequence design, RNAinverse is an algorithm that aims
to find RNA sequences that can adopt a desired RNA secondary
ViennaRNA Web Services 309
structure, while RNAxs is an efficient tool for the design of optimal

siRNA sequences based on a set of target sequences [11]. The
ViennaRNA group was among the first to develop efficient tools
for prediction of noncoding RNAs [12, 13]. For the purpose of
noncoding RNA gene finding, the SCA server screens alignments
of RNA sequences for signs of evolutionarily conserved secondary
structures [14]. The RNAz approach is more sophisticated,
employing a machine learning approach to predict noncoding
RNAs based on both thermodynamic stability and evolutionary
conservation [13, 15, 16]. With Bcheck, the ViennaRNA web ser-
vices also offer a specialized tool for RNAseP gene finding [17].
2 Materials
2.1 Hardware Use of the ViennaRNA web services does not require installation
and Software of particular software other than a common web browser that sup-
Requirements ports JavaScript. We recommend, however, to use modern web
browsers such as Mozilla Firefox, Google Chrome, or Safari since
these browsers natively support rendering of image files in SVG
format and do not require installation of additional plug-ins.
2.2 Input Data Tools from the ViennaRNA web services require input of DNA or
Formats RNA sequences in either FASTA or ClustalW format, two widely
used formats for storing sequence information. The user can paste
sequences into the corresponding field or upload files via a file
upload dialog. Unless DNA thermodynamic parameters are specifi-
cally selected, sequences will automatically be converted to RNA
sequences (replacing T residues by U) upon submission of the job.
The programs of the ViennaRNA package do not have any intrin-
sic restriction on the length or number of sequences that can be
processed; certain restrictions do apply for the web services,
though. Limitations have been chosen gracefully and are stated on
the web pages of the individual services.
3 Methods
3.1 Secondary The RNAfold server offers an interface to the most basic secondary
Structure Prediction structure prediction program within the ViennaRNA web services.
with the RNAfold Web In short, the RNAfold algorithm uses a set of thermodynamic
Server energy parameters in a dynamic programming framework to
calculate the minimum free energy (MFE) and its associated sec-
3.1.1 Description ondary structure of a single RNA molecule. In addition, the parti-
and Scope of the Program tion function of the equilibrium ensemble of all secondary
structures is computed which in turn allows to infer base pair prob-
abilities and reliability measures for the MFE prediction.
3.1.2 Description The input sequence for the RNAfold web server can be uploaded
of Input Data and Program from the client computer as FASTA file or simply typed/pasted
Options into the input field of the web interface. Without changing the
default options, the server computes the MFE and partition func-
tion for the ensemble of canonical structures, i.e., structures with-
out helices of size 1. However, structures having such isolated base
pairs can be included in the prediction by unchecking the “avoid
isolated base pairs” option. Since a weak GU pair at the end of a
helix is likely to unfold, especially if the helix branches off a large
loop, exclusion of such conformations usually yields more stable
structures. Therefore, for some predictions, it might be a good
idea to use the appropriate checkbox in the basic options. The
auxiliary constraint folding feature of RNAfold allows for an even
bigger restriction of the secondary structure space. Here, a pseudo-
dot-bracket notation consisting of structure constraint symbols
allows to specify whether a nucleotide must be unpaired (symbol:
x), should be paired with another unspecified base (symbols “|”,
“<”, “>”), or has a predefined pairing partner (symbols “(”,“)”).
It has to be noted that the web server allows for MFE predic-
tion only; however, relying on this single representative of the
ground state may be misleading. Unless the sequence length is
very small, the partition function and the ensemble properties
derived from the partition function (1) provide a reliability mea-
sure of the prediction and (2) take into account that an RNA mol-
ecule in thermodynamic equilibrium is in constant flux between
several distinct low-energy structures and the ground state.
3.1.3 Description Upon successful computation of the results, the comprehensive

of the Program Output prediction results page not only displays the thermodynamic prop-
erties of the sequence but also comprises an interactive graphical
representation of the predicted secondary structure. An example of
the results page is given in Fig. 2.
The top of the results page lists the minimum free energy and
its associated secondary structure in dot-bracket notation and
ensemble properties like the Gibbs free energy of the ensemble,
frequency of the MFE structure within the ensemble, and the
ensemble diversity. Along with the MFE structure, an additional
representative of the equilibrium ensemble is given by the centroid
structure, i.e., the structure with the smallest distance to the
Boltzmann-weighted set of all structures. By definition, this struc-
ture comprises all base pairs with a probability pij > 0.5. While the
frequency of the MFE structure reflects its equilibrium probability,
the ensemble diversity gives information about the folding poten-
tial of the nucleic acid sequence. Lower values indicate that the
ensemble is dominated by a single, low free energy structure,
whereas several diverse structures with almost equal stability give
rise to a larger value. A complete picture of the pairing potential is
given in the dot plot which depicts the base pair probabilities in the
Fig. 2 Sample RNAfold output depicting the minimum free energy secondary structure, the centroid structure,
and graphical representations thereof
form of a matrix representation. Here, the probabilities are indi-

cated in the upper triangle of the matrix by a black square with its
size proportional to the pairing probability. In contrast to this
ensemble-wide insight, the lower half of the matrix reflects the
base pairs present in the MFE structure. The dot plot can be either
viewed/downloaded as PDF or PostScript file or be converted to a
bitmap image format. It is worthwhile to note that the PostScript
image is a simple text file from which the numerical values of the
base pair probabilities can be extracted quite easily. While the dot
plot representation is information rich, it may be difficult for an
untrained user to easily draw conclusions from it. Therefore, the

structure representatives given in the results can be annotated by
the base pair probability or positional entropy by a color code
ranging from blue to red indicating low and high values, respec-
tively. For the probability annotation, base pairs in the dot-bracket
string are coded by their respective equilibrium probabilities, while
unpaired nucleotides are coded by their probability of being
unpaired. Thus, structure parts being coded by red color are more
reliable than those with blue or green color. A similar measure,
which, however, takes all pairing possibilities of a nucleotide into
account, is the positional entropy. Again, red color indicates high
reliability of the prediction, whereas blue or green colors indicate
parts of the structure with low reliability.
The default graphical output consists of two interactive sec-
ondary structure plots of both structure representatives as well as a
mountain plot which subsumes all thermodynamic data of the pre-
diction. While the user can select the type of color-coded data in
the interactive plots, each graphical representation can also be
downloaded in either vector graphic format such as PDF or
PostScript or as a bitmap graphic via the image converter. Finally,
the mountain plot superpositions MFE structure, centroid
structure, and base pair probabilities in a single plot. In this two-
dimensional plot, the x-axis displays the sequence position while
the y-axis, the height of the mountain, is given by the number of
base pairs enclosing the respective position. This means that for the
MFE and centroid structure, plateaus represent unpaired nucleo-
tides, i.e., loops, whereas slopes represent the 5′ or 3′ side of heli-
ces, depending on whether they face upward or downward,
respectively. Additionally, the mountain plot is accompanied by a
plot of the positional entropy.
3.2 Prediction For certain classes of functional RNA molecules, secondary struc-
of Consensus ture elements are often evolutionarily more conserved than the
Secondary Structures underlying primary sequence. The accuracy of single-sequence
with the RNAalifold RNA secondary structure prediction tools is limited, irrespective of
Web Server the approach or method employed [18]. Consensus structure pre-
diction of a set of (aligned) sequences is thus a commonly used way
3.2.1 Description to increase the predictive power of structure prediction tools and
and Scope of the Program to spot evolutionarily conserved, often functional structural motifs.
In addition to thermodynamic folding, consensus structure
prediction relies on signals of sequence variation. In particular,
consistent mutations (one base in a base pair is changed with
respect to a reference sequence) and compensatory mutations
(both bases are changed, e.g., a G-C base pair in one sequence and
an A-U base pair in the other) are counted as supporting evidence
for a conserved secondary structure, while base pairs that can only
be formed in a single or few sequences contradict the formation of
a conserved structural motif.
RNAalifold follows the align first, fold later strategy of consen-

sus structure prediction. That means it takes a multiple sequence
alignment as input and predicts the consensus structure by mini-
mizing a pseudo-free energy composed of the mean free energy of
the structures of the single sequences and the number of consistent
and compensatory mutations [6].
3.2.2 Description The input for RNAalifold consists of multiple sequence alignments
of Input Data and Program in FASTA or ClustalW format. Information on the phylogenetic
Options tree underlying the alignment is currently not considered by the
algorithm. It must be noted that the quality of the structure pre-
diction is directly influenced by the quality of the alignment. For
average pairwise sequence identities above 75 %, alignments gener-
ated by simple sequence-only-based alignment tools such as
ClustalW (http://www.ebi.ac.uk/Tools/msa/clustalw2) are suffi-
cient for good results. For a set of highly divergent sequences, we
recommend using an alignment program that considers both
sequence and secondary structures such as LocARNA (http://rna.
informatik.uni-freiburg.de/LocARNA).
RNAalifold has two ways of scoring evolutionary conservation
of base pairs. The first approach simply counts occurrences of
canonical base pairs and counterexamples for each base pair. The
algorithm then only considers base pairs with less than three coun-
terexamples. The other scoring scheme uses RIBOSUM scores,
which are substitution matrices for RNA sequences, to compute an
alignment-like score for each possible base pair. The main advan-
tage of the RIBOSUM score is that it does not penalize counterex-
amples as harshly as the simple scoring scheme. For alignments with
many sequences, this feature will allow RNAalifold to infer a con-
sensus structure even if there are some counterexamples, possibly
due to faulty alignment choices. The new version of RNAalifold [6]
also introduced an alternate way of dealing with gap characters in
the aligned sequences, with gaps not contributing to the energy
computation. We recommend the default setting which is the new
RNAalifold version with RIBOSUM scoring, but other variants can
be chosen as well.
Many of the parameters of RNAfold apply to RNAalifold as
well, but there are two parameters unique to RNAalifold. The
parameter “Weight of covariance term” controls the relative
weighting of energy contributions to contributions from structural
conservation supporting base pairs. The default value of 1 implies
equal contribution, while a value below 1 will down-weigh the
conservation supporting score.
On the other side, “Penalty for non-compatible sequences”
controls the impact of counterexamples. If this value is set to 0,
non-supporting base pairs will not be penalized. Both parameters
are accessible via the advanced options field.
3.2.3 Description The RNAalifold output page provides the consensus secondary struc-
of the Program Output ture in dot-bracket notation and the computed minimum free energy,
which actually is a pseudo-energy with contributions coming from
both sequence-averaged free energy minimization and base pair
conservation bonus energies. Each of the individual contributions is
listed separately. The textual representation of the secondary struc-
ture is assisted by graphical output in the form of a structure-
annotated alignment and a visualization of the secondary structure
highlighting consistent and compensatory mutations (Fig. 3).
3.3 RNA-RNA RNAcofold is an algorithm for prediction of the hybrid structure

Interaction Prediction: formed between two RNA molecules [8]. The approach is simple:
RNAcofold it concatenates the two sequences and computes the MFE struc-
ture as if it was for a single sequence. The only difference to single-
3.3.1 Description
molecule secondary structure prediction is that all structural
and Scope of the Program
features containing the end of the first and the start of the second
molecule (called “cut point”) are treated as exterior loops. The
approach is versatile enough to predict multiple interaction sites,
while it is fast enough to be computed for molecules of several
thousand nucleotides.
The drawback of this approach is, however, that it cannot pre-
dict structures that would be considered as pseudo-knots, i.e.,
crossing base pairs, in the single structure case. This excludes many
intermolecular structures known in nature, e.g., the kissing hairpin
structural motif. On the other hand, this ansatz makes it possible
to easily compute many of the features known from single-sequence
folding, mainly via exploiting partition function-related informa-
tion. In detail, partition function calculations can be used to pre-
dict equilibrium concentrations of the dimers dependent on some
initial monomer concentrations.
3.3.2 Description RNAcofold takes two RNA sequences termed as sequence A and
of Input Data and Program sequence B as its input and will, by default, not only compute
Options the partition function and base pair probabilities of the heterodi-
mer (i.e., the structure formed by a dimer of the two input
sequences, hence denoted as AB) but also of the homodimers
(AA and BB), which are the structures that two molecules of
sequence A or sequence B will form. Binding of two molecules is
always a concentration-dependent process. Therefore, the
RNAcofold server offers two ways of predicting concentrations
of dimers. It either varies the monomer concentrations using
predefined values or uses a list of concentrations provided by the
user. In the latter case, two values separated by a white space
have to be provided by line.
3.3.3 Description The RNAcofold output page, by default, displays the heterodimer
of the Program Output structure in dot-bracket notation with an “&” sign separating the
two sequences. In addition, dot plots of the heterodimer structure
Fig. 3 Sample RNAalifold output. (a) Visualization of the consensus secondary structure predicted by RNAalifold.
The number of different types of base pairs observed for a particular pair of positions in the alignment is color
coded. Pale colors indicate contradicting evidence, meaning that not all sequences in the alignment can form
the particular base pair. Red indicates base pairs that show conservation at nucleotide level in all sequences,
while base pairs colored in green, for example, are supported by compensatory/consistent mutations; in this
case, three different types of base pairs are observed. (b) Structure-annotated multiple sequence alignment.
The same color-coding scheme as discussed above applies here
(AB) as well as the two monomer structures (A, B) and the two
homodimer structures (AA, BB) are provided. In the dot plots of
the dimers, a cross sign is used to indicate the cut point that sepa-
rates the two molecules. The output page also contains ensemble
free energies of all five species, as well as the energy that is gained
by forming the heterodimer compared to the two monomers
denoted as “Delta G for heterodimer binding.”
If the option “Vary concentration” was selected, concentration
dependency plots showing the relative concentrations for each
dimer (AA, BB, AB) and the monomers (A, B) will be generated.
If the user specified concentrations, results will be provided in form
of a table.
3.4 RNA-RNA RNAup takes another approach for the prediction of intermolecu-
Interaction Prediction: lar RNA-RNA interactions [9]. It basically models RNA-RNA
RNAup interaction as a two-step process, which resembles a simplified
approach of what is actually happening in vivo when RNA-RNA
3.4.1 Description
hybrids form. In the first step, the interaction sites have to be made
accessible (i.e., single stranded) in order to engage in intermolecu-
lar base pairing. In the second step, the supposed interaction sites
of two RNA molecules can then form base pairs between each
other. The RNAup algorithm follows this model by first comput-
ing the opening energy for every sequence window of length w.
Computation of interaction energies for all intervals with a maxi-
mum length of W is then accomplished in a second step. Lastly,
results from these two steps are merged to obtain the total interac-
tion energy, which is minimized to find the best site of interaction.
This approach allows any form of structure around the interaction
sites, thus being able to also predict for example the kissing hairpin
structural motif. However, the drawback of this approach is that
only a single interaction site can be taken into account. Therefore,
RNAup is mainly used for the prediction of interactions of short
RNAs on long target RNAs such as bacterial sRNAs on mRNAs.
3.4.2 Description As RNAcofold, RNAup takes two sequences as an input. Unique

of Input Data and Program to RNAup are two parameters controlling aspects of the interac-
Options tion prediction. The default option of the parameter “length of the
unstructured region” is set to 4, meaning that RNAup will calculate
the probabilities of being unpaired for stretches of 4 consecutive
nucleotides. Note that via a simple transformation of these proba-
bilities, one can also easily calculate the energy needed to open
existing secondary structures. The other parameter “maximal
length of the region of interaction” controls the size of the interac-
tion region that is reported, set to 25 nt by default. RNAup reports
probabilities of being unpaired by considering contributions from
all loop types. In the advanced options field, the user can specifi-
cally select that, for example, also probabilities of being unpaired in
a hairpin loop (H), interior loop (I), external loop (E), or multi-
loop (M) are being reported.
3.4.3 Description The RNAup output page will report the location of the interaction
of the Program Output sites in both sequences and the interaction secondary structure in
dot-bracket notation. The interaction free energy is composed of
multiple contributions (energy gained from duplex formation and
energies needed to open existing secondary structures in the two
molecules), which are separately listed. The output is also accom-
panied by two plots depicting the interaction free energy (red) and
the energy needed to open existing secondary structure along the
longer of the two sequences. The original RNAup output text file
is also provided, which lists detailed energy values for each nucleo-
tide and, if selected, detailed energy contributions of individual
loop types.
3.5 RNA Design While RNA secondary structure prediction is a computationally

with the RNAinverse well-defined problem with the solution guaranteed to be optimal
Web Server given the set of energy parameters, the inverse problem of finding
a sequence that adopts a given secondary structure can only be
3.5.1 Description
solved in terms of heuristic approaches. RNAinverse implements
such a heuristic approach of finding sequences that can fold into a
given secondary structure and is thus a valuable tool for designing
RNA sequences. The RNAinverse algorithm starts with a randomly
selected nucleotide sequence and keeps changing nucleotides until
the RNA sequence adopts the desired fold. The algorithm has two
modes of operandi. In the MFE mode, RNAinverse searches for
sequences that adopt the desired structure as their minimum free
energy structure, while in the partition function mode, the fre-
quency of the structure in the ensemble of potential secondary
structures a sequence can adopt is maximized. While finding
sequences that adopt the desired structure usually succeeds within
a few iterations of the algorithm, it has to be noted that the longer
and complex the desired structure is, the search becomes more dif-
ficult, and the algorithm might not be able to find a sequence
meeting the criteria. In that case, the sequence with the minimal
edit distance to the desired structure of the sequences generated is
reported.
3.5.2 Description Input for the RNAinverse server is the desired secondary structure
of Input Data and Program in dot-bracket notation (see Subheading 3.10. for more details) and
Options optionally an RNA sequence string of equal length. If no sequence
is provided, RNAinverse will generate a random sequence. Note
that when specifying the sequence, the letter “N” can be used to
indicate a random base. Certain restrictions also apply to the sec-
ondary structure. A hairpin loop has to consist of at least three
unpaired bases, and crossing base pairs, i.e., pseudo-knots, are not
allowed. A sample input of a sequence-structure pair is given below:
.((((...))))...((((...)))).
NANNNNNNNNNUNNNNNNNNNNNNNNN
In this example, the desired secondary structure consists of

two small hairpin structures with each four base pairs. The first
base pair in the first stem is set to be an A-U pair. As noted before,
depending on the initial (random) sequence provided, the algo-
rithm may fail to find a secondary structure meeting the desired
criteria. Instead of running RNAinverse several times, we advise
the user to set the parameter “number of sequences to be gener-
ated” to a higher number than 1, thereby increasing chances that a
valid sequence is found.
3.5.3 Description The results page lists the generated sequences and the minimum
of the Program Output free energy of the predicted structure. Links to the RNAfold server
that automatically paste the sequence into the input field are pro-
vided for each sequence. This allows the user to explore the pre-
dicted secondary structure and folding potential of the sequences
in more detail. We do recommend to perform this analysis, espe-
cially for sequences listed under “Results for thermodynamic
ensemble prediction.” The procedure of optimizing the frequency
in the thermodynamic ensemble may lead to deviation of the mini-
mum free energy structure from the desired structure. When mul-
tiple sequences were generated, results are presented in sortable
tables, which allows convenient sorting, e.g., by free energy.
3.6 Exploring Some RNA molecules are known to take a very long time or even
Folding Kinetics never adopt their ground-state structure after transcription. These
with the Barriers RNAs usually get trapped in a so-called metastable conformation,
Server either due to co-transcriptional folding or while refolding from a
denatured state. Approaches like RNAfold that consider the RNA
3.6.1 Description
molecule to be in thermodynamic equilibrium can therefore not
be used to predict this behavior since they exclude folding kinetics
by design. However, when all secondary structures an RNA mol-
ecule can adopt are known, they can be interpreted as points in a
landscape with their corresponding free energy marking the eleva-
tion. A move set, e.g., inserting/removing a base pair, then defines
the neighborhood of any given structure. Finally, folding kinetics
is given by how RNA molecules explore this landscape which in
turn can be modeled as a Markov process. Given an initial popula-
tion density of the structure space, this allows to mathematically
solve the problem of the population density for an arbitrary point
in time. However, the number of secondary structures an RNA
molecule can fold into grows exponentially with its sequence
length making the solution impractical to compute even for very
short RNAs of about 50 nt. Thus, a widely used approach to cir-
cumvent this limitation is to heavily shrink the state space by
lumping secondary structure states together in the so-called macro
states. One possibility of macro-state generation is to use barrier
trees where all structures that reach the same local minimum via
consecutive steepest descent moves belong to the same macro state.
Structures connecting the macro states are considered saddle

points and represent the energy barriers separating the local min-
ima. Taken together, the local minima, the saddle points, and the
energy barriers define the leaves, internal nodes, and edge lengths
between the nodes of a so-called barrier tree, respectively. Still, this
approach relies on the enumeration of the secondary structures
and thus is applicable for RNA molecules up to a length of some
100 nt only.
The barriers web server employs the programs RNAsubopt
[19], barriers [7], and treekin [20] in order to generate a set of
about ten million low free energy structures, compute the barrier
tree and transition rate matrix for the Markov process, and solve
the master equation of the system, respectively.
3.6.2 Description A single RNA sequence with at most 100 nt in length may be used
of Input Data and as input for the barriers server. In order to keep the barrier tree
Program Options simple, the user may set the maximum number of leaves (default
50). Furthermore, local minima that do not exhibit an energy bar-
rier higher than a certain user-defined threshold can be merged.
This allows for keeping the number of leaves in the barrier tree
small even for RNAs with high diversity of low-energy structures.
The number of secondary structures generated by RNAsubopt can
be limited again by omitting structures with isolated base pairs,
which is the default, or removing all structures where a weak GU
pair terminates a helix.
3.6.3 Description Among the results of the barriers web server calculations are the
of the Program Output number of secondary structures obtained by RNAsubopt as well as
the resulting barrier tree in graphical form. Additionally, the refold-
ing path between any of the ten lowest free energy macro states can
be displayed with an animated SVG image. In this image, the
energy profile in the form of a height plot and a circular and a
regular structure representation are shown for each intermediate
structure along the refolding path.
The folding kinetics section of the output consists of the
downloadable transition rate matrix of the Markov process and
four options that control the kinetic simulation. First, the initial
structure, i.e., the structure that has a population density equal to
100 % at the start of the simulation, can be chosen from the ten
lowest free energy macro states. The time span of the simulation,
which is measured in arbitrary time steps, can be set via the “Start
time” and “Stop time” values. Usually, the default values should
suffice; however, for RNAs which populate a metastable state for
too long, the stop time may have to be set to a value considerably
higher than one million. In any case, the “Stop time” has to be set
such that the plot displays macro-state number 1 with highest den-
sity at the end of the simulation. If no further changes appear in the
population density over time, the system has reached equilibrium.
To keep the simulation plot clean from local minima which are
sparsely populated, the treekin options allow for ignoring all macro
states which do not exceed a predefined threshold.
3.7 Noncoding RNA Detection of functional RNA structures and noncoding RNA (ncRNA)
Gene Prediction genes is a challenging task. Simply folding the sequence of interest
with the RNAz Server and assessing its minimum free energy is not a sufficiently good
and Related Tools approach, since the MFE is highly dependent on the G + C content
of the sequence and the length of the sequence. Nevertheless,
3.7.1 Description
functional RNA structures often show elevated thermodynamic
stability when compared to a set of randomized sequences of the
same length and base composition [13]. Moreover, functional
RNAs are also often found to be evolutionarily conserved [12].
Aiming for accurate prediction of specifically the class of thermo-
dynamically stable and evolutionarily conserved RNA structures,
the ViennaRNA group developed RNAz [13, 15] and the RNAz
server for easy and user-friendly access to the RNAz toolbox [16].
In detail, the RNAz approach measures thermodynamic stability
in terms of a normalized z-score of folding energies. It indicates how
many standard deviations the structure of a given sequence is more
or less stable than that of random sequences of the same length and
base composition. Negative z-scores thus indicate that a sequence is
more stable than expected by chance. The second component is the
evaluation of the structural conservation, which requires that the
input for RNAz is an alignment of at least two sequences. Structural
conservation is measured by the structure conservation index (SCI),
which evaluates structural conservation indirectly through folding
energies rather than on the structure level itself (see ref. 14 for
details). In theory, a SCI above 1 indicates structural conservation
supported by compensatory mutations. The SCI is, however,
strongly influenced by the alignment quality with respect to second-
ary structures, and as a rule of thumb, a SCI close to or above the
average pairwise identity of the sequences in the alignment can be
considered as good. While this is valuable information to know when
evaluating individual RNAz predictions in detail, we generally advise
the user to assess predictions via the RNA class probability, which is
calculated by support vector machine classification combining ther-
modynamic stability and structural conservation.
3.7.2 Description Detection of functional RNA structures with the RNAz server is a
of Input Data and multistep process. As stated before, RNAz requires multiple
Program Options sequence alignments as input. Alignments can be provided in vari-
ous formats such as ClustalW or MAF (http://genome.ucsc.edu/
FAQ/FAQformat.html) but have to consist of at least two sequences
and should not be less than 50 nt in length. Multiple alignments in
ClustalW have to be separated by a header line starting with
“CLUSTAL W.” After successful upload of the alignments to the
server, the user will be redirected to the analysis options page (a
screenshot is shown in Fig. 4). The RNAz version hosted at the
Fig. 4 Screenshot of the RNAz analysis options page. For each field, detailed help is available by clicking on
the question mark icons
server is limited to analyzing alignments with at most six sequences

and a maximal alignment length of 200 nt. Longer alignments can,
however, be uploaded and will be preprocessed with the parameters
set on the analysis options page. In detail, long alignments are sliced
into smaller alignments given a user-defined window size and step
size. Default values have been proven to be adjusted well but may
require to be set to different values when screening, for example,
for ncRNAs of a certain size. The RNAz pipeline will also filter
alignments and remove sequences that have too many repeat-
masked letters (denoted by lowercase letters) or gaps. If there are
more than six sequences left after filtering, optimized algorithms in
the RNAz package will automatically select a set of appropriate
sequences according to user-controlled values, e.g., the value speci-
fied in the target mean pairwise identity field. Finally, on the last
page, the user can select between various output options and leave
an e-mail address for notification once computation is completed.
3.7.3 Description The RNAz output page is clearly structured providing links to
of the Program Output detailed analyses for each alignment analyzed. In particular, for
each alignment screened, the SCI, z-score, and RNA class proba-
bility are reported among other features. Each prediction is accom-
panied by a structure-annotated alignment and visualization of the
consensus secondary structure (see Subheading 3.2.3 for detailed
description of these plots). We would like to draw attention of the
reader to the excellent help pages of the RNAz server, which dis-
cuss many of the analysis options and the RNAz output in great
detail. In default mode, RNAz screens both the alignment in sense
and antisense direction for putatively conserved functional struc-
tures. Considering the hit with the higher probability of the two
screened windows is a first good choice. For a more detailed analy-
sis, we refer the user to specialized tools such as RNAstrand ([21];
http://www.bioinf.uni-leipzig.de/Software/RNAstrand). Results
are kept on the server for 30 days and can be accessed by the URL
outlined on the results page. RNAz is also available as a stand-
alone program and can be obtained from http://www.tbi.univie.
ac.at/~wash/RNAz. For additional information on noncoding
RNA gene detection, we refer the interested reader to ref. 22.
3.7.4 Related Tools
for Noncoding RNA While RNAz considers both thermodynamic stability and evolu-
Prediction tionary conservation, the structure conservation analysis (SCA)
server exclusively evaluates multiple sequence alignments with
respect to conserved structural elements [14]. This is done by cal-
culating various measures like the SCI, base pair distance, or tree-
editing distances and comparing the values obtained from the
alignment of choice against a set of randomized alignments gener-
ated by shuffling alignment columns while preserving gap patterns
[12]. The output page features a graphical representation of the
calculated scores together with z-scores and empirical p-values for
statistical assessment. Note that for similarity measures like the
SCI, a positive z-score indicates structural conservation, while for
distance measures like the base pair distance, a negative z-score is
indicative of structural conservation.
The Bcheck web server is a specialized tool for finding RNaseP
genes [17]. The input simply consists of a FASTA sequence one
wants to screen for a potential RNaseP gene, while the output page
contains the predicted subsequence together with secondary struc-
ture annotation derived from fitting the RNaseP covariance model.
3.8 Other Services So far, we have described the most commonly used programs of
Hosted at the ViennaRNA package and tools for ncRNA detection. The
the ViennaRNA Web ViennaRNA web services also offer access to a series of other spe-
Services cialized tools.
RNApredator is a tool for prediction of target interactions

of bacterial sRNAs with mRNA molecules on a genome-wide
scale [10]. The algorithm makes use of RNAPlex, a fast approach
for computing RNA-RNA interactions considering accessibility
of the target site, which allows genome-wide screens to be per-
formed within a few minutes. The RNApredator output page
outlines the predicted interactions in detail including enrich-
ment evaluation in gene ontology terms.
RNAxs is an algorithm that aims at the optimal design of
siRNAs targeting selected mRNA molecules [11]. Among other
design criteria like seed matching, RNAxs also considers target site
accessibility. For a detailed protocol on how to use the RNAxs web
server, we refer the interested reader to ref. 23.
A very recent addition to the ViennaRNA web services is the
CMCompare web server [24]. It allows comparison and quality
assessment of Infernal RNA family models (http://rfam.sanger.
ac.uk/). In addition to these programs, the database AREsite is
also hosted on the ViennaRNA web services [25]. AREsite is an
online resource for the investigation of AU-rich elements (AREs)
in vertebrate mRNA UTR sequences. In addition to localization of
AREs, AREsite also reports local accessibility values and evolution-
ary conservation profiles.
3.9 For the Advanced Interfaces to the core programs of the ViennaRNA package all
User: Additional have an advanced options panel. Clicking on it will expand the
Parameter Options field, showing options to control energy parameters, dangling
ends, and folding temperature. Default RNA energy parameters
are those of the Turner 2004 model [26]. Usage of the Turner
1999 model is only recommended when reevaluating the folding
of RNA sequences from publications that used the Turner 1999
energy model. Folding of DNA sequences is also possible, which
can be done by selecting “DNA parameters” (Matthews model,
2004). Folding of DNA/RNA hybrids is currently not imple-
mented. Turner 2004 parameters were measured at 37 °C, which
is also set as default temperature. Other temperatures can be
entered and energy parameters will be extrapolated accordingly,
but please note that for temperatures far from 37 °C, results will
increasingly become unreliable. Dangling ends are nucleotides that
stack onto the ends of helices. Programs of the ViennaRNA pack-
age have several options for including dangling ends in the compu-
tations. The default setting is that dangling end energies are
considered for both sides of a helix in any case. Other options are
to turn off dangling end energy contributions or to extend the
dangling end model to also include coaxial stacking of adjacent
helices in a multi-loop. For some programs, it is also possible to
specify that the RNA sequence should be treated as a circular RNA
molecule. As circular RNA molecules cannot adopt all conforma-
tions of their linear counterparts due to tension in the exterior
loop, a switch to correct the prediction for this case is available.
3.10 Dot-Bracket Programs of the ViennaRNA web services generate or require as

Notation and Other input RNA secondary structures in dot-bracket notation. The dot-
Formats for Storing bracket notation, or sometimes referred to as Vienna format, is a
RNA Secondary simple string notation for RNA secondary structures. An opening
Structures bracket corresponds to the 5′ base of a base pair, while a closing
bracket denotes the 3′ base in the base pair. An unpaired base is
indicated by a dot. This notation is intuitive since it follows math-
ematical rules for parenthesizing. An example of the dot-bracket
notation for the secondary structure of a tRNA showing the typical
cloverleaf fold is given below:
GGGCUAUUAGCUCAGUUGGUUAGAGCGCACCCCUGAUAAGGGUGAGGUCGCUGAUUCGAAUUCAGCAUAGCCCA
(((((((..((((.........)))).(((((.......))))).....(((((.......)))))))))))).
The dot-bracket notation is also widely supported by programs

other than the ViennaRNA package; nevertheless, results pages of
the ViennaRNA web services also provide output in ConnecT
(CT) format for interchange with other programs.
3.11 Insights into Underlying the secondary structure prediction algorithms of the
Energy Contributions ViennaRNA package is the so-called loop-based energy model or
with RNAeval Web nearest-neighbor model. The principle of this model is that any
Server RNA secondary structure can uniquely be decomposed into a set
of loops. In this model, stacked base pairs that build the helices of
RNA secondary structures are also represented as loops, as a special
case of interior loops with no unpaired nucleotide between two
base pairs. For users interested in the energy contributions of
individual structural elements to the total minimum free energy,
we recommend the RNAeval server. It takes as input a
sequence-structure pair and lists the individual energy contribu-
tions according to the loop-based energy model.
3.12 Obtaining Images displayed on the results pages are either in a bitmap format
High-Quality Images such as PNG or in a vector format such as SVG. We do not recom-
for Publication mend to use images in PNG format for figures in publications, but
advise the user to instead download the corresponding file in PDF
or EPS format, which preserves the quality of the image when
scaled. Images in these file formats can easily be postprocessed and
edited with free image tools such as Inkscape. If this is not a viable
option for the user, results pages also provide links named “IMAGE
CONVERTER” to obtain PNG or GIF files at higher, user-speci-
fied resolutions than those displayed on the results pages.
3.13 The ViennaRNA A stand-alone version of the ViennaRNA package can be obtained
Package: Stand-Alone from http://www.tbi.univie.ac.at/RNA. The ViennaRNA pack-
Version age has been developed and tested under the operating system
Linux and therefore is best operated under Linux or a Unix-like
environment. For most use cases, off-the-shelf hardware is sufficient.

For details on installation and execution of programs, we refer the
interested reader to another article in the Methods in Molecular
Biology series [27]. Note that at the end of each results page there
is an equivalent command line call outlined to conduct the same
analysis with a local installation of the ViennaRNA package.
3.14 Related Web The Nucleic Acids Research web server index (http://www.oxford-
Services journals.org/nar/webserver/subcat/10/90) lists many tools for
the purpose of RNA secondary structure prediction and related
tasks and offers a good starting point for exploring other tools.
Among the service listed, we would like to specifically high-
light the Freiburg RNA tools (http://rna.informatik.uni-freiburg.
de), which is a comprehensive set of programs that provides among
other services also excellent tools for the prediction of RNA-RNA
interactions [28].
Acknowledgments
This work was supported in part by the University of Basel, the

Austrian FWF project “SFB F43 RNA regulation of the transcrip-
tome,” and the University of Leipzig.
References
1. Hofacker IL, Fontana W, Stadler P, Bonhoeffer 8. Bernhart SH, Tafer H, Mückstein U, Flamm
S, Tacker M, Schuster P (1994) Fast folding C, Stadler PF, Hofacker IL (2006) Partition
and comparison of RNA secondary structures. function and base pairing probabilities of RNA
Monatsh Chem 125:167–188 heterodimers. Algorithm Mol Biol 1:3
2. Lorenz R, Bernhart SH, Höner Zu Siederdissen 9. Mückstein U, Tafer H, Hackermüller J,
C, Tafer H, Flamm C, Stadler PF, Hofacker IL Bernhart SH, Stadler PF, Hofacker IL (2006)
(2011) ViennaRNA Package 2.0. Algorithm Thermodynamics of RNA-RNA binding.
Mol Biol 6:26 Bioinformatics 22:1177–1182
3. Hofacker IL (2003) Vienna RNA secondary 10. Eggenhofer F, Tafer H, Stadler PF, Hofacker
structure server. Nucleic Acids Res 31: IL (2011) RNApredator: fast accessibility-
3429–3431 based prediction of sRNA targets. Nucleic
4. Gruber AR, Lorenz R, Bernhart SH, Neuböck Acids Res 39:W149–W154
R, Hofacker IL (2008) The Vienna RNA web- 11. Tafer H, Ameres SL, Obernosterer G,
suite. Nucleic Acids Res 36:W70–W74 Gebeshuber CA, Schroeder R, Martinez J,
5. Hofacker IL, Fekete M, Stadler PF (2002) Hofacker IL (2008) The impact of target site
Secondary structure prediction for aligned accessibility on the design of effective siRNAs.
RNA sequences. J Mol Biol 319:1059–1066 Nat Biotechnol 26:578–583
6. Bernhart SH, Hofacker IL, Will S, Gruber AR, 12. Washietl S, Hofacker IL (2004) Consensus
Stadler PF (2008) RNAalifold: improved con- folding of aligned sequences as a new measure
sensus structure prediction for RNA align- for the detection of functional RNAs by com-
ments. BMC Bioinform 9:474 parative genomics. J Mol Biol 342:19–30
7. Flamm C, Hofacker IL, Stadler PF, Wolfinger 13. Washietl S, Hofacker IL, Stadler PF (2005) Fast
MT (2002) Barrier trees of degenerate land- and reliable prediction of noncoding RNAs.
scapes. Z Phys Chem 216:155 Proc Natl Acad Sci U S A 102:2454–2459
14. Gruber AR, Bernhart SH, Hofacker IL, 22. Washietl S (2010) Sequence and structure anal-
Washietl S (2008) Strategies for measuring ysis of noncoding RNAs. Methods Mol Biol
evolutionary conservation of RNA secondary 609:285–306
structures. BMC Bioinform 9:122 23. Hofacker IL, Tafer H (2010) Designing opti-
15. Gruber AR, Findeiß S, Washietl S, Hofacker mal siRNA based on target site accessibility.
IL, Stadler PF (2010) RNAz 2.0: improved Methods Mol Biol 623:137–154
noncoding RNA detection. Pac Symp 24. Eggenhofer F, Hofacker IL, Höner Zu
Biocomput 2010:69 Siederdissen C (2013) CMCompare web-
16. Gruber AR, Neuböck R, Hofacker IL, Washietl server: comparing RNA families via covari-
S (2007) The RNAz web server: prediction of ance models. Nucleic Acids Res 41:
thermodynamically stable and evolutionarily W499–W503
conserved RNA structures. Nucleic Acids Res 25. Gruber AR, Fallmann J, Kratochvill F, Kovarik
35:W335–W338 P, Hofacker IL (2011) AREsite: a database for
17. Yusuf D, Marz M, Stadler PF, Hofacker IL the comprehensive investigation of AU-rich
(2010) Bcheck: a wrapper tool for detecting elements. Nucleic Acids Res 39:D66–D69
RNase P RNA genes. BMC Genomics 11:432 26. Mathews DH, Disney MD, Childs JL,
18. Rivas E (2013) The four ingredients of single- Schroeder SJ, Zuker M, Turner DH (2004)
sequence RNA secondary structure prediction: Incorporating chemical modification con-
a unifying perspective. RNA Biol 10:59–70 straints into a dynamic programming algo-
19. Wuchty S, Walter F, Hofacker IL, Schuster P rithm for prediction of RNA secondary
et al (1999) Complete suboptimal folding of structure. Proc Natl Acad Sci U S A 101:
RNA and the stability of secondary structures. 7287–7292
Biopolymers 49:145–165 27. Hofacker I, Lorenz R (2014) Predicting RNA
20. Wolfinger MT, Svrcek-Seiler AW, Flamm C, structure: advances and limitations. Methods
Hofacker IL, Stadler PF (2004) Efficient com- Mol Biol 1086:1
putation of RNA folding dynamics. J Phys 28. Smith C, Heyne S, Richter AS, Will S,
Math Gen 37:4731 Backofen R (2010) Freiburg RNA tools: a
21. Reiche K, Stadler PF (2007) RNAstrand: read- web server integrating INTARNA, EXPARNA
ing direction of structured RNAs in multiple and LOCARNA. Nucleic Acids Res 38:
sequence alignments. Algorithm Mol Biol 2:6 W373–W377
Chapter 20
Exploring the RNA Editing Potential of RNA-Seq Data

by ExpEdit
Mattia D’Antonio, Ernesto Picardi, Tiziana Castrignanò,
Anna Maria D’Erchia, and Graziano Pesole
Abstract
Revealing the impact of A-to-I RNA editing in RNA-Seq experiments is relevant in humans because RNA
editing can influence gene expression. In addition, its deregulation has been linked to a variety of human
diseases. Exploiting the RNA editing potential in complete RNA-Seq datasets, however, is a challenging
task. Indeed, no dedicated software is available, and sometimes deep computational skills and appropriate
hardware resources are required. To explore the impact of known RNA editing events in massive transcrip-
tome sequencing experiments, we developed the ExpEdit web service application. In the present work, we
provide an overview of ExpEdit as well as methodologies to investigate known RNA editing in human
RNA-Seq datasets.
Key words RNA editing, A-to-I editing, Deep sequencing, Bioinformatics, Genomics, RNA-Seq
1 Introduction
RNA editing is a posttranscriptional mechanism whereby genetic

information is enzymatically overwritten at specific sites to gener-
ate functional transcripts, different from their corresponding
genomic templates [1]. In mammals and, in particular, in humans,
the A-to-I conversion is the most frequent type of RNA editing
catalyzed by the ADAR (adenosine deaminase acting on RNA)
family of enzymes. Inosines are commonly recognized as guano-
sines by cell molecular machineries as well as sequencing enzymes.
RNA editing has a relevant biological role in preserving the
cellular homeostasis even though several aspects are yet unknown.
Thanks to bioinformatics methods, thousands of editing changes
have detected and collected in freely available specialized databases
as DARNED [2] and RADAR [3]. These extraordinary collec-
tions can be employed to provide RNA editing snapshots in whole
transcriptome experiments through the RNA-Seq technology.
Indeed, a large amount of RNA-Seq data is now publicly available
327
328 Mattia D’Antonio et al.
through the SRA repository at NCBI. In this context, we have

developed ExpEdit, a web application for exploring RNA editing
in humans at known or user-specified sites supported by RNA-Seq
experiments [4].
A great benefit of web applications, such as ExpEdit, is the
possibility to analyze RNA-Seq data requiring neither technical
competences nor expensive computational resources (usually a
computing server or a computational cluster). In addition, results
are provided in organized interfaces for easy and intuitive
visualizations.
In this work, we provide an explanation of ExpEdit as well as
methodologies to efficiently explore known human RNA editing
in RNA-Seq datasets.
2 Materials
Results reported in this paper are based on raw sequencing reads

obtained from the study published by [5] and stored into NCBI
SRA under the accession SRP002274. It consists of RNA-Seq data
from the human brain obtained through the Illumina Genome
Analyzer II sequencer. We picked three of available six paired-end
lanes (SRR039628/29, SRR039630/31 and SRR039632/33).
Each lane comprises on average 7.8 million reads with an individ-
ual length of 50 bp.
2.1 Required The management of large datasets generated from deep sequenc-
Software and ing experiments typically requires powerful hardware resources.
Hardware Resources Hundred GBs of free disk space are needed to store inputs, outputs
and intermediate temporary files. Bioinformatics software is often
expensive in terms of computational resources and requires a large
amount of RAM (16–24 GB at least) and several CPUs/cores to
be executed in a reasonable time.
To handle several inputs as different lanes from the same
experiment, analysis workflows need to be parallelized on multiple
servers or serialized to analyze one input a time in full automation.
Web applications overcome these hardware requirements, and the
final user only needs a web browser and a stable Internet
connection.
Input files can be directly uploaded or transferred providing a
web link. In the first case, they have to be stored locally, and the
upload process could be time consuming depending on file size
and connection speed. Alternatively, inputs can also be transferred
providing a web link to the location of data stored on remote serv-
ers. Using web links, the transfer is independent on the user con-
nection speed. A further benefit of web applications is that analysis
results can be downloaded by the user and stored locally to perform
off-line investigations.
RNA Editing Detection by ExpEdit 329
For our aim and thus the investigation of known RNA editing
events in RNA-Seq datasets, the user needs a standard desktop or
laptop computer with a stable Internet connection and an updated
browser like the Internet Explorer, Firefox, or Safari.
ExpEdit service is available at the following web page http://
bioinformatics.cineca.it/expedit, and a free registration is required
to upload data and start computational analyses.
3 Methods
ExpEdit is designed to integrate in-house developed scripts as well

as open source analysis tools into a single pipeline. It automatically
handles and manages all required computational tasks. Every sub-
mission and retrieval operation is performed through an interactive
web-based graphical user interface (GUI) verified on the main
popular browsers. The GUI is written in PHP: Hypertext
Preprocessor (PHP) language using HyperText Markup Language
(HTML), and jQuery, combined with HTML5 and CSS3 stan-
dards, to enable a better user interaction.
ExpEdit takes as input short-read datasets produced by
Illumina sequencing platforms in several standard file formats as
FASTQ [6], SRA [7], or BAM [8]. The user can also upload these
files in compressed archives (several compressed formats are also
managed, such as .zip, .tar, .gzip, or .bz2) or provide ftp/http links
to facilitate and speed up the upload process. In case of ftp/http
links, the Wget application is used to retrieve input files.
The ExpEdit analysis pipeline can be summarized in three
main steps: (1) quality assessment, (2) genome alignment, and (3)
RNA editing calling. The first two steps are skipped if BAM inputs
are provided.
Initially, ExpEdit performs an overall quality check (quality
control or QC) of input data because read quality is critical for reli-
able analyses and may sensibly affect downstream results. For each
FASTQ file, ExpEdit runs the FastQC program generating HTML
report pages with detailed QC results organized in tables and
graphs. Such reports include relevant information about raw reads
such as quality score distributions, nucleotide distributions, or
overrepresented sequences.
The ExpEdit pipeline implements also an additional applica-
tion for quality control: the NGS QC Toolkit. In contrast to
FastQC, it can remove contaminants as primers or adaptors and
trim read regions with low quality (generally with a Phred score
lower than 20) [9].
After the quality assessment, ExpEdit proceeds with the map-
ping onto the reference human genome. This is a critical step
because erroneous alignments can seriously affect RNA editing
detection and quantification. Genome mapping is performed by
means of TopHat [10, 11] program based on Bowtie [12]. TopHat

handles both single and paired-end reads and can perform spliced
alignments as required for human RNA-Seq experiments.
Alternatively, pre-aligned reads obtained by other software can
be uploaded in ExpEdit as SAM/BAM format. Users are also com-
pletely free to download TopHat alignments, refine them off-line,
and reload modified SAM/BAM files.
At the end of genome mapping, ExpEdit starts the RNA editing
analysis launching in-house scripts, developed in Python program-
ming language that makes use of the SAMtools package. In particu-
lar, unique read alignments are converted from SAM/BAM format
to the pileup format, and genomic positions are explored site by
site. During the parsing of pileup file, genomic positions are com-
pared with those stored in specialized databases as DARNED [2].
Known positions subject to RNA editing are then printed out as
well as the nucleotide distribution of supporting reads and the
RNA editing level. A specific filter to exclude bases with a prefixed
Phred-like quality score (by default fixed to 20) is also imple-
mented. In addition, read nucleotides supporting a specific
genomic position can be filtered according to the strand to take
into account the effect of antisense transcription (recommended
for strand-oriented reads). Individual positions are also annotated
including gene name and RNA regions (CDS, intron, or UTR) or
info about single nucleotide polymorphisms (SNPs).
Finally, ExpEdit automatically produces output tables that can
be dynamically sorted, filtered, and exported as Excel files for off-
line downstream analyses.
ExpEdit can also explore RNA editing at custom genomic
positions by uploading a text file containing a list of candidates.
3.1 Hands On: Using Every ExpEdit action can be executed through an interactive GUI
the ExpEdit GUI available at http://bioinformatics.cineca.it/expedit/, and what-
ever popular web browser can be used to perform a complete anal-
ysis. Hereafter, we provide guidelines to create an ExpEdit study/
project, upload data, run an entire analysis, and inspect results.
3.1.1 Creating a New The first step to submit a dataset to ExpEdit is the creation of a
Study/Project new project/study that will contain one or more input files as well
as analysis results. To register a new study, a few information are
required such as a title, a description and an access level (private,
group, or public) (Fig. 1).
3.1.2 Uploading The ExpEdit upload engine offers several options: web upload,
Input Files web link, and Dropbox. The FTP protocol is planned but not yet
implemented. Web Upload supports up to 12 GB file size on 64-bit
operating systems and up to 2 GB on 32-bit operating systems. It
is implemented in JavaScript and allows the user to upload several
Fig. 1 Creation of a new ExpEdit project. Through the web interface, the user can create new projects by
providing minimal information
files at once, handling a fully automated upload queue. The user

can follow the upload progress and interact with the system adding
or removing files also during the transfer. Alternatively, the user can
choose to upload data providing a web link (HTTP, HTTPs, and
FTP protocols are allowed). In this case, one or more links have to
be provided, and the system will manage the download using an
internal queue. An e-mail will notify the completion of downloads.
A third option consists of the use of the Dropbox plug-in (Fig. 2).
The ExpEdit upload facility supports several input formats
such as text-based raw sequences (i.e., .fastq or .fq), pre-aligned
data (i.e., .bam or .sam), compressed reads (i.e., .sra), and com-
pressed archives (i.e., .zip, .gz, .tar, or .bz2).
For paired-end RNA-Seq reads, each pair needs to be specified
using a drag-and-drop tool provided by the GUI. The same tool
can be used to reorder lanes (if needed). Finally, a unique label has
to be assigned to each file or pair of files. Such label will be used
instead of filename in the results pages.
Once input files have been imported into the study, users will
be allowed to request for a new analysis.
3.1.3 Creating a New To create a new analysis, one or more files have to be selected. Then,
Analysis a name and a description can be provided. Before launching ExpEdit,
several parameters can be tuned to customize the analysis, even
though default parameters are suitable for many experiments.
ExpEdit includes four categories of parameters:
1. Common to many pipeline modules such as the human refer-
ence database (hg18 and hg19 assemblies are allowed).
2. For quality check and filtering.
Fig. 2 Uploading files in ExpEdit. Input files can be added to an existing project by the web interface or by
providing a web link or by importing data from a Dropbox account
3. For reads mapping through main TopHat options and allowing

the alignment on both genome and transcriptome by providing
a set of gene model annotations and/or known transcripts in
GTF2 or GFF3 format. Different sequencing libraries are
admitted in order to handle strand-oriented reads. Standard
parameters such as the max number of allowed mismatches or
multihits, the gap max length, and the min and max intron
length can also be specified.
4. To detect RNA editing sites allowing the selection of known
databases or the upload of a custom list of RNA editing
candidates (the list must contain a site per row and three tab-
separated fields including genomic region, coordinate and
strand). A Phred-like quality score for filtering out supporting
nucleotides below a predefined value (by default equal to 20)
can be provided. Read nucleotides can also be filtered according
to the strand to take into account the effect of antisense
transcription (recommended for strand-oriented reads).
Figure 3 shows a screenshot of the ExpEdit web form to create
an analysis and tune parameters.
Fig. 3 Creation of a new analysis. The analysis workflow can be configured tuning several parameters
3.1.4 Monitoring Once all parameters have been customized, the analysis can be
an Analysis submitted to the queue system. The execution process is fully
automated with a complex dispatching architecture integrated
with a TORQUE Resource Manager. The user can follow the anal-
ysis progress through a monitor page, which displays the list of
pipeline steps and the intermediate output results (these files can
be visualized and/or downloaded, also during the analysis execu-
tion). For each step, the running status and the used algorithm and
its version are shown (Fig. 4). In case of execution fault, the run-
ning status is marked as error. At the completion of each step of the
pipeline, all generated output files are listed and displayed in the
monitoring page. In this way, the user can visualize or download all
intermediate files. The command line and the estimated running
time per task are also provided.
3.1.5 Analysis Results At the end of the pipeline, all results are parsed and stored into
a dedicated and optimized database. In addition, an e-mail
containing a link to the available results will be automatically sent
to the user.
Fig. 4 Monitoring an analysis. During the execution, each computational task can be inspected to monitor the
job status and intermediate output files
Results are organized in two sections (Fig. 5). The “quality

checks” section refers to results from the quality control analysis
obtained by FastQC and NGS QC Toolkit. The “editing sites” sec-
tion reports a summary of all detected RNA editing sites and gives
an overview of results. By clicking on these summaries, a detailed
list of sites is displayed (see Subheading 3.1.6) (Fig. 5). Otherwise,
the user can create one or more operations of intersection and
union (see Subheading 3.1.7).
3.1.6 Result Details Final ExpEdit results are provided in a detailed table that can be
filtered, sorted and downloaded (Fig. 6). Every site in the table is
identified by a genomic position (chromosome, coordinate, and
strand) annotated with relevant biological information. Given a
reference base, ExpEdit reports the total number of reads support-
ing the specific site and the number of reads supporting each
Fig. 5 Visualization of summary results. After each run, ExpEdit shows simple summary tables (quality checks
on the top and detected editing sites on the bottom)
Fig. 6 Visualization of detailed results. RNA editing site is shown in dynamic tables
nucleotide. The editing percentage is defined as the ratio between

the number of Gs over the number of As + Gs. An editing percent-
age equal to 100 means that all reads support the editing event.
A percentage equal to 0 means that RNA editing is unlikely.
Results can be sorted in any order (ascending and descending)
and filtered with custom thresholds to facilitate the identification
of functionally significant variants. Filters can be combined to
produce complex queries, and resulting tables can be exported as
textual/Excel files for off-line downstream analyses.
3.1.7 Intersections When two or more analyzed samples are available, union and inter-
and Unions section operations can be performed. A union operation may be
helpful to merge together results obtained from biological repli-
cates, while intersections may be useful to filter out aspecific results.
Unions and intersections can be combined together to build com-
plex operations, involving multiple lanes.
3.2 Case of Study In this section, we provide an example of ExpEdit analysis using
raw sequencing data obtained from the study published by [5] and
stored into NCBI SRA database under the accession SRP002274.
Input reads can be uploaded in ExpEdit using the web link
facility and the following URLs:
http://ftp-trace.ncbi.nlm.nih.gov/sra/sra-instant/reads/
ByRun/sra/SRR/SRR039/SRR039628/SRR039628.sra
SRA archives will be automatically extracted by the ExpEdit
preprocessing facility and converted into regular FASTQ files.
Now a new analysis can be created and executed by selecting
the human genome assembly hg19/build37 and tuning parame-
ters in the following way:
– Quality check and filtering: default values for cutoffs, length
(70) and quality (20). No primer/adaptor library has to be
selected.
– Genome alignment: the paired-end library has to be config-
ured as unstranded with a mate inner distance of 200 bp and a
mate standard deviation of 20. Default values should be used
for other parameters, i.e., max-multihits (1), segment-

mismatches (2), max-intron-length (500,000), min-intron-
length (70), min-anchor (8), read-edit-dist (2), read-gap-length
(2), segment-length (25), and splice-mismatches (0).
– Detection of RNA editing sites: Phred quality filter, 20, no
strand-based filter. No custom list of editing candidates has to
be provided.
The first analysis step will consist in the run of FastQC version
0.10.1 that will be executed in about 13 min. All inputs should
show the same trend with qualities decreasing at 3′ end and Phred
score dropping below the threshold of 20 in the last 4–5 positions.
Sequence duplication level should be between 34 and 35 %, with-
out specific overrepresented sequence.
In the second step, NGS QC Toolkit version 2.3 will be exe-
cuted in about 35 min and the output should report statistics simi-
lar to FastQC.
The mapping of high-quality reads to the human genome by
TopHat version 2.0.9 will be executed in about 3 h. Each lane
should show an alignment rate above 90 %, and TopHat should
report a rate of concordant alignment of 81.6, 81.7 and 81.1 % for
each sample.
Finally, the ExpEdit computational core for RNA editing call-
ing will be executed in about 4 h. Results obtained from this step
will be parsed and stored into a MySQL database.
The complete pipeline should take less than 8 h. Input files (six
files in three lanes for a total file size of 11 GB) will produce at the
end 130 GB of intermediate output files.
Final tables should reveal on average 800 RNA editing sites
with coverage higher than or equal to 10 and about 320 positions
with an RNA editing percentage higher than 0.
Since all lanes are biological replicates of the same brain sam-
ple, the union operation can be executed to merge together these
results.
All tables can be downloaded for off-line inspection by Excel
or browsed to look at specific RNA editing events.
Acknowledgments

Università e Ricerca (MIUR), PRIN 2009 and 2010, and Consiglio
Nazionale delle Ricerche, Flagship Project Epigen, Medicina
Personalizzata, and Aging Program 2012–2014.
References
1. Gott JM, Emeson RB (2000) Functions and 7. Leinonen R, Sugawara H, Shumway M,
mechanisms of RNA editing. Annu Rev Genet Collaboration., International Nucleotide
34:499–531 Sequence Database (2011) The sequence read
2. Kiran A, Baranov PV (2010) DARNED: a archive. Nucleic Acids Res 39(Database issue):
DAtabase of RNa EDiting in humans. Bioin- D19–D21
formatics 26(14):1772–1776. doi:10.1093/ 8. Li H, Handsaker B, Wysoker A, Fennell T, Ruan
bioinformatics/btq285, Epub 2010 Jun 14 J, Homer N, Marth G, Abecasis G, Durbin R,
3. Ramaswami G, Li JB (2014) RADAR: a rigor- 1000 Genome Project Data Processing Subgroup
ously annotated database of A-to-I RNA edit- (2011) The sequence alignment/map format
ing. Nucleic Acids Res 42(D1):D109–D113. and SAMtools. Bioinformatics 25(16):2078–
doi:10.1093/nar/gkt996 2079. doi:10.1093/bioinformatics/btp352
4. Picardi E, D’Antonio M, Carrabino D, 9. Patel RK, Jain M (2012) NGS QC toolkit: a
Castrignanò T, Pesole G (2011) ExpEdit: a toolkit for quality control of next generation
webserver to explore human RNA editing in sequencing data. PLoS One 7(2):e30619.
RNA-Seq experiments. Bioinformatics doi:10.1371/journal.pone.0030619
27(9):1311–1312. doi:10.1093/bioinformat- 10. Trapnell C, Pachter L, Salzberg SL (2009)
ics/btr117 TopHat: discovering splice junctions with
5. Au KF, Jiang H, Lin L, Xing Y, Wong WH RNA-Seq. Bioinformatics 25(9):1105–1111.
(2010) Detection of splice junctions from doi:10.1093/bioinformatics/btp120
paired-end RNA-seq data by SpliceMap. 11. Kim D, Pertea G, Trapnell C, Pimentel H, Kelley
Nucleic Acids Res 38(14):4570–8. R, Salzberg SL (2011) TopHat2: accurate align-
doi:10.1093/nar/gkq211 ment of transcriptomes in the presence of inser-
6. Cock PJ, Fields CJ, Goto N, Heuer ML, Rice tions, deletions and gene fusions. Genome Biol
PM (2010) The Sanger FASTQ file format for 14(4):R36. doi:10.1186/gb-2013-14-4-r36
sequences with quality scores, and the Solexa/ 12. Langmead B, Salzberg S (2012) Fast gapped-
Illumina FASTQ variants. Nucleic Acids Res read alignment with Bowtie 2. Nat Methods
38(6):1767–71. doi:10.1093/nar/gkp1137 9(4):357–359. doi:10.1038/nmeth.1923
Chapter 21
A Guideline for the Annotation of UTR Regulatory Elements

in the UTRsite Collection
Matteo Giulietti, Giorgio Grillo, Sabino Liuni, and Graziano Pesole
Abstract
Gene expression regulatory elements are scattered in gene promoters and pre-mRNAs. In particular, RNA
elements lying in untranslated regions (5′ and 3′UTRs) are poorly studied because of their peculiar fea-
tures (i.e., a combination of primary and secondary structure elements) which also pose remarkable com-
putational challenges. Several years ago, we began collecting experimentally characterized UTR regulatory
elements, developing the specialized database UTRsite. This paper describes the detailed guidelines to
annotate cis-regulatory elements in 5′ and 3′ UnTranslated Regions (UTRs) by computational analyses,
retracing all main steps used by UTRsite curators.
Key words UTR, cis-Regulatory elements, RNA secondary structure, Sequence alignments, Perl,
Motif pattern
1 Introduction
In the post-genomic era, the identification of gene regulatory ele-

ments and the understanding of their biological role in modulating
gene expression represent a big challenge. Gene regulatory ele-
ments are generally located in noncoding regions of eukaryotic
genomes and act at transcriptional and/or posttranscriptional
level. Many DNA elements in promoter regions are known to con-
trol gene transcription by interacting with transcription factors and
several databases and software to classify and identify these ele-
ments have been developed [1]. At posttranscriptional level,
instead, the control of gene expression rely on RNA elements that
are scattered in whole pre-mRNA sequences and are involved in
many processes such as splicing, RNA folding, capping, polyade-
nylation, nucleocytoplasmic transport, translation efficiency,
subcellular localization, stability and determine the fate of mature
mRNA in the cell [2]. Regarding the splicing, we recently devel-
oped SpliceAid [3, 4] and SpliceAid-F [5] databases that collect
data about human splicing factors and their experimentally assessed
339
340 Matteo Giulietti et al.
pre-mRNA binding sites. For regulatory elements lying in 5′ and

3′ UnTranslated Regions (5′UTRs, 3′UTRs), we have developed
two specialized database, UTRdb and UTRsite, collecting eukary-
otic UTR sequences and experimentally characterized UTR-
specific regulatory signals, respectively [6]. In spite of transcription
and splicing factor binding sites, RNA elements in UTRs are much
less investigated. Indeed, their biological activities often require a
combination of primary and secondary structure elements that
are hard to predict correctly. Computational methods and available
tools for predicting RNA secondary structure are reviewed in [7, 8].
This paper describes detailed computational approaches and guide-
lines to characterize and annotate UTR regulatory cis-elements
following same steps adopted by UTRsite curators. In addition,
we show the annotation process through an example concerning
the characterization of the DNA polymerase beta stem loop II ele-
ment [9].
2 Materials
2.1 Hardware The annotation workflow described in this paper can be executed
and Software on a regular computer without specific hardware resources.
However, it must have a stable Internet connection to download
executables or to access dedicated Web services. Although what-
ever operating system as Linux, Mac OS X, or Windows is suitable,
the Perl interpreter needs to be preinstalled as well as the BioPerl
package. An updated Perl interpreter for Windows and Mac OS X
can be downloaded from www.activestate.com/activeperl. Linux
users can install a copy via apt or yum tools. BioPerl package can be
downloaded from www.bioperl.org.
2.2 Web Servers PubMed at www.ncbi.nlm.nih.gov/pubmed will be used to search

and Databases scientific literature. BLAST tool [10] at http://blast.ncbi.nlm.nih.
gov/Blast.cgi will be employed to retrieve all orthologous
sequences of UTR elements under investigation. DIALIGN [11]
at http://bibiserv.techfak.uni-bielefeld.de/dialign/ is a multiple
sequence alignment (MSA) tool that differently from other MSA
tools, is useful to detect local homologies in sequences with low
overall similarity. LocARNA [12] at http://rna.informatik.uni-
freiburg.de:8080/LocARNA/Input.jsp is a program to predict a
consensus secondary structure from a set of unaligned RNAs by
computing multiple alignments (see Note 1). PatSearch tool [13,
14] at http://itbtools.ba.itb.cnr.it/patsearch is a program able to
search for specific patterns in nucleotide sequences. It is useful to
detect sequence regions involved in secondary structure or to build
position-weight matrices. UTRsite [6] is a specialized database of
functional elements in 5′ or 3′ UTRs, available at http://utrsite.
ba.itb.cnr.it.
Annotation of UTR Regulatory Elements 341
3 Methods
3.1 Literature Search The UTRsite entries describe the known regulatory elements
located in 5′UTR or 3′UTR or both, whose functional role and
structure have been experimentally assessed and reported in litera-
ture. Therefore, PubMed database is the starting point for UTRsite
data collection and to identify new regulatory elements. The litera-
ture search is a crucial and time-consuming process.
The literature information related to UTR regulatory elements
can be obtained using a variety of PubMed queries providing key-
words like “UTR” or “untranslated regions” or “5′UTR” or
“3′UTR” and different combinations among them (by means of
Boolean operators AND, OR, and BUT NOT). Limits can also be
activated to reduce the number of records and false positives,
increasing the search specificity. Papers describing the same func-
tional element are examined in order to extract all available infor-
mation about primary sequence, secondary structure, RNA
localization (5′ or 3′UTR), biological role, RNA-binding proteins
and experimental procedures. In general, the functional character-
ization of UTR regulatory elements includes gene expression
assays using recombinant or modified UTRs, site-specific muta-
genesis, RNase protection, and chemical probing experiments.
Here, we present and explanatory example focusing on the
DNA polymerase beta stem loop II regulatory element (POLB)
which has been identified in a PubMed search using aforemen-
tioned terms [9, 15–17]. Related literature describes this element
as a 3′UTR cis-acting element involved in the posttranscriptional
regulation of the DNA Polymerase beta gene. It is 42-nt long and
has a stem-loop secondary structure that is fundamental for the
interaction between the gene transcript and the Hax-1 protein
which, in turn, is responsible for at least two mechanisms: localiza-
tion in the perinuclear space and stabilization of the otherwise
unstable mRNA.
3.2 Detection Once the UTR regulatory element has been identified from litera-
and Extraction ture and the relative sequence has been retrieved, the NCBI
of Orthologous BLAST tool should be used to search for all available orthologous
Sequences sequences. We suggest to perform BLAST searches using as query
the whole mRNA sequence including the element (or its RefSeq
identifier) or the UTR sequence alone or simply the extracted
motif sequence. Multiple queries are essential because the length
of the motif alone is generally short and, thus, BLAST may provide
many not significant and undesirable alignments. All detected
database hits must be carefully checked in order to define a reliable
set of orthologous sequences, all containing the regulatory ele-
ments under investigation, and then investigate its conservation
profile at level of primary and secondary structure that will be used
to obtain an optimal definition of the regulatory pattern.
In the case of DNA polymerase beta stem loop II element, we

used as query the Rattus norvegicus RefSeq id NM_017141, cor-
responding to the sequence used in [9] to experimentally assess the
activity of the UTR element. In the NCBI BLAST tool, we selected
the blastn algorithm, and refseq_rna as database. Other parameters
were unchanged. After the search, all significant hits can be selected
and downloaded in a FASTA-formatted text file (see Note 2).
Given the list of Accession IDs obtained by the BLAST search,
UTRs can be extracted from RefSeq mRNAs using the in house
Perl script UTR_extractor.pl according to the following code:
#!/usr/bin/perl
use Bio::DB::GenBank;
if (! $ARGV[0]) {print "usage: UTR_extractor.pl [input
list file] [5UTR or 3UTR] \n";}
open (FILEIN, "<", "./$ARGV[0]");
open (FILEOUT, ">", "./output.txt");
@IDs = <FILEIN>;
foreach $id (@IDs) {
$gb = new Bio::DB::GenBank;
if (($id=~/_/) and ($id=~/\./)) {$seq = $gb->get_
Seq_by_version($id);} #RefSeq version (eg.
XM_003257165.2)
if (($id=~/_/) and ($id!~/\./)) {$seq = $gb->get_
Seq_by_acc($id);} #RefSeq (eg. XM_003257165)
if (($id!~/_/) and ($id!~/\./)) {$seq = $gb->get_
Seq_by_gi($id);} # GI (eg. 441611727)
for my $feat ($seq->get_SeqFeatures) {
if ($feat->primary_tag eq "gene") {
$gene_seq = $feat->seq->seq();
$gene_start = $feat->location->start;
$gene_end = $feat->location->end;}
if ($feat->primary_tag eq "CDS") {
$CDS_start = $feat->location->start;
$CDS_end = $feat->location->end;}
}
if ($ARGV[1] eq "5UTR") {
$UTR5=substr ($gene_seq, $gene_start-1, $CDS_start-
$gene_start);
print “>gi|".$seq->primary_id()."|".$seq->display_
id()."|5'UTR of ".$seq->desc()."\n$UTR5\n";
print FILEOUT ">gi|".$seq->primary_id()."|".$seq-
>display_id()."|5'UTR of ".$seq->desc()."\
n$UTR5\n";}
if ($ARGV[1] eq "3UTR") {
$UTR3=substr ($gene_seq, $CDS_end, $gene_end-
$CDS_end);
print ">gi|".$seq->primary_id()."|".$seq->display_
id()."|3'UTR of ".$seq->desc()."\n$UTR3\n";
print FILEOUT ">gi|".$seq->primary_id()."|".$seq-
>display_id()."|3'UTR of ".$seq->desc()."\
n$UTR3\n";}
}
The above script can be executed by the command:

$ perl UTR_extractor.pl [input list file] [5UTR or 3UTR]
where [input list file] is the file containing a list of identifier
(RefSeq accessions or GIs) and [5UTR or 3UTR] is a flag to spec-
ify which UTR end to extract. The script is able to retrieve
sequences from GenBank using BioPerl and provide UTRs in a
FASTA-formatted text file.
3.3 Alignment In the next step, retrieved UTR sequences needs to be multi-aligned.
of Orthologous For this task we suggest the tool DIALIGN paying attention to the
Sequences content of the input file in order to exclude mRNAs that do not have
an annotated UTR (for example mRNAs from predicted RefSeq).
In addition, it could be useful to include in the input the sequence
of the element obtained by literature to facilitate its localization in
orthologous sequences and improve the final multi-alignment.
In case of very divergent sequences, the alignment software
could produce biased results. So to limit false alignments, we suggest
working in a hierarchical way aligning first sequences belonging to
close related species and, then, adding further sequences one a time.
Although alignment programs are very accurate, manual refinement
is almost always needed to obtain a reliable multiple alignment.
In our example on DNA polymerase beta stem loop II regula-
tory element we used DIALIGN with default parameters. Only 30
sequences out of about 100 orthologous mRNAs derived from
BLAST searches could be aligned with the rat element. This is
mainly due to the presence of mRNAs without annotated full-
length 3′UTRs. Some sequences, however, have to be discarded
for insufficient sequence similarity to the rat element.
The multiple alignment is also important to deduce the taxo-
nomic range as showed in Fig. 1. In this case, all 30 sequences
belong to the class Mammalia.
3.4 Inferring Multiple alignments generated by DIALIGN can be used to pre-

the Secondary dict potential secondary structures and verify the correspondence
Structure between a consensus secondary structure and the one proposed by
the literature. To this task, the LocARNA software should be
employed. It takes as input a multiple alignment and tries to infer
a consensus secondary structure.
Following our example on DNA polymerase beta stem loop II
element, the inferred consensus secondary structure differs from
the one predicted in literature. Indeed, the latter consists in 13 bp
stem structure (Fig. 2a), while the former includes only a 9 bp
stem and an asymmetric internal loop (Fig. 2b).
3.5 Generation Once a reliable multiple alignment is obtained, taking also into
of the Motif Pattern account the secondary structure of the UTR element, a motif pat-
and Assessment of its tern can be modelled using the PatSearch software (see Note 3)
Sensitivity/Specificity and its specific syntax rules. First, the multiple alignment has to be
Fig. 1 Multiple alignment generated by DIALIGN and manually refined. The first column reports the GenBank GI
identifier and the organism. The character “|” in the alignment is used to distinguishing stems from loops.
Mismatched nucleotides in stems are underlined. Colored nucleotides are used to facilitate the alignment
visualization
checked in order to mark conserved and variable positions. In case

of variable positions, the IUPAC nomenclature should be used
(e.g.: Y means C or T; B means C, G, or T, and so on).
Each pattern can be composed by individual pattern units
identified by p1, p2, pn. Individual units can be strings like “p1 =
ACGYYAB” or complex pattern unit like “p2 = ACGYYAB[1,2,3]”
allowing mismatch, deletions and insertions. A pattern unit can
also include a range descriptor like “1…5”. For example, the pat-
tern “AACU 1…5 GGAU” will recognize all sequences starting
with AACU, followed by 1–5 uncharacterized nucleotides and
ending with GGAU. The range descriptor is quite useful to describe
alignments gaps or not conserved nucleotides. In addition,
PatSearch program includes important rules to take into account
RNA secondary structures through pairing rules as “r1 = {AU,
UA,CG,GC,GU,UG}” allowing also GU base pairing. The pattern
“p1 = ACCUAAGCC 3…10 r1 ~ p1”, for instance, facilitate the
identification of sequences composed by the string ACCUAAGCC
Fig. 2 The secondary structure of the rat DNA polymerase beta stem loop II element
(a) and the inferred consensus structure from the multiple alignment (b)
followed by 3–10 bases acting as a loop and GG(C/U)UU(A/G)

GGU as a second side of the stem (see Note 4). Complex PatSearch
rules can also be defined. For a complete description of PatSearch
syntax, please refer to [13, 14].
In our example on the DNA polymerase beta stem loop II
regulatory element, the following pattern can be defined:
r1={AU,UA,CG,GC,UG,GU}
UUAUUK[0,1,0] p1=CY UA[0,0,1] p2=ASCUUUGC UVY
r1~p2[1,0,0] CUUU[1,2,0] r1~p1[1,0,0] UGKUYU
where
UUAUUK[0,1,0]: is the first unpaired tract, where a deletion of
1 nt is allowed;
p1 = CY: is the first side of the first stem tract;
UA[0,0,1] : is the first side of the asymmetric internal loop, where
an insertion of 1 nt is allowed;
p2 = ASCUUUGC: is the first side of the second stem tract;
UVY: is the external loop;
r1 ~ p2[1,0,0]: is the second side of the second stem tract, where
a mismatch is allowed;
CUUU[1,2,0]: the second side of the asymmetric internal loop,
where a mismatch and a deletion of up to 2 nt are allowed;
r1 ~ p1[1,0,0]: is the second side of the first stem tract, where a
mismatch is allowed;
UGKUYU: is the second unpaired tract.
The correctness and statistical significance of PatSearch

patterns can be assessed using shuffled datasets maintaining the
same nucleotide composition of the original multiple alignment
and applying a simple chi-square test. If the pattern is statistically
significant, it can be used to search for similar patterns in other
sequences and results could be good candidates for further func-
tional investigations.
Using the above pattern for the DNA polymerase beta stem
loop II regulatory element against all mammalian UTR sequences
stored in UTRdb dataset (comprising over 260,000 sequences),
PatSearch detects 24 UTRs with p < 0.0001.
3.6 Submit Significant patterns and new UTR elements can be submitted to
the Annotated UTR the UTRsite Web service (see Note 5). Each entry in UTRsite has
Element to UTRsite several sections including all relevant info characterizing the ele-
ment as those detected using the above methodology.
The UTRsite entry corresponding to the DNA polymerase
beta stem loop II regulatory element is shown in Fig. 3 and can be
directly inspected at the following Web page http://utrsite.ba.itb.
cnr.it/index.php/UTRSite%20signal/Signal/action/view/
frmUTRSiteID/U0049.
4 Notes
1. Potential alternatives to LocARNA are: (a) RNAalifod [18] at

http://rna.tbi.univie.ac.at/cgi-bin/RNAalifold.cgi that pre-
dicts a consensus secondary structure from a set of pre-aligned
sequences; (b) Infernal [19] at http://infernal.janelia.org/
that builds consensus RNA profiles called covariance models
(CMs), which is not available as Web service but needs to be
downloaded and locally installed; (c) RNAG [20] at http://
ccmbweb.ccv.brown.edu/cgi-bin/rnag.pl, a Gibbs sampler for
predicting RNA secondary structure for unaligned sequences.
2. The mRNA sequences can be downloaded directly from
BLAST result page. Alternatively, the retrieval of FASTA-
formatted mRNAs from GenBank is feasible by providing a list
of IDs (accessions or GIs) to the EFetch E-utility of NCBI
Entrez through an URL. For example: http://eutils.ncbi.nlm.
nih.gov/entrez/eutils/efetch.fcgi?db=nucleotide&rettype=fa
sta&retmode=text&id=NM_123, XM_456,NM_789, …
where “db = nucleotide” indicates the NCBI database, “ret-
type = fasta” specifies the output sequence format and
“id = NM_123, XM_456,NM_789,…” represent the list of
IDs separated by commas.
Fig. 3 UTRsite entry (Accession ID U0049) fir the DNA polymerase beta stem loop II regulatory element
3. PatSearch executables for Linux and Mac OS X can be required

to Graziano Pesole (graziano.pesole@uniba.it) or Giorgio
Grillo (giorgio.grillo@ba.itb.cnr.it).
4. We suggest testing the correctness of PatSearch syntax per-
forming the pattern searching against all sequence of the mul-
tiple alignment.
5. To submit new UTR elements, users can follow the above
methodology and then contact the UTRsite staff by e-mail
following links inside the main Web page.
References
1. Wang C, Zhang MQ, Zhang Z (2013) 11. Morgenstern B (2004) DIALIGN: multiple
Computational identification of active enhanc- DNA and protein sequence alignment at
ers in model organisms. Genomics Proteomics BiBiServ. Nucleic Acids Res 32:W33–W36
Bioinformatics 11:142–150 12. Smith C, Heyne S, Richter AS et al (2010)
2. Keene JD (2007) RNA regulons: coordination Freiburg RNA tools: a web server integrating
of post-transcriptional events. Nat Rev Genet INTARNA, EXPARNA and LOCARNA.
8:533–543 Nucleic Acids Res 38:W373–W377
3. Piva F, Giulietti M, Burini AB et al (2012) 13. Grillo G, Licciulli F, Liuni S et al (2003)
SpliceAid 2: a database of human splicing fac- PatSearch: a program for the detection of pat-
tors expression data and RNA target motifs. terns and structural motifs in nucleotide
Hum Mutat 33:81–85 sequences. Nucleic Acids Res 31:3608–3612
4. Piva F, Giulietti M, Nocchi L et al (2009) 14. Pesole G, Liuni S, D’Souza M (2000)
SpliceAid: a database of experimental RNA PatSearch: a pattern matcher software that
target motifs bound by splicing proteins in finds functional elements in nucleotide and
humans. Bioinformatics 25:1211–1213 protein sequences and assesses their statistical
5. Giulietti M, Piva F, D’Antonio M et al (2013) significance. Bioinformatics 16:439–450
SpliceAid-F: a database of human splicing fac- 15. Grzybowska EA, Zayat V, Konopinski R et al
tors and their RNA-binding sites. Nucleic (2013) HAX-1 is a nucleocytoplasmic shut-
Acids Res 41:D125–D131 tling protein with a possible role in mRNA
6. Grillo G, Turi A, Licciulli F et al (2010) processing. FEBS J 280:256–272
UTRdb and UTRsite (RELEASE 2010): a 16. Nowak R, Siedlecki JA, Kaczmarek L et al
collection of sequences and regulatory motifs (1989) Levels and size complexity of DNA
of the untranslated regions of eukaryotic polymerase beta mRNA in rat regenerating
mRNAs. Nucleic Acids Res 38:D75–D80 liver and other organs. Biochim Biophys Acta
7. Mathews DH, Moss WN, Turner DH (2010) 1008:203–207
Folding and finding RNA secondary structure. 17. Konopinski R, Nowak R, Siedlecki JA (1996)
Cold Spring Harb Perspect Biol 2:a003665 Alternative polyadenylation of the gene tran-
8. Hamilton RS, Davis I (2011) Identifying and scripts encoding a rat DNA polymerase beta.
searching for conserved RNA localisation sig- Gene 176:191–195
nals. Methods Mol Biol 714:447–466 18. Hofacker IL (2007) RNA consensus structure
9. Sarnowska E, Grzybowska EA, Sobczak K prediction with RNAalifold. Methods Mol
et al (2007) Hairpin structure within the Biol 395:527–544
3′UTR of DNA polymerase beta mRNA acts 19. Nawrocki EP, Eddy SR (2013) Infernal 1.1:
as a post-transcriptional regulatory element 100-fold faster RNA homology searches.
and interacts with Hax-1. Nucleic Acids Res Bioinformatics 29:2933–2935
35:5499–5510 20. Wei D, Alpert LV, Lawrence CE (2011)
10. Johnson M, Zaretskaya I, Raytselis Y et al RNAG: a new Gibbs sampler for predicting
(2008) NCBI BLAST: a better web interface. RNA secondary structure for unaligned
Nucleic Acids Res 36:W5–W9 sequences. Bioinformatics 27:2486–2493
Chapter 22
Rfam: Annotating Families of Non-Coding RNA Sequences

Jennifer Daub, Ruth Y. Eberhardt, John G. Tate, and Sarah W. Burge
Abstract
The primary task of the Rfam database is to collate experimentally validated noncoding RNA (ncRNA)
sequences from the published literature and facilitate the prediction and annotation of new homologues in
novel nucleotide sequences. We group homologous ncRNA sequences into “families” and related families
are further grouped into “clans.” We collate and manually curate data cross-references for these families
from other databases and external resources. Our Web site offers researchers a simple interface to Rfam and
provides tools with which to annotate their own sequences using our covariance models (CMs), through our
tools for searching, browsing, and downloading information on Rfam families. In this chapter, we will work
through examples of annotating a query sequence, collating family information, and searching for data.
Key words Noncoding RNA, Rfam, Multiple sequence alignment, Secondary structure, Covariance
model, Infernal, Families, Clans
1 Introduction
The Rfam database (http://rfam.sanger.ac.uk) is a collection of

noncoding RNA sequences, grouped into “families” and “clans” on
the basis of nucleotide sequence and secondary structure homology.
This collection is not restricted to one type of ncRNA or one model
organism but includes many different types of ncRNA genes, cis-
regulatory elements, and intronic elements, all of which are anno-
tated in as many species as possible. Rfam release 11.0 (Aug 2012)
contains 2,208 ncRNA families (comprising 1,881 ncRNA genes,
317 cis-regulatory elements, and 10 catalytic introns) and 106 clans
(encompassing 321 families). These 2,208 families annotate over 6
million sequence regions in the public nucleotide databases.
Rfam relies on an underlying sequence database known as
“Rfamseq” which we derive from the European Nucleotide Archive
(ENA) [1]. This database is searched for homologous sequences
and annotated with matching Rfam families. For release 11.0,
Rfamseq contains the entire STD (standard annotated assembled
sequence) and WGS (whole genome shotgun sequencing) divi-
sions of ENA (EMBL release 110), approximately 89 million
349
350 Jennifer Daub et al.
sequences (over 274.9 × 109 bp), and includes genome sequences

for 3,110 unique species (101 eukaryotic, 1,747 bacterial, 123
archaeal, and 1,139 viral genomes).
1.1 Families An Rfam family is a collection of nucleotide sequences that we

believe have conserved sequence and secondary structure and is
derived from a common ancestor. Each family is represented by a
multiple sequence alignment annotated with a conserved second-
ary structure (referred to as the “seed”) and a probabilistic model
(covariance model, CM).
The seed alignment is a minimal set of sequences, which are
chosen as representative members of the family. Membership of the
seed is curated by Rfam; however this alignment is derived from at
least one experimentally validated example from the literature. The
seed secondary structure annotation is obtained from the source
literature or predicted using software tools such as WAR (Webserver
for aligning structural RNAs) [2] and mfold [3]. The Infernal soft-
ware [4] suite is used to generate the CM from the seed and used
to search the Rfamseq databases for new homologues. All predicted
homologues in Rfamseq are collated with the seed sequences into
a single large alignment termed the “full.” We refer our readers to
two comprehensive reviews on the use of covariance models for
annotating ncRNA [5, 6] and other Rfam publications for more
details on the process of building and annotating Rfam families [7].
Each family is assigned a stable accession number of the form
RFXXXXX, a family identifier (ID), a short description, and a
ncRNA type classification. These terms are important and can be
used to identify families of interest throughout the Web site. Rfam
collates and generates a range of information about each family,
which is found in the nine different tabs belonging to each family
page (described in more detail in Subheading 3.2.1). An overview
of the information available for families is provided in Fig. 1.
1.2 Clans Rfam has implemented the use of “clans” to group-related families
since Rfam release 10.0 in 2010 [8]. Families within a clan are
proposed to be derived from a common ancestor or be related in
sequence and structure, but are maintained as separate families due
to distinct functions.
Evidence for relationship between families is established using
computational tools (SCOOP (Simple Comparison Of Outputs
Program) [9], PRC (Profile comparer) [10]), and curation-based
approaches (literature searches and curation from specialized data-
bases). Within a clan, the same sequence may be a member of more
than one family, which is contrary to our usual curation rule that mul-
tiple families cannot annotate the same sequence region.
Clans are given stable accessions of the format CLXXXXX and
identifiers, as for families. For each clan, we provide a curation
summary that includes supporting literature, and we collate sec-
ondary structure images, PDB (Protein Data Bank) mappings, and
alignments for each of the families in the clan. Users can access
these using the four tabs provided on each clan page (Fig. 2).
Rfam: Annotating Non-Coding RNA 351
Fig. 1 Overview of family pages
Fig. 2 Overview of clan pages

2 Materials
2.1 Computational The Web site is compatible with all recent browsers and makes use of
Software several Java applets, such as Jalview for viewing sequence alignments
and AstexViewer for visualizing three-dimensional structures.
2.2 Example In this chapter, we provide a guide to using the Rfam Web site,
Sequence based on an example sequence which is a short region from the
genome of Clostridium tetani E88 (AE015927.1/1226050-
1226250), annotated as belonging to family RF00162, in clan
CL00012. This sequence can also be found on our ftp site: ftp://
ftp.sanger.ac.uk/pub/databases/Rfam/RNA_Methods_Mol_
Biol/C_tetani_AE015927_fragment.fa
>AE015927.1/1226050-1226250 Clostridium tetani E88, com-
plete genome
AAATAAAAAAACACTTCGATAAGAAGTGTTTATAGCAAAA
TAAACTTTTCTTATCTTTCAGGAAATTTCCTGCTGGA
GTTAGCACCTTTATACTTTGTGTATAGGTTGCTGGGT
TTCATAGGGCCAGTCCCTCAACCGCTCTTGATAAGAT
ATTTTAAATTATTTATTCAATTTAAAGCA
TATTATATATTTAATTATTTC
3 Methods
There are multiple ways to access data using the Rfam Web site;
Figure 3 summarizes some of the suggested routes to finding
information about a family, clan, or sequence of interest.
3.1 Sequence You can access the main sequence search page using the SEARCH
Searching link in the navigation bar, which appears at the top of every page in
the Rfam Web site. The interactive sequence search can only be
3.1.1 Searching
used for nucleotide sequence of less than 10,000 bases. You can see
with a Nucleotide
documentation on the restrictions on sequences under the More…
Sequence
link. If you have multiple sequences or if you prefer to retrieve
results as plain text, you can use the Batch search facility, which
runs searches offline and returns results via email.
In the interactive sequence search, the query sequence is scanned
against a library of Rfam family sequences using WU-BLAST with
an E-value threshold of 1.0 (see Note 1). Any matches are then
searched against the corresponding CM using the hand-curated bit
score gathering threshold (GA) for the family (see Note 2).
Paste the example sequence into the query box and press
Submit—your results should be returned within a few seconds.
The results page summarizes the number of hits and the query
sequence used. You are also provided with a URL to bookmark if
you wish to return to this result later.
Fig. 3 Searching Rfam
Our example sequence returns matches to two families

RF00162 (family ID “SAM”) and RF01725 (family ID “SAM-I-
IV-variant”). Consider the coordinates of the matches (the start
and end values) and the strand. The model for RF00162 aligns to
nucleotides 151-48 of our query sequence on the minus (−) strand,
and the RF01725 model aligns to nucleotides 153-47 also on the
minus strand, so both models are aligned to the same/overlapping
region of our query sequence.
If we consider the quality of these matches using the bit scores
and E-values (see Note 2), we can see that our query sequence
scores much more highly against the RF00162 model than
RF01725 (103.6 and 24.2 bits, respectively, and E-values of 3e−26
and 0.0025, respectively). We may therefore conclude that our
sequence is a close homologue of sequences in RF00162 and a
distant homologue of those in RF01725.
Fig. 4 Sequence search results
After running a search, you can view the alignment of your

query sequence to the model and the consensus sequence, as taken
from the output of CMalign (Fig. 4a), by clicking on the Show but-
ton in the far right column. There are six lines in this alignment:
#NC, #SS, #CM, #MATCH, #SEQ, and #PP. These few lines con-
dense large amounts of information from the Infernal output, and
users are referred to the Infernal manual for a full and detailed
explanation of this and all possible characters [11].
Briefly, in our example case, the #SS line shows the predicted
secondary structure of the query sequence in dot bracket notation.
The different closed bracket notations (“()” and “<>”) are used by
Infernal to indicate different structural features. In this example
the model contains three stem loops closed by another stem of
base pairing. Finally, “_” characters mark hairpin loops, and “,”
characters mark single-stranded residues in multifurcation loops.
The #CM line shows the consensus sequence of the covariance
model. The highest scoring residue at each position from the seed
alignment is shown. Uppercase residues are highly conserved; low-
ercase residues are weakly or not conserved. In the example model,
note the level of sequence conservation in the base-paired regions
and the increased variation in the last two loop regions. The #SEQ
line shows our query sequence and coordinates as aligned to the
model. “−” characters indicate deletions in the query sequence
with respect to the model. In the example sequence, these occur in
the single-stranded, less well-conserved regions.
The #MATCH line is the important indicator of alignment bit
score. Each position in the alignment is scored according to the
consensus sequence. For base-paired regions, if the observed pair
in the query sequence is the highest-scoring possible pair, both
residues are shown in uppercase; if a pair has an acceptable com-

pensatory pair, both residues are annotated by “:” characters.
Spaces indicate positions with a negative contribution to the align-
ment score. A similar convention is implemented for single-
stranded regions, with the highest scoring residues noted in
uppercase, acceptable scoring residues indicated with “+,” and
spaces indicating negatively scoring positions. Looking at our
example, we can see that the majority of positions are scored with
uppercase, or positively scoring characters, and our sequence is a
very good match to the RF000162 model.
The #NC line marks the positions of significant negatively
scoring pairs with a “v” character. In our example, base pairing
between residues 151 and 48, which is a predicted A:A base pair-
ing, is flagged up.
The #PP line is used to annotate the expected accuracy of each
aligned residue. “*” characters indicate 95–100 % accuracy, 9 indi-
cates 85–95 % accuracy, and so on with 0 indicating 0–5 % accu-
racy. You can see in the example the majority of the sequence is
aligned with 95–100 % accuracy, and the lower confidence region
occurs in the less well-conserved loop region.
If we now view the alignment of our query to the RF01725,
we can clearly see that there is much poorer confidence in the
alignment (#PP line) and poorer scoring in the #MATCH line
(Fig. 4b).
3.1.2 Searching Using Rather than searching for matches to ENA sequences, the Rfam
a Sequence Accession Web site allows you to quickly look up the sequence using its ENA
accession. Note that the sequence must exist in the ENA version
from which the current Rfamseq derives. You can use the Look up
sequence box under the sequence search form, or you can use the
Jump to field found on every tabbed page in the Rfam Web site (see
Notes 3 and 4).
Using the Look up sequence box, we can see that there are 181
annotations in Rfam for our example sequence, AE015927.1 (this
number is for the full sequence rather than the fragment that we
searched in Subheading 3.1.1). The results are displayed in a tabu-
lar format, ordered by default according to the order of coordi-
nates on the query sequence. You can sort the table on the different
columns of data by clicking on the table headers (see Note 5).
Browsing down the results to the match with coordinates
1226200-1226097, we find the SAM annotation that we identi-
fied by searching the sequence above. If we now sort the table by
Rfam familyID and scroll down to the SAM riboswitches, we can
see that Rfam annotates four different instances of the SAM family
(RF00162) on this sequence. Note the bit scores and coordinates
of the different annotations (three are annotated on the minus
strand and one on the positive).
3.2 Finding Family Entering the accession RF00162 or the family ID SAM in the
Information Jump to box will take us directly to the summary tab of the family
page for this specific family (see Fig. 1). The accession, identifier,
3.2.1 Family Pages
and description are displayed prominently at the top of all family
pages. If the family is in a clan, the other members of the clan will
also be displayed in the header of the summary page. We can see
immediately that family RF00162 belongs to clan CL00012 (clan
ID “SAM”), which contains two other families: SAM-I-IV-variant
(RF01725) and SAM-IV (RF00634).
Since 2007, Rfam has linked every family to the article in
Wikipedia that best describes that family [12]. The Wikipedia arti-
cle is displayed in the main body of the Summary tab and should
provide the user with preliminary background information, links,
and literature sources that are relevant to this ncRNA family.
If you browse through the different tabs (Fig. 1), you can col-
late pertinent information on the family itself and the sequences in
the full alignment, such as, under the Alignments tab, the number
of sequences in the seed and full alignments (433 and 4,757,
respectively). In the same tab, you can view or download the seed
or full alignments and sequences in several formats, with the option
of having sequences annotated using sequence accessions or species
identifiers.
You can browse multiple graphical representations of the pre-
dicted secondary structure and sequence conservation for the seed,
under the Secondary structures tab. In addition to the stem-loop
representations, we also provide R-chie diagrams [13] and a link for
viewing the structure interactively using the VARNA applet [14].
You can find the predicted phylogenetic trees for seed and full
alignments in the Trees tab, while under the Species tab, you can
browse, download, and align different taxonomic subsets of
sequences from the full alignment (see Subheading 3.2.2).
Any PDB sequence mappings we have made to this family can
be viewed in the Structures tab. Note that multiple PDB sequence
regions may be mapped to the same ENA sequence region. In the
case of RF00162, for example, we map 20 fragments from differ-
ent PDB sequences and chains onto nine regions from five differ-
ent ENA sequences. You can view these sequence fragments in the
context of the three-dimensional PDB structure using the
AstexViewer applet [15], which allows you to manipulate and
rotate the structure to view the sequence region and toggle
between different views of surface transparency.
The Database references tab provides external database cross-
references, such as links to Gene Ontology (GO) and Sequence
Ontology (SO) [16, 17]. Under the Curation tab, you will find
important information about the curation of the Rfam family,
including our family type classification, family source references,
authors, and family building parameters, and you can also down-
load the covariance model. Note that our example family has been
given the type classification “Cis-reg; riboswitch.” Also under the

Curation tab, you will see the GA for this family is 44 bits. The bit
score from our earlier query sequence search (Subheading 3.1.1)
was 103.6 bits, indicating that our query sequence is well above
the GA and a very good homologue of this family. At the bottom
of the curation page, you can also view the distribution of all of the
bit scores we obtain for every sequence in this family when we
search Rfamseq. Again you can see that our query sequence, with
a bit score of 103.6, lies to the far right of this range, indicating a
very good match to the model.
Browsing through the links on the Summary page to the two
other families in this clan, SAM-I-IV-variant (RF01725) and
SAM-IV (RF00634), you will find they are also classified as Cis-
reg; riboswitch. If you view the distribution of bit scores for
RF01725 under the Curation tab, you will see that the match
between this family and our query sequence (see above;
Subheading 3.1.1) is 24.2 bits, which is just above the trusted bit
score cutoff.
Following the link to the clan page for CL00012 (or using this
clan accession in the Jump to box) will take you to the main clan
page, where we also provide the curation summary, with appropri-
ate source references and a collation of the alignments and PDB
mappings for all three families in this clan. You can use the Secondary
structures tab here to view the seq-cons stem-loop secondary struc-
ture images for any two clan members, displayed side by side.
3.2.2 Species Two alternative representations of the species distribution in the

Distribution and full alignment are provided under the Species tab in the family page:
Downloading a Subset the “sunburst” visualization (the default) and an interactive tree
of Sequences representation. Only the sunburst tool is described here.
Briefly, the sunburst contains segments colored according to
taxonomic kingdom. Details on how it is generated can be found
under the More… link. Looking at the sunburst for our example
family RF00162, we can see immediately that it is predominantly
bacterial, with a significant number of unclassified sequences. As
the mouse is moved over the nodes within the sunburst, a tooltip
is displayed, showing the relevant taxonomic level, the number of
sequences, and the number of species for that node. The full taxo-
nomic lineage of the current node is displayed in the sunburst con-
trol panel at the top right. The size of each segment of the sunburst
can be scaled according either to the number of species or the
number of sequences in the family within this taxonomic node.
Toggle between these two options using the radio button to see
the difference this makes to the representation.
Using the sunburst, we can also select and download sequences
in the family. We can select sequences from our example species,
Clostridium tetani, by clicking the Clostridium tetani segment in
the outer ring (species level). You can zoom into the sunburst
using the size slider to make it easier to find the right node. When
you identify C. tetani, the tooltip will indicate “Clostridium tetani
[species], 1 sequence, 1 species.” After clicking to select the node,
the segment turns red, and the selected sequences are registered at
the bottom of the control panel on the right. You can download
these sequences in FASTA format or have them aligned to the
CM. Select Align to CM for our C. tetani selection and save the
resulting Stockholm formatted alignment of the four SAM annota-
tions indicated in Subheading 3.1.2.
The sunburst tool allows you to choose any arbitrary set of
species: you can select or exclude sequences for any taxonomic
division or combination of divisions you choose. For example, you
may choose to download all of the regions for all of the Clostridia
taxonomic class (425 sequences from 265 species) or for all of the
phylum Firmicutes (2,218 sequences from 677 species).
Alternatively, you could choose to select all of the phylum
Firmicutes excluding all the Clostridia apart from C. tetani.
For comparison, if you browse to the sunburst for family
RF01725, you can clearly see the difference in the taxonomic dis-
tribution between these two families. If you locate the Firmicutes,
you will see that there are only five sequences from five species
present in this phylum in RF01725.
3.3 Text Searching Under the main SEARCH tab, select the Entry type tab, you are
presented with a hierarchical representation of the Rfam family
3.3.1 ncRNA Type
types. There are three top level categories to this structure, under
(Type Search)
which all families are classified: gene, cis-reg, and intron. Use the
More… link to read more details about our classifications and what
they mean within Rfam.
Our example family is of the type “Cis-reg; riboswitch;.” This
means that our family is a cis-regulatory element of the subtype
riboswitch. To view all the families that Rfam annotates in the
“Cis-reg; riboswitch;” category, simply check the box next to ribo-
switch and press Submit. In the results page, you will see we have
26 families in this category. We already know our example family
RF00162 is in a clan with two others in this list, but some of these
other riboswitch families may also be of interest. This simple list of
checkboxes is very flexible and allows you to select/deselect any
combination of ncRNA categories depending on what data you are
looking for.
3.3.2 Annotations by Under the main SEARCH link, select the Taxonomy tab. The oper-
Species, Class, and Phylum ation of this taxonomy search box is relatively complex and is
(Taxonomy Searching) explained in more detail under the More… link. Briefly, to view all
of the family annotations that Rfam makes to Clostridium tetani
(regardless of sequence accession), you can simply add Clostridium
tetani to the search box, and you will be taken to a summary of all
families that we annotate in any sequence from this species. You
can also check the lower box on this page to identify families that
are unique to this species or taxonomic group. In the case of

Clostridium tetani, we annotate 32 different families. Note that
this search will only identify the families which we have identified
in this species, but will not list the individual or multiple annota-
tions of each family or each sequence. You would need to peruse
each family to see the individual annotations to this species. See
Subheading 3.2.2 for downloading taxonomic subsets of sequences
for a single family or use the Rfam BioMart (Subheading 3.5).
You can also widen the search to a more generic taxonomic
level and, for example, use the taxonomic class “Clostridia” in this
search box. This will identify all of the Rfam families in any species
that belong to this taxonomic class (in this case, we find 93 differ-
ent families). Increasing the taxonomic range to phylum level
“Firmicutes,” we find 180 families are identified in species in this
phylum. Again you should bear in mind that this result will not list
the individual annotations per species, and you would need to pur-
sue individual families to identify these species and sequences.
3.3.3 Genome If you know the sequence accession of your genome of interest
Annotation, by Species (e.g., the Clostridium tetani ENA accession is AE015927.1), you
Name or by Sequence can use the Jump to box to go directly to genome pages for that spe-
Accession (Jump cies (in the case of higher eukaryotes, you would use a chromosome
to and Browse) sequence accession, e.g., CM000209.2, for mouse chromosome 1).
If you do not know the genome accession or chromosome
accession for your species of interest, you can browse the list of
complete genomes that are annotated by Rfam, via the BROWSE
link in the navigation bar (see Subheading 3.4).
The genome page gives the total number of annotations (in
this case 181), the number of different families (32) found on
sequences from that genome, and some basic information and sta-
tistics on the genome itself. To view the annotations, visit the chro-
mosome tab. Since this is a bacterial genome and a complete
sequence, there is only one chromosome sequence. In the case of
higher eukaryotes, you would find the multiple chromosomes
listed here separately. From the chromosome tab, you can choose
to view the annotations or download them in GFF format.
3.3.4 PubMed ID, Author, All of our curated family and clan data are indexed, so you can use
PDB Structure, and Others multiple criteria to identify families of interest. The Keyword search
(Keyword Search) box appears in the top right corner of the every page in the Rfam
Web site, as well as under Keyword tab in the main search page. The
keyword search allows you to perform a simple text search against the
textual data from several different sections of the Rfam database:
● Rfam: accessions, identifiers, descriptions, type, comment lines
from Rfam families
● Wikipedia: text from Wikipedia articles linked to Rfam families
● Literature: literature reference titles, authors, and PubMed
identifiers
● PDB: sequence mappings

● Clans: accessions, identifiers, and family associations
If your keyword is found for a family across any of these cate-
gories, the family will be listed in the results. Because of the vol-
ume and diversity of textual information in Rfam, a text search can
return a large number of hits. Furthermore, the keyword search
will add wildcards to the end of your keyword, so the shorter your
keyword/string, the wider the possible set of matches and the
larger the number of matching families. This may result in matches
that do not interest you (sensitivity versus specificity), but it can
also be an extremely useful way to quickly identify families of inter-
est. Results are ordered according to the number of sections of the
database in which the search string appears, so that families for
which the string appears in multiple sections will be listed first, giv-
ing a crude indication of the likely relevance (or ubiquity) of the
search terms.
In the curation tab of the family page for our example family,
RF00162, you will see that “Henkin T” is credited as one of the
authors of the family. We can find other families that may be rele-
vant by using this name as a search string in the keyword search.
Searching with “Henkin” will return seven families, of which the
top result is RF00162, since the author’s name appears in three
sections of the database for that family (see Note 6). However,
there are four other families that have a match in the “literature”
section, and if you peruse the family information for each of these
families, you will find this author under the Database references tab.
Two of these four families are also riboswitches. In the remaining
two families, the reference to this author can be found only in the
associated Wikipedia article (most likely in the references or further
reading section).
As another example, we may wish to look up families/
sequences relating to a known structure in the PDB, e.g., 3NPB
(the crystal structure of YbxF bound to the SAM-I riboswitch
aptamer). You can use this PDB accession as a keyword search
term, and in this case, you will be taken directly to the family
RF00162 since that is the only family that maps to this specific
PDB sequence. In general, if there is only a single matching family
in the results of a keyword search, you will be redirected to the
family page immediately, rather than being presented with a table
of results with only a single row.
3.4 Browsing As an alternative to directed searching, you can use the BROWSE
link in the navigation toolbar to view our families listed under five
different criteria: families (listed by family ID), via annotated
genomes (listed by species name), as clans (listed by clan ID), fami-
lies with structures (listed by family ID), and families with Wikipedia
articles (listed by Wikipedia article title).
3.5 BioMart BioMart is an open-source data management system [18], provid-

ing a powerful Web-based interface to complete databases such as
Rfam. The Rfam BioMart can be accessed via the link in the navi-
gation toolbar. The BioMart provides a user-friendly means of per-
forming sophisticated queries on our database and offers much
more flexibility than the searches which are built into the Rfam
Web site. All of the text search examples above could equally well
be performed using the BioMart filters, with the advantage that
you have greater control over the content and format of the results.
3.6 RESTful Interface We provide a RESTful interface to enable users to interact pro-
grammatically with the services provided by Rfam. A detailed
description of this is beyond the scope of this chapter, but full
documentation, including sample queries and scripts, is available
via our help pages.
3.7 FTP Our ftp site (ftp://ftp.sanger.ac.uk/pub/databases/Rfam/) con-

tains all of our release data and database files for the current and
previous releases of Rfam. It is important to note that many of
these files are very large. A detailed description of the ftp site can
be found in the Rfam Web site help documentation.
4 Notes
1. This chapter is based on Rfam release 11.0. In subsequent

Rfam releases and with the adoption of Infernal 1.1, the
BLAST prefilters will be replaced with faster and more sensi-
tive HMMER3-based filters.
2. The bit score (also known as the log-odds score) is the quality
score generated for every sequence that the Infernal software
tries to match against the model. It is the log-odds ratio of the
probability that the sequence was generated by the CM versus a
random model of background sequence. Very simply, it is a mea-
sure of how well the sequence matches the model, with higher
bit scores indicating higher confidence in the match to the
model. When searching against Rfamseq and building a family,
we manually curate a bit score threshold (termed the gathering
threshold or cutoff (GA)) that in our opinion best discriminates
between “real” homologues and the background distribution of
false hits in the database. Sequences scoring above this threshold
are considered members of the family, those below it are excluded.
3. Each Infernal sequence search and resultant bit score has an
associated E-value. The E-value is the statistical significance
of the hit, i.e., the number of hits expected to score this high
in a database of this size (measured by the total number of
nucleotides) if the database contained only nonhomologous
random sequences [11].
4. The Jump to box: this field appears on the homepage and in

the sidebar of all tabbed pages, such as the family, clan, and
genome pages. It provides quick access to the data pages in the
site, allowing you to enter any of a wide range of accessions or
identifiers and be taken directly to the appropriate page. This
can be extremely useful when you know exactly what you are
looking for. The Jump to field will only accept valid, accurate
identifiers. It does not perform a search like the keyword
search and does not accept multiple values or search terms.
5. In cases where the ENA sequence is a whole genome sequence,
the Jump to box will perform differently from the Look up
sequence box, directing you to the page for the genome rather
than simply collating the list of annotations (see
Subheading 3.3.3 for using the Jump to box to view a genome).
Furthermore, the Jump to box will perform differently depend-
ing on whether the sequence accession version is used. The
example AE015927.1 will take you to the genome pages,
while AE015927 will take you to the collated annotations.
6. The majority of tabular data is presented in tables that are sort-
able by the data in each of the different columns. Clicking on
a column header will perform an alphanumerical sort on all of
the data in the table based on the data in the chosen column.
7. If you include the author’s initial in this search using “Henkin
T” as opposed to “Henkin,” you will obtain the same set of
seven families in the results, but with slightly different scoring
profiles across the sections of the database.
Acknowledgments
The Rfam database is funded by the Wellcome Trust (grant num-

ber WT077044/Z/05/Z).
References
1. Cochrane G, Alako B, Amid C, Bower L, 4. Nawrocki EP, Kolbe DL, Eddy SR (2009)
Cerdeno-Tarraga A, Cleland I, Gibson R, Infernal 1.0: inference of RNA alignments.
Goodgame N, Jang M, Kay S et al (2013) Bioinformatics 25(10):1335–1337
Facing growth in the European Nucleotide 5. Gardner PP (2009) The use of covariance
Archive. Nucleic Acids Res 41(Database models to annotate RNAs in whole genomes.
issue):D30–D35 Brief Funct Genomic Proteomic 8(6):
2. Torarinsson E, Lindgreen S (2008) WAR: 444–450
Webserver for aligning structural RNAs. 6. Nawrocki EP, Eddy SR (2013) Computational
Nucleic Acids Res 36(Web server issue): identification of functional RNA homologs in
W79–W84 metagenomic data. RNA Biol
3. Zuker M (2003) Mfold web server for nucleic 10(7):1170–1179
acid folding and hybridization prediction. 7. Burge SW, Daub J, Eberhardt R, Tate J,
Nucleic Acids Res 31(13):3406–3415 Barquist L, Nawrocki EP, Eddy SR, Gardner
PP, Bateman A (2013) Rfam 11.0: 10 years of visualizing RNA secondary structures. Nucleic
RNA families. Nucleic Acids Res 41(Database Acids Res 40(12):e95
issue):D226–D232 14. Darty K, Denise A, Ponty Y (2009) VARNA:
8. Gardner PP, Daub J, Tate J, Moore BL, Osuch interactive drawing and editing of the RNA
IH, Griffiths-Jones S, Finn RD, Nawrocki EP, secondary structure. Bioinformatics 25(15):
Kolbe DL, Eddy SR et al (2011) Rfam: Wikipedia, 1974–1975
clans and the “decimal” release. Nucleic Acids 15. Hartshorn MJ (2002) AstexViewer: a visuali-
Res 39(Database issue):D141–D145 sation aid for structure-based drug design.
9. Bateman A, Finn RD (2007) SCOOP: a simple J Comput Aided Mol Des 16(12):871–881
method for identification of novel protein 16. Ashburner M, Ball CA, Blake JA, Botstein D,
superfamily relationships. Bioinformatics Butler H, Cherry JM, Davis AP, Dolinski K,
23(7):809–814 Dwight SS, Eppig JT et al (2000) Gene ontol-
10. Madera M (2008) Profile comparer: a program ogy: tool for the unification of biology. The
for scoring and aligning profile hidden Markov Gene Ontology Consortium. Nat Genet
models. Bioinformatics 24(22):2630–2631 25(1):25–29
11. Infernal User Guide. ftp://selab.janelia.org/ 17. Mungall CJ, Batchelor C, Eilbeck K (2011)
pub/software/infernal/Userguide.pdf Evolution of the Sequence Ontology terms
12. Daub J, Gardner PP, Tate J, Ramskold D, and relationships. J Biomed Inform 44(1):
Manske M, Scott WG, Weinberg Z, Griffiths- 87–93
Jones S, Bateman A (2008) The RNA 18. Zhang J, Haider S, Baran J, Cros A, Guberman
WikiProject: community annotation of RNA JM, Hsu J, Liang Y, Yao L, Kasprzyk A (2011)
families. RNA 14(12):2462–2464 BioMart: a data federation framework for
13. Lai D, Proctor JR, Zhu JY, Meyer IM (2012) large collaborative projects. Database 2011:
R-CHIE: a web server and R package for bar038
Chapter 23
ASPicDB: A Database Web Tool for Alternative

Splicing Analysis
Mattia D’Antonio, Tiziana Castrgnanò, Matteo Pallocca,
Anna Maria D’Erchia, Ernesto Picardi, and Graziano Pesole
Abstract
Alternative splicing (AS) is a basic molecular phenomenon that increases the functional complexity of
higher eukaryotic transcriptomes. Indeed, through AS individual gene loci can generate multiple RNAs
from the same pre-mRNA. AS has been investigated in a variety of clinical and pathological studies, such
as the transcriptome regulation in cancer. In human, recent works based on massive RNA sequencing
indicate that >95 % of pre-mRNAs are processed to yield multiple transcripts. Given the biological rele-
vance of AS, several computational efforts have been done leading to the implementation of novel algo-
rithms and specific specialized databases. Here we describe the web application ASPicDB that allows the
recovery of detailed biological information about the splicing mechanism. ASPicDB provides powerful
querying systems to interrogate AS events at gene, transcript, and protein levels. Finally, ASPicDB includes
web visualization instruments to browse and export results for further off-line analyses.
Key words Alternative splicing, Transcriptome, Transcription, Alternative isoforms, Database, Web
tool, Transcript prediction, Bioinformatics, Transcriptomics
1 Introduction
There has been a growing interest toward alternative splicing (AS)

in recent years, as demonstrated by large-scale studies revealing
that the majority (>95 %) of human multi-exonic genes undergo
alternative splicing events [1–3] and showing AS as the main
molecular phenomenon responsible for the increased complexity
of eukaryotic transcriptomes. Splice sites have a great biological
value because their alternative usage can be used as a robust molec-
ular signature of biodiversity, human pathologies including cancer,
and to trace cell developmental stages [4].
To investigate AS in eukaryotes, we have developed ASPicDB
[5, 6], a specialized database designed to store reliable alternative
splicing annotations predicted using ASPic [7, 8] algorithm and its
recent evolution PIntron [9].
365
ASPicDB has also been conceived to handle AS biological

information in different ways. Indeed, customized web interfaces
enable complex queries to retrieve AS annotations at transcript,
gene, and protein levels. Several interactive tables and graphical
overviews allow a punctual inspection of results. Database search is
also improved through sequence-based queries using BLAST.
Along the text, the term ASPicDB will be interchangeably
used to describe both the relational database and the related web
application.
2 Materials
ASPicDB is freely available at the following address: www.caspur.

it/ASPicDB.
This resource is a relational database built with MySQL and
featuring an ad hoc web interface that allows the user to fully
retrieve/visualize any information collected in the database. Web
applications are particularly convenient to wrap databases, since
they can significantly simplify and enhance the query and visualiza-
tion phases. A customized web interface can help the user to query
the dataset without worrying about the data organization and
other technical issues. Moreover, a graphical user interface can
show the results in a more organized and clear fashion and allows
the integration of data with further information, statistics, graphi-
cal representations, annotations, cross-linking, and other useful
tools to help the user to browse and interpret the dataset.
Since ASPicDB is available through a web interface, the user
only needs a regular computer or laptop with an updated web
browser to run queries and visualize results. Specific dataset can be
exported as flat files (see Subheading 3.5 below) for further analysis
and cross-validation with different tools. In this case a standard
text editor could be required. Depending on file size, text manipu-
lation could be time and memory expensive. For text files greater
than 10 MB we recommend to avoid spreadsheet-based tools.
3 Methods
ASPicDB stores AS annotations predicted using ASPic [7, 8] algo-

rithm and its recent evolution PIntron [9]. Such algorithm deduces
all nonredundant sets of alternative isoforms for a given gene,
providing as input its genomic sequence and related expressed
sequences (mostly ESTs and full-length cDNAs).
Data contained in ASPicDB can be divided in two main sub-
sets: predictions calculated by the ASPic/PIntron algorithm and
annotations associated with predictions.
ASPicDB for Alternative Splicing Analysis 367
ASPic/PIntron predictions are based on multi-alignments of

expressed sequence tags (ESTs) [7] from UniGene clusters [10] of
a variety of organisms (Homo sapiens, Mus musculus, Rattus nor-
vegicus, Bos taurus, Gallus gallus, and Arabidopsis thaliana).
Predicted full-length isoforms are automatically annotated employ-
ing several external databases such as GenBank [11], RefSeq [12],
UniProtKB/Swiss-Prot [13], Protein Data Bank (PDB) [14],
Pfam [15], and UniRef90 [16].
Current ASPicDB release also integrates info from the
SpliceAid-F database [17] collecting known splicing factors and
their binding sites.
ASPicDB can be queried through three different search forms:
(1) the simple search form allowing basic queries on the most com-
mon fields, such as specific genes; (2) the advanced search form
designed to perform complex queries on all available predictions
and annotations, allowing the interrogation of dedicated tracks at
gene, transcript, exon, introns, or protein levels; and (3) the
BLAST [18] form permitting sequence queries on the entire col-
lection of transcripts or proteins.
We provide usage guidelines for each form in the following
subsections.
3.1 Simple The simple search form comprises three different panels as shown
Search Form in Fig. 1. The user can decide to perform a query providing a gene
ID or a gene ontology term or whatever keywords of interest.
Fig. 1 Overview of the simple query form

In the first panel, the following gene identifiers are

supported:
● HGNC [19] (e.g., NSMCE1)
● UniGene [10] (e.g., Hs.481720)
● RefSeq [12] (e.g., NM_152713)
● Entrez Gene [20] (e.g., 3248)
● MIM [21] (e.g., 202300)
The central panel allows the selection of one or more genes
matching specific query keywords. The match is mainly restricted
to gene descriptions extracted from the GenBank database.
Finally, the third panel allows the interrogation of genes based
on certain ontology terms [22] related to cellular components,
molecular functions, or biological processes. The form admits also
term substrings (such as calcium ion) in association with one of the
three already mentioned semantic areas.
Search Results are presented in tabular form including different
fields associated with the found gene such as accession ID, descrip-
tion, organism, genomic coordinates, transcript number, alterna-
tive proteins, and total number of splicing events. A specific link
allows the access to the graphical visualization about predictions
and annotations.
3.2 Advanced The advanced search form allows the combination of different
Search Form search criteria in order to increase the specificity of the interroga-
tion. Five different panels are available to perform queries on dif-
ferent biological aspects, and each panel contains a set of common
and specific features.
Every module includes the following three fields: Organism,
Accession, and Coordinates. Only the Organism field is mandatory,
and, actually, queries on multiple organisms are not feasible.
Through the Coordinates fields, users can select genomic regions
in which the search can be performed.
In addition to all common filters, in the gene section (Fig. 2),
users can query:
● Keywords matching gene descriptions (e.g., “apoptosis”)
● RNA regions in which splicing events are located (i.e., 5′ UTR;
CDS; 3′ UTR)
● Type of splicing events (e.g., exon skipping, intron retention,
etc.)
● Specific donor-acceptor splice sites (e.g., GT-AG)
● Intron class (e.g.. U2, U12)
● Number of alternative proteins and transcripts (range)
● Number of independent splicing events (range)
Fig. 2 Overview of the advanced query form (Gene Tab)
● Presence of nonsense-mediated decay (NMD) transcripts

In contrast, in the transcript sections, users can search alterna-
tive isoforms of a certain gene. Although several fields are shared
with the gene search form, the transcript section allows also to:
● Retrieve a specific transcript by its signature [23].
● Filter by the presence of annotated poly(A) tails and poly(A)
signals (according to the PolyA_DB [24] standard).
● Filter by CDS, 5′ UTR, and 3′ UTR length.
● Filter by the number of exons.
The exon retrieval form enables the user to spot single exons
that match to a set of personalized criteria, such as:
● Minimal and maximal exon length
● Exon class (initial, internal, terminal)
● Features of flanking introns (donor-acceptor sites, intron class)
● Number of ESTs supporting flanking introns
● Presence of poly-A tail or poly-A signal annotated in the exon
To retrieve particular splice sites (introns), the user can cus-
tomize the following filters:
Specific donor-acceptor sites
● Intron class (U12, U2, not determined)
● Number of ESTs supporting the intron
● Intron length (interval)
● Minimal and maximal repeat site length in the intron sequence
● Presence of ambiguous splicing sites (i.e., the intron boundar-
ies cannot be precisely defined due to an ambiguous alignment
between cDNA and gDNA)
Finally, the protein search form is divided in two sections: the
first one allows to search proteins with structural or functional
properties; the second one is able to look for genes encoding alter-
native proteins matching user-defined criteria.
The first section allows to query for:
● Pfam class
● Protein class (e.g., globular, transmembrane)
● Number of functional domains
● Cellular localization for globular proteins (e.g., nucleus, cyto-
plasm, mitochondria)
● Presence of GPI anchors and signal peptides
The second panel contains a set of checkboxes to select fea-
tures requested to be different in the alternative encoded proteins
for the resulting genes. For instance, by selecting “subcellular
localization,” the application will fetch all genes that encode for
alternative proteins that differ in subcellular localization (e.g.,
some proteins localized in nucleus and others in cytoplasm).
3.3 BLAST An integrated BLAST search enables sequence queries to interrogate

Search Form the full set of ASPicDB alternative transcripts and proteins.
Alignment results are listed in a table reporting the five most
significant matches, in which every row provides additional infor-
mation such as the matched transcript identifier, the sequence
query length, and the matching score. Results are also linked to
the corresponding ASPicDB entries.
This facility could be very useful in combination with experi-
mental techniques to investigate specific transcripts or embedded
features in diverse conditions.
3.4 Graphical ASPicDB allows graphical representation of results facilitating the

Visualization inspection of annotations in an easy and intuitive way.
of Results Query results, obtained through the above-described forms,
are linked to detailed web pages showing all available information
for each extracted entry (i.e., gene, transcript, protein, exons, and
introns). The main result page contains eight subsections. Of these,
three provides a different dynamic image with mouseover tooltips.
The Gene Information table contains an overview of genomic
characteristics for the gene under investigation (accession IDs,
description, coordinates, the number of ESTs contained in the
UniGene cluster). Furthermore, the number of transcripts, proteins,
and splicing events is also summarized along with a set of web links to
orthologous genes (processed in ASPicDB for different organisms),
input and output raw files, and additional public data from wide-
spread databases (e.g., ASAP2 [25], ASTD [26], AceView [27]).
In the Gene Structure view, the user can visualize the predicted
gene structure. The image displays predicted introns and exons
other than the reference transcript from RefSeq database.
Figure 3 reports an example of structure view for the KRT82
(keratin 82) human gene. A further but more complex example is
for the TP63 human gene shown in Fig. 4, because it contains
several alternative transcripts. Introns, as the number 13 in Fig. 4,
Fig. 3 Structure view for the human keratin 82 (KRT82) gene

Fig. 4 Structure view for the human tumor protein p63 (TP63)
that do not have canonical splicing sites (black dots at ends) are
represented by orange dotted lines and are labeled “fuzzy introns.”
The Predicted Transcripts section provides a graphical repre-
sentation of the predicted transcripts. Only the most reliable
introns (i.e., those supported by a perfectly aligned EST with
canonical donor and acceptor sites or those supported by two or
more ESTs) are used for the assembly process. Figure 5 shows a
“Transcript View” snapshot for the ADSL human gene where
exons (thick-colored strips) and introns (black junction regions)
are clearly identifiable, along with other functional regions such as
5′ UTR (yellow strips) and 3′ UTR (sky blue). Interestingly, mass-
spectrometry identified peptide sequences [28] supporting specific
coding exons or splice junctions are represented in the top of the
Transcript View as small green boxes.
The mouseover on specific exons or introns activates the
tooltip with additional information, while a click opens a popup
page containing the genomic sequence of the selected element.
Furthermore, two icons to the right of each transcript are linked to
the genomic sequence of the transcript and the corresponding pro-
tein (if the transcript has a coding sequence).
A 3′ terminal arrow marks polyadenylated transcripts that may
also contain a poly(A) site (AAUAAA or its accepted variants), and
a hyphen marks the position of mapping CAGE [29] tags which
identify transcription initiation sites. In Fig. 6, five of the alterna-
tive transcripts of the BRAF gene are shown. Since the fourth
doesn’t contain an annotated CDS, its exons are displayed in gray
Fig. 5 Visualization of predicted transcripts for the adenylosuccinate lyase (ADSL) encoding gene
Fig. 6 Visualization of alternative transcripts predicted for B-Raf proto-oncogene serine/threonine kinase (BRAF)
(unannotated RNA), and obviously it’s not possible to download

the associated protein sequence. The figure also shows the presence
of a poly(A) tail at the 3′ end of the second and fifth transcript.
A CDS can be terminated with a Premature Termination
Codon, represented with a black rhombus. Potential splicing regu-
latory element binding sites (fetched from SpliceAid-F [17] data-
base) are displayed as red vertical bars.
The Predicted Proteins section shows an overview of all the
functional domains of the alternative predicted proteins including
Pfam’s positions, transmembrane (TN), and coiled-coil (CC)
domains. TN domains are represented by an orange stylized icon,
while CC domains with a helix. Figure 7 shows an example of
the ARAF protein structure. A mouseover tooltip displays addi-
tional details on the selected domain such as coordinates, length,
and P value.
The Predicted splice sites section shows the multiple alignment
of the genomic sequence against the mRNA and EST, specifically
at the boundaries of all the predicted introns. This feature can be
very useful when double-checking the predicted data, since it
allows to verify which information is used for prediction, and how
the alignments have been made.
In the Transcript Table a complete list of predicted alternative
transcripts is provided as well as their variations with respect to a
Fig. 7 Visualization of predicted functional domains for A-Raf proto-oncogene, serine/threonine kinase (ARAF)
reference transcript (the first one in the table) usually corresponding

to one of the transcripts collected in RefSeq. The purpose of this
section is to integrate the Predicted Transcripts section with addi-
tional details. Each transcript is annotated with the length, the num-
ber of exons, and the coding sequence (CDS) coordinates. In case
of truncated transcript coding sequences, coordinates are not com-
pletely reliable and it’s possible to find a CDS longer than the one
reported. In such cases coordinates are annotated with a sign < at
the start point and/or a sign > at the end point.
The table also shows information on the predicted product
and in particular if the CDS coordinates totally or partially match
with the ones in the reference transcript and if its shares start/stop
codons. The Variant Type column lists all predicted events for the
transcript, encoded with a conventional code [23] to summarize
the information.
Every splicing event is indicated with a keyword, such as new
(new exon), skip (exon skip), IR (intron retention), init (alterna-
tive initiation), or term (alternative termination).
Every keyword is followed by the element affected by the
event. For instance, E3 is the third exon or I9 the ninth intron in
the reference RefSeq. If there is an additional element (e.g., exon)
present with respect of the reference RefSeq element, a letter is
added after the position of the last element in the reference, e.g., if
a new exon is located downstream exon 1 in the reference, the vari-
ant is labeled “new (E1A).” Furthermore, the event localization is
included: 5’UTR, CDS, 3’UTR, or in an intersection region
5’UTR_CDS and CDS_3’UTR.
In this example, skip(E2),5’UTR; term(E6c),CDS_3’UTR,
two splicing events are described. The second exon (in 5’UTR) is
skipped and the transcript terminates with the third (c) variant of
the sixth exon, overlapping the CDS and 3’UTR.
The Intron Table lists all predicted introns and their main char-
acteristics, in particular, the relative and absolute coordinates, the
length, and the number of supporting ESTs. Donor-acceptor sites
are reported with relative scores, encoded with the number of mis-
matching bases on the first and last 15 bases of, respectively, the
downstream and upstream exon. The introns included in the refer-
ence transcript are printed in red.
The user can find in the Protein Table every predicted protein
with annotation coming from different annotation software. This
section lists additional details about the predicted alternative pro-
teins, such as:
● Type: the protein class (i.e., globular on transmembrane) as
predicted by Ensembl [30]
● Subloc: subcellular localization (i.e., nucleus, cytoplasm, mito-
chondria) from BaCelLo [31]
● PDB: link to the Protein Data Bank [14]
● UniProt: link to the UniProt Bank [13]
● SPEP: presence of signal peptides (predicted by SPEPlip

software [32])
● GPI prediction from PredGPI [33]
● TMDOM: coordinates of transmembrane domains
(Ensembl [30])
3.5 Exporting The possibility of exporting search results is crucial to allow further
Results analysis and cross-validation with different tools. ASPicDB includes
different download options and formats.
Search result tables contain a link to the download section,
where it is possible to explore the element (i.e., gene, transcript,
exon, intron, or protein) sequences. Every element has different
download options, e.g., an intron can be extended with exonic
sequences preceding and following the donor-acceptor sites (exten-
sion length is customizable by the user). For each transcript, exon
sequences are downloadable, as well as its alternative protein.
Since the result sets can be considerably large, the extraction of
the sequence data can be time-consuming. To avoid long waiting
times, the user can just leave an email address, where a web link
with the data will be sent, upon query completion.
A second layer of data export is available in the different Result
Visualization section. In Gene Information all coding and tran-
script sequences are available for download, as well as the predic-
tion result in a GTF file (Gene Transfer File). The GTF contains
the list of predicted transcripts, exons, and introns. For each tran-
script all its exons and introns are listed, with a report containing
the EST supporting each intron.
Moreover, in Predicted Transcripts, it’s possible to download
sequences for each exon, intron, transcript, and encoded proteins.
Proteins can also be fetched in the Protein Table. In Predicted Splice
sites an overall export process is available: users can download all
the multiple EST alignments against genome splicing sites.
Acknowledgments

Università e Ricerca (MIUR): PRIN 2009 and 2010 and Consiglio
Nazionale delle Ricerche: Flagship Project Epigen, Medicina
Personalizzata and Aging Program 2012–2014.
References
1. Wang ET, Sandberg R, Luo S, Khrebtukova I, 2. Pan Q, Shai O, Lee LJ, Frey BJ, Blencowe BJ
Zhang L, Mayr C, Kingsmore SF, Schroth GP, (2008) Deep surveying of alternative splicing
Burge CB (2008) Alternative isoform regula- complexity in the human transcriptome by
tion in human tissue transcriptomes. Nature high-throughput sequencing. Nat Genet
456:470–476 40(12):1413–1415
3. Hallegger M, Llorian M, Smith CW (2010) 15. Pfam database. ftp://ftp.sanger.ac.uk/pub/

Alternative splicing: global insights. FEBS J databases/Pfam/Tools/
277(4):856–866 16. Uniref 90. http://www.uniprot.org/help/
4. Barash Y, Calarco JA, Gao W, Pan Q, Wang X, uniref
Shai O, Blencowe BJ, Frey BJ (2010) 17. Giulietti M, Piva F, D’Antonio M, D’Onorio
Deciphering the splicing code. Nature 465: De Meo P, Paoletti D, Castrignanò T, D’Erchia
53–59 AM, Picardi E, Zambelli F, Principato G,
5. Castrignanò T, D’Antonio M, Anselmo A, Pavesi G, Pesole G (2013) SpliceAid-F: a data-
Carrabino D, D’Onorio De Meo A, D'Erchia base of human splicing factors and their RNA-
AM, Licciulli F, Mangiulli M, Mignone F, binding sites. Nucleic Acids Res 41(Database
Pavesi G, Picardi E, Riva A, Rizzi R, Bonizzoni issue):D125–D131
P, Pesole G (2008) ASPicDB: a database 18. Altschul SF, Gish W, Miller W, Myers EW,
resource for alternative splicing analysis. Lipman DJ (1990) Basic local alignment
Bioinformatics 24(10):1300–1304 search tool. J Mol Biol 215(3):403–410
6. Martelli PL, D’Antonio M, Bonizzoni P, 19. Gray KA, Daugherty LC, Gordon SM, Seal
Castrignanò T, D’Erchia AM, D’Onorio De RL, Wright MW, Bruford EA (2013)
Meo P, Fariselli P, Finelli M, Licciulli F, Genenames.org: the HGNC resources in
Mangiulli M, Mignone F, Pavesi G, Picardi E, 2013. Nucleic Acids Res 41(Database
Rizzi R, Rossi I, Valletti A, Zauli A, Zambelli F, issue):D545–D552. doi:10.1093/nar/
Casadio R, Pesole G (2011) ASPicDB: a data- gks1066, Epub 2012 Nov 17
base of annotated transcript and protein vari- 20. Maglott D, Ostell J, Pruitt KD, Tatusova T
ants generated by alternative splicing. Nucleic (2005) Entrez Gene: gene-centered informa-
Acids Res 39(Database issue):D80–D85 tion at NCBI. Nucleic Acids Res 33(Database
7. Bonizzoni P, Mauri G, Pesole G, Picardi E, issue):D54–D58
Pirola Y, Rizzi R (2009) Detecting alternative 21. Hamosh A, Scott AF, Amberger JS, Bocchini
gene structures from spliced ESTs: a computa- CA, McKusick VA (2005) Online Mendelian
tional approach. J Comput Biol 16:43–66 Inheritance in Man (OMIM), a knowledge-
8. Bonizzoni P, Rizzi R, Pesole G (2005) ASPIC: base of human genes and genetic disorders.
a novel method to predict the exon-intron Nucleic Acids Res 33(Database issue):
structure of a gene that is optimally compatible D514–D517
to a set of transcript sequences. BMC 22. Ashburner M, Ball CA, Blake JA, Botstein D,
Bioinformatics 6:244 Butler H, Cherry JM, Davis AP, Dolinski K,
9. Pirola Y, Rizzi R, Picardi E, Pesole G, Della Dwight SS, Eppig JT, Harris MA, Hill DP,
Vedova G, Bonizzoni P (2012) PIntron: a Issel-Tarver L, Kasarskis A, Lewis S, Matese
fast method for detecting the gene structure JC, Richardson JE, Ringwald M, Rubin GM,
due to alternative splicing via maximal pair- Sherlock G (2000) Gene ontology: tool for the
ings of a pattern and a text. BMC unification of biology. The Gene Ontology
Bioinformatics 13(Suppl 5):S2. doi:10.1186/ Consortium. Nat Genet 25(1):25–29
1471-2105-13-S5-S2 23. Riva A, Pesole G (2009) A unique, consistent
10. Pontius JU, Wagner L, Schuler GD (2003) identifier for alternatively spliced transcript
UniGene: a unified view of the transcriptome. variants. PLoS One 4(10):e7631
In: McEntyre J, Ostell J The NCBI Handbook. 24. Zhang H, Hu J, Recce M, Tian B (2005)
Bethesda, MD: National Center for PolyA_DB: a database for mammalian mRNA
Biotechnology Information polyadenylation. Nucleic Acids Res
11. Benson DA, Cavanaugh M, Clark K, Karsch- 33(Database issue):D116–D120
Mizrachi I, Lipman DJ, Ostell J, Sayers EW 25. Kim N, Alekseyenko AV, Roy M, Lee C (2007)
(2013) Genbank. Nucleic Acids Res The ASAP II database: analysis and compara-
41(Database issue):D36–D42 tive genomics of alternative splicing in 15 ani-
12. refseq. http://www.ncbi.nlm.nih.gov/pubmed/ mal species. Nucleic Acids Res 35(Database
24259432 issue):D93–D98
13. Boutet E, Lieberherr D, Tognolli M, Schneider 26. Stamm S, Riethoven JJ, Le Texier V,
M, Bairoch A (2007) UniProtKB/Swiss-Prot. Gopalakrishnan C, Kumanduri V, Tang Y,
Methods Mol Biol 406:89–112 Barbosa-Morais NL, Thanaraj TA (2006)
14. Berman HM, Westbrook J, Feng Z, Gilliland ASD: a bioinformatics resource on alternative
G, Bhat TN, Weissig H, Shindyalov IN, splicing. Nucleic Acids Res 34:D46–D55
Bourne PE (2000) The protein data bank. 27. Thierry-Mieg D, Thierry-Mieg J (2006)
Nucleic Acids Res 28(1):235–242 AceView: a comprehensive cDNA-supported
gene and transcripts annotation. Genome Biol Mongin E, Pettett R, Pocock M, Potter S,
7 (Suppl 1), S12, 1–14 Rust A, Schmidt E, Searle S, Slater G, Smith J,
28. Desiere F, Deutsch EW, King NL, Nesvizhskii Spooner W, Stabenau A, Stalker J, Stupka E,
AI, Mallick P, Eng J, Chen S, Eddes J, Ureta-Vidal A, Vastrik I, Clamp M (2002) The
Loevenich SN, Aebersold R (2006) The Ensembl genome database project. Nucleic
PeptideAtlas project. Nucleic Acids Res Acids Res 30(1):38–41
34(Database issue):D655–D658 31. Pierleoni A, Martelli PL, Fariselli P, Casadio R
29. Kawaji H, Kasukawa T, Fukuda S, Katayama S, (2006) BaCelLo: a balanced subcellular
Kai C, Kawai J, Carninci P, Hayashizaki Y localization predictor. Bioinformatics 22(14):
(2006) CAGE basic/analysis databases: the e408–e416
CAGE resource for comprehensive promoter 32. Fariselli P, Finocchiaro G, Casadio R (2003)
analysis. Nucleic Acids Res 34(Database SPEPlip: the detection of signal peptide and
issue):D632–D636 lipoprotein cleavage sites. Bioinformatics
30. Hubbard T, Barker D, Birney E, Cameron G, 19(18):2498–2499
Chen Y, Clark L, Cox T, Cuff J, Curwen V, 33. Pierleoni A, Martelli PL, Casadio R (2008)
Down T, Durbin R, Eyras E, Gilbert J, PredGPI: a GPI-anchor predictor. BMC
Hammond M, Huminiecki L, Kasprzyk A, Bioinformatics 9:392–395. doi:10.1186/
Lehvaslaiho H, Lijnzaad P, Melsopp C, 1471-2105-9-392
Chapter 24
Analysis of Alternative Splicing Events in Custom Gene

Datasets by AStalavista
Sylvain Foissac and Michael Sammeth
Abstract
Alternative splicing (AS) is a eukaryotic principle to derive more than one RNA product from transcribed
genes by removing distinct subsets of introns from a premature polymer. We know today that this process
is highly regulated and makes up a large part of the differences between species, cell types, and states. The
key to compare AS across different genes or organisms is to tokenize the AS phenomenon into atomary
units, so-called AS events. These events then usually are grouped by common patterns to investigate the
underlying molecular mechanisms that drive their regulation. However, attempts to decompose loci with
AS observations into events are often hampered by applying a limited set of a priori defined event patterns
which are not capable to describe all AS configurations and therefore cannot decompose the phenomenon
exhaustively.
In this chapter, we describe working scenarios of AStalavista, a computational method that reports all
AS events reflected by transcript annotations. We show how to practically employ AStalavista to study AS
variation in complex transcriptomes, as characterized by the human GENCODE annotation. Our exam-
ples demonstrate how the inherent and universal AStalavista paradigm allows for an automatic delineation
of AS events in custom gene datasets. Additionally, we sketch an example of an AStalavista use case includ-
ing next-generation sequencing data (RNA-Seq) to enrich the landscape of discovered AS events.
Key words Gene expression, RNA processing, Alternative splicing, AS event, Definition of alternative
splicing, Transcriptome annotation, RNA-seq, Splicing nomenclature, AS code, Bioinformatics
1 Introduction
Alternative splicing (AS) is the mechanism enabling eukaryotic

cells to define and to regulate the set of introns that is removed
from nascent RNA molecules during transcription. AS plays a key
role in gene expression and transcriptome diversity, and the phe-
nomenon still remains to be fully characterized [1]. One way to
investigate AS is to compare and to characterize the RNA content
of cell samples (e.g., from different experimental conditions, tis-
sues, treatments, etc.) in order to identify variations between tran-
scripts and consequently the underlying AS events. On the way to
achieve efficient and user-friendly tools, computational methods of
379
380 Sylvain Foissac and Michael Sammeth
AS analysis are constantly coping with advances in transcriptomics,

in particular, recent innovations in the field of so-called next-
generation sequencing (NGS) technologies, which require high-
throughput pipelines to analyze large data volumes.
AStalavista (alternative splicing and transcriptional landscape
visualization tool altogether) provides a computational method
and software application for the automatic characterization of AS
landscapes in custom transcriptome data [2]. The method auto-
matically identifies AS events by comparing all transcripts of a given
annotation file and reporting an exhaustive yet nonredundant list
of the resulting variations. To allow for the generic characterization
of—possibly hitherto undiscovered—morphologies of AS events,
so-called AS patterns, AStalavista employs a generic nomenclature,
the alternative splicing code (the “AS code”).
In this chapter, we present the AStalavista method and its
notation to describe AS events and their corresponding patterns.
Employing publicly available data, we demonstrate the simplicity
and universality of the approach for custom gene datasets: we first
provide a concrete example of an AStalavista run on the transcript
annotation of the complete human genome, and then we use the
method to study transcriptome variation depicted by a NGS exper-
iment of the ENCODE project.
2 Materials
2.1 Installing The AStalavista package can be accessed from the website
AStalavista http://astalavista.sammeth.net
where also the source code is available from the “Download” sec-
tion. To install, just unzip/untar the bundle once downloaded, for
instance, by the command
tar xzf AStalavista-3.2.tgz
The command creates a folder named “AStalavista” (with the
corresponding version number) that contains all the files from the
tarball: documentation, license information, and program files.
AStalavista is executed by wrapper scripts located in the “bin”
subfolder within the installation directory, either “astalavista.bat”
(for Windows platforms) or “astalavista” (for UNIX-based operat-
ing systems, including OSX).
The “--help” flag provides a brief description of the software,
its version, and the available options: flags to handle verbosity,
parallelization, output behavior, and available tools (“--list-tools”;
see below).
AStalavista-3.2/bin/AStalavista--elp
AStalavista v3.2 (Flux Library: 1.22)
-------Documentation & Issue Tracker-------
Analysis of Alternative Splicing Events 381
Flux Wiki (Docs): http://sammeth.net/confluence

Flux JIRA (Bugs): http://sammeth.net/jira
Please feel free to create an account in the
public
JIRA and reports any bugs or feature requests.
-------------------------------------------
The Flux library comes with a set of tools.
You can switch tools with the -t option. The
general options
change the behaviour of all the packaged tools.
General Flux Options:
…
List of available tools
…
2.2 Obtaining The full GENCODE transcriptome annotation dataset [3] can be
the GENCODE downloaded from the project’s ftp site by the command:
Annotation wget
ftp://ftp.sanger.ac.uk/pub/gencode/
release_12/gencode.v12.annotation.gtf.gz ./
and subsequently, the downloaded gzip archive is decompressed by
gunzip Gencode.v17.annotation.gtf.gz
2.3 Retrieving In our NGS examples, we employed publicly available RNA-Seq

RNA-Seq Alignments data from the ENCODE project [4] as downloaded from the
CRG’s (Centre for Genomic Regulation, Barcelona) RNA-Seq
Dashboard:
http://genome.crg.es/encode_RNA_dashboard/hg19/
The CRG dashboard allows navigating hundreds of RNA-Seq
experiments produced by the ENCODE consortium to profile the
human transcriptome of different cell types, subcellular compart-
ments, and RNA populations. Employing the STAR mapper [5],
reads of some of the datasets have been split-mapped to the human
genome in order to identify the position of the introns. For our
example, we picked one of these mapping files downloaded from
the URL:
http://genome.crg.es/~jlagarde/encode/pre-DCC/wgEn-
codeCshlLongRnaSeq/20130218_promocell_batches1-2_
minus_batch1sFASTQ/SID38226_SID38227_SJ.bed
The aligned RNA-Seq data has been obtained from long RNAs
that have been extracted from HPAEC (Human Pulmonary Artery
Endothelial Cells) and sequenced on an Illumina HiSeq 2000
machine using a strand-specific protocol. However, the ENCODE
dashboard provides data from many other experimental condi-
tions, which could be combined or compared with this example
dataset.
3 The AStalavista Method
3.1 AS Event Computer-aided comparison of transcriptome data is a natural way

Definition to analyze AS, and many studies have followed that approach [6].
and Nomenclature In order to make conclusions on the effects of AS, the idea is to
identify in the first place common differences in the exon-intron
structure of processed RNA molecules. This can be straightfor-
wardly achieved by comparing the transcript sequences with each
other, either directly or by using their localization within the
context of the corresponding genomic sequence, if available. The
“mapping” of transcripts onto a genomic reference allows to iden-
tify the position of all exons and introns, producing what we call a
transcriptome annotation.
3.1.1 AS Events Given an annotation, each transcript can be uniquely identified by

the type and the genomic position of all of its exonic boundaries.
Here we consider three types of exonic boundaries, we refer to as
“sites”: the transcription start site (TSS) at the 5′ extremity of the
transcript, the cleavage site (CVS) at the 3′-end of the transcript—
often coupled with the polyadenylation site—and splice sites that
delineate introns in the premature transcript; splice sites are fur-
thermore categorized as donor or acceptor sites according to their
localization at the 5′- or at the 3′-end of an intron, respectively.
Alternative splicing occurs when a splice site is used in the splicing
process of one transcript but not of another transcript of the same
locus. An AS event intuitively comprises one or multiple of such
alternative splice sites; moreover, since transcription and splicing
are known to co-occur and interact in eukaryotic cells [7, 8], some
AS events also include alternative TSSs or CVSs in addition to
alternative splice site(s).
More precisely, when comparing the exon boundaries of two
or more transcripts from a genomic locus, the AStalavista approach
first identifies constitutive (splice) sites as exon boundaries employed
by all transcripts that overlap the corresponding genomic position;
consequently, sites that are not used by all transcripts of the locus
are considered as variable. In a further step, AStalavista delineates
AS events as a maximal sequence of consecutive variable sites, with
at least one alternative splice site. Based on this definition, AS
events either contain exclusively alternative splice sites delimited by
common sites at both ends (so-called internal AS events) or they
extend until the transcript extremities and contain variable TSSs
and/or CVSs.
By this generic event definition of AStalavista, resulting events
can describe variations between two or more different transcript
structures. The number of transcripts that are simultaneously
compared naturally plays an important role in the process of assess-
ing whether a certain site is variable or constitutive. For the sake of
convenience, we will consider here pairwise AS events only, but we
Table 1
Examples of distinct AS events with the corresponding AS patterns
Traditional
name Exon-intron structure AS code Example
A Exon 0, 1-2^ PTBP1 polypyrimidine
skipping tract-binding protein 1,
ninth exon
B Alternative 1-, 2- PTBP1 polypyrimidine

acceptor tract-binding protein 1,
ninth exon
C 0, 1-2^3-4^ FBLIM1 filamin-binding

LIM protein 1,
fourth + fifth exon
D 1^4-, 2^3- ZWINT ZW10 interacting

kinetochore protein,
sixth − seventh exon
would like to point interested readers to the concept of complete

alternative splicing events for a more advanced characterization of
AS loci [9].
Table 1 shows some AS patterns along with examples as
they are observed in the GENCODE annotation of the human
transcriptome. The alternative exon 9 in the PTBP1 (polypyrimi-
dine tract binding protein 1) gene has been reported to be entirely
skipped (“exon skipping,” A) or included with different acceptor
flanks (“alternative acceptor,” B). One of the multiple possibilities
to process the pre-mRNA of the FBLIM1 (Filamin-binding LIM
protein 1) gene (C) excludes simultaneously the subsequent exons
4 (188 nt) and 5 (103 nt), such that overall the reading frame is
not disturbed (292 nt corresponds to 97aa). (D) Similarly, varia-
tions on either side of the sixth intron in the ZWINT locus regu-
late the optional inclusion of 47 amino acids into the resulting
protein. As we will show in the next section, the complete set of AS
events can be automatically identified and reported by the
AStalavista software [10].
3.1.2 The AS Code AS events are commonly classified in different categories based on
the observed differences in the exon-intron structure between
the respective transcripts. Typically, most studies employ the terms
“exon skipping,” “alternative donor/acceptor,” “intron retention,”

and (sometimes) “mutually exclusive exons” to describe AS patterns.
However, a nomenclature built exclusively by this set of predefined
terms allows to describe only a couple of AS configurations, which
due to their simplicity are observed quite frequently but constitute
only a small fraction of the spectrum of all event patterns found in
complete transcriptome annotations. An exhaustive characterization
of the AS landscape from arbitrary transcript structures requires a
more flexible notation that permits to identify all events that share
the same “topography.”
For this reason, AStalavista uses a generic notation system that
can univocally describe any AS event. This nomenclature assigns an
“AS code” to every sequence of alternative splice sites, such that
events with qualitatively identical variations in their exon-intron
structure receive the same AS code. This AS code represents each
variable site by a number and a symbol, according to its relative
position within the event and to its type. More precisely, the code
of a certain AS event is generated as follows: first, all variable sites
that are included in the event are sorted according to their position
in the directionality of transcription, from 5′ to 3′. Each site is
assigned a number according to this order (1, 2, 3, etc.) and a
symbol based on its type: “[“for TSS, ”]” for CVS, “^” for donor,
and “-” for acceptor. Then, sites from the same transcript are
represented next to each other by consecutive pairs of such number-
symbol tuples (in the corresponding 5′ → 3′ order), forming a
string that represents the AS event. In the absence of variable sites
in one of the transcripts of the event (e.g., to describe the string for
a transcript skipping an exon), the digit “0” is employed as a cor-
responding string. The AS code then is a comma-separated concat-
enation of thus obtained strings, one from each transcript variant,
ordered by their numbers. Some examples for AS events along
with their corresponding AS code are shown in Table 1.
3.2 Processing The ENCODE project intends to establish a repertoire of func-

AS Events tional elements in the human genome [11]. To this aim, a tremen-
from Transcriptome dous amount of experimental data has been (is being) produced
Annotations: and analyzed by hundreds of scientists worldwide. As part of the
The GENCODE Example results from this effort, the GENCODE annotation provides the
community with an expertized and thoroughly curated set of genes
and transcript positions in the genomic sequence. The GENCODE
annotation describes the exon-intron structure of thousands of
alternative transcripts and therefore allows us to profile AS exten-
sively and accurately. Therefore, as a first example for the applica-
tion, we will employ AStalavista to characterize the AS landscape of
the entire human transcriptome as described by the GENCODE
annotation.
3.2.1 Running The AStalavista program package is subdivided in different tools; a

AStalavista list of available tools can be obtained by requesting help via the
on the GENCODE corresponding command option
Annotation
./bin/AStalavista –list-tools
which by the current version (AStalavista-3.2) outputs the
summary:
List of available tools
scorer - Splice site scorer
sortBED - Sort a BED file. If no output
file is specified, result is printed to standard out
sortGTF - Sort a GTF file. If no output file
is specified, result is printed to standard out
asta - AStalavista event retriever
subsetter - Extract a random subset of lines
from a file
The tool “asta” performs the extraction of alternative splicing,
for which additional parameters and their default values in turn are
obtained by
./bin/AStalavista -t asta -h
Parameters can either be provided on the command line or in
a separate parameter file (specified by flag “-p”). A mandatory
parameter is the path to the file with the annotation that is to be
analyzed (parameter “IN_FILE”, or command line parameter
“-i”). As an example, an AStalavista run on the v12 of the
GENCODE transcript collection is executed by
./bin/AStalavista -t asta -i Gencode.
v12. annotation.gtf
AStalavista employs all exons (i.e., lines identified by the fea-
ture “exon” in the third column) from the gtf file, which are to
be sorted by their genomic positions and require a “transcript_id”
identifier for correct gene clustering. If the file is not sorted,
AStalavista will automatically sort the file before processing.
As a default setting, AStalavista reports only alternative splicing
events [10] that are not linked to variable transcription initiation
and/or cleavage sites, so-called internal AS events (ASI). External
AS events (ASE) that include TSSs and/or CVSs can also be
included in the list of reported event types (value ASE for the
parameter EVENTS or command line flag “-e”):
v12.annotation.gtf -e ASI,ASE
Per default, AStalavista further on shows exclusively pairwise
AS events, i.e., events with exactly two variants (dimension 2);
events of a higher dimension are projected to events between all
pairs of their variants. To change the order of dimension for the

reported AS events, parameter EVENTS_DIMENSION (flag -d)
has to be set to a corresponding integer value or “-1” for output-
ting the so-called complete AS events [9]:
v12.annotation.gtf -d -1
3.2.2 Interpreting AStalavista produces from the input annotation in gtf format a file
the Results with events that are equally stored as gtf features: start (column 4)
and end (column 5) of the event are the genomic coordinates of
the delimiting common sites [10], respectively, the first/last vari-
able site in the case of ASE events; the nature of these delimiting
sites is detailed in the “flanks” attribute, which adopts the same
“AS code” notation as used for “splice chain” and “structure”
(see below):
chr1 Gencode_v12 as_event
12227 12721 . + . tran-
script_id "ENST00000518655.2,ENST00000456328.2/
ENST00000515242.2"; gene_id "chr1:11869-14412W";
flanks "12227^,12721^"; structure "1-,2-"; splice_
chain "12595-,12613-"; dimension "2_2"; degree "4";
The attribute “degree” summarizes the number of variable
sites in the event, which is naturally increasing with longer events
or events of more variants. Additionally, the attribute “dimension”
provides information about the reported event with respect to the
underlying complete event: the dimensionality is denoted as a
string of the form “X_Y” where X is the number of variants in the
reported event and Y is the number of events in the corresponding
complete event. All events with X = Y are complete, and those with
X < Y are not (X > Y is not possible by definition).
The “transcript_id” attribute describes a comma-separated list
of the transcript identifiers for each of the variant structures of the
event; if there is more than one transcript supporting a certain vari-
ant of the event, the corresponding identifiers are furthermore
separated by a “/” separator. The attribute “splice_chain” describes
the genomic coordinates of each variable site in the event, with the
variants in the same comma-separated order as the identifiers of
“transcript_id”.
As introduced previously, the AS code nomenclature specifies
after the genomic coordinate of each site also the site type by a
certain symbol: “^” denotes a splice donor, “-” a splice acceptor,
and “[“a transcription start and”]” a cleavage site. Together with
the name of the chromosome/contig, the value of “splice_chain”
can be considered as a unique identifier of the event, because it
describes every event univocally by the morphology of its sites as
localized by corresponding genomic coordinates. Therefore, also
events delimited by the same flanks (start/end tuples) can still
differ by their splice chains.
1-2^3-,4- 1-2^4-,3-
1,621 1,475
33
6,5 0,1-2^3-4^
3
The
9,286 AS Code
Landscape
(Gencode v12)
1-,2-
0,1-2^
10
,2 1^,2^
75
0,1^2-
322
34,
6,06
1
5,552
Fig. 1 AS landscape of internal events of the Gencode (v12) annotation
By replacing genomic coordinates reported for the “splice_chain”

by relative positions of the corresponding sites within the event,
the attribute “structure” abstracts events to their respective pat-
terns that are named by the corresponding AS code; examples for
AS codes of common events are: “0,1-2^” for exon skipping,
“1^,2^” for alternative donors, “1-,2-” for alternative acceptors,
“'0,1^2-” for intron retention, and “'1-2^,3-4^” for mutually
exclusive exons.
Figure 1 depicts all internal AS events found in the Gencode
(v12) transcriptome annotation, grouped by their respective AS
code. Altogether, the landscape of different AS morphologies
reveals that traditional events (blue, yellow, green, and red slices)
in our days hardly make up 37 % of the spectrum spanned by all AS
patterns—the majority of events (gray slice) cluster into 813 differ-
ent, more complex events. Note that, depending on the position of
the event, it might further be possible to introduce alternative
TSSs or CVSs in several AS shapes to further amplify the landscape
(external AS events).
The result of a default AStalavista runs as shown in Fig. 1, for
example, GENCODE annotation retrieves 92,743 internal AS
events (ASI) that cluster in 817 different patterns (respectively
670,104 events clustering in 17,741 patterns when including
ASE events). Expectedly, the most abundant event patterns are of
degree 2 which corresponds to the minimum number of variable
sites required for an AS event: “0,1-2^” (exon skipping) ranks first,

“1-,2-” (alt. acceptors) ranks second, “1^,2^” (alternate donors)
ranks third, and “0,1^2-” (intron retention) ranks fourth most
abundant pattern. The reasons for the differences between these
patterns lie in splicing mechanism and constraints of coding
sequences, which constitute the main part of messenger RNA
transcripts [10].
Although the event complexity specified by “degree” usually
anticorrelates with the observed abundance, there are exceptions.
For instance, the skipping of multiple exons is mechanistically facil-
itated and therefore more often observed than other events of the
corresponding degree: “0,1-2^3-4^” ranks before “1^,2^” and
“0,1^2-”, “0,1-2^3-4^5-6^” and also before many other patterns
of degree 4, etc. Moreover, only ~47 % of pairwise events found on
the GENCODE annotation are complete; from another point
of view, that means that the true complexity of more than half of
the events is underestimated when looking at them exclusively by
pairwise variant comparisons.
3.3 Discovery AStalavista can also analyze NGS data from RNA sequencing
of Novel AS Events by experiments (RNA-Seq). The main conceptual difference between
RNA-Seq Experiments transcriptome annotations, like GENCODE and RNA-Seq data, is
that the latter only provide partial information on the transcrip-
3.3.1 Employing
tomes: RNA-Seq reads are limited to parts of expressed genes,
AStalavista for NGS Data
because current NGS technologies still cannot reproduce sequences
of entire RNA molecules and transcripts thus are fragmented
before sequencing. Although RNA-Seq reads correspond to frag-
ments and cannot be considered as full-length transcripts, they
can provide valuable information about AS by potentially novel
splice junctions they harbor: if the exon-exon junction of a spliced
transcript is captured in a read, the position of the corresponding
intron can be revealed by aligning the read sequence on the
genomic reference.
Dedicated NGS alignment methods—as implemented in soft-
ware like GEM [12] or STAR [5]—allow to generate such spliced
alignments where introns correspond to gaps flanked by exonic
regions. These experimentally derived evidences about splicing
events can be analyzed by AStalavista, either separately or in com-
bination with a reference annotation, to identify novel alternative
splicing events that are not captured by the reference gene set.
An interesting and original aspect of the latter approach is that no
transcript assembly is required for the analysis; indeed, AStalavista
considers reads as independent pieces of information and does not
take any assumptions which of them might originate from the same
and which of them have been obtained from independent RNA
molecules, respectively. In the next section, we employ this strat-
egy to illustrate how AStalavista can enrich a repertoire of AS
events by processing RNA-Seq data.
3.3.2 Processing We converted the format of the bed file with mapped reads
RNA-Seq Data downloaded earlier from the CRG dashboard, and produced an
AStalavista-compatible gtf file by the commands:
bed = SID38226_SID38227_SJ.bed; gtf = ${bed%.
bed}.gtf; cat $bed | awk -v bed = $bed -v OFS = "\t"
'{print $1,bed,"exon",$2,$2,$5,$6,".","gene_id
\"SJ"NR"\"; transcript_id \"SJ"NR"\";";print
$1,bed,"exon",$3,$3,$5,$6,".","gene_id
\"SJ"NR"\"; transcript_id \"SJ"NR"\";"}' > $gtf
Finally, we concatenated thus obtained gtf representation of
the NGS mappings with the GENCODE annotation from our
previous example:
cat Gencode.v12.annotation.gtf $gtf >
Gencode_rnaseq.gtf
The joint dataset then was processed with AStalavista in order
to compare correspondingly obtained AS events with the results
derived from the GENCODE annotation alone:
./bin/AStalavista -t asta -i Gencode_
rnaseq.gtf
3.3.3 Identifying Novel After processing the NGS-enriched GENCODE annotation with
AS Events AStalavista, we characterized events involving RNA-Seq reads that
were not detected using the annotation alone. We found 342 such
novel internal AS events (ASI). Figure 2 shows that the distribution
of patterns among these additionally found events follows closely the
one described by GENCODE events, albeit some rank positions are
permuted: evidence by RNA-Seq split mappings indicates that among
the simple AS patterns, alternative exon boundaries (89 donors and
79 acceptors events) are most underestimated by the current
GENCODE annotation, followed by alternatively skipped exons
(53 novel alternative exons); furthermore, RNA-Seq data pinpoints
177 novel complex events of 34 distinct patterns (Fig. 2).
Figure 2 also summarizes the number and nature of AS events
that are revealed by RNA-Seq introns; many describe the skipping of
(multiple) exons: 53 “0,1-2^” (1 exon), 21 “0,1-2^3-4^”
(2 exons), 11 “0,1-2^3-4^5-6^” (3 exons), etc. Albeit our analysis
focuses on a small set of high-confidence events, we observe nearly
half of the exons that predicted to be skipped (20 out of 53) exhibit
a frame-preserving length, i.e., the size of the exon is a multiple of 3.
These events include, for instance, the third exon in the TMCO1
(transmembrane domain protein) locus, where also additional EST
(Expressed Sequenced Tags) evidence (BP308568) and comple-
mentary computational gene models (ENST00000476173) support
the skipping of the corresponding 60 nt exon; as another example,
we find NGS evidence for the skipping of the 36 nt long exon num-
ber 10 in the TRIP (thyroid hormone receptor interacting protein)
gene, which so far has not been discovered by EST evidence.
1-2^3-,4-
0,1-2^3-4^
Novel Events
(RNA-Seq) 10
21
17
7
1-,2-
79
0,1-2^
1^,2^ 0,1^2-
53
5
86
Fig. 2 Distribution of novel AS events found by additional RNA-Seq evidence
As in other analyses that employ NGS data, potential artifacts

might impair the results, and some of the events might originate
from errors in the sequencing or the alignment process. As in the
case of reference annotations, AStalavista processes the exons and
introns provided by NGS data and all variations in the exon-intron
structure are reported. Reassuringly, however, we find among
these novel events several traces of constraints created by coding
sequences, for instance, a nonnegligible proportion of exons that
are predicted to be skipped by RNA-Seq evidence would not break
the coding frame. Also, there is independent evidence from cDNA
and EST data that supports some of the NGS-derived events.
Finally, it is worth noting that most of the NGS evidence for exon
skipping predicts alternative exons longer than 100 bp, which are
unlikely to originate from misalignments or gaps in the alignment
process.
Certainly, those AS events can be subject to further ranking
and/or filtering procedures, for instance, by considering their
overlap with the annotated coding sequence, by taking the number
of reads supporting each of the splice junctions in the event and
their corresponding alignment quality (e.g., gaps) into account,
etc. However, the example demonstrates that already a straight-
forward application of AStalavista to different RNA-Seq datasets
allows to analyze cell-type or condition-specific transcriptomes, as

by adding the split-mapped reads to a reference annotation and
comparing the correspondingly obtained AS events for novel
configurations.
An interesting specificity of the approach is that—unlike most
of other RNA-Seq data processing tools—AStalavista does not
attempt to assemble the reads into longer contigs nor even into
transcribed fragments of the genomic space (the so-called trans-
frags). Although transfrag-based methods have been producing
better results in terms of accuracy along the years, transcriptome
reconstruction from RNA-Seq data is still a highly challenging task
in the field [3]. Whether methods use a genomic reference, as in
our example, or whether they rather compare reads with each other
(“de novo” assembly), the risk of merging some unrelated reads to
a common contig cannot be completely avoided [13], generating
some artificial hybrid that does not exist in vivo. Moreover, tran-
scriptome assembly methods are prone to errors when dealing with
repeated sequences and/or incomplete data, thus processing RNA-
Seq data by AStalavista right after the read mapping step improves
the general reliability of the analysis. Under the assumption that
genomic read alignments are usually a more reliable source of
information than transcript structures reconstructed upon them,
AStalavista is not concluding that reads with a common subse-
quence originate from identical transcripts but rather focuses on
splicing differences they (might) harbor: by the atomary definition
of delineating variations strictly within the common genomic space
ensures that any of the reported AS events is supported by two real
sequences.
Another advantage of the AStalavista approach is its efficiency
in terms of processing time. The example shown in this chapter
took about 5 min on a 2 GHz Intel system, occupying <1 GB of
RAM, which makes it suitable for high-throughput processing.
The rather simplistic examples we provide here can straightfor-
wardly be extended to the application of AStalavista for processing
additional datasets, e.g., for comparing different reference annota-
tions, transcriptomes from different cell compartments, and/or
different experimental conditions or data preprocessing strategies.
References
1. Kornblihtt AR, Schor IE, Alló M (2013) 3. Harrow J, Frankish A, Gonzalez JM (2012)
Alternative splicing: a pivotal step between GENCODE: the reference human genome
eukaryotic transcription and translation. Nat annotation for The ENCODE Project. Genome
Rev Mol Cell Biol 14:153–165 Res 22:1760
2. Foissac S, Sammeth M (2007) ASTALAVISTA: 4. Dunham I, Birney E, Lajoie BR, Sanyal A
dynamic and flexible analysis of alternative (2012) An integrated encyclopedia of DNA
splicing events in custom gene datasets. Nucleic elements in the human genome. Nature
Acids Res 35:W297 489:57
5. Dobin A, Davis CA, Schlesinger F (2013) 9. Sammeth M (2009) Complete alternative

STAR: ultrafast universal RNA-seq aligner. splicing events are bubbles in splicing graphs.
Bioinformatics 29:15 J Comput Biol 16:1117
6. Mironov AA, Fickett JW, Gelfand MS (1999) 10. Sammeth M, Foissac S, Guigó R (2008) A gen-
Frequent alternative splicing of human genes. eral definition and nomenclature for alternative
Genome Res 9:1288–1293. doi:10.1101/gr. splicing events. PLoS Comput Biol 4:e1000147.
9.12.1288 doi:10.1371/journal.pcbi.1000147
7. Tilgner H, Knowles DG, Johnson R (2012) 11. Djebali S, Davis CA, Merkel A, Dobin A,
Deep sequencing of subcellular RNA frac- Lassmann T, Mortazavi A, Tanzer A et al (2012)
tions shows splicing to be predominantly Landscape of transcription in human cells. Nature
co-transcriptional in the human genome 489:101–108. doi:10.1038/nature11233
but inefficient for lncRNAs. Genome Res 12. Marco-Sola S, Sammeth M, Guigó R, Ribeca P
22:1616 (2012) The GEM mapper: fast, accurate and
8. Boireau S, Maiuri P, Basyuk E, de la Mata M, versatile alignment by filtration. Nat Methods
Knezevich A, Pradet-Balade B, Bäcker V, 9:1185. doi:10.1038/nmeth.2221
Kornblihtt A, Marcello A, Bertrand E (2007) 13. Mundry M, Bornberg-Bauer E (2012) Evalu-
The transcriptional cycle of HIV-1 in real-time ating characteristics of de novo assembly soft-
and live cells. J Cell Biol 179:291–304. ware on 454 transcriptome data: a simulation
doi:10.1083/jcb.200706018 approach. PLoS One 7:e31410
Chapter 25
Computational Design of Artificial RNA Molecules

for Gene Regulation
Alessandro Laganà, Dario Veneziano, Francesco Russo,
Alfredo Pulvirenti, Rosalba Giugno, Carlo Maria Croce, and Alfredo Ferro
Abstract
RNA interference (RNAi) is a powerful tool for the regulation of gene expression. Small exogenous
noncoding RNAs (ncRNAs) such as siRNA and shRNA are the active silencing agents, intended to target
and cleave complementary mRNAs in a specific way. They are widely and successfully employed in func-
tional studies, and several ongoing and already completed siRNA-based clinical trials suggest encouraging
results in the regulation of overexpressed genes in disease.
siRNAs share many aspects of their biogenesis and function with miRNAs, small ncRNA molecules
transcribed from endogenous genes which are able to repress the expression of target mRNAs by either
inhibiting their translation or promoting their degradation. Although siRNA and artificial miRNA mole-
cules can significantly reduce the expression of overexpressed target genes, cancer and other diseases can
also be triggered or sustained by upregulated miRNAs.
Thus, in the past recent years, molecular tools for miRNA silencing, such as antagomiRs and miRNA
sponges, have been developed. These molecules have shown their efficacy in the derepression of genes
downregulated by overexpressed miRNAs. In particular, while a single antagomiR is able to inhibit a single
complementary miRNA, an artificial sponge construct usually contains one or more binding sites for one
or more miRNAs and functions by competing with the natural targets of these miRNAs. As a consequence,
natural miRNA targets are reexpressed at their physiological level.
In this chapter we review the most successful methods for the computational design of siRNAs,
antagomiRs, and miRNA sponges and describe the most popular tools that implement them.
Key words RNAi, siRNA, shRNA, antagomiR, miRNA, Sponge, Gene expression
1 Introduction
The discovery of the first short ncRNA capable of acting as endog-

enous regulator of gene expression was made by Ambros et al. in
1993, while investigating the role of lin-4 in the developmental tim-
ing of C. elegans [1]. That accomplishment revealed only a glimpse
of the much broader reality uncovered by Fire and Mello 5 years
later when they reported the capability of exogenous double-
stranded RNAs to silence genes in a specific manner, disclosing the
393
394 Alessandro Laganà et al.
mechanism we know today as RNA interference (RNAi) [2]. In

1999, a similar process was discovered to take place in plants, as
short RNA sequences (~20–25 nt) were found to be able to bind
their mRNA targets through perfect base complementarity [3].
Since then, a revolution has occurred in the field of RNA biology,
thanks to the characterization of RNAi as an innate biological pro-
cess through which the expression of specifically targeted genes can
be modulated and/or silenced, ushering in a new world of research
and potential therapeutic applications. Specifically, RNAi is a natu-
rally occurring mechanism resulting in a sequence-specific, post-
transcriptional downregulation of gene expression induced by
double-stranded RNA (dsRNA) homologous to the target gene.
This regulation can occur at different levels of the gene expression
process, including transcription, mRNA processing, and translation.
The small RNA molecules responsible for RNAi originate from
endogenous and exogenous double-stranded precursors and play a
specific role in guiding effector protein complexes toward their
nucleic acid targets by partial or full complementarity bonds [4].
Although several classes of regulatory ncRNAs have been iden-
tified, the most relevant ones can essentially be represented by two
groups, according to their origin, structure, associated effector
proteins, and function: small interfering RNAs (siRNAs), princi-
pally of exogenous origin in animals, and microRNAs (miRNAs),
which are endogenous genome products. These molecules, appar-
ently present only in eukaryotes and in some viruses, are in fact the
most abundant regulatory molecules in terms of both phylogeny
and physiology.
siRNA is a class of double-stranded RNA molecules, 20–25
base pairs in length, whose role appears to be involved in the
defense of the cell and maintenance of genome integrity through
the silencing of exogenous nucleic acids and undesired transcripts
(such as transposons and repetitive elements) [4, 5]. Although
their posttranscriptional gene-silencing role in plants and in some
simple animal species was shown to be of endogenous origin, siR-
NAs in animals, instead, are mainly exogenous molecules obtained
from perfectly complementary long double-stranded precursors
coming from viruses and transposons [3]. As it was first demon-
strated in nematodes [2] and later in mammalians [6], siRNA-
mediated RNAi can be obtained either by simple dsRNAs of about
21 nucleotides (nt) with two-nucleotide 3′ overhang or by stably
expressed short hairpin RNAs (shRNAs), which are processed into
siRNAs [7, 8]. Compared to unprocessed siRNA, this last approach
undoubtedly presents many advantages, such as long-lasting silenc-
ing effects, cost-effectiveness, as well as easy delivery methods.
Generally, shRNA is transcribed in cells from a DNA template as a
single-stranded RNA molecule (about 50–100 bases). The com-
plementary regions are spaced by a hairpin loop, therefore the
name “short hairpin” RNA [9]. Intracellular presence of siRNAs
Computational Design of RNAi Molecules 395
which are perfectly complementary to their target mRNAs is of

crucial importance for the induction of RNAi and leads to mRNA
degradation.
While siRNAs are mainly exogenous molecules, miRNAs
instead are a class of RNAi inducers which derive from partially
complementary double-stranded hairpin precursors of endogenous
origin. Once processed, they are small single-stranded RNAs (20–
22 nt long) able to modulate posttranscriptional gene silencing
through repression, and at times degradation, of specific mRNA
target molecules [10]. It has been estimated that miRNA-coding
genes represent 1 % of the total gene population, being the biggest
class of regulatory molecules. They are present in plants, higher
eukaryotes, and in some viruses. miRNAs are often encoded in
clusters by genes usually located in introns and, more rarely, in
exons of protein-coding genes, as well as in intergenic regions
[10]. RNA polymerases produce primary miRNA transcripts (pri-
miRNAs) from microRNA genes [11–13]. pri-miRNAs are
approximately >100 nt long and are subsequently processed into
~70-nt precursor miRNAs (pre-miRNAs) by a microprocessor
complex comprised of the enzyme Drosha and a subunit DCGR8
[14–16]. Pre-miRNAs are then exported into the cytoplasm by the
exportin 5 protein [17, 18]. From here on, in spite of their origi-
nating difference, common features, such as the length of their
mature products and their sequence-specific inhibitory functions,
suggest that siRNA and miRNA have similar processing and com-
mon mechanisms. Their double-stranded precursors (shRNAs and
pre-miRNAs) are indeed both cleaved by Dicer, a ribonuclease III-
type protein [9, 19–21], into short 19–21 duplexes having two
symmetric nucleotide overhangs at the 3′ end and a 5′ phosphate
along with a 3′ hydroxyl group. After cleavage, Dicer, together
with protein complexes TRBP and PACT, loads the obtained
duplexes into a nuclease-containing multiprotein complex referred
to as the RNA-induced silencing complex (RISC). Once the duplex
is loaded, the Ago2 protein component of the RISC then cleaves
one of the two strands of the duplex, which is thus, by convention,
considered the sense “passenger” strand. The antisense strand,
which remains loaded into the thus activated RISC, is instead
called “guide” since it acts as an adapter for the complex to mRNA
targets and allows it to carry out the RNAi mechanism [22–26]. In
miRNAs, the strand which most commonly plays the role of the
guide is called “mature miRNA,” while the other one is called
“miRNA*” [27]. Animal 3′ UTR sequences often present miRNA
binding sites in multiple copies, while, conversely, most miRNAs in
plants, as well as siRNAs, bind their targets in their coding regions
with perfect complementarity. Another fundamental difference
between miRNAs and siRNAs consists, in fact, in the type of bind-
ing, which is considered a key factor in their regulatory function:
miRNAs bind their target with partial complementarity, allowing
bulges and loops in duplexes. However, a key feature in their target

recognition is represented by the perfect base pairing with the tar-
get in positions 2–8 of the miRNA guide, which is known as the
seed region. The presence of mismatches in the central part of the
duplex is usually associated to translational repression, which seems
to be the default mechanism of miRNA-mediated RNAi. The
cleavage of perfectly paired duplexes, which is the default RNAi
mechanism in the case of siRNAs, is instead considered for miR-
NAs an additional feature leading to the same effect on the protein
level.
The biological role of these molecules is currently being inten-
sively elucidated, and their involvement in fundamental processes,
such as apoptosis, metabolism, cell proliferation, and organism
development, has been widely demonstrated.
Due to the fast and cost-effective way of disrupting genes’
functions provided by RNAi-based gene knockdown techniques,
great and rapid progress has been made in recent years, and siRNAs
have become a standard tool routinely used in molecular genetics
and functional genomics laboratory [28]. Moreover, a large num-
ber of RNAi-based potential therapeutic agents are actively being
explored, while several RNAi-based therapies against several dis-
eases such as viral infections, inflammatory diseases, and cancers
have already reached preclinical and even clinical stage in develop-
ment [29–31].
In light of all this, great interest has thus eventually arisen in
loss-of-function studies in vivo in order to further investigate the
precise molecular function of miRNAs in mammals, which is still
unknown on the greater part. Thus in 2005, Krutzfeldt et al.
devised a novel class of chemically engineered oligonucleotides
able to silence endogenous miRNAs in vivo, which they termed
“antagomiRs” [32]. These new molecules were also shown to per-
form efficient and stable loss-of-function phenotypes for specific
miRNAs by lentivirus-mediated delivery in cultured cells [33]. In
2007, other miRNA inhibitors, called “miRNA sponges,” were
devised [34]. They consisted in transcripts expressed from strong
promoters and essentially containing multiple, tandem binding
sites to an miRNA of interest. Once vectors encoding these com-
petitive miRNA inhibitors were transiently transfected into cul-
tured cells, miRNA targets were shown to be derepressed just as
strongly as previously accomplished, thus suggesting a valid alter-
native to antagomiRs.
These results clearly show that these molecular constructs may
represent a valid strategy for silencing miRNA in diseases, such as
cancer, and further investigate potential therapeutic applications.
In this chapter we will give an overview of the most successful
approaches and algorithms regarding RNAi design techniques and
briefly describe the most popular tools that implement them.
2 Materials
The methods described in this chapter are implemented in tools

publicly available online. They can be executed on any personal
computer equipped with Internet connection and a browser and
don’t require any particular resource.
3 Methods
RNAi represents today a well-established approach for gene silenc-

ing in loss-of-function studies and genetic screens. Since its discov-
ery in 1998, the main focuses of RNAi computational research
have been the discovery of siRNA design rules and the develop-
ment of siRNA efficacy and specificity prediction models.
Nevertheless, a considerable number of siRNA design tools and
siRNA databases are freely available online and widely employed
for gene knockdown experiments.
The other side of artificial RNAi is represented by molecular
tools for miRNA silencing, such as antagomiRs and sponges.
In this section we summarize the main rules for the design of
siRNA, antagomiRs, and sponges and provide a brief introduction
to several tools and databases which are freely available online.
3.1 Design From a computational point of view, siRNA design is the process
of Functional siRNA of choosing a functional binding site on a target mRNA sequence,
which will correspond to the sense strand of the siRNA under
design (typically 21–23 nt long) [35]. The siRNA antisense
sequence is obtained as the complement to the sense strand.
Symmetric 3′ overhangs, usually dTdT, are added to improve sta-
bility of the duplex and to facilitate RISC loading, ensuring equal
ratios of sense and antisense strands incorporation [6, 36, 37].
Other overhang sequences are acceptable, but some combinations,
such as GG, should be avoided. The efficacy of RNAi is mostly
determined by sequence-specific factors which affect the stability
of the duplex ends. siRNA duplexes often have asymmetric loading
of the antisense versus sense strands [38, 39]. The strand whose 5′
end is thermodynamically less stable is preferentially incorporated
into the RISC.
Elbashir et al. suggest to choose the 23-nt sequence motif
AA(N19)TT as binding site (N19 means any combination of 19
nucleotides), where (N19)TT corresponds to the sense strand of
the siRNA, while the complement to AA(N19) corresponds to the
antisense strand (see Fig. 1a) [6].
Many different features associated to functional siRNAs have
been identified in the past years and some of them are now widely
accepted as standard rules in siRNA design and are implemented in
Fig. 1 siRNA design rules. (a) An example of target region in an mRNA sequence and the corresponding siRNA
duplex with 3′ overhangs. (b) Specific positional rules for siRNA design. The darker cells represent positions
on the siRNA antisense (AS) and sense (S) strands. The light gray cells contain specific rules for the corre-
sponding positions. For each set of rules, references are given (Ref). The striped gray background indicates
inconsistencies of the rules, due to the different experiments that they come from
the majority of design tools. They can be classified into four differ-
ent categories: (1) general binding rules, (2) nucleotide composi-
tion rules, (3) specific positional rules, and (4) thermodynamics
rules. Table 1 summarizes categories (1), (2), and (4), while rules
in category (3) are represented in Fig. 1b.
General binding rules refer to factors such as the position of
the binding sites in the target transcript. For example, the target
region should preferably be between 50 and 100 nt downstream of
the start codon, and the middle of the coding sequence should be
avoided. Another rule suggests pooling of four or five siRNA
duplexes per target gene, in order to ensure a stronger repression.
Another class of rules concerns the siRNA nucleotide compo-
sition. A major feature, implemented by every design tool, is the
G/C content, which should typically be in the range of 30–55 %,
although values as low as 25 % or as high as 79 % are still associated
to functional siRNAs.
Other features in this category include the presence/absence
of particular motifs in the antisense strand and the absence of
internal repeats.
Specific positional rules are the most numerous and regard the
selection of nucleotides to prefer or avoid in specific positions of
either the sense or the antisense strand of the duplex. For example,
Table 1
Rules for siRNA design
Design rule References

General targeting rules
Select the target region preferably 50–100-nt downstream of the start codon [21]
Avoid to target the middle of the coding sequence of the target gene [64]
Pooling of four or five siRNA duplexes per gene [64]
Nucleotide composition rules
Antisense strand with higher information content [65]
Absence of any GC stretch >9 nt long [66]
At least five A/U residues in the 5′ terminal one-third of the antisense strand [49, 66, 67]
A higher “A/U” content in the 3′ end than that in the 5′ end (sense strand) [68]
G/C content ranges: 32–58 %, 30–52 %, 32–79 %, 36–53 %, 35–73 %, 25–55 % [21, 41, 68–71]
Absence of internal repeats [41]
Presence of motifs “AAC,” “UC,” “UG,” “AAG,” “AGC,” “UCU,” “UCCG,” [49, 70–72]
“CUU,” “CU,” “GUU,” “UCC,” “CG,” “AUC,” “GCG,” “UUU,” “ACA,”
“UUC,” “CAA” in antisense strand
Avoid motifs “CUU,” “CUA,” “GUU,” “GU,” “GAU,” “ACGA,” “GCC,” [49, 70–72]
“GUGG,” “CCC,” “GGC,” “CCG,” “GGG,” “CAG,” “GAG,” “GCA,”
“AUA,” “CUG,” “AG,” “GG,” “GGA” in antisense strand
High content of “U” in antisense strand [72]
Low content of “G” in antisense strand [72]
Thermodynamics rules
Total hairpin energy <1 [69]
Antisense 5′ end binding energy <9 [69]
Sense 5′ end binding energy in range 5–9 [69]
Middle binding energy <13 [69]
Energy difference <0 [69]
Energy difference within −1 and 0 [69]
Significant ΔG difference between positions 1 and 18 [67]
High ΔG in positions 1–4, 5–8, and 13–14 in the antisense strand [70]
Low ΔG in positions 18–19 in the antisense strand [70]
Avoid folding of siRNA [70]
The rules are classified in four different categories, three of which are summarized in this table (see also Fig. 1b).
For each rule, or group of rules, references are given
the antisense sequence should always have an A/U base at its 5′

end. This is associated to a weaker thermodynamic stability which
facilitates incorporation of the strand into the RISC, as already
discussed above. Other rules suggest to avoid G/C nucleotides at
the 5′ end of the sense strand or to have either a G or a C in posi-
tion 4 of the sense strand.
Finally, thermodynamics rules refer to the global or local
energy of the duplex and are partly related to the choice of nucleo-
tides in specific positions such as the 5′ end of the antisense strand,
as already mentioned, or in other regions such as the middle of the
duplex. Other thermodynamic features associated to siRNA effi-
cacy include the structural accessibility of the target site and the
fact that folding of siRNAs should be avoided.
These and other rules are implemented in a considerable num-
ber of siRNA design tools available online, which will be described
in the next subsection. One thing that is worth mention is that all
the rules were obtained from distinct experiments performed in dif-
ferent conditions by different labs; thus inconsistencies and contra-
dictions are inevitable, such as the one highlighted in Fig. 1b. This
is one major drawback as discussed at the end of next section.
3.2 siRNA Design Table 2 provides a list of tools for the automated design of siRNA
Online Resources and shRNA sequences. Most tools have user-friendly interfaces
which don’t require any additional specifications aside from the
target sequence.
OptiRNAi 2.0 is a fast tool which predicts 21–23-nt RNAi
target sites on a user-provided sequence using the criteria described
by Elbashir et al. in 2001 and Reynolds et al. in 2004 [36, 40, 41].
The program generates a list of up to ten siRNA target sites for
each of which a score indicates how well it matches the considered
features. The tool doesn’t return the actual siRNA antisense
sequence, which has to be manually derived as the reverse comple-
ment of the binding sites.
As for OptiRNAi, siDirect 2 also accepts just the target
sequence as input. It returns a table with a list of potential binding
sites (including the 2-nt overhang) [42]. For each site, the corre-
sponding siRNA duplex strands are given, together with the melt-
ing temperature (Tm) of the seed-target duplex, as a measure of
thermodynamic stability. The seed-target duplex is formed between
the region 2–8 of the siRNA guide strand (from the 5′ end) and its
target mRNA site.
Other details provided include the list of potential off-target
genes for the guide and passenger strands and a graphical view of
the siRNA binding sites in the target sequence.
siRNA scales are another design tool which accepts as input a
target sequence and returns a list with all possible 19-nt long
siRNA sequences and the predicted percentage of target mRNA
copies present in the cells after siRNA-directed cleavage as a
Table 2
Tools for siRNA design
Tool Param sh as Therm Off Link

OptiRNAi 2.0 × × × × × http://rnai.nci.nih.gov
siDirect 2 × × ✓ ✓ ✓ http://sidirect2.rnai.jp
siRNA scales × × ✓ × × http://gesteland.genetics.utah.edu/siRNA_scales
siExplorer × × ✓ × ✓ http://rna.chem.t.u-tokyo.ac.jp/cgi/siexplorer.htm
RFRCDB-siRNA × × ✓ × ✓ http://www.bioinf.seu.edu.cn/siRNA/index.htm
OligoWalk × × ✓ ✓ × http://rna.urmc.rochester.edu/cgi-bin
/server_exe/oligowalk/oligowalk_form.cgi
Sfold ✓ × ✓ ✓ × http://sfold.wadsworth.org
siMAX ✓ × ✓ ✓ ✓ http://www.operon.com/products/siRNA/
sirna-overview.aspx
DSIR ✓ ✓ ✓ × ✓ http://biodev.cea.fr/DSIR/
siRNA scan ✓ × ✓ × ✓ http://bioinfo2.noble.org/RNAiScan.htm
RNAxs ✓ × ✓ ✓ ✓ http://rna.tbi.univie.ac.at/cgi-bin/RNAxs
i-Score ✓ × ✓ ✓ × http://www.med.nagoya-u.ac.jp/neurogenetics/i_
Score/i_score.html
siVirus ✓ × × × ✓ http://sivirus.rnai.jp
Details and links are given for each tool. Param: it is possible to provide parameters and constraints as input. sh: the tool
also returns the complete shRNA sequence. as: both antisense and sense siRNA sequences are returned. Therm: the tool
takes into account thermodynamic features. Off: the tool performs or has the option for the computation of off-target
sequences for the designed siRNAs
measure of efficiency [43]. Both sense and antisense strands are

returned. Users can also specify to show only siRNAs with high
predicted efficiency (%mRNA ≤ 30).
siExplorer and RFRCDB-siRNA are other siRNA design tools
which accept a target sequence as input and returns a list of poten-
tial binding sites and siRNAs ranked by their experimental or pre-
dicted efficiency [44, 45]. In particular, siExplorer also returns GC%
content and, for each sequence, provides a link to perform a BLAST
search for off-target evaluation. Moreover, charts with the distribu-
tion of prediction scores and the distribution of the top 10 binding
sites on the target sequences are visualized. Users can choose to
show the top 10, 20, or 50 results that can also be downloaded in
Excel format. RFRCDB-siRNA also provides tested sequences in
addition to the predicted ones. The tool, indeed, performs a data-
base search on an experimentally validated siRNA database in order
to find possible matches with the user-provided sequence.
OligoWalk is another siRNA design tool [46]. It returns a
list of siRNA sequences with their probability of being efficient.
For each siRNA, a thermodynamics analysis is performed. In par-

ticular, the target structure is computed before and after oligo
binding, and details about the energy and the Tm of the duplex,
together with other energy values, are provided.
Sfold is a suite of RNA folding prediction tools, which include
a program for the design of siRNA sequences based on thermody-
namics features [47]. The tool allows interactive computation for
up to 250-nt-long target sequences and batch processing for lon-
ger sequences. In the latter case, a notification is sent by e-mail
when results are available. In addition to the target sequence, users
can also provide further structural constraint information, i.e.,
force/prevent pairing of specific base pairs.
Several files are returned as output, containing detailed infor-
mation on the predicted siRNAs and their interaction with the tar-
get, such as the siRNA duplex GC content and thermodynamics
score, the total stability of the duplex and the differential stability
of its ends, the target accessibility score, the average internal stabil-
ity and disruption energy of the binding site, and the sum of prob-
abilities of unpaired target bases. Other details regarding secondary
structure probabilities are also given and various filters are avail-
able. Finally, the complete probability profile, the regional proba-
bility profile, and the siRNA internal stability profile are provided
as graph charts.
Other tools with more sophisticated interfaces allow the speci-
fication of parameters and constraints about the siRNA to be
designed and its target sequence.
siMAX is a design resource whose input consists of the target
sequence together with a limited number of parameters such as GC
content range, minimum and maximum distance from start and
stop codons, and the mRNA motif to look for (e.g., AA(N19)TT
or AA(N19)NN) [48]. A filter can be enabled in order to avoid
stretches of bases of the same kind. Users can also select a species
among human, mouse, or rat, as a BLAST parameter for the off-
target analysis. Results include the siRNA sense and antisense
strands, the distance of the binding sites from start and stop
codons, the GC content, the results of the BLAST analysis, and the
details about the secondary structure of the siRNA.
The tool DSIR allows the specification of a few prediction
parameters as well, such as the siRNA length and the score
threshold [49]. Filters for nucleotide stretches and immunostimu-
latory motifs can also be enabled. Users can choose to design either
simple double-stranded siRNAs or shRNA sequences. The tool
returns a list of candidate siRNAs (both strands) together with
their scores, the complete shRNA sequences (if chosen), and the
option to perform a BLAST search on some or all of the predicted
siRNAs for the evaluation of off-target effects. Results are export-
able in different formats.
siRNA scan is another tool that allows users to specify several
design options, other than the length and GC content of the
siRNA, such as the 5′ terminal base of the antisense strand, the

minimum number of A/U base pairs in seven terminal bases of the
antisense strand, and the 5′ terminal base of the sense strand [50].
Stretches of 4 or 9 G/C nucleotides in a row can be avoided, and
the number of mismatches in the BLAST similarity search can also
be specified.
RNAxs is another tool based on thermodynamics features,
concerning local target accessibility in particular [51]. Users can
specify design parameters such as the accessibility thresholds, the
folding energy, the sequence and energy asymmetry, and the cus-
tom sequence constraints. However, default values which have
shown to give an optimal separation of functional and nonfunc-
tional siRNAs are already pre-chosen. The output consists of the
candidate siRNA sequences, together with accessibility, asymme-
try, and self-folding scores. Accessibility plots of the binding sites
are also shown, and a BLAST search for similarity can be easily
performed by clicking on the provided links.
The tool i-Score, instead, is a sort of “consensus-based tool,”
since, in addition to its original score, it also provides nine different
designing scores based on different rule sources or other design
tools, such as DSIR [52]. For a given target sequence, it returns
the complete list of possible siRNAs together with the ten scores,
duplex energy, GC content, and the length of the longest GC
stretch, highlighting the top ten miRNAs according to i-Score,
s-BiopredSi score, and DSIR score.
Finally, we introduce siVirus, a tool for antiviral siRNA design
[53]. The system allows the design of siRNA for HIV-1, HCV,
SARS, and influenza virus. For each virus, users can select multiple
viral subtypes and the target regions in the viral genome. The out-
put consists of a list of siRNA target sites (with the 2-nt overhang),
mapped to one or more of the selected genomic regions. For each
siRNA, the predicted efficacy according to three different set of
rules is given, together with predicted off-target hits and conserva-
tion percentage in the selected sequences.
Unfortunately, as mentioned at the end of the previous sec-
tion, a major issue concerning tools such as the ones described
above is represented by the inconsistencies among the current
siRNA design rules, mostly due to the heterogeneity of the siRNA
data [35]. The Max-Planck Institute devised a principle aimed at
identifying all key features relevant to miRNA design. Nevertheless,
this effort has shown to yield many noneffective siRNAs which
have shown to have a high false-positive rate [54].
A recent meta-rules ensemble strategy which integrates several
factors, meta-design rules, and filter criteria has shown to report a
98 % rejection of false-positive miRNAs, showing great improve-
ment over traditional state-of-the-art siRNA design programs [55].
In addition to such strategy, the integration of heterogeneous data
sources can greatly alleviate inconsistency issues among siRNA
design tools.
Table 3
siRNA databases
Database Data Link

NCBI RNAi siRNA/shRNA sequences http://www.ncbi.nlm.nih.gov/projects/genome/rnai/
MIT/ICBP Human and mouse siRNA http://web.mit.edu/sirna/
siRNA DB
HuSiDa Human siRNA http://itb.biologie.hu-berlin.de/~nebulus/sirna/
siRecords Mammalian siRNA http://sirecords.biolead.org/sirecords
VIRsiRNAdb Antiviral siRNA/shRNA http://crdd.osdd.net/servers/virsirnadb/
For each resource, the kind of data and the link are given
3.3 siRNA Databases Several sources of siRNA sequences are publicly available online.
Here we provide a brief overview of some of them, which are listed
in Table 3.
NCBI’s RNAi resource page allows easy access to the RNAi
probes (siRNA/shRNA), stored in the NCBI database. For each
probe, details about the sequence, the targets, and the hairpin, in
case of an shRNA, are given. Queries are submitted through the
standard NCBI interface, which allows results filtering by auto-
matically adding the keywords “gene silencing” to the query.
The MIT/ICBP siRNA Database is a comprehensive database
which stores and distributes information on validated siRNAs and
shRNAs.
Currently the database contains siRNA and shRNA sequences
against over 100 genes from three different sources: (1) sequences
designed and tested by MIT researchers, (2) sequences designed
by Qiagen and tested by Natasha Caplen’s group at the NCI, and
(3) sequences designed by Greg Hannon and Steve Elledge and
tested by the ICBP and CGAP programs at the NCI. The database
can be searched by keywords (e.g., target gene name) or browsed
by gene name and siRNA ID. The results include links to NCBI
probe pages. The website also has a section for the submission of
new validated reagents. Sequences are available for human and
mouse.
HuSiDa is a database that contains sequences of published
functional siRNA molecules targeting human genes and important
technical details of the corresponding gene silencing experiments
[56]. The database is searchable by different terms, such as gene
name, cell line, transfection methods, siRNA source, and siRNA
sequence.
siRecords archives experimentally tested siRNA inferred from
literature [57]. Different data are available for each siRNA, such as
its sequence and the alignment with the target gene, the cell types
or tissues in which it was tested, the forms of the siRNA agents
(e.g., chemically synthesized oligos, vector-transfected shRNA,

etc.), and the methods applied to test its efficacy (e.g., Western
blot, RT-PCR, etc.). A 4-level efficacy score is assigned to each
RNAi experiment based on the data provided by the authors in the
original papers. In particular, an experiment is rated as “very high”
if the gene product is reduced by more than 90 %, “high” if the
gene product is reduced by 70–90 %, “medium” if 50–70 % repres-
sion is achieved, and “low” if less than 50 % of the gene product is
reduced.
Finally, VIRsiRNAdb is a curated database of experimentally
validated viral siRNA and shRNA-targeting genes of 42 human
viruses including influenza, SARS, and hepatitis viruses [58].
Currently, the database provides detailed experimental information
about 1,358 siRNAs/shRNAs and can be browsed by virus, virus
family, gene, and Pubmed ID. It is also searchable by different
keywords. For each siRNA, detailed information are shown, includ-
ing sequence, virus subtype, target gene, GenBank accession,
design algorithm, cell type, test object, test method, and efficacy.
A section of the database, called EscapeDb, provides information
about siRNAs for which viral escape is known, such as the target
site mutations.
As previously mentioned, a fundamental issue in siRNA design
efficacy and efficiency is represented by the lack of an effective
means which would merge all heterogeneous data into the same
framework, allowing results based on different data to be
comparable.
Different solutions to this nontrivial problem have been
devised, but their description is beyond the scope of this chapter.
One solution worth mention, though, is the one proposed by Liu
et al., which consists in considering each siRNA data source as a
“task” and aiming at the development of a joint efficacy model for
all the siRNA data sets simultaneously rather than focusing on sin-
gle data sets in order to derive design rules and efficacy prediction,
combining the different results at the end [35].
In conclusion, the main issues with siRNA design are essen-
tially represented by inconsistency among the design rules and
improper integration of the cross-platform siRNA data. In a more
detailed analysis, aspects such as the lack of a complete feature set
and an inadequate consideration of the specificity of target mRNAs
could also very well be considered as major causes of the aforemen-
tioned problems.
3.4 AntagomiRs: A better understanding of the precise molecular function of miR-

Composition NAs in mammals requires loss-of-function studies in vivo to shed
and Resources light on a landscape of processes and mechanisms which are to date
still largely unknown. To this end, specific and efficient silencers of
endogenous miRNAs are necessary tools in aiding research.
In 2005, the Krutzfeldt et al. studied the biological significance of
silencing miRNAs in vivo with chemically modified, cholesterol-

conjugated oligoribonucleotides which they termed “antagomiRs,”
disclosing a landscape of therapeutical applications regarding
miRNA involved in a disease [32].
The composition of these synthetic RNA analogues basically
consists in a hydroxyprolinol-linked cholesterol solid support and
2′-OMe phosphoramidites to make them more resistant to degra-
dation. They essentially reproduce the antisense strand of the
endogenous miRNA they inhibit; thus there are no actual design
rules applied.
To our knowledge, due to the simple nature of their composi-
tion, there are no tools available online for antagomiR design;
rather biomedical companies provide researchers with the possibil-
ity to chemically synthesize single-stranded, modified RNAs which
specifically inhibit endogenous miRNA function after transfection
into cells by binding to them and causing the miRNA to be subse-
quently degraded.
Of relevant importance as a resource for antagomiRs, the data-
base antagomiRBase presents a collection of 53 putative antagomiR
sequences for a set of 22 human miRNAs which were used as tem-
plate in the design of the putative antisense sequences, using GC
content and secondary structures of the stem-loop sequences of
the miRNAs as parameters, along with the prediction of the free
energy of the unbound antagomiRs [59]. The database presents
the following information for each miRNA-antagomiR pair: the
position of the guide or passenger strand of the miRNA to be tar-
geted in its stem-loop sequence, the actual target and antagomiR
sequences, respectively, and the binding energy of the hybrid
duplex. A tool is also provided to allow the user to specify a 20–25-
nt sequence which is used to query the database, and in case a
match in the antagomiR sequences is found, it returns its second-
ary structure along with all its miRNA targets.
Generally, antagomiRs are efficient as a means to control the
expression of specific miRNA molecules. Nevertheless, a valid
alternative that could also provide the great advantage of silencing
an entire family of miRNAs simultaneously is represented by a
group of longer ncRNAs called “sponges,” which we will discuss in
the next paragraph.
3.5 Design Ebert et al. first introduced miRNA sponges in 2007, as an alterna-
of Effective miRNA tive to chemically modified antisense oligonucleotides for miRNA
Sponges inhibition [34]. Sponges contain multiple binding sites for endog-
enous miRNAs and function by “absorbing” and distracting them
from their natural targets, thus representing a useful tool to probe
miRNA functions in a variety of experimental systems.
Sponges can be easily cloned into expression vectors and tran-
siently transfected into cultured cells in order to efficiently derepress
miRNA targets. They can also be delivered by virus-based vectors,
in order to ensure their stable expression and create continuous

miRNA loss of function in cell lines and transgenic organisms [60].
MiRNA binding sites in sponges are usually specific to the
miRNA seed region, allowing inhibition of a whole miRNA family.
This can represent an advantage over the use of antagomiRs which
are highly specific for a single miRNA, being their function depen-
dent upon full complementary match to the miRNA. A single
sponge can thus efficiently replace many antagomiRs, with the
consequent reduction of potential off-target effects.
Sponges can also be designed to inhibit multiple miRNAs at once.
This powerful feature makes them an efficient solution for loss-of-
function studies over the traditional knockout model based on miRNA
gene deletion, allowing the inhibition of entire genomic and/or func-
tional miRNA clusters, in addition to families. Moreover, the deletion
of a single miRNA gene which is part of a cluster could affect the other
miRNAs of the cluster, while a sponge represents an efficient and easy
way to avoid this side effect and still assure selective inhibition.
Up to date, there are no tools available for the automatic
design of single or multiple miRNA sponges; thus we are going to
describe some of the design methodologies employed so far in a
few successful application.
Ebert et al. constructed PolII- and PolIII-generated sponges.
PolII sponges were constructed by inserting multiple miRNA
binding sites into the 3′ UTR of a destabilized GFP reporter gene
driven by the CMV promoter. PolIII sponges were constructed by
sub-cloning the miRNA binding region from the GFP construct
into a vector containing a U6 snRNA promoter with 5′ and 3′
stem-loop elements. MiRNA binding sites were either bulged or
perfectly complementary to the miRNAs. In the first case, a bulge
at positions 9–12 of the binding site was introduced in order to
prevent cleavage and degradation of the sponge. These sites were
separated by 4-nt spacers, while perfect sites had no spacers
(Fig. 2b). Both CMV and U6 sponges with 4–7 bulged binding
sites produced stronger derepressive effects than sponges with two
perfect binding sites. Fluorescence in situ hybridization showed
that U6 sponges mainly localized to the nucleus, thus making
CMV constructs a better choice. Experiments showed that sponges
could selectively inhibit different miRNAs and that a sponge
designed for a certain miRNA could also derepress targets of the
other miRNAs of the same family. Moreover, the authors sug-
gested 6 as the highest number of functional binding sites, as
sponges with more than 6 sites showed a marginal increase in activ-
ity above 6 sites. However, they argued that sponges expressed at
lower levels could benefit the presence of additional sites.
In subsequent works different types of constructs have been
proposed for the expression of miRNA sponges, but from now on
we will focus only on the design rules, that being the purpose of
this review.
Fig. 2 miRNA sponge constructs. (a) Basic sponge with six miRNA binding sites separated by 4-nt spacers.
(b) Perfect miRNA binding site on a sponge. (c) Bulged miRNA binding site on a sponge. (d) Prototype decoy
consisting of a short hairpin molecule where the loop exposes a binding site for an miRNA. (e) Tough decoy (TuD)
with two exposed miRNA binding sites. (f) Synthetic tough decoy (S-TuD) consisting of two fully 2′-O-methylated
RNA strands exposing an miRNA binding site each
Haraguchi et al. reported optimal conditions for the design of

TuD RNAs (tough decoy RNAs), efficient sponges with structur-
ally accessible and indigestible miRNA binding sites [61]. The pro-
totype decoy consisted of a short hairpin molecule where the loop
exposed a binding site for an miRNA (Fig. 2c). The length of the
stem is critical for the efficient transport of stem-loop structures
into the cytoplasm by Exp-5. Experiments determined that the
optimal stem length, associated to higher inhibitory effects, was
18 bp. Indeed, stems longer than 18 bp had a reduced binding
affinity to Exp-5, while longer stems could be easily digested by
Dicer in the cytoplasm. Starting from this prototype, the authors
investigated several structural modifications in order to optimize
the inhibitory potency of decoys. Experiments showed that the
optimal TuD RNA consisted of a bulged stem-loop structure
where both sides of the bulge were miRNA binding sites flanked
by 3-nt linkers and the two stems separated by the bulge were
18 nt and 8 nt long, respectively (Fig. 2d). The optimal binding
sites had a 4-nt insert between nucleotides 10 and 11, in order to
avoid cleavage of the decoy. In a subsequent work, the same authors
introduced S-TuD (synthetic TuD), a modification of TuD which
consists of two fully 2′-O-methylated RNA strands exposing an
miRNA binding site each [62]. Following the hybridization of the
two paired strands, the resultant S-TuD forms a secondary structure

which resembles the corresponding TuD RNA molecule (Fig. 2e).
In this work, the authors found that internal base pairing between
the two miRNA binding sites on the two strands of a S-TuD can
negatively affect the structural accessibility for the miRNA and
reduce the inhibitory effect. In light of this, they refined the design
rules and suggested that an optimal S-TuD molecule would feature
two miRNA binding sites which are perfectly complementary to
the target miRNA sequence and which don’t form any base pairing
regions longer than 9 nt. If this is not possible, the introduction of
a single mutation or a 4-nt insertion in the middle region of the
binding sites, as for the TuD previously described, is sufficient in
many cases to abolish the base pairing without significantly affect-
ing the affinity to the target miRNA.
These design rules proposed by Haraguchi et al. are the most
sophisticated rules described so far for the design of effective
miRNA sponges. However, the construction of simple RNA mol-
ecules featuring several binding sites for one or more miRNAs,
separated by 2–4-nt spacers, is still the most widely used approach.
This was confirmed by a recent work of Kluiver et al., in which the
authors developed a methodology for the rapid generation of
miRNA sponges by making use of simple constructs with up to 20
perfect or bulged miRNA binding sites, as described in the earlier
work by Ebert et al. [63].
Nevertheless, it must be noted that despite the optimization of
the sponge construct, different application contexts could yield
different degrees of inhibition, making the verification of the suc-
cess of a sponge treatment more challenging than that of genetic
miRNA deletion. Thus, it is still under investigation whether
in vivo sponge expression can effectively provide a valid alternative
to genetic knockouts of miRNA families [60].
References
1. Lee RC, Feinbaum RL, Ambros V (1993) The 6. Elbashir SM, Harborth J, Lendeckel W, Yalcin
C. elegans heterochronic gene lin-4 encodes A, Weber K, Tuschl T (2001) Duplexes of
small RNAs with antisense complementarity to 21-nucleotide RNAs mediate RNA interfer-
lin-14. Cell 75:843–854 ence in cultured mammalian cells. Nature
2. Fire A, Xu S, Montgomery M, Kostas S, Driver 411:494–498
S, Mello C (1998) Potent and specific genetic 7. Brummelkamp TR, Bernards R, Agami R
interference by double-stranded RNA in (2002) A system for stable expression of short
Caenorhabditis elegans. Nature 391:806 interfering RNAs in mammalian cells. Science
3. Tomari Y, Zamore PD (2005) Perspective: 296:550–553
machines for RNAi. Genes Dev 19: 8. Paddison PJ, Caudy AA, Bernstein E, Hannon
517–529 GJ, Conklin DS (2002) Short hairpin RNAs
4. Chu C-Y, Rana TM (2007) Small RNAs: regu- (shRNAs) induce sequence-specific silencing
lators and guardians of the genome. J Cell in mammalian cells. Genes Dev 16:948–958
Physiol 213:412–419 9. Bernstein E, Caudy AA, Hammond SM,
5. Kutter C, Svoboda P (2008) miRNA, siRNA, Hannon GJ (2001) Role for a bidentate ribo-
piRNA: knowns of the unknown. RNA Biol nuclease in the initiation step of RNA interfer-
5:181–188 ence. Nature 409:363–366
10. Bartel DP (2004) MicroRNAs: genomics, bio- siRNA during RISC activation. Cell 123:
genesis, mechanism, and function. Cell 116: 621–629
281–297 25. Gregory RI, Chendrimada TP, Cooch N,
11. Borchert GM, Lanier W, Davidson BL (2006) Shiekhattar R (2005) Human RISC couples
RNA polymerase III transcribes human MicroRNA biogenesis and posttranscriptional
microRNAs. Nat Struct Mol Biol 13: gene silencing. Cell 123:631–640
1097–1101 26. Maniataki E, Mourelatos Z (2005) A human,
12. Monteys AM, Spengler RM, Wan J, Tecedor ATP-independent, RISC assembly machine
L, Lennox KA, Xing Y, Davidson BL (2010) fueled by pre-miRNA. Genes Dev 19:
Structure and activity of putative intronic miR- 2979–2990
NAs promoters. RNA 16:495. doi:10.1261/ 27. Okamura K, Phillips MD, Tyler DM, Duan H,
rna.1731910 Chou Y-T, Lai EC (2008) The regulatory
13. Ozsolak F, Poling LL, Wang Z, Liu H, Liu XS, activity of microRNA* species has substantial
Roeder RG, Zhang X, Song JS, Fisher DE influence on microRNA and 3′ UTR evolu-
(2008) Chromatin structure analyses identify tion. Nat Struct Mol Biol 15:354–363
miRNA promoters. Genes Dev 22:3172–3183 28. Gunsalus KC, Piano F (2005) RNAi as a tool
14. Zeng Y, Yi R, Cullen BR (2005) Recognition to study cell biology: building the genome-
and cleavage of primary microRNA precursors phenome bridge. Curr Opin Cell Biol 17:3–8
by the nuclear processing enzyme Drosha. 29. Kim DH, Rossi JJ (2007) Strategies for silenc-
EMBO J 24:138–148 ing human disease using RNA interference.
15. Lee Y, Ahn C, Han J, Choi H, Kim J, Yim J, Nat Rev Genet 8:173–184
Lee J, Provost P, Rådmark O, Kim S et al 30. Hadj-Slimane R, Lepelletier Y, Lopez N,
(2003) The nuclear RNase III Drosha initiates Garbay C, Raynaud F (2007) Short interfering
microRNA processing. Nature 425:415–419 RNA (siRNA), a novel therapeutic tool acting
16. Gregory RI, Yan K-P, Amuthan G, on angiogenesis. Biochimie 89:1234–1244
Chendrimada T, Doratotaj B, Cooch N, 31. de Fougerolles A, Vornlocher H-P, Maraganore
Shiekhattar R (2004) The Microprocessor J, Lieberman J (2007) Interfering with dis-
complex mediates the genesis of microRNAs. ease: a progress report on siRNA-based thera-
Nature 432:235–240 peutics. Nat Rev Drug Discov 6:443–453
17. Lund E, Güttinger S, Calado A, Dahlberg JE, 32. Krützfeldt J, Rajewsky N, Braich R, Rajeev
Kutay U (2004) Nuclear export of microRNA KG, Tuschl T, Manoharan M, Stoffel M
precursors. Science 303:95–98 (2005) Silencing of microRNAs in vivo with
18. Yi R, Qin Y, Macara IG, Cullen BR (2003) ‘antagomirs’. Nature 438:685–689
Exportin-5 mediates the nuclear export of pre- 33. Scherr M, Eder M (2007) Gene silencing by
microRNAs and short hairpin RNAs. Genes small regulatory RNAs in mammalian cells.
Dev 17:3011–3016 Cell Cycle 6:444
19. Hamilton AJ, Baulcombe DC (1999) A spe- 34. Ebert MS, Neilson JR, Sharp PA (2007)
cies of small antisense RNA in posttranscrip- MicroRNA sponges: competitive inhibitors of
tional gene silencing in plants. Science small RNAs in mammalian cells. Nat Methods
286:950–952 4:721–726
20. Zamore PD, Tuschl T, Sharp PA, Bartel DP 35. Liu Q, Zhou H, Zhu R, Xu Y, Cao Z (2014)
(2000) RNAi: double-stranded RNA directs Reconsideration of in silico siRNA design from
the ATP-dependent cleavage of mRNA at 21 a perspective of heterogeneous data integra-
to 23 nucleotide intervals. Cell 101:25–33 tion: problems and solutions. Brief Bioinform
21. Elbashir SM, Lendeckel W, Tuschl T (2001) 15:292. doi:10.1093/bib/bbs073
RNA interference is mediated by 21- and 36. Elbashir SM, Martinez J, Patkaniowska A,
22-nucleotide RNAs. Genes Dev 15:188–200 Lendeckel W, Tuschl T (2001) Functional
22. Matranga C, Tomari Y, Shin C, Bartel DP, anatomy of siRNAs for mediating efficient
Zamore PD (2005) Passenger-strand cleavage RNAi in Drosophila melanogaster embryo
facilitates assembly of siRNA into Ago2- lysate. EMBO J 20:6877–6888
containing RNAi enzyme complexes. Cell 37. Strapps WR, Pickering V, Muiru GT, Rice J,
123:607–620 Orsborn S, Polisky BA, Sachs A, Bartz SR
23. Miyoshi K, Tsukumo H, Nagami T, Siomi H, (2010) The siRNA sequence and guide strand
Siomi MC (2005) Slicer function of Drosophila overhangs are determinants of in vivo duration
Argonautes and its involvement in RISC for- of silencing. Nucleic Acids Res 38:4788–4797
mation. Genes Dev 19:2837–2848 38. Khvorova A, Reynolds A, Jayasena SD (2003)
24. Rand TA, Petersen S, Du F, Wang X (2005) Functional siRNAs and miRNAs exhibit strand
Argonaute2 cleaves the anti-guide strand of bias. Cell 115:209–216
39. Schwarz DS, Hutvágner G, Du T, Xu Z, Yatsuya H, Qiao S et al (2007) Thermodynamic

Aronin N, Zamore PD (2003) Asymmetry in instability of siRNA duplex is a prerequisite for
the assembly of the RNAi enzyme complex. dependable prediction of siRNA activities.
Cell 115:199–208 Nucleic Acids Res 35:e123
40. Cui W, Ning J, Naik UP, Duncan MK (2004) 53. Naito Y, Ui-Tei K, Nishikawa T, Takebe Y,
OptiRNAi, an RNAi design tool. Comput Saigo K (2006) siVirus: web-based antiviral
Methods Programs Biomed 75:67–73 siRNA design software for highly divergent
41. Reynolds A, Leake D, Boese Q, Scaringe S, viral sequences. Nucleic Acids Res 34:
Marshall WS, Khvorova A (2004) Rational W448–W450
siRNA design for RNA interference. Nat 54. Gong W, Ren Y, Xu Q, Wang Y, Lin D, Zhou
Biotechnol 22:326–330 H, Li T (2006) Integrated siRNA design based
42. Naito Y, Yoshimura J, Morishita S, Ui-Tei K on surveying of features associated with high
(2009) siDirect 2.0: updated software for RNAi effectiveness. BMC Bioinform 7:516
designing functional siRNA with reduced 55. Mysara M, Garibaldi JM, Elhefnawi M (2011)
seed-dependent off-target effect. BMC MysiRNA-designer: a workflow for efficient
Bioinform 10:392 siRNA design. PLoS One 6:e25642
43. Matveeva O, Nechipurenko Y, Rossi L, Moore 56. Truss M, Swat M, Kielbasa SM, Schäfer R,
B, Saetrom P, Ogurtsov AY, Atkins JF, Herzel H, Hagemeier C (2005) HuSiDa–the
Shabalina SA (2007) Comparison of human siRNA database: an open-access database
approaches for rational siRNA design leading for published functional siRNA sequences and
to a new efficient and transparent method. technical details of efficient transfer into recipi-
Nucleic Acids Res 35:e63 ent cells. Nucleic Acids Res 33:D108–D111
44. Katoh T, Suzuki T (2007) Specific residues at 57. Ren Y, Gong W, Zhou H, Wang Y, Xiao F, Li
every third position of siRNA shape its effi- T (2009) siRecords: a database of mammalian
cient RNAi activity. Nucleic Acids Res 35:e27 RNAi experiments and efficacies. Nucleic
45. Jiang P, Wu H, Da Y, Sang F, Wei J, Sun X, Lu Acids Res 37:D146–D149
Z (2007) RFRCDB-siRNA: improved design 58. Thakur N, Qureshi A, Kumar M (2011)
of siRNAs by random forest regression model VIRsiRNAdb: a curated database of experi-
coupled with database searching. Comput mentally validated viral siRNA/shRNA.
Methods Programs Biomed 87:230–238 Nucleic Acids Res 40:D230–D236
46. Lu ZJ, Mathews DH (2008) OligoWalk: an 59. Ganguli S, Mitra S, Datta A (2011)
online siRNA design tool utilizing hybridiza- Antagomirbase: a putative antagomir database.
tion thermodynamics. Nucleic Acids Res Bioinformation 7:41
36:W104–W108 60. Ebert MS, Sharp PA (2010) MicroRNA
47. Ding Y, Chan CY, Lawrence CE (2004) Sfold sponges: progress and possibilities. RNA
web server for statistical folding and rational 16:2043. doi:10.1261/rna.2414110
design of nucleic acids. Nucleic Acids Res 61. Haraguchi T, Ozaki Y, Iba H (2009) Vectors
32:W135–W141 expressing efficient RNA decoys achieve the
48. Schramm G, Ramey R (2005) siRNA design long-term suppression of specific microRNA
including secondary structure target site pre- activity in mammalian cells. Nucleic Acids Res
diction. Nat Methods 2 37:e43
49. Vert JP, Foveau N, Lajaunie C, Vandenbrouck 62. Haraguchi T, Nakano H, Tagawa T, Ohki T,
Y (2006) An accurate and interpretable model Ueno Y, Yoshida T, Iba H (2012) A potent
for siRNA efficacy prediction. BMC Bioinform 2′-O-methylated RNA-based microRNA
7:520 inhibitor with unique secondary structures.
50. Xu P, Zhang Y, Kang L, Roossinck MJ, Mysore Nucleic Acids Res 40:e58
KS (2006) Computational estimation and 63. Kluiver J, Gibcus JH, Hettinga C, Adema A,
experimental verification of off-target silencing Richter MKS, Halsema N, Slezak-Prochazka I,
during posttranscriptional gene silencing in Ding Y, Kroesen B-J, van den Berg A (2012)
plants. Plant Physiol 142:429–440 Rapid generation of MicroRNA sponges for
51. Tafer H, Ameres SL, Obernosterer G, MicroRNA inhibition. PLoS One 7:e29275
Gebeshuber CA, Schroeder R, Martinez J, 64. Hsieh A, Bo R, Manola J, Vazquez F, Bare O,
Hofacker IL (2008) The impact of target site Khvorova A, Scaringe S, Sellers W (2004) A
accessibility on the design of effective siRNAs. library of siRNA duplexes targeting the phos-
Nat Biotechnol 26:578–583 phoinositide 3-kinase pathway: determinants
52. Ichihara M, Murakumo Y, Masuda A, Matsuura of gene silencing for use in cell-based screens.
T, Asai N, Jijiwa M, Ishida M, Shinmi J, Nucleic Acids Res 32(3):893
65. Peek AS (2007) Improving model predictions 69. Chalk AM, Wahlestedt C, Sonnhammer ELL
for RNA interference activities that use sup- (2004) Improved and automated prediction of
port vector machine regression by combining effective siRNA. Biochem Biophys Res
and filtering features. BMC Bioinform 8:182 Commun 319:264–274
66. Ui-Tei K, Naito Y, Takahashi F, Haraguchi T, 70. Klingelhoefer JW, Moutsianas L, Holmes C
Ohki-Hamazaki H, Juni A, Ueda R, Saigo K (2009) Approximate Bayesian feature selection
(2004) Guidelines for the selection of highly on a large meta-dataset offers novel insights
effective siRNA sequences for mammalian and on factors that effect siRNA potency.
chick RNA interference. Nucleic Acids Res Bioinformatics 25:1594–1601
32:936–948 71. Liu Q, Zhou H, Cui J, Cao Z, Xu Y (2012)
67. Shabalina SA, Spiridonov AN, Ogurtsov AY Reconsideration of in-silico siRNA design
(2006) Computational models with thermo- based on feature selection: a cross-platform
dynamic and composition features improve data integration perspective. PLoS One
siRNA design. BMC Bioinform 7:65 7:e37879
68. Amarzguioui M, Prydz H (2004) An algo- 72. Wang L, Huang C, Yang JY (2010)
rithm for selection of functional siRNA Predicting siRNA potency with random for-
sequences. Biochem Biophys Res Commun ests and support vector machines. BMC
316:1050–1058 Genomics 11:S2
INDEX
A E
Adenosine to inosine (A-to-I) editing..................... 189, 231, Energy model ............................ 21, 23, 67, 80, 124, 323, 324
232, 237 Environmental DNA (e-DNA)................................257–277
Alternative Expressed isoforms ...........................................................178
isoforms ...................................................... 173, 366, 369 Extended secondary structure .......................... 64, 66, 68, 78,
splicing .......................................137, 141, 142, 164, 170, 86, 110, 111
173–188, 287, 293, 299, 365–376, 379–391
Alternative splicing (AS) F
code .............................................380, 383–384, 386, 387 Families ............65, 86, 87, 107, 214, 280, 349–362, 407, 409
event ................................................... 379, 380, 382–391 Free energy minimization .......................................3–15, 314
AntagomiR ............................................... 396, 397, 405–407
G
B
Gene
Base pair classification ...................................... 102, 105, 108 annotated .................................................... 144, 147, 247
Bioinformatics .............................. 17, 29, 32, 33, 69, 80, 123, coding ................................................................. 208, 373
163, 164, 166, 180, 244, 263, 275, 280, 327–330 expression ............................. 50, 123, 137, 141, 142, 145,
164, 189, 209, 219, 255, 339, 341, 379, 393, 394
C
profiling ........................................................137, 141
Cis-regulatory elements ....................................................349 structure ....................... 138, 174, 177, 178, 180, 182, 371
Clans ................................................ 349–352, 356–360, 362 Genomics ................................. 4, 39, 74, 164, 173–175, 178,
CLIP-seq.................................................................. 295, 296 179, 182, 183, 185, 188, 190, 192–194, 199, 201,
Common/consensus secondary structures ....................17–34 203, 208, 214, 238, 240, 254, 255, 259, 287, 293,
Comparative methods ..........................................................4 298, 327, 330, 332, 334, 339, 366, 368, 371, 372,
Computational 374, 381, 382, 384–388, 391, 396, 403, 407
biology .............................................................. 3, 67, 262 Graph drawing ...................................................................64
tools .................................................... 209, 219, 307, 350
Covariance model ..............................322, 346, 350, 354, 356 H
Covariation ............................................21, 22, 24–27, 87, 90 Half-read seeding .....................................................149, 153
Heuristic ...................... 21, 32, 33, 53, 61, 123–132, 216, 317
D
High-throughput
Database .............................................................. 4, 5, 34, 41, methods ......................................................................219
43, 50, 67, 102, 103, 107–109, 111, 120, 124, 166, sequencing ............. 23, 149, 219, 220, 244, 245, 279, 295
167, 169, 192, 194, 199, 201, 202, 204, 210,
213–215, 217, 218, 220–223, 224, 226, 234, I
243, 260, 272, 280–285, 287–289, 323, 327,
Immunoprecipitation ...................81, 219, 220, 294, 295, 300
330–333, 339–342
Infernal ..............................................323, 346, 350, 354, 361
Deep sequencing ......................................141, 165, 214, 220,
221, 222, 231–241, 328
J
Differential expression ............................. 143, 164, 168–171,
244, 245, 253, 254, 295, 300 Junction .................................... 113, 114, 117, 142, 147–150,
DNA-seq .................................. 138, 148, 190, 191, 193–196, 153–159, 294, 297, 372, 388
199, 201, 203, 288, 323 reads.....................................147–150, 154, 155, 157–159
DOI 10.1007/978-1-4939-2291-8, © Springer Science+Business Media New York 2015
413
RNA BIOINFORMATICS
414 Index
M R
Maximum expected gain (MEG) estimators ............... 20, 21, REDItools ........................................ 190–192, 194, 199–204
26–32, 34 Rfam ............................................. 4, 5, 41, 67, 68, 76, 77, 82,
Meta-barcoding ........................................................257–277 83, 87–91, 97, 98, 214, 281, 349–362
Metagenomics ...................................257, 258, 260, 262, 287 Ribonucleic acid (RNA)
Metatranscriptomics .................................................279–289 alignment ..........................................................40–42, 87
Microbiome ...................................... 258, 259, 261–262, 266 binding protein ....................294–296, 298, 300, 302, 341
Micro RNA (miRNA) bioinformatics ............................................. 17, 33, 69, 80
modeling ............................................. 207, 208, 211, 217 editing..................................189–204, 232, 237, 327–337
prediction............................................................207–215 folding prediction ..................................... 10–11, 14, 402
target free energy parameters ................................................6, 7
prediction.............................. 208–217, 219, 221–223 functional annotation....................................................39
recognition .................................... 210, 231, 232, 396 motifs....................11, 13, 61, 83, 103, 107–109, 114, 115
Minimum free energy .................................... 4, 15, 110, 125, processing .....................................................................95
213, 308–311, 314, 317, 318, 320, 324 secondary structure
Modtools ............................................................................49 prediction ....................................3–15, 17–35, 43, 50,
Motif pattern ............................................................343–346 52, 61, 69, 103, 109, 126–127, 307, 308, 312, 317,
Multiple sequence alignment........................4, 18, 19, 26, 28, 325, 340, 346
32, 34, 65, 68, 80, 85, 87, 88, 91, 313, 315, 320, relationship ...........................................................3, 9
322, 340, 350 structure comparison ....................................................40
Mutual information ................................................ 21, 22, 24 untranslated sequences..................................................49
visualization ............................................................64–65
N RIP-seq ....................................................................293–302
ncRNAs. See Non-coding RNAs (ncRNAs) RNA interference (RNAi) .........214, 394–397, 400, 404, 405
Next generation sequencing (NGS) ........................ 138, 209, RNAProfile ..................................................................49–61
219, 221, 225, 243, 258, 293, 380 RNA–RNA
Next-generation sequencing (NGS)-Trex.................243–255 interaction................................... 33, 85–87, 97, 123–132
Non-canonical motifs .........................................................72 prediction............................................ 124–126, 128, 130
Non-coding RNAs (ncRNAs) ......................... 17, 24–26, 34, RNA-seq .................. 137–145, 147–161, 163–171, 190–199,
39, 67, 87, 123, 137, 394, 406 201, 203, 213, 214, 221, 223–226, 243–255, 279,
detection .....................................................................322 280, 294–301, 327–337, 381, 388–391
workflow .............................................................243–255
P S
Perl ................................... 6, 93, 97, 166, 174, 180, 232–234,
Secondary structure ............................... 3–15, 17–35, 39–43,
236, 238–240, 265, 270, 271, 340, 342, 343
45, 49–61, 63–99, 101–104, 107, 109–115,
Photoactivatable-Ribonucleoside-Enhanced
123–127, 209, 220, 225, 233, 239, 307–315, 317,
Cross-linking and Immunoprecipitation
318, 322, 324, 325, 340, 341, 343, 345, 346, 349,
(PAR-CLIP).................................. 214, 215, 220,
350, 354, 356, 357, 402, 406, 409
221, 222, 226, 295, 296, 301
Seeding extension ..................................... 149, 150, 155–156
Phylogenetic tree ................... 21, 27, 260, 288, 289, 313, 356
Sequence
Post-transcriptional regulation ........................... 50, 123, 341
alignments ............................ 4, 18, 19, 21, 26, 28, 32, 34,
Probabilistic model ............................. 21–23, 25–33, 35, 350
41, 43, 44, 65, 68, 80, 85, 87, 91, 137, 157, 166–167,
Pseudo-knots ............................................. 21, 31, 35, 64–66,
209, 287, 313, 315, 320, 322, 340, 350, 352
69, 71, 74, 78, 79, 80, 85–91, 110, 314, 317
design ................................................. 6, 12–14, 308, 404
Pyrosequencing........................................ 259, 261, 264–266,
Short hairpin RNAs (shRNAs) ............................... 394, 395,
268, 269, 271, 287, 289
400–402, 404, 405
Python ........................................ 74, 80, 84, 88, 92, 111, 169,
16S rRNA ...........................................87, 259, 261, 282, 283
174, 180, 190, 191, 202, 263, 264, 273–276, 330
phylogenetic analysis ..................................................280
Small interfering RNAs (siRNA) ............................ 123, 309,
Q
323, 394–405
Quality control ..........................137–145, 155, 282, 329, 334 Software package .................................4, 6, 68, 173, 244, 295
RNA BIOINFORMATICS
Index
415
Spliced alignment .................................... 173, 174, 176–178, Transcriptome
187, 188, 252, 330, 388 annotation............................148, 295, 381, 382, 384–388
SpliceMap ........................................ 148–153, 155, 157–161 profile ................................................................. 244, 245
Splicing nomenclature ...................................... 380, 382–384 Transcriptomics ........................................ 209, 214, 219, 380
Structure-informed multiple sequence Transcript prediction ................................ 367–368, 372–376
alignments ........................................................87
U
T
Untranslated region (UTR) ....................50, 54, 58, 143–145,
Taxonomy ..........................................265, 275, 288, 358–359
189, 216, 217, 219, 254, 323, 330, 339–347
Tertiary structure ......................................... 39, 72, 101, 103,
104, 107–109, 114–119
W
Transcription ......................................17, 280, 300, 301, 318,
330, 332, 339, 340, 372, 379, 382, 384–386, 394 Web tool ...................................................................365–376

RNA Bioinformatics

Uploaded by

Copyright:

Available Formats

RNA Bioinformatics

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

RNA Bioinformatics

Uploaded by

Copyright:

Available Formats

What are the main topics covered in the document?

What are the main topics covered in the document?

What are some of the main methods discussed for analyzing RNA?

What are some of the main methods discussed for analyzing RNA?

Methods in

Molecular Biology 1269

Ernesto Picardi Editor

For further volumes:

ISSN 1064-3745 ISSN 1940-6029 (electronic)

Library of Congress Control Number: 2014958476

© Springer Science+Business Media New York 2015

Printed on acid-free paper

Humana Press is a brand of Springer

Bari, Italy Ernesto Picardi

PART I RNA SECONDARY AND TERTIARY STRUCTURES

PART II ANALYSIS OF HIGH-THROUGHPUT RNA SEQUENCING DATA

14 Using Deep Sequencing Data for Identification of Editing Sites

PART III WEB RESOURCES FOR RNA DATA ANALYSIS

ADRIANA ALBERTI • Genoscope—National Sequencing Center, Evry, France

ALFREDO FERRO • Department of Clinical and Molecular Biomedicine, University

LAURENT NOÉ • Laboratoire d’Informatique Fondamentale de Lille (LIFL), UMR CNRS

DARIO VENEZIANO • Department of Molecular Virology, Immunology and Medical Genetics,

RNA Secondary and Tertiary Structures

Free Energy Minimization to Predict RNA Secondary

The RNA molecule, once considered as an intermediate step

sequence, of complementary base-pairings that are formed between

few examples that illustrate the use of free energy minimization to

software package instead of using a webserver directly or for more

Free energy minimization to predict RNA secondary structures is

letters A,C,G,U or use the FASTA format. For example, in the

comparison”. This will result in Fig. 2, where in the upper triangu-

that is in the near-neighborhood of the wildtype will originate from

in the first two figures. Changing the number of mutations to 2 in

folding prediction by energy minimization. Recently, an extension

this type of sequence design, put forth in [30], relies on pro-

1. All free energy minimization methods rely on thermodynamic

energy range above the minimum free energy without a

thermodynamic parameters improves predic- RNA polymerase is required for hepatitis c

RNA Secondary Structure Prediction

Key words Common/consensus secondary structures, Comparative methods, Multiple sequence

Functional non-coding RNAs (ncRNAs) play essential roles in

secondary structure prediction [13, 48] or to find functional RNAs

Problem 1 (RNA Common Secondary Structure Prediction of

Evaluation Procedure 1 (For Problem 1)

Tool References Description

Softwarea Used informationb Gainc Prob. dist.d

Several pieces of information are generally utilized in tools and

2.1.3 Experimental Recently, experimental techniques to probe RNA structure by

with computational approaches. If available, such experimental

2.4 Phylogenetic Because most secondary structures of functional ncRNAs are

3.2 Choice The probabilities p(θ | A) provide a distribution of common RNA

3.2.1 RNAalipffold Model The RNAalipffold model is a probabilistic version of RNAalifold

where E(θ, A) is the averaged free energy of RNA sequences in the

3.2.3 Averaged Predictions based on RNAalipffold and Pfold models tend to be

p (q | A) = w1 × p (pfold) (q | A) + w2 × p (alifold) (q | A) + w3 × p (ave) (q | A), (8)

where w1, w2, and w3 are positive values that satisfy w1 + w2 + w3 = 1 ,

Fig. 4 Example of a predicted common secondary structure from the Centroid

(Subheading 3.2.3) based on the McCaskill or CONTRAfold model.

ì1 if y is the exactly same as q

3.3.3 The CONTRAfold- In conventional RNA secondary structure predictions, another