Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (2024)

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (1)

BioinformaticsGenomics and post-genomics

Frédéric Dardel and François Képès

translated byNoah Hardy

This work has been published with the help of the French Ministère de laCulture – Centre national du livre

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (2)

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (3)


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (4)

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (5)

BioinformaticsGenomics and post-genomics

Frédéric Dardel and François Képès

translated byNoah Hardy

This work has been published with the help of the French Ministère de laCulture – Centre national du livre

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (6)

First published in French as Bioinformatique. Génomique et post-génomique © 2002 ÉcolePolytechnique

Translated into English by Noah Hardy

English language translation copyright © 2006 John Wiley & Sons LtdThe Atrium, Southern Gate, Chichester,West Sussex PO19 8SQ, EnglandTelephone (+44) 1243 779777

Email (for orders and customer service enquiries): [emailprotected] our Home Page on or

All Rights Reserved. No part of this publication may be reproduced, stored in a retrieval system ortransmitted in any form or by any means, electronic, mechanical, photocopying, recording, scanningor otherwise, except under the terms of the Copyright, Designs and Patents Act 1988 or under theterms of a licence issued by the Copyright Licensing Agency Ltd, 90 Tottenham Court Road, LondonW1T 4LP, UK, without the permission in writing of the Publisher. Requests to the Publisher should beaddressed to the Permissions Department, John Wiley & Sons Ltd, The Atrium, Southern Gate,Chichester, West Sussex PO19 8SQ, England, or emailed to [emailprotected], or faxed to (+44) 1243 770620.

Designations used by companies to distinguish their products are often claimed as trademarks. Allbrand names and product names used in this book are trade names, service marks, trademarks orregistered trademarks of their respective owners. The Publisher is not associated with any product orvendor mentioned in this book.

This publication is designed to provide accurate and authoritative information in regard to the subjectmatter covered. It is sold on the understanding that the Publisher is not engaged in renderingprofessional services. If professional advice or other expert assistance is required, the services of acompetent professional should be sought.

Other Wiley Editorial OfficesJohn Wiley & Sons Inc., 111 River Street, Hoboken, NJ 07030, USA

Jossey-Bass, 989 Market Street, San Francisco, CA 94103-1741, USA

Wiley-VCH Verlag GmbH, Boschstr. 12, D-69469 Weinheim, Germany

John Wiley & Sons Australia Ltd, 33 Park Road, Milton, Queensland 4064, Australia

John Wiley & Sons (Asia) Pte Ltd, 2 Clementi Loop #02-01, Jin Xing Distripark, Singapore 129809

John Wiley & Sons Canada Ltd, 6045 Freemont Blvd, Mississauga, Ontario, Canada L5R 413

Wiley also publishes its books in a variety of electronic formats. Some content that appears in printmay not be available in electronic books.

Library of Congress Cataloging-in-Publication Data

Bioinformatique. EnglishBioinformatics : genomics and post-genomics / edited by Frédéric Dardel and François Képès;

translated into English by Noah Hardy.p. cm.

Includes bibliographical references and index.ISBN-13: 978-0-470-02001-2 (cloth : alk. paper)ISBN-10: 0-470-02001-6 (cloth : alk. paper)

1. Bioinformatics. 2. Genomics. I. Dardel, Frédéric. II. Képès, François. III. Title. QH324.2.B558 2006572.80285 – dc22


British Library Cataloguing in Publication Data

A catalogue record for this book is available from the British Library

ISBN-13 978-0-470-02001-2 ISBN 0-470-02001-6

Typeset in 101/2 /121/2pt Sabon by SNP Best-set Typesetter Ltd., Hong KongPrinted and bound in Great Britain by Antony Rowe Ltd., Chippenham, WiltsThis book is printed on acid-free paper responsibly manufactured from sustainable forestry in which at least two trees are planted for each one used for paper production.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (7)


Preface to the French edition viiPreface to the English edition ix

1 Genome sequencing 1

1.1 Automatic sequencing 11.2 Sequencing strategies 41.3 Fragmentation strategies 81.4 Sequence assembly 121.5 Filling gaps 141.6 Obstacles to reconstruction 161.7 Utilizing a complementary ‘large’ clone library 181.8 The first large-scale sequencing project: The Haemophilus

influenzae genome 191.9 cDNA and EST 20

2 Sequence comparisons 25

2.1 Introduction: Comparison as a sequence prediction method 252.2 A sample molecule: the human androsterone receptor 262.3 Sequence hom*ologies – functional hom*ologies 272.4 Comparison matrices 282.5 The problem of insertions and deletions 332.6 Optimal alignment: the dynamic programming method 342.7 Fast heuristic methods 382.8 Sensitivity, specificity, and confidence level 462.9 Multiple alignments 50

3 Comparative genomics 61

3.1 General properties of genomes 613.2 Genome comparisons 673.3 Gene evolution and phylogeny: applications to annotation 75

4 Genetic information and biological sequences 85

4.1 Introduction: Coding levels 854.2 Genes and the genetic code 854.3 Expression signals 87

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (8)

4.4 Specific sites 914.5 Sites located on DNA 914.6 Sites present on RNA 964.7 Pattern detection methods 96

5 Statistics and sequences 107

5.1 Introduction 1075.2 Nucleotide base and amino acid distribution 1075.3 The biological basis of codon bias 1125.4 Using statistical bias for prediction 1135.5 Modeling DNA sequences 1165.6 Complex models 1205.7 Sequencing errors and hidden Markov models 1235.8 Hidden Markov processes: a general sequence analysis tool 1275.9 The search for genes – a difficult art 127

6 Structure prediction 131

6.1 The structure of RNA 1316.2 Properties of the RNA molecule 1326.3 Secondary RNA structures 1346.4 Thermodynamic stability of RNA structures 1386.5 Finding the most stable structure 1446.6 Validation of predicted secondary structures 1496.7 Using chemical and enzymatic probing to analyze folding 1506.8 Long-distance interactions and three-dimensional structure prediction 1526.9 Protein structure 1556.10 Secondary structure prediction 1586.11 Three-dimensional modeling based on hom*ologous protein structure 1616.12 Predicting folding 166

7 Transcriptome and proteome: macromolecular networks 169

7.1 Introduction 1697.2 Post-genomic methods 1707.3 Macromolecular networks 1827.4 Topology of macromolecular networks 1937.5 Modularity and dynamics of macromolecular networks 1997.6 Inference of regulatory networks 206

8 Simulation of biological processes in the genome context 211

8.1 Types of simulations 2138.2 Prediction and explanation 2138.3 Simulation of molecular networks 2158.4 Generic post-genomic simulators 226

Index 233


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (9)

Preface to the French edition

This book is directly based on a course that we teach at the École Polytech-nique. We thank all our colleagues and friends there who have made its exis-tence and publication possible:

Sylvain Blanquet, Chairman of the Biology Department, who first thought ofcreating this interdisciplinary course some ten years ago, when such a projectwas highly innovative. He has followed and supported the development of theproject in the context of the genomic revolution.

Jean-Marc Steyaert, our colleague at the Computer Science Department,where he initiated the teaching of bioinformatics, and in which he remainsactive. His critical attention, constant theoretical and methodological contribu-tions, and increasing involvement in biological problematics, have contributedin an essential way to bringing this book about, as well as influencing its contents.

Philippe Dessen, who participated in some of the very early stages in theteaching of bioinformatics at the École Polytechnique while he was there; ourcontacts with him over the years have been invaluable.

Finally, going from the stage of course-notes to a published book would nothave been possible without the decisive contributions of Jean-Paul Coard, ofÉditions de l’École Polytechnique, as well of those of Jean-Claude Mathieu,Véronique Lecointe, Martine Maguer, and Frédéric Zantonio, involved in thetechnical production.

Frédéric DardelFrançois Képès

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (10)

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (11)

At the suggestion of Vincent Schachter, to whom we are very grateful, wedecided to produce an English version of our book. We have worked closelywith Noah Hardy, to ensure the accuracy of the translation, and have updatedall chapters with new material where necessary. Chapter 8 was rewritten entirelyin English. We hope that this edition will enable many more readers to enjoyour book. We would like to thank Joan Marsh and her colleagues at Wiley fortheir help in producing this edition.

Frédéric DardelFrançois Képès

Preface to the English edition

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (12)

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (13)

1Genome sequencing

1.1 Automatic sequencing

The dideoxyribonucleotide method, developed during the early 1980s inEngland, at the Cambridge University laboratory of Fred Sanger, is today uni-versally employed to sequence DNA fragments. It is based on the use of DNApolymerase to elongate a single strand of DNA, starting from a primer, utiliz-ing another DNA strand as the template. The DNA polymerase elongates thestrand in the presence of four deoxyribonucleotide monomers (dATP, dTTP,dGTP, and dCTP) and a dideoxyribonucleotide analog (ddNTP), which acts asthe chain terminator (Figure 1.1). Specific incorporation of the analog by DNApolymerase yields a mixture of fragments that selectively terminate at positionscorresponding to each nucleotide (As, in the example below).

The principle of the dideoxyribonucleotide (‘dideoxy’) method is illustratedin Figure 1.2. Four parallel reactions are carried out, one with each ddNTP, theDNA fragments obtained being separated by electrophoresis. A fluorescenttracer is used to identify fragments synthesized by the polymerase so as to dis-tinguish them from template DNA. The tracer is attached to one extremity ofthe fragment, either at the 5′-end of the sequencing primer or at the 3′-end ofthe dideoxynucleotide terminator. Modern automatic sequencers utilize an insitu detection system during electrophoresis, in which a laser beam emitting inthe fluorophore absorption spectrum is passed through the gel (Figure 1.3). Amigrating DNA fragment in the path of the laser beam then emits a fluorescentsignal detected by a photodiode on the other side of the gel. The signal is am-plified and transmitted to a computer programmed with special software foranalyzing it.

Under favorable conditions, this technique permits reading up to 1,000nucleotides per sequenced fragment, and an average of 500 to 800 nucleotidesduring routine experiments. Two dideoxy methodologies coexist at present: oneemploys a single photophore, and the other uses four, each with a distinct emis-sion spectrum. In the first system, the four mixtures, corresponding to the fourddNTPs, are introduced into different electrophoresis gel wells. Analysis is

Bioinformatics: Genomics and post-genomics, Frédéric Dardel and François Képès© 2006 John Wiley & Sons, Ltd

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (14)

accomplished by comparing the migration rates of the fragments in the fourresulting lanes.

In the second system, each of the four sequencing reactions uses a differentfluorophore that modifies the corresponding ddNTP. After the four polymer-ization reactions have taken place, the resulting DNA fragments are mixed andintroduced into the same gel well. Constituent nucleotides are identified accord-ing to the emission properties of the fluorescent tracers exposed to the laserbeam using selective color filters, after which a single gel lane is analyzed.

The four-fluorophore technique is a bit more expensive, since it requiressomewhat more varied chemistry. However, it has the advantage of being betteradapted to high-throughput systems, since more samples are analyzed on thesame gel. In the latest generation of automatic sequencers, the classical rectan-gular polyacrylamide gel is replaced by a reusable capillary tube, whereas theseparation and detection principles remain unchanged. This technique reducesthe time required for an experiment from several hours to a few minutes, alsominimizing preparation time. In principle, the highest-performance multi-capillary machines can process up to 1,000 samples per day, equivalent to 0.5 Mbases of raw sequence per day per machine.

Massive high-throughput sequencing centers today often use several dozensuch machines with robots that control the sequencing reactions automaticallyexecuting pipetting, mixing, and incubation steps, thereby minimizing the risk














O- N N














O- N N









Figure 1.1 Dideoxynucleotide structureReplacement of the 3′-OH group in the dideoxynucleotide (ddNTP) by a 3′-H group prevents formation of a phosphodiester link at its 3′-end. The modified nucleotides have a normal 5′-triphosphate side, thus may be incorporated into the chain by DNA polymerase. Since A-T and G-C pairing rules are followed during ddNTP incorporation, the ddATP will be incorporated whereverthere is a T facing it on the template strand.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (15)


DNA Polymerase

DNA Polymerase

DNA Polymerase

Figure 1.2 Principle of the Sanger sequencing methodIn the presence of a template DNA strand and the four dNTPs, DNA polymerase can elongate acomplementary DNA strand starting from an oligonucleotide primer, which hybridizes to the template strand. When a dideoxynucleotide is incorporated by the polymerase, it acts as a chainterminator, blocking further elongation. This incorporation is entirely random, proceeding at a ratethat is a function of the ratio of the dideoxynucleotide concentration to that of the correspond-ing deoxynucleotide (here it is [ddATP] / [dATP] = 1 / 400).

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (16)

of human error (see Figure 1.4). The preparation of DNA templates remainsthe most difficult step to automate, although significant progress has been madein this respect.

1.2 Sequencing strategies

The sequencing methodologies described above fail to address major difficultiesthat need to be considered when operating a large-scale sequencing program:

• Only DNA fragments of between 500 and 1,000 nucleotides may besequenced;

• A sequencing primer that is complementary to the template is required forthe DNA polymerase to begin synthesizing.

Fortunately, these two obstacles may be simultaneously overcome by frag-menting the DNA that is to be sequenced into segments of size compatible to


Figure 1.3 Automatic sequencing using 1- and 4-fluorophore sequencersSamples introduced into the wells (top) are separated by electrophoresis on a polyacrylamide-ureagel. The 5′-CAATCCCGGATGTTT sequence is read from bottom to top.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (17)


Figure 1.4 Advanced automatic multicapillary DNA sequencer for simultaneous sequencing of 96samples. An automatic injection system executes several consecutive separations without manualintervention (®Applied Biosystems).















Figure 1.5 Example of a sequencing profileIntensity of the signal detected by the photodiodes as a function of separation time. Each coloris associated with one of the four separation reactions (A, G, C, and T).

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (18)

that of the sequencing system yield (~103 base pairs) and by inserting them intoan appropriate vector (plasmid or virus). The vector is selected according toseveral criteria:

• It must be able to replicate autonomously in a convenient host cell (usuallyEscherichia coli);

• It must bear one or several gene markers that permit selection of cells thatcontain it (antibiotic resistance, for example);

• Its nucleotide sequence must be known;

• It must contain restriction endonuclease sites that permit cloning by inser-tion of foreign DNA fragments.

In practice, small bacterial plasmids are generally used. The DNA to besequenced is fragmented and ligated into the vector, which is then propagatedin host cells. The clone cell lines (derived from a single initial cell by successivedivision), each containing a different recombinant vector with the same insertedDNA fragment are then isolated. A library of DNA fragments may thus be con-stituted by collecting a large number of these clone cell lines, and used for furtherstudy (see Figure 1.7).

In order to determine the DNA sequence of such a fragment, the correspond-ing cell line is cultured and its DNA extracted for sequencing by the dideoxynu-cleotide method. Since the nucleotide sequences located on each side of thevector clone site are known, they are used as the primer (see Figure 1.6). Theseprimers are independent of the DNA inserted into the vector and may be usedto sequence any fragment; they are therefore called universal primers. Because


Replication Select


Sequencing primers

Reverse sequencingDirect sequencing


Figure 1.6 Sequencing in a vector starting with universal primers.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (19)


Figure 1.7 Strategy for constructing a DNA library.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (20)

such primers are constant, it is very easy to incorporate fluorescent tracersneeded during oligonucleotide synthesis into them. The fluorescent primers thusproduced may be used in most sequencing procedures.

1.3 Fragmentation strategies

In sequencing a long stretch of DNA – especially a complete genome – it is essen-tial that it be cut into fragments of a size compatible with the sequencing tech-nology. This poses two additional questions:

• Which cutting strategy should be employed?

• How can the complete sequence be reconstituted from the pieces?

These two questions are intricately related, since the reassembly method issensitively dependent on how the fragmentation is accomplished. Twoapproaches have mainly been used: random fragmentation and segmentationafter mapping.

Random fragmentation

In random fragmentation, the full length of the DNA to be sequenced is cut intosmall pieces of optimal sequencing size (~1,000 base pairs). A high cutting fre-quency (one site per 200–250bp) restriction enzyme may be used for thispurpose, under conditions of partial digestion (10–20 percent) in order to gen-erate 1,000- to 2,000-bp fragments. Alternatively, ultrasound may be used tobreak the DNA into small pieces, since the mechanical constraints induced byultrasonic vibrations in DNA in solution are sufficient to rupture the long phos-phodiester chain.

The mechanical (ultrasound) method results in more random breaks than theenzymatic method but necessitates an additional step to repair the extremitiesof the DNA fragments produced, since breaks produced by ultrasound treat-ment do not occur at the same level in the two DNA strands. This may requireparing the extended extremities of single strands, so that the resulting fragmentsmay be inserted into blunt cloning sites in the sequencing vector.

The basic postulate of the random or shotgun method is that if enough clonesare analyzed, the entire original DNA sequence will be covered. Assuming thatfragmentation and cloning are really random processes and that the DNAsequence is sufficiently large compared with that of individual clones (which isgenerally the case for a full genome), the probability that a given DNA nucleotidestudied not be covered by random sequencing is a Poisson distribution:


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (21)

where N is the total number of nucleotides sequenced in the set of clones andL the total length of the DNA studied. N/L is the coverage rate, which is therate of data redundancy. In order to obtain a 99 percent sequencing rate, thatis, p0 = 0.01, it is necessary to sequence a number of clones equal to 4.6 times(log 0.01 ≈ −4.6) the length of the DNA studied.

In the case of a genome or a very long DNA fragment, it is thus practicallyinevitable that gaps remain in the sequence, which must be filled using someapproach other than the random shotgun method. It is also possible to statisti-cally evaluate the length and average number of such gaps:

where n is the average length of each fragment sequenced (~500 nucleotides).The following is an example of the results for a bacterial genome (L ≈ 106 bp)and for the genome of a higher organism, such as a mammal or a plant (L ≈109 bp) with a coverage rate of factor 6 (an average value for this type ofproject), which yields 99.75 percent of the sequenced nucleotides:

The random approach raises two important points:

• It is impossible to cover the entire genome without greatly increasing thenumber of clones to be sequenced; to be nearly certain of covering theentire bacterial genome in the above example would require that p0 <<10−6, at least 14 times the coverage rate. From the practical point of view,it is more economical to accept a coverage rate of between 4 and 6 andthen fill the few dozen remaining gaps ad hoc (see Table 1.1 below).

• Assembly of the puzzle of the set of fragments may require systematic side-by-side comparison of all the sequences obtained. For k sequences, this

Number of gaps N neN L= −


Average length of each gap Ln N=

Total length of gaps Le N L= −

p e0 = −N L,


Table 1.1

Bacteria (1Mbp) Mammals (1Gbp)

Number of sequences 12,000 12,000,000Number of remaining gaps 30 29,750Average gap size 200 200

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (22)

amounts to k(k − 1)/2, which is around 108 comparisons for a bacterialgenome and 1014 for a mammalian genome. While today’s computers canhandle the first number of comparisons, the second number remains a formidable challenge.

Segmentation after mapping

The complexity of reconstruction becomes seriously difficult with very largegenomes, which is why some groups have resorted to a two-level approach insuch cases. The shotgun method is used only to reduce the number of clonesrequired to cover the genome. A DNA library is then established in the follow-ing three types of vector, which can accept much larger fragments:

• Cosmids are hybrids of bacterial and bacteriophage plasmids grown in E.coli; they accept up to 30–40kbp of inserted DNA.

• BACs (bacterial artificial chromosomes) are very large plasmids con-structed starting from the replication origin of the bacterial chromosome.They can accept inserted DNA fragments of around 100–130kbp.

• YACs (yeast artificial chromosomes) are analogs of the preceding two typesof vectors, but derive from chromosomes of a lower eukaryote cell, suchas brewer’s yeast, Saccharomyces cerevisiae. YACs accept DNA fragmentsof the order of 1,000kbp. Unlike the two other kinds of vector, which useE. coli as the host, YACs must be maintained in yeast cells.

These three types of vector can be employed to construct practically com-plete DNA libraries from given genomes, using a much smaller number of clones than plasmids intended for sequencing use. A human genome library con-sisting of 33,000 YACs that can accept inserted DNA fragments of averagelength around 1Mbp has been constructed in a joint project between theGénéthon (located in the Paris suburb of Evry) and the CEPH (Centre d’étudedu polymorphisme humain, in Paris). The coverage rate of this clone library isthus practically 10 times the human genome, which consists of around 3.5 ×109 bp.

However, these large clone libraries cannot be directly exploited for sequenc-ing purposes. A map must be constructed that positions the YAC, BAC, andcosmid vectors used in appropriate order on the chromosomes of the genomebeing studied. YAC positioning was accomplished in 1995 at Généthon/CEPH.The International Human Genome sequencing consortium later produced andmapped a library of 350,000 BACs, which provided similar tenfold coverage ofthe human genome. Since they bear smaller inserts, more BACs were required.


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (23)

This map presents two advantages:

1. Used in conjunction with a genetic map, it allows isolation of genes asso-ciated with genetic diseases (Généthon’s main goal) by localizing them ona given YAC.

2. It can be used as a framework in a sequencing program. Once YACs orBACs are positioned, they just have to be sequenced in an orderly manner,using the random strategy for each. The complexity of reconstruction isthen analogous to that of a bacterial genome, since the construct size is onthe order of one mega base pair. Considering the coverage rate (10×) ofthe YAC or BAC libraries, it is ‘enough’ to sequence a fraction of them tocover the entire human genome. This approach is a direct transposition ofthe classical ‘divide and conquer’ strategy used in computer science. Forthe human genome, the International Consortium sequenced 30,000 of the350,000 BACs that made up its original library, enough to provide a ‘tilingpathway’ (see Figure 1.8) that completely covered the chromosomes. Inthe end, BACs were preferred to YACs, since they are smaller and easierto propagate in bacterial cells.

Mapping or positioning the primary library (cosmids, BACs, and YACs) israther time-consuming. It is carried out using a combination of techniques, such


Figure 1.8 Two-step sequencing strategy with intermediate mapping.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (24)

as comparing restriction enzyme digestion profiles, DNA-DNA hybridization,and genetic marker identification. For example, the restriction profiles of over-lapping YACs contain common fragments. Systematic search for similaritiesamong profiles throughout the entire YAC library identifies candidates.

It is then possible to verify the overlapping of two YACs by DNA-DNAhybridization. After the YAC� DNA restriction fragments have been separatedby electrophoresis, they are transferred from the gel onto a membrane, wherethey are immobilized by cross-linking. The two strands are then separated bydenaturing in an alkaline pH. The membrane is then incubated in the presenceof the denatured YAC� DNA marked with a radioactive or fluorescent tracer.If the two YACs have common sequences, the marked YAC� DNA locally pairswith that of the YAC� DNA, in which case hybridization is said to have takenplace between the two YACs. The tracer is seen to be fixed on the membrane,thereby revealing DNA fragments common to both YACs. If the restrictionenzyme cutting profiles coincide accidentally and do not correspond to anoverlap, there is no hybridization, thus no tracer fixation. Systematically con-ducted, this analysis results in a gene map.

1.4 Sequence assembly

The sequenced fragment assembly method requires that possible overlaps beidentified first, so as to detect clones with common DNA sequences. If twoclones are found to overlap, they are merged to form what is called a contig, aterm designating a set of fragments connected to each other by overlappingsequences that are either identical or very similar (within the limits of sequenc-


Figure 1.9 Example of the comparison of restriction enzyme digestion profiles of two overlap-ping YAC clones.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (25)

ing error). The next step consists in step-by-step comparison of each new frag-ment with contigs that have already been identified (see Figure 1.10). This com-parison must take into account the two possible relative orientations of the twosequences: If the fragment overlaps one contig, that contig is extended; other-wise a new contig consisting of this fragment alone is created. When a fragmentsimultaneously overlaps two contigs, both are fused with the fragment. At anytime during a major sequencing project, the data correspond to a set of severalcontigs, whereas ideally, only one contig covering the whole sequence, remainsat the end of the project.

While a contig is being assembled, a consensus sequence associated with it isdefined according to the alignment of its constituent fragments. The consensussequence compares the positions being read and checks their agreement, reveal-ing any differences or ambiguity due to data errors (unread or incorrect nucleotideinterpretation). Differences and interpretation ambiguities may be resolved byfurther data analysis and if necessary, by additional sequencing. This verificationstep is indispensable but time-consuming, since it is at least partly manual.


Figure 1.10 Iterative contig assemblyThe arrows indicate the orientation of sequenced DNA strands.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (26)

Overlap identification

Using the algorithms described in Chapter 2, overlaps may be located by 1 × 1alignment of each new sequence added to already assembled contigs. If the align-ment score is above a given threshold, the two sequences are considered to beoverlaps. However, this ‘brute force’ method is costly in terms of calculationtime, since the alignment algorithms are O(nm), where n and m are the lengthsof the two compared sequences. If k fragments are to be assembled, the algorithm is O(k2). This is prohibitive for very large genomes, where k > 106

base pairs. In order to simplify this problem, note that overlapping sequencesare usually identical (or nearly so; except for a few rare errors) throughout the entire common region. In 1982, this observation led Roger Staden of Cambridge University to propose a more efficient strategy, which has since been improved. It consists in creating a table of 4n n-uplets of possiblenucleotides (n being of the order of 6 to 12). A list of fragments containingcommon n-uplets is compiled for each entry in the table, which is prepared inlinear time O(k). Two overlapping fragments will have a great number of common n-uplets, i.e., all those that correspond to the common region.Applying this criterion, it is possible to identify candidate overlap fragmentssimply by looking up fragments that have several common n-uplets in the table.Overlapping may then be verified by applying a classical alignment method. Thisapproach differs from the ‘brute force’ method in that the alignment algorithmis used only in cases in which overlapping is highly probable. The cost of thismethod is thus approximately a linear function of the number of gels to be analyzed.

The heuristic strategy developed by Staden can nevertheless be ineffective incertain cases, either owing to insufficient sequence data quality, resulting infailure to identify overlaps, or because the sequence analyzed contains severalrepetitions of a given motif, which can introduce contig fragment connectionerrors at the repetition sites. Several other methods avoid this obstacle by ana-lyzing all possible overlaps according to criteria that permit evaluation ofoverlap quality (alignment scores). A graph of all possible connections amongfragments is then drawn, in which the best pathway (‘minimal cost pathway’)is determined. However, although these methods (based on the Dijkstra algo-rithm) guarantee that the alignment obtained is optimum overall, they are con-siderably more costly in terms of calculation time.

1.5 Filling gaps

As discussed above, in large-scale, random ‘shotgun’ sequencing projects it ispractically inevitable that DNA regions remain that are not covered by clones,which poses two questions:


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (27)

• How to position contigs that are disconnected from each other, whichamounts to ordering and orienting various contigs on the sequencedgenome;

• How to complete the sequence of each gap.

This phase of the project can be rather laborious, since a combination of adhoc methods must usually be mobilized in order to fill all the gaps. The threeprincipal methods are integration of genetic map data; PCR across the gaps;and use of another DNA library, generally one containing larger fragments(cosmids, BACs, YACs, etc).

Integration of genetic maps

For certain model organisms, (E. coli, yeast, Drosophila, mouse, etc), geneticmaps exist that indicate the relative positions of various chromosome loci. Thesemaps, obtained by classical methods for measuring the co-transmission frequen-cies of several genetic markers, provide precise indications of the order of theassociated genes, as well as a qualitative idea of the distances that separate them(measured in centimorgans by geneticists). For example, more than a thousandgenetic loci of the colibacillus E. coli had been identified and mapped before sys-tematic sequencing of its genome was undertaken. Some of these genes had beencloned and sequenced for specific purposes, as the need arose. If a gene associ-ated with a known function and located in a genome is identified within asequence, it is then possible to plot the site of the contig that contains it.

PCR amplification of missing regions

When the number of contigs is not too great, PCR amplification may be usedto fill in some of their gaps. As may be seen in Table 1.1, the number of antici-pated gaps in bacterial genomes is of the order of a few dozen. Oligonucleotideswhose sequences match the 3′-extremities of the contigs obtained are then synthesized and chromosomal DNA amplification is attempted, using all paircombinations of the matching oligonucleotides as the primers. Such PCR amplification yields positive results only if the two primers match the sequencesused on the complementary strands, which must be separated by no more thana few thousand nucleotides. In other words, it is possible to use PCR to amplifythe corresponding chromosomal DNA segment when the extremities of the twocontigs are separated by a gap of less than ~5kbp. This method thus not onlypositions two neighboring contigs, but uses PCR to determine the sequence ofa missing DNA fragment.


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (28)

The PCR approach is very appealing, since it both solves the contig-ordering/orientation problem and determines the missing sequences. However,it runs into difficulty in several cases, especially when the gaps are too large forPCR to yield results. In other cases, the sequence repetitions at the ends ofvarious contigs may lead to junction ambiguities, due to the appearance of falsepositives during PCR. Finally, when the number of gaps to be filled-in is toogreat, this method becomes difficult to implement. Indeed, if there are N contigs,therefore 2N PCR primers, 2N(N-1) PCR amplifications must be carried out.Consulting the data in Table 1.1, this amounts to between several hundred anda few thousand reactions for a bacterium, which may be carried out by exist-ing automatic sequencers, but would be more than a billion PCR reactions fora mammalian genome, which remains unrealistic today.

1.6 Obstacles to reconstruction

Repetitive sequences

Most genomes contain some repetitive sequences. In bacteria and lower eukary-otes such as yeast, a very large fraction of chromosomal DNA is ‘useful’, that


Figure 1.11 Simplified example of gap-filling by PCR, involving two contigs and two gaps on a circular chromosome. PCR amplifications of chromosomal DNA by primers 1 and 3 on one hand and primers 2 and 4 on the other yield a positive result, whereas the other combinations(1–4 and 2–3) fail. Thus, by sequencing the DNA fragments obtained with PCR, one can completethe project.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (29)

is, corresponds to coding regions that are transcribed into RNA and translatedinto proteins. Nevertheless, certain sequences are repeated several times at different places in the genome. For example, identical multiple copies of thegenes that code for ribosomal RNA are generally present in the genomes of such organisms over stretches of DNA several thousand base pairs long. Sevensuch copies are found in E. coli, and more than a hundred in brewer’s yeast, Saccharomyces cerevisiae. Only a small fraction of the genomes of highereukaryotes (around 2 percent for humans) is actually coding DNA, and numerous repetitive sequences, amounting to quite a significant portion of thegenome, have been identified, the origins and exact functions of which remainobscure.

These repetitive sequences are a nuisance in genomic sequencing, since theyconsiderably complicate the task of aligning and reconstructing the genome. Aclone containing a repetitive sequence indicates potential alignments with otherclones that contain copies of the repetitive sequence. Thus, in addition toauthentic alignments with its real neighbors, spurious alignments are obtained.The inevitable ambiguities that result risk introducing assembly errors duringreconstruction. These problems become all the more difficult as the repetitivesequences become long and increasingly similar.


Sequencing methods systematically require that DNA fragments be cloned in asequencing vector. The recombinant vectors are then introduced into a host cell in order to constitute the DNA library to be sequenced. Owing to the toxicity of certain sequences to the host organism into which the vector is introduced, such exogenous sequence fragments sometimes cannot be stablymaintained in the vector. For example, some DNA sequences can lead to theexpression of an RNA segment or a protein that is toxic to the host cell, or thatwill titrate and sequester a factor essential to the host. The result is death of thecell bearing these vectors. Another possibility is that an inserted sequence willinterfere with the vector’s replication, in which case the vector can no longer betransmitted to the two daughter cells during cell division, therefore will not beperpetuated down the lineage of the host cell, rapidly disappearing in the suc-cession of cell divisions. In both these cases, the corresponding genomic DNAfragment is not represented in the library, and is said to be ‘unclonable’. Theresult of this technical difficulty is a sampling bias, which inevitably leads to agap in the reconstruction of the contigs. However, if the gap is of reasonablesize, it can be filled by PCR amplification, since it is possible to sequence theamplified fragment without the necessity of cloning it.


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (30)

1.7 Utilizing a complementary ‘large’1 clone library

In order to verify the order of clones within contigs, and to help fill the remain-ing gaps with PCR, especially in cases of very large genomes, the need to resortto an additional library of larger DNA fragments (typically 20 to 100kbp incosmids and BACs) is practically unavoidable. Of course, in order to cover thewhole genome, the number of clones required to constitute this library will bemuch smaller, since they bear DNA fragments that are 10 to 100 times largerthan the library constructed in the sequencing vector.

The size of DNA fragments inserted into BAC and cosmid vectors may bevery simply estimated by electrophoresis. The 500 nucleotides at each end ofthe large fragment are sequenced, and these two sequences are sought in thecontigs obtained, referring to the first clone library. If the two ends are foundin the same contig, it is possible to verify whether the distance between them inthat contig is compatible with the size of the inserted DNA determined by elec-trophoresis. If the two sequences belong to two unconnected contigs, it is pos-sible to position them with respect to each other, so as to estimate the size ofthe gap separating them, then to fill it.

By verifying the coherence of the distances between the ends of two DNAfragments inserted into a BAC or cosmid vector during assembly, this strategy


1 This is an oversimplification. It is not the clones that are large, but the DNA fragments inserted intothe vectors borne by the cloning lines that constitute the library.

Figure 1.12 Utilization of the extreme ends of a ‘large clone’ to verify and guide the assemblyof contigs and to fill gaps.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (31)

eliminates most of the difficulties concerning repetitive sequences. The distancesbetween the ends of large clones may even be directly introduced into the recon-struction algorithm in the form of constraints. This approach also considerablysimplifies the problem of using PCR to fill gaps between contigs by reducing thenumber of primer combinations to be tried. Finally, in very unfavorable casesin which the gaps are too large to be accessible to PCR, the ‘large’ clone thatcovers the missing region may be utilized to try to sequence it directly, using itsDNA as the template and synthesizing ad hoc its sequencing primers, which arecomplementary to the nucleotide sequences on either side of the gap. The onlydifference between this approach and classical sequencing is that instead of the universal primers described above, specific primers are used that allowsequencing to start at the edge of the missing zone rather than at the edge ofthe inserted sequence. The size of the gap may be progressively reduced until itis accessible by PCR.

1.8 The first large-scale sequencing project: The Haemophilus influenzae genome

Haemophilus influenzae is a small gram-negative bacterium whose natural hostis the human upper respiratory tract, in which it can cause mucosal infections,otitis, sinusitis, and meningitis, particularly in young children. Its genome con-sists of a single circular chromosome 1.83 million base pairs long, which is rel-atively small compared with the genome of its cousin, E. coli, which exceeds4.3 million base pairs. Its small genome and the therapeutic interest of this bac-terium as a target made H. influenzae the candidate of choice for the first com-plete automated genome sequencing project, carried out in 1995 at TIGR (TheInstitute of Genome Research), in the United States.

The following is the comprehensive H. influenzae sequencing program under-taken at TIGR:

• Genomic DNA was mechanically cleaved by ultrasound and the ends ofthe fragments produced were ‘equalized’ by nuclease treatment.

• Fragments between 1.6 and 2.0kbp long were selectively purified by electrophoresis and ligated into a plasmid bearing a resistance marker toampicillin, an antibiotic in the penicillin family.

• The recombinant plasmids obtained were transformed in E. coli and theresulting colonies cultured in a selective medium with the antibiotic, thenpurified and isolated. The library thus obtained contained 19,687 E. coliclones, each bearing a recombinant vector that had incorporated a frag-ment of the H. influenzae genome.


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (32)

• The double-stranded DNA of the 19,687 plasmids was prepared in 96-well plates, permitting 96 parallel purifications, using semi-roboticmethods. The set of 24,304 sequences read represented 11.6 millionnucleotides, a genome coverage rate greater than a factor of 6. The recon-struction was achieved using an automatic program that allowed identifi-cation of 42 contigs (connected sequence blocks) separated by ‘gaps’.

• These gaps were filled by often laborious ad hoc methods: molecularhybridization, PCR, and recloning, starting from another DNA library thatcontained larger fragments (15–20kbp vs. 1.6–2.0 for the library usedduring random sequencing). It was thereby possible to determine the com-plete sequence of 1,830,137 basepairs, with an error rate estimated to bearound 1/10,000. The cost excluding overheads was estimated to bearound $0.50 per nucleotide in the final sequence, a total of $0.9 million.

1.9 cDNA and EST

All the sequencing methods described above are meant for large DNA frag-ments, or even for entire genomes. Although they provide overall information,they are relatively cumbersome to implement. Only a small fraction of the verylarge genomes of higher eukaryotes actually codes for proteins (2–5 percent).In addition to the large non-coding regions that separate them, genes are ofteninternally divided by a large number of introns, which are removed from themature messenger RNA by RNA splicing. Thus, the coding part of the genomeis dispersed and amounts to only a minor part of the whole genome. This iswhy some research groups, in addition to using the overall approach, have con-centrated on the essential information contained in coding segments, which iseasier to extract first.

The latter approach consists in isolating mature messenger RNA (i.e., minusits introns), then synthesizing the complementary DNA (cDNA). The protocolis illustrated in Figure 1.13.

• Extracted and purified from tissues, mRNA bears a poly-A tail at the 3′-end,which permits it to be purified by chromatography on resins into which poly-T has been chemically cross-linked (the poly-T pairs with the poly-A).

• RNA/DNA reverse transcription is carried out using poly-T as the primer.In addition, the viral reverse transcriptase used for this purpose possessesthe property of hydrolyzing the template DNA while polymerization2 isproceeding.


2 This specific hydrolysis activity of an RNA strand in an RNA:DNA heteroduplex is called Ribonuclease H activity.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (33)

• When transcription of the first DNA strand has terminated, polymeriza-tion starts again in the other direction, this time using DNA as the tem-plate. The polymerization exploits a hairpin structure at the end of the firststrand for this purpose, which permits priming the second strand.

The resulting double-stranded DNA is a copy of the messenger RNA knownas cDNA (c for complementary to RNA). cDNA is different from genomic DNAin that it is devoid of introns. The interest in cloning and sequencing cDNAs isthat it reveals which subset of genes is actually transcribed into mRNA in agiven cell. It is thus possible to identify the transcription profiles of each celltype in a complex organism made up of differentiated cells that constitute spe-cialized tissues and organs.

Another advantage of this approach is that, in conjunction with the genomicDNA sequence, it permits precise identification of intron and exon boundaries,thus of splicing sites. In addition, due to alternative splicing, and according tothe tissue type and context, the same gene can give rise to several different messenger RNAs, therefore to several protein variants. The transcriptome (byhom*ology with genome) is the set of all the messenger RNAs that can be transcribed from the chromosomes of a given cell. Transcriptome-related infor-mation that would be difficult to obtain from the genomic sequence alone isaccessible by automatic cDNA sequencing, among other ways.

Two quite different strategies may be employed during cDNA constructionand analysis: First, it is possible to attempt to obtain the longest cDNAs, in

1.9 cDNA AND EST 21

Figure 1.13 Construction of a cDNA segment starting from 3′-polyadenylated mRNA.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (34)

order to cover the entire open reading-frame (ORF) that corresponds to the genebeing studied. This requires considerable caution while extracting the messen-ger RNAs. Any degradation of the mRNA will result in the production of anincomplete cDNA. Obtaining a complete cDNA allows determination of theentire ORF sequence, from which the sequence of the corresponding proteinmay be deduced. Cloning the cDNA bearing the ORF in an appropriate geneticvector will then permit production of the recombinant protein in a heterologousprokaryote or eukaryote host. Since the cDNA is devoid of introns, it can betranslated in any cell, independent of any splicing machinery.

Alternatively, the cDNA – even if incomplete – may be used to determine the sequences of its own ends, which constitute the ‘signatures’ allowingunequivocal identification of the corresponding gene. This strategy has beenused in a massive and systematic way in attempting to identify all the genestranscribed in a given type of cell. These cDNA sequence fragments are calledESTs (Expressed Sequence Tags). To date, more than six million ESTs have beensequenced for the human, the most-studied organism, and the number ofsequenced ESTs for eight other species exceeds 400,000: mouse, rat, cow,zebrafish, chicken, and frog in the animal kingdom; wheat and corn in the plantrealm. In numerous cases, several ESTs correspond to the same gene, and it ispossible to apply the reconstruction methods (contig assembly) described aboveto reconstitute the corresponding messenger RNA sequence.


Adams M.D., et al. (1993). Rapid cDNA sequencing (expressed sequence tags) from adirectionally cloned human infant brain cDNA library. Nat Genet 4: 373–380.

Adams M.D., et al. (2000). The genome sequence of Drosophila melanogaster. Science287: 2185–2195.

Bonfield J.K., et al. (1995). A new DNA sequence assembly program. Nucleic Acids Res23: 4992–4999.

Broder S., Venter J.C. (2000). Whole genomes: the foundation of new biology and med-icine. Curr Opin Biotechnol 11: 581–585.

Cohen D., et al. (1993). A first-generation physical map of the human genome. Nature366: 698–701.

Dear S., Staden R. (1991). A sequence assembly and editing program for efficient man-agement of large projects. Nucleic Acids Res 19: 3907–3911.

Fleischmann R.D., et al. (1995). Whole-genome random sequencing and assembly ofHaemophilus influenzae Rd. Science 269: 496–512.

International human genome sequencing consortium (2004). Finishing the euchromaticsequence of the human genome. Nature 431: 931–945.

Kent W.J., Haussler D. (2001). Assembly of the working draft of the human genomewith GigAssembler. Genome Res 11: 1541–1548.


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (35)

Lander E.S., et al. (2001). Initial sequencing and analysis of the human genome. Nature409: 860–921.

Lander E.S., Waterman M.S. (1988). Genomic mapping by fingerprinting random clones:a mathematical analysis. Genomics 2: 231–239.

McPherson J.D., et al. (2001). A physical map of the human genome. Nature 409:934–941.

Myers E.W., et al. (2000). A whole-genome assembly of Drosophila. Science 287:2196–2204.

Olson M., et al. (1989). A common language for physical mapping of the humangenome. Science 245: 1434–1435.

Osoegawa K., et al. (2001). A bacterial artificial chromosome library for sequencing thecomplete human genome. Genome Res 11: 483–496.

The C. elegans sequencing consortium (1998). Genome sequence of the nematode C.elegans: a platform for investigating biology. Science 282: 2012–2018.

Venter J.C., et al. (2001). The sequence of the human genome. Science 291: 1304–1351.Venter J.C., et al. (1998). Shotgun sequencing of the human genome. Science 280:

1540–1542.Weber J.L., Myers E.W. (1997). Human whole-genome shotgun sequencing. Genome

Res 7: 401–409.Weissenbach J., et al. (1992). A second-generation linkage map of the human genome.

Nature 359: 794–801.


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (36)

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (37)

2Sequence comparisons

2.1 Introduction: Comparison as a sequence prediction method

From its inception, the science of biology has been concerned with compari-sons of the various manifestations of life – cells, organs, and organisms. Thestudy of nucleic acid and protein sequences is no exception to this rationale,and comparing their alignments has become an essential approach of the bioinformatician.

The accelerating growth of genome sequencing programs is generating a hugequantity of data. Information derived from genomic DNA sequences is used topredict regions that are transcribed into messenger RNA. The translation ofopen reading frames (ORFs) into transcribed sequences and comparison withcDNA fragments and expressed sequence tags (ESTs) permits deduction of theamino acids that constitute the proteins encoded by these genes. Bacterialgenomes encode between 1,000 and 10,000 different proteins (slightly morethan 4,000 for Escherichia coli), and eukaryote genomes between 5,000 and50,0001 (slightly more than 12,000 for Drosophila). These gene sequences mustthen be annotated; an attempt must be made to identify their functions in cells,which is easy in some cases, since a similar sequence from a related species maybe located in a sequence database. For example, certain chimpanzee proteins are nearly identical to corresponding human proteins. However, proteins newly identified by genome sequencing often do not have obvious hom*ologs among already known and characterized proteins. In such cases, further analysis must be undertaken in order to reveal faint resemblances, in the attemptto ascribe a function to a protein under study. Today, systematic comparisonusing computer-based analytic methods in conjunction with a posterioriinterpretation by biologists plays an essential role in the decryption of geneticinformation.

1 Excluding variants resulting from alternative RNA splicing and protein maturation.

Bioinformatics: Genomics and post-genomics, Frédéric Dardel and François Képès© 2006 John Wiley & Sons, Ltd

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (38)

2.2 A sample molecule: the human androsterone receptor

The nuclear receptor for the human steroid sex hormone androsterone will beused in this chapter as a sample protein sequence to illustrate questions andapproaches. The function of the androsterone receptor protein is to stimulatetranscription of certain genes in response to a hormone signal. Androsteroneenters the target cell nucleus, where it binds to its receptor, part of which then binds to a specific DNA sequence located upstream from the transcription pro-moters of the genes to be stimulated. This in turn elicits RNA polymerasebinding, thereby activating transcription. The amino acid sequence of that part of the androsterone receptor which binds to the DNA is given below (amino acids 550 through 620, of a total of 919 in the complete hormoneprotein).


The three-dimensional (3D) structure of the corresponding protein domainbound to its DNA target is illustrated in Figure 2.1.


Figure 2.1 Structure of a steroid hormone nuclear receptor (dimer) bound to DNA.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (39)

2.3 Sequence hom*ologies – functional hom*ologies

The basic postulate of all biological sequence analysis is:


The above is true for proteins whose amino acids and sidechain chemicalfunctions are conserved, especially those located in the active site. To a lesserdegree, this is also true for DNA, where such regions as transcription promotersand binding sites for specific proteins generally present significant similarities.

The objective of the bioinformatician is to detect such similarities, using com-puter science methods to draw biological conclusions:

• If two molecules of known function resemble each other, it is reasonableto conclude that they at least partly share their mechanism of action.

• Similarity in the sequences of an unknown and a known protein suggeststhe function of the unknown protein.

• The stronger and more extensive the sequence resemblance, the moreprobable the protein function hom*ology.

• Identifying the most highly conserved or most similar parts of a proteinindicates important regions of the protein, thus the probable location ofits active site.

In order to illustrate the nature of such similarities, Figure 2.2 presents threeexamples of protein sequences that are aligned with the same part of the andros-terone receptor sequence:

In the first case (A), the hom*ology is patent, since most amino acids are con-served. This resemblance is not very surprising, since it concerns the nuclearreceptor for progesterone, another steroid sex hormone. These two proteinshave quite analogous cell functions. In the second case, the resemblance is lessclear, and the number of strictly conserved residues decreases. It was even nec-essary to introduce two gaps into the androsterone receptor sequence to con-tinue aligning it with the sequence of the other receptor. Again however, theresemblance is significant, since the protein involved is also a nuclear receptor,whose ligand this time is thyroid hormone, which although not a steroid sexhormone, explains why the resemblance is more distant. The similarity observed


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (40)

is interesting from the biological point of view, since it suggests that an evolu-tionary relationship exists among various nuclear receptors. The third exampleis meant to illustrate the pitfalls of this approach, since there is no functionalrelationship between the bacterial protein sequence and the androsterone recep-tor. Indeed, bacteria have no nuclei, hence no nuclear receptors for hormones,and do not produce hormones. The resemblance between the two sequences isevidently accidental.

The sporadic occurrence of false positives is statistically inevitable during asystematic search of very large data banks. Some way of evaluating the qualityof alignments is needed to distinguish between significant alignments and back-ground noise, in the attempt to quantify their pertinence.

2.4 Comparison matrices

The approach used for this purpose is sequence alignment scoring. The simplestway to do this is to count the number of identical residues in two alignedsequences and normalize by sequence length, which yields a percentage identity.The identity rates of the three examples used in Figure 2.2 are, respectively 82percent, 42 percent, and 27 percent. This method is well adapted for use withDNA sequences. In fact, the four nucleotides A, T, G, and C play equivalentroles in the structure and function of the DNA molecule. However, thisapproach is a bit too coarse when applied to protein sequences, since it doesnot take into account two important considerations:

• The different relative frequencies of the twenty amino acids: Certain aminoacids, such as alanine (A) and leucine (V), are abundant, whereas others,





C) Alignment with a bacterial ferrodoxin protein sequence (Proteus vulgaris)

Figure 2.2 Comparison of the amino acid sequence of the human androsterone receptor DNA-binding region (upper lines) with the sequences of three other proteins. Strict conservation is indicated by vertical lines. Replacement by amino acids of similar chemical function is indicatedby a plus sign. Dashes indicate positions at which insertions were necessary in order to pursuealignment.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (41)

such as tryptophan (W) and cysteine (C), are relatively rare (~9 percenteach for A and V and 1.5 percent for W and C.) Thus, conservation of aleucine is statistically less significant than conservation of a tryptophan inan alignment of two sequences.

• The chemical structures and functions of some amino acids, while not iden-tical, are very similar (Figure 2.3), and therefore related. This must betaken into account when calculating the scores of two amino acids located opposite each other in a sequence alignment (as indicated by a plus signin Figure 2.2).

Comparison tables, or matrices, are therefore used to attribute a score to analignment of any two nucleotides or amino acids. For protein sequence align-ments, 20 × 20 matrices M are used to evaluate all combinations of amino acidpairs. The value of the coefficient M(a, b) indicates the quality of alignmentbetween two amino acids a and b. It is then possible to calculate an overall scorefor the alignment of two sequences of length L (ak and bk represent the kth aminoacids of each sequence):














Lysine (K)

Arginine (R)


Figure 2.3 Example of hom*ologous amino acids. All three bear -NH2 or -NH3 groups at the endsof straight carbon chains. Lysine and arginine are positively charged at physiological pH.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (42)

Calculation of the percentage sequence identity is a special case of thisapproach, in which the matrix M is reduced to the identity matrix. It is one ofthe most frequently used for DNA sequences.

Several types of matrices have been proposed for protein sequences. Some arebased on a priori estimations of physicochemical similarities. However, this issomewhat arbitrary and rarely used anymore. Others are based on the proba-bility of mutation over evolution. For example, PAM (Probability of AcceptableMutation) matrices assign a score based on the alignment of protein sequenceswith identical functions in related species. The frequency at which amino acida is replaced by amino acid b in another species is noted. Knowing the evolu-tionary distance between the two species2, it is possible to deduce the mutationprobability p(a → b) per unit of time. Using the M(a, b) matrix coefficient, thelogarithm of this probability normalized by the frequencies of a and b is:

For example, if matrix M10 for an evolutionary period of ten million years iscalculated using logarithms, it is possible to compute M20 = M10

2 and M100 =M10

10, for 20 and 100 million years, respectively, by simple matrix multiplica-tion. Such PAM matrices have been widely used over the past twenty years.However, they also suffer from bias, owing to a choice of protein families thatwas initially too restrictive for use in calculating the probabilities p(a → b).These matrices have recently been replaced by others based on the alignment ofblocks of highly conserved sequences known as a ‘BLOSUM’ (block substitu-tion matrix.) These contiguous, gap-free blocks may be found in the sequencedatabase (see Figure 2.4). This approach postulates that blocks correspond tohighly conserved elements in the three-dimensional structure of the correspon-

M a bp a bp a p b

, log( ) = →( )( ) ( )






1 0 0 0

0 1 0 0

0 0 1 0

0 0 0 1

score M a bk kk


= ( )=

∑ ,1


2 The period that has transpired since the two species diverged from a common ancestor. Among othercriteria, estimation of the duration of this period is based on paleontological analysis.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (43)

ding proteins, which it has been possible to verify in cases of known 3D struc-ture. The amino acid substitutions observed in these blocks thus indicatereplacements that are ‘acceptable’ from the point of view of protein structureand function.

This may be verified in the left column of the block in Figure 2.4, which con-tains eight tyrosines (Y) and one phenylalanine (F). These two amino acids arevery similar; both are aromatic, differing only by the presence of an additionalhydroxyl (-OH) group on the tyrosine. It is possible to calculate the probabili-ties of the substitution of one amino acid for another by considering allunordered amino acid pair combinations that appear in the columns of theblock. For example, the first column contains eight Y and one F, yielding eight(Y, F) pairs and twenty-eight (Y, Y) pairs. Compiling all the blocks and usingall twenty amino acids, we obtain numerous occurrences of each of the 210 pos-sible pairs. The M(a, b) = M(b, a) term of the BLOSUM matrix is then calcu-lated in a manner analogous to PAM matrices, taking the base 2 logarithm3 ofthe ratio of the frequency of the (a, b) pair to the frequencies of a and b, writtenas p(a) and p(b).

If p(a, b) > p(a) ·p(b), the (a, b) substitution is overrepresented, indicatingthat amino acids a and b are readily interchangeable, thus probably hom*ologs.The M(a, b) term is therefore positive. If M(a, b) = 0, replacement of a by b isneutral, and if M(a, b) is negative, the substitution is unfavorable.

Several BLOSUM matrices have been calculated using alignment blocks con-structed according to more or less stringent identity criteria among the varioussequences. In thorough analysis, the BLOSUM62 matrix, based on sequence

M a bp a b

p a p b, log

,( ) = ( )( ) ⋅ ( )2



Figure 2.4 Example of a sequence block. The sequences (left) are all nuclear receptor fragments.Strictly conserved amino acids are shaded in gray. The receptor ligand and species from which itderives are indicated on the right.

3 Since log base 2 is used, the coefficients of matrix M is expressed in bits.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (44)

blocks of greater than 62 percent identity, yields the best empirical results andis the matrix most frequently used for searching data banks. It is presented inFigure 2.5, in which one can see that the diagonal coefficients, which correspondto the rarest residues (C and W), are the largest, indicating that their strict con-servation is what is most significant.

By comparing PAM and BLOSUM matrix coefficients, it may be seen thatwhile their general tendencies are similar, there are real differences betweenthem. BLOSUM matrices are more tolerant of substitution among polar orhydrophilic amino acids, mostly found on the surfaces of proteins, and morestringent with respect to other types of modifications involving apolar orhydrophobic amino acids. The latter are usually present within the 3D folds ofthe protein, and their substitution often has major structural repercussions.BLOSUM matrices are thus more useful in identifying proteins whose 3D struc-tures are hom*ologous. It has been possible to verify this empirically, using


9 -1 -1 -3


0 -3 -3 -3 -4 -3 -3 -3 -3 -1 -1 -1 -1 -2 -2 -2

4 1 -1 1 0 1 0 0 0 -1 -1 0 -1 -2 -2 -2 -2 -2 -3

5 -1 0 -2 0 -1 -1 -1 -2 -1 -1 -1 -1 -1 0 -2 -2 -2

7 -1 -2 -2 -1 -1 -1 -2 -2 -1 -2 -3 -3 -2 -4 -3 -4

4 0 -2 -2 -1 -1 -2 -1 -1 -1 -1 -1 0 -2 -2 -3

6 0 -1 -2 -2 -2 -2 -2 -3 -4 -4 -3 -3 -3 -2

6 1 0 0 1 0 0 -2 -3 -3 -3 -3 -2 -4

6 2 0 -1 -2 -1 -3 -3 -4 -3 -3 -3 -4

5 2 0 0 1 -2 -3 -3 -2 -3 -2 -2

5 0 1 1 0 -3 -2 -2 -3 -1 -2

8 0 -1 -2 -3 -3 -3 -1 2 -2

5 2 -1 -3 -2 -3 -3 -2 -3

5 -1 -3 -2 -2 -3 -2 -3

5 1 2 1 0 -1 -1

4 2 3 0 -1 -3

4 1 0 -1 -2

4 -1 -1 -3

6 3 1

7 2













Figure 2.5 BLOSUM62 matrix calculated using over 2,000 sequence blocks. Matrix coefficientsare multiplied by 2 and rounded-off to the nearest integer. The units are thus half-bits. Squarescontaining positive or null coefficients are shaded.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (45)

test-sequence sets corresponding to proteins of known structure. This is not sur-prising, since the blocks of aligned sequences used usually consist of helices andsheets, the structural elements of these proteins.

2.5 The problem of insertions and deletions

The matrices described in the preceding paragraph permit direct evaluation ofthe quality of the alignment of gap-free sequences. However, when comparingcomplete sequences of proteins or long DNA fragments, ‘jumps’ must often beintroduced into one of the sequences being studied. This is referred to as aninsertion or a deletion with respect to the reference sequence.

In proteins, insertions and deletions are generally found in loops located onthe protein surface. In these regions, which are exposed to aqueous solvent,modification of the length of the peptide chain has little effect on the 3D struc-ture of the protein, compared with insertions and deletions located within thefolds of the protein.

Insertions and deletions must be taken into account in calculating the scoresof the alignments that contain them. The simplest approach consists in extend-ing the definition of matrices, adding coefficients for the alignment of anucleotide or an amino acid with a gap (represented by a dash (‘−’) in the align-ments, as seen in Figures 2.2 and 2.6). Thus new terms are introduced into thematrix, M(a, −), corresponding to the cost of alignment of a gap ‘−’ with aresidue a. Since there is no rigorous statistical method for defining the value(cost), it is chosen empirically. A value ‘∆’ which is significantly lower than thelowest alignment score of the two amino acids or nucleotides, is generally usedfor all amino acids:

For example, with the BLOSUM62 matrix in Figure 2.5, a cost ∆ of −6 to −10is applied for the insertion of a gap, whereas the lowest score for the matrix is −4. This extension of the notion of matrix score therefore permits quantitative evaluation of the alignment of two sequences, ‘weighing’ them in comparisonwith other alignments. A problem biologists have in analyzing two proteinsequences is to find the best alignment; that is, the one whose score evaluated using the matrix is the highest.

∀ −( ) = ( )a M a constant, , ∆



^^ ^ deletion insertion

Figure 2.6 Examples of deletion and insertion in a test sequence.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (46)

2.6 Optimal alignment: the dynamic programming method

There are a great number of different ways to align two sequences of length nwith insertions/deletions. If certain trivially equivalent alignments are elimi-nated, it is possible to express this number as:

This number, the central coefficient of the binomial, increases approximatelyas 22n. For two sequences of 20 nucleotides or amino acids, this already amountsto 137 billion combinations. It is therefore unthinkable to use an approach ofthe ‘brute force’ type, which would consist in testing all of them. Instead, a verygeneral algorithmic method known as dynamic programming is used, whichfinds the optimal alignment in time O(n2). Let there be two sequences S1 and S2

of respective lengths m and n. The principle of the dynamic programmingmethod is to progressively construct optimal alignments of longer and longersub-sequences of S1 and S2, using the results obtained with the shorter sub-sequences. For this purpose, a table T of dimensions m × n is constructed. The value of cell T(i, j) indicates the score of the best alignment of the i first S1

amino acids with the j first S2 amino acids. The table is constructed recursively.Beginning by initializing at T(0, 0), which corresponds to an empty alignment(zero), the table is filled in progressively from left to right and from top tobottom.

Three types of alignment should be considered when filling in the T(i, j) cell,that is, the best score for sub-sequences of lengths i and j:

• Those that align S1(i) with S2(j). In this case, the score of the alignment ofS1(i) with S2(j) is added to the best score obtained for sequences of lengthi − 1 and j − 1: T(i − 1, j − 1) + M(S1(i), S2(j)).

• Those that align S1(i) with a gap. In this case, the score of the alignmentof S1(i) with a gap or with constant ∆ (above) is added to the best scoreobtained for sequences of length i − 1 and j: T(i − 1, j) + ∆.

A nn

n( ) =



……Xi ……Xi ……-

……Yj ……- ……Yj

1 2 3

Figure 2.7 Possible types of alignment for sub-sequences of lengths i, j.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (47)

• Those that align S2(j) with a gap. In this case, the score of the alignmentof S2(j) with a gap or with constant ∆ (above) is added to the best scoreobtained for sequences of length i and j − 1: T(i, j − 1) + ∆.

To calculate T(i, j), it suffices to take the maximum of these three scores:

Completing table T in the lower right-hand corner box, T(m, n), we obtainthe best alignment score between S1 and S2. To find the alignment(s) whichobtain(s) this score4 we must recall which of the above alternative terms is thehighest; that is, which among the three expressions �, �, and � is (are, in caseof a tie) was (were) used to calculate T(i, j). These possibilities represent dif-ferent possible paths through the matrix T, as presented in Figure 2.8.

This approach determines which path in matrix T yields the overall maximalalignment of sequences S1 and S2. Figure 2.9 gives an example of a calculationusing matrix T, with the resulting maximal path, for the alignment of a frag-ment of the androsterone receptor with a fragment of the thyroid hormonereceptor, as it was presented in Figure 2.2.

Application of the dynamic programming method to alignment is known asthe algorithm of Needleman & Wunsch, who first applied it to sequences ofbiological molecules. Given comparison matrix M and the cost of insertion/deletion ∆, the Needleman & Wunsch algorithm finds the minimal cost align-ment. For two sequences of length m and n, this algorithm has complexity O(m,

T i j

T i j M S i S j

T i j

T i j

, max

, ,



( ) =− −( ) + ( ) ( )( )

−( ) +−( ) +

1 1



1 2 �



S (i)1

S (j)2


T(i-1,j) T(i,j)


Figure 2.8 The three terms used in calculating T(i, j) determine a matrix path.

4 Several alignments can achieve the maximal score. Each time that the maximum number taken to cal-culate T(i, j) is attained by at least two of the three possible alternative terms, ‘branching’ occurs in thematrix T return pathway. The number and multiplicity of branchings determines the number of equiv-alent maximal alignments.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (48)

n) in terms of calculation time and memory space. Table T(i, j) must first becompleted, then followed in reverse, in order to reconstitute an optimalpathway.

The Needleman & Wunsch method permits aligning sequences in an overallmanner, attempting to construct the best alignment of the two sequences alongtheir full lengths. This approach is justified when two proteins are hom*ologousthroughout their entire sequences. However, very often the hom*ology is local;concerning only a part of each of the two sequences being compared. In thecase of the androsterone nuclear receptor, the sequence alignments presented inFigure 2.2 concern just the DNA binding zone, which consists of only arounda hundred amino acids. This protein also contains domains that affect otherfunctions, including an androsterone binding domain and a regulatory domain,as well as domains responsible for interactions with other proteins. Sometimesthe hom*ology between two proteins is limited to one or several domains thatcorrespond to the functions they have in common. The Needleman & Wunschalgorithm is clearly not good for detecting this type of similarity in two


















0 -6 -12 -18 -24 -30 -36 -42 -48 -54 -60 -66 -72 -78

-6 6 0 -6 -12 -18 -24 -30 -36 -42 -48 -54 -60 -66

-12 0 8 5 -1 -7 -13 -19 -25 -31 -37 -43 -49 -55

-18 -6 2 13 7 1 -5 -11 -17 -23 -29 -35 -41 -47

-24 -12 -4 7 13 7 1 -5 -11 -17 -23 -29 -35 -41

-30 -18 -10 1 7 12 6 0 -6 -12 -18 -24 -27 -33

-36 -24 -18 -5 1 6 14 8 2 -1 -7 -13 -19 -25

-42 -30 -19 -11 -5 0 8 12 13 7 4 -2 -8 -14

-48 -36 -25 -17 -11 -6 2 8 12 13 7 2 -4 -10

-54 -42 -31 -23 -17 -12 -4 2 6 10 11 6 6 0

-60 -48 -37 -29 -23 -18 -10 -4 1 6 9 13 7 3

-66 -54 -43 -35 -29 -24 -16 -10 -5 0 5 7 10 4

-72 -60 -49 -41 -34 -28 -22 -16 -10 -5 0 3 5 9

-78 -66 -55 -47 -40 -34 -28 -22 -16 -11 -6 7 2 3

-84 -72 -61 -53 -46 -39 -34 -28 -22 -16 -11 1 6 1

-90 -78 -67 -59 -52 -45 -40 -34 -28 -22 -17 -5 0 15


Figure 2.9 Example of matrix T(i, j), The scores are indicated in cells; arrows indicate each timea given pathway is followed, in order to obtain the maximum score. The resulting maximal align-ment is indicated to the right of the matrix. The value of ∆, the cost of alignment with a gap, is−6. The first line and the first column, T(0, j) and T(i, 0), are respectively initialized by i × ∆ andj × ∆, effectively corresponding to alignments that begin with i or j gaps.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (49)

sequences. Indeed, the scores obtained would be severely penalized by ‘non-hom*ologies’ outside the conserved region.

To avoid this difficulty, Smith & Waterman proposed a very simple modifi-cation to the Needleman & Wunsch algorithm: score ‘debt’ incurred in the non-hom*ologous regions of table T(i, j) must be abolished in order to locate locallysimilar regions. This principle consists in annulling the score as soon as itbecomes negative; otherwise the calculation remains identical. There is then aquadruple alternative:

Figure 2.10 is the same as Figure 2.9 with the calculation of T(i, j) modified.No longer satisfied with looking at the score in the bottom-right cell of the

matrix, as with the Needleman-Wunsch algorithm, we now look at all cells and

T i jT i j M S i S j

T i j

T i j

, max, ,



( ) =− −( ) + ( ) ( )( )

−( ) +−( ) +

1 1



1 2



















0 0 0 0 0 0 0 0 0 0 0 0 0 0

0 6 0 0 0 0 0 0 0 0 0 3 0 0

0 0 8 5 0 0 0 0 2 1 2 0 1 0

0 0 2 13 7 1 0 0 2 3 3 0 0 0

0 0 0 7 13 7 1 0 0 1 2 1 0 0

0 0 0 0 7 12 6 0 0 0 0 1 3 0

0 0 1 1 1 6 14 8 2 5 1 0 0 0

0 0 5 3 0 0 8 12 13 7 10 4 0 0

0 0 0 5 1 0 2 8 12 13 7 8 2 0

0 0 0 0 4 0 0 2 6 10 11 6 6

0 0 0 0 0 2 0 0 1 6 9 13 7 9

0 0 0 0 0 0 0 0 0 0 5 7 10 4

0 0 0 0 1 1 0 0 0 0 0 3 5 9

0 3 0 0 0 0 0 0 0 0 0 7 2 3

0 0 3 0 1 1 0 0 0 0 0 1 6 1

0 0 0 0 0 0 0 0 0 0 0 0 0 15




Figure 2.10 Alignments using the local method of Smith & Waterman. This algorithm finds theoverall alignment, but also detects others, such as the one indicated on the right, whose score is12.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (50)

find the one with the highest score. Using that score, we can identify the most hom*ologous segment in the two sequences, as was done using the globalalignment method. It is also possible to examine all cells whose scores exceeda threshold value chosen by the user. This finds alternative alignments, like theone indicated in Figure 2.10, which sometimes can reveal interesting biologicalproperties, such as imperfect repetitions of a pattern in the sequence beingstudied.

Dynamic programming has yielded numerous variants and improvements inthese methods. The principle, which is used in most programs accessible online via the web, involves calculating the cost of insertions/deletions. In the twoabove-described approaches for an insertion or deletion of length n, the cost is n × ∆. This linear cost in terms of n is simple to compute, but not very realistic from the biological point of view. The introduction of a discontinuityin the alignment is the penalty. However, the length of this discontinuity is notcritical. In protein sequences, these breaks in alignment most often occur inpeptide chain loops exposed on the surface of the 3D molecule. Indeed, thelength of these loops may be relatively variable without at all disturbing the heart of the molecule, where practically no insertions or deletions are found. To take this into account, we use an affine cost for insertions/deletions of length n:

The parameter a is the cost of opening the discontinuity, and b is the cost of extending it. Again, choosing a and of b is relatively empirical; a < b, anda < min {M(a, b)}, where M is the comparison matrix (BLOSUM, PAM, etc.)For example, with the BLOSUM62 matrix, one commonly uses a = −10 and b = −2, or neighboring values. Applying this cost function does not question theuse of dynamic programming; it just makes the procedure of calculating thematrix T(i, j) slightly more complicated.

2.7 Fast heuristic methods

The two variants of dynamic programming just described are efficient andoptimal, but their costs in terms of calculation time and memory are quadratic;specifically, O(n, m), where n and m are the respective lengths of the twosequences being studied. This cost is very reasonable when comparing twosequences of several hundred to several thousand nucleotides or amino acids,but becomes high when carrying out a systematic search of sequence databasescontaining around 109 to 1010 residues. However, this latter type of search isbecoming one of the dominant bioinformatics applications. Indeed, massive

cost n= +a b


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (51)

genomics sequencing programs lead to the identification ‘by the kilometer’ ofthousands of genes whose functions are most often unknown. Each new genemust be compared with the databases in order to try to find a known hom*olog,which allows attribution of a possible function. This need has led to the development of much more rapid heuristic methods that provide high-quality (although usually not optimal) alignment solutions at very reasonablecost.

The diagonal strip method: FASTA and FASTP

The inspiration for this first heuristic approach is found in a graphic represen-tation of the hom*ology between two sequences S1 and S2, known as a dot-plot.If their respective lengths are n and m, they describe a matrix of n × m pixels.Each pixel of coordinates i, j of the image generated is either black or white,according to whether the value of the comparison matrix, M(S1(i), S2(j)) is aboveor below a given threshold. In other terms, if S1(i) and S2(j) are identical orsimilar, the corresponding pixel will be colored. When two sequences S1 and S2

share hom*ologous regions, the regions appear as diagonal lines on the image(cf Figure 2.11).

When two sequences share regions that present hom*ologies, we notice thepresence of very large diagonal segments localized within a narrow band. This





hom*ology diagonal

Figure 2.11 Dot-plot comparison of the sequence of the androsterone nuclear receptor (abscissa)with that of another nuclear receptor of unknown function (ordinate). The positions of the func-tional domains of the androsterone receptor are indicated above. The most significant alignmentappears as a series of diagonal lines, which are highlighted in the figure. The unknown receptoris shorter than the androsterone receptor and does not seem to have a regulatory domain.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (52)

property has been used to develop an overall heuristic alignment method foridentifying the diagonals. Two slightly different variants of this method exist,one for protein sequences, called FASTP (for FAST Protein alignment) and theother for DNA sequences, called FASTA (for FAST nucleic Acid alignment).

To improve asymptotic efficiency with respect to dynamic programmingmethods, it is of course necessary to avoid calculating the entire dot-plot, whichwould clearly be O(m, n) for two sequences of respective length m and n. FASTAand FASTP simplify this step by the following:

1. First, tabulate the lists of k-tuples of nucleotides or amino acids appear-ing in sequence S1 and of the positions at which they appear. Typically, thevalue of k will be 4 to 6 for nucleic acids and 2 for proteins, because ofthe smaller size of the alphabet used (4 nucleotides compared with 20amino acids). For example, for the following DNA sequence,


we obtain the following list of 4-tuples and their positions:

AAAA 13, 14, 15 CGAT 7 GAAA 12 TGGA 10AAAT 16 CGCG 5 GATG 8 TGGC 2ATGG 1, 9 GCGA 6


This list is constructed in linear time as a function of the length n of S1.

2. Create a score table for the n + m − 1 diagonals of the matrix. All valuesin this table are initialized at zero.

3. Sweep through the S2 sequence looking for k-tuples tabulated for S1. Eachtime an S1 k-tuple is found in S2, locate the diagonal(s) on which it appears.Each of them is identified in an unequivocal manner by the difference i − j, where i and j are the respective positions of the common k-tuple inS1 and S2.

4. Once the totality of S2 has been analyzed, look for a maximum hom*ologyband like the one highlighted in Figure 2.11, using the score table of thediagonals.

5. Construct the alignment limited for this diagonal band, attaching thevarious k-tuples end-to-end. It is possible to refine the alignment by usingthe Needleman & Wunsch method to fill in the remaining gaps.


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (53)

The behavior of this method crucially depends on the k parameter. Increas-ing its value accelerates the search, since the mean number of occurrences of ak-tuple in a sequence of given length diminishes as k increases. Thus, there willbe far fewer cases to treat during step 3, which is one of the costliest in termsof time. The price to be paid is a loss in the sensitivity of the method, since thereis a risk of missing weak hom*ologies for which there are only a few conservedk-tuples. If the identity rate between two DNA sequences is 60 percent, out of100 nucleotides, an average of 13 quadruplets will be conserved (0.64 = 0.129),versus only 1 if the identity rate falls to 35 percent5.

Applying these simplifications renders FASTA and FASTP very efficient inaligning relatively conserved sequences. But for weakly hom*ologous proteinsequences, the limiting number of common k-tuples often requires resorting tothe value k = 1, which causes FASTP to lose much of its efficiency. In addition,using a comparison matrix such as BLOSUM62 to account for replacements bychemically close amino acids further complicates the task. In order to search allprotein databases systematically, FASTP therefore remains a useful but – sinceit is rather slow – limited tool. This limitation is not likely to be lifted by theconstant increase in computer power, since it is offset by an at least equallyrapid increase in the size of the databases to be searched.

K-tuples: a general search method

The FASTA strategy, which consists in tabulating a list of k-tuples that overlapin a sequence, is a very general strategy for seeking partial or total sequenceoverlaps originally proposed by J. Ninio and J.-P. Dumas in 1982. For example,it is also used in contig reconstruction algorithms during genome assembly (seethe chapter on sequencing). Much longer k-tuples are used for this latter appli-cation, since we are looking for sequence identities, not hom*ologies. When twoclones overlap, one would expect that the sequence fraction they have incommon be identical (within the limits of sequencing errors). Thus, we gener-ally try to use much longer k-tuples that are present only once in the entiregenome. According to the hypothesis of the equal distribution of the fournucleotides (for a discussion of DNA composition, see Chapter 5), we take k >log4 L, where L is the genome length, which is 11 nucleotides for a bacterialgenome and 16 for the human genome.


5 Applying this simplistic calculation supposes that conserving a given position in the sequence is inde-pendent of the position of its neighbors, which is generally not the case. Fortunately for this method,identity regions often form compacted zones separated by more variable regions. This reflects the factthat the selective pressure of evolution is not uniform, but preferentially exercised at sites that are essen-tial for the functioning of the molecule. This explains why FASTA and FASTP ‘work’ so well in spite ofthe drastic simplification represented by k-tuples.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (54)

In principle, k-tuples may be tabulated either for the sequence being studied,as FASTA does it, or for the entire sequence database. The latter approach wasused very successfully when sequence databases first became available online.The k-tuple table, which must be recalculated each time the database is updated,is a gigantic index of the positions of the various k-tuples. Thus, for example,for the FFKR tetrapeptide, the table contains a list of pointers pointing towardall the sequences that contain it, with the corresponding position in the following sequence:











Figure 2.12 Example of an alignment produced by FASTP between the androsterone receptorsequence (‘Query’) and the retinoic acid receptor sequence RXRB (‘Match’).

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (55)

This strategy seems attractive, since simply reading the index table providesthe list of candidates that are hom*ologous to the sequence being studied.However, this is rarely done these days, for three reasons:

• Database size increases much more rapidly today, and the quantity of workrequired to retabulate the occurrence of k-tuples at each update becomesprohibitive.

• In order to vary the k parameter, as many index tables as desired k valuesmust be compiled, which makes the calculations more complicated. Inaddition, there will be few k-tuples for low k-values, and each index tableentry will contain a great number of references to the database, whichmakes the process very long.

• Because of their large size, sequence databases cannot completely fit intothe main memories of computers, and are therefore generally stored inexternal disks. Consulting database sequences from the table of k-tuplesis an operation that requires direct access (random) to the database entries,which is very inefficient. In the inverse strategy utilized by FASTA, sweep-ing through the base to look for k-tuples of the sequence being studied isa sequential operation, which is much less costly in terms of access time.

For these same reasons, the BLAST algorithm which we will now consider,also involves tabulation of the sequence studied; not that of the database.

hom*ology by pieces: the ultra-rapid BLAST method

In order to further improve efficiency, another approach was developed specif-ically for use in aligning protein sequences. Like the Needleman & Wunsch algo-rithm, FASTP is an overall alignment method. The hom*ology sought often maybe weak and localized, hence difficult to detect by the two methods mentionedabove. BLAST (Basic Local-Alignment Search Tool) is a heuristic method specif-ically developed to compare an unknown protein sequence with a set ofsequences found in protein databases. BLAST detects locally hom*ologous shortsegments (a few dozen amino acids) at a linear cost proportional to the size ofthe database searched. The main application of BLAST is the use of known



... ...




FFKR androsteronereceptor,position582 proteinX,position


→→ → → →→



Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (56)

proteins to rapidly detect significant hom*ologies, in the attempt to attribute a function to the protein being studied. BLAST uses amino acid comparisonmatrices such as BLOSUM62 to evaluate the pertinence of these alignments and to attribute a ‘plausibility’ score to the obtained alignment in evaluatingthe probability that the observed resemblance is due to chance6.

Like FASTA and FASTP, the BLAST principle is based on the use of k-tuplesof amino acids in the sequence analyzed in order to reduce complexity. BLASTstill uses long k-tuples (at least four amino acids) to maintain high selectivity;that is, so as not to have to deal with too many chance occurrences of thesesame k-tuples when sweeping through the sequence database. In order toprevent this high selectivity from imposing a loss of sensitivity, a list of k-tuplessufficiently hom*ologous to each k-tuple is compiled. For this purpose, a com-parison matrix is used that allows calculation of a score corresponding to thealignment without inserting the two k-tuples.

Taking the example of the androsterone receptor with k = 4, the FFRKquadruplet mentioned above appears in position 582. Aligning it with itself, theBLOSUM62 matrix yields a score of 22:


score: 6655 total: 22

Looking for all tetrapeptides whose alignment scores are superior to a thresh-old H, a user-defined parameter, for example, with H = 17, the following hom*ol-ogous peptides for FFKR result:

Score = 22 Score = 19 Score = 18 Score = 17



This operation is repeated for all the k-tuples of sequence S, which allowsconstruction of a list L of k-tuples that are ‘close’ to S1. For any k-tuple w of


6 Since databases today contain several dozen billion nucleotides and amino acids, the sporadic appearance of local alignments that are credible to the eye but still due to chance is practically inevitable.More thorough critical analysis by the biologist is thus usually necessary when the hom*ology score islow.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (57)

L there exists a k-tuple v of S1 such that the score (v, w) ≥ H. The position atwhich the k-tuple appeared in sequence S also figures in list L. The position ofthe FFKR pattern in the androsterone receptor sequence is associated with allthe quadruplets in the example in the above table.

The list L of k-tuples is then used to scour the sequence databases. Each timea k-tuple belonging to L is located, a hom*ology begins that guarantees aminimum score for H. An attempt is then made to extend the alignment of S,using the sequence found around the hom*ologous k-tuple. The BLAST strategyis to seek maximal hom*ology segments; the alignment is prolonged progres-sively, first on one side, then the other, using the comparison matrix to calcu-late the scores. The method employed proceeds according to the followingscheme:

max_score ← initial alignment score of the two k-tuplesmax_length ← kDo

Prolong the alignment of a residue on each sequenceCalculate the resulting score, extended_score

If extended_score > max_score thenmax_score ← extended_score; increment max_length

While extended_score ≥ max_score-tolerance

The tolerance factor introduced in checking the loop allows elongation of the alignment even if the score drops slightly below the current maximum, inthe hope that when elongating later segments, favorable coincidences of aminoacids will be encountered that raise the score. If the score descends below themax_score-tolerance, BLAST stops and recovers the maximal score valueand the maximum length encountered. This operation is carried out for bothsides of the initial k-tuple, so as to obtain local alignment of the maximum scorebetween the two sequences, known as the Maximum Scoring Pair (MSP).Finally, BLAST displays all local alignments whose scores exceed a threshold setby the user.

The cost in terms of calculation time for a BLAST search sensitively dependson the various parameters of the heuristic algorithm: k, the length of the k-tuples H, the hom*ology score threshold, and to a lesser degree, the toleranceutilized for extension, starting from the k-tuples found. For example, if the Hthreshold is raised, it will reduce the number of k-tuples in list L, since only the most hom*ologous k-tuples will be retained. The number of coincidencesbetween the database examined and list L will be reduced proportionally, accelerating the calculation. Regrettably, when H is increased, there is also therisk of allowing certain segments whose hom*ologies are more distant to escape,yielding lower scores. A compromise must be found between the sensitivity of


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (58)

the method and its effectiveness. After much statistical and empirical testing,the following parameter values are now conventionally used to search proteinsequence databases with BLOSUM62 or PAM250:

Using these parameters, a BLAST search is around one order of magnitudefaster than using FASTP, and two to three orders of magnitude faster than withthe dynamic programming algorithms of Smith & Waterman or Needleman & Wunsch. As is the case for FASTA and FASTP, even though today there arenumerous variants and improvements for BLAST, all start from the basic prin-ciple described above. The most recent specifically allow fusing several com-patible local alignments into a single alignment, taking short insertions anddeletions into account to a limited degree.

2.8 Sensitivity, specificity, and confidence level

Using FASTA, FASTP, or BLAST to search sequence databases almost alwaysyields a series of more or less high-score alignments. In view of this result, thebiologist must consider several questions:

• How much confidence can one have in the alignment result produced bythe program? In more rigorous terms, is it possible to evaluate the prob-ability that the result obtained is random?

• Has the program identified all the sequences that are hom*ologous7 to thesequence sought?

• For what percentage of the sequences found is the hom*ology really signif-icant from the biological point of view?

These three questions concerning the level of confidence, sensitivity, andspecificity of the heuristic algorithm used are partially related to each other anddepend on the parameters applied. Sensitivity and specificity are somewhatantagonistic, since if one relaxes the stringency of the parameters, for example,by lowering the BLAST hom*ology threshold H, there is a greater risk of obtain-ing false positives. It was thus crucial to devise a tool capable of evaluating thelikelihood of an alignment. A statistical model of sequence alignment withoutinsertions and deletions developed by the authors of BLAST provides the means

K H= = =4 17 20tolerance


7 By hom*ologous sequences is meant all those for which the Smith & Waterman algorithm yields a localhom*ology score equal or superior to a threshold used by the heuristic algorithm.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (59)

for conducting this analysis. However, a detailed presentation of it would extendbeyond the framework of the present discussion; therefore only its main resultsare presented here:

Briefly, a comparison matrix (BLOSUM62, for example) is used to comparea sequence of length n with a database sequence of length m. Take the case inwhich the sequence studied presents no ‘biological’ hom*ology with the databasesequences, in order to look only at ‘fortuitous’ similarities. We are interested inthe distribution of the number N(s) of alignments whose scores exceed the values. Since it is roughly a Poisson distribution, the following expression describesthe probability of an accidental alignment of score greater than s:

E(s) is the expected value of N(s) where l and K are positive constants thatdepend on the frequencies of the amino acids in the databank and on the coef-ficients of the comparison matrix. One can either calculate l and K explicitly8

or adjust them from a real distribution of the scores obtained with a sufficientlylarge sample of non-hom*ologous sequences. The resulting probability thus hasthe expression:

Its shape as a function of the threshold s may be seen in Figure 2.13, in whichit drops rather sharply over a given score value.

Based on its score, it is possible to determine the probability that each align-ment produced by BLAST is purely accidental. The best ‘biological’ alignmentshave infinitesimal probabilities of random statistical occurrence (10−20 to 10−100),generally leaving no doubt of their significance. At around 10−1 to 10−5, theremay be a degree of ambiguity, in what is known as a twilight zone, in whichcase a posteriori validation by the biologist is generally necessary.

Application: BLAST search for hom*ologs of the androsterone receptor

As stated above, the main function of BLAST is to carry out systematic searchesof databases. We have undertaken the same exhaustive search for hom*ologs ofthe human nuclear androsterone receptor, which is 919 amino acids long. To

p score e K m n e>( ) = − − −s ls1 . . .

p score e E>( ) = − − ( )s s1


8 For those who are curious, l is the strictly positive solution of the equation

with p(a) being the frequency of amino acid a and M(a, b) being the coefficient of the comparison matrix(BLOSUM62). There is no closed form for K, which is the sum of the terms of a converging series thatdepends on the same p(a) and M(a, b). Since the convergence is rapid, it is possible to directly calculatea value close to K by summing the first terms.

p a p b ea b

M a b( ) ( )∑ ( )



amino acids


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (60)

do this, we selected a non-redundant9 and documented database of well-characterized protein sequences known as SWISSPROT. Its May 2005 versioncontained 180,652 entries, for a total of 65 million amino acids.

During analysis of the database, BLAST found more than 1.8 million ‘posi-tive’ k-tuples for which it attempted extension, for 4,536 of which it was possible to prolong the alignment. Of these alignments, 268 satisfied the selectedminimal score criteria, among which 45 implicated protein sequences of humanorigin, including (of course) the sequence of the androsterone receptor (cf Figure2.14). Some of these sequences code for known nuclear receptors involved inresponses to various stimulatory molecules, such as steroid sex hormones,vitamin D, corticoids, and thyroid hormones. This functional relationship is ana posteriori verification of the postulate stated at the beginning of this chapter:

More interestingly, BLAST analysis also identifies at least as many highlyhom*ologous protein sequences of previously unknown function. Their high levelof similarity, evaluated by the alignment score, would seem to indicate that theyare nuclear receptors. They are known as ‘orphan receptors’ (see an example inFigure 2.15). In most cases, the molecule that plays the role of cofactor, suchas androsterone or corticosteroids, is unknown. The discovery of orphan recep-tors was made possible by bioinformatics analysis; today they pose numerousquestions, as well as suggesting various therapeutic perspectives.











Figure 2.13 Value of the probability of an alignment score higher than a threshold s as a func-tion of that same threshold. The abscissa is expressed in arbitrary units.

9 Today, protein sequences are most often obtained by in silico translation of DNA open reading frames.Numerous copies of some DNA sequences are present in databases (genomic sequences, cDNA sequences,EST, etc . . .) derived from several tissues. The SWISSPROT database is ‘curated’ so as to render it non-redundant. It thus contains only one entry per protein, plus comments concerning the protein’s activityand functional domains. This sorting and annotation work explains the reduced size of the SWISSPROTdatabase compared with DNA sequence databases.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (61)

• What are the cofactors of these receptors?

• What are their functions? With which metabolic and/or physiologicalresponse(s) are they associated?

• Is it possible to modulate their functions using exogenous molecules?Could these molecules constitute a new class of medicines? Most known


sequence hom*ology⇔functional hom*ology

Sequences producing significant alignments: Score(bits)

E Value


Figure 2.14 Human protein sequences identified by BLAST as presenting hom*ologies with theandrosterone nuclear receptor. The columns on the right indicate alignment scores obtained usingthe BLOSUM62 matrix and the mathematical expectance E(s) of the number of random alignmentsof a higher score than that value.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (62)

nuclear receptors have turned out to be interesting therapeutic targets, andthe discovery of new receptors enlarges the perspectives accordingly.

In order to identify the functions of orphan receptors, and in general, thefunctions of proteins newly identified using bioinformatics, a new methodologyhas been devised: reverse genetics. Instead of looking for genes associated witha phenotype, then mapping and isolating them in order to derive their sequence,we start with bioinformatics sequence analysis in an attempt to discern the func-tion or phenotype associated with an interesting candidate. For example, wecan inactivate or modify the corresponding mouse gene, if it exists, then try todetermine the consequences. One can also look for molecules capable of bindingto a protein of unknown function, in order to study their effect in the animal,and eventually in humans.

2.9 Multiple alignments

Until now, we have not considered the problem of aligning segments pairwise.When carrying out a database search, several sequences that present similaritieswith the sequence being studied are often encountered. To compare all thesehom*ologous sequences among themselves simultaneously, it is natural to try toalign N sequences together, in the manner shown in Figure 2.4. This problemof multiple alignment has many interesting biological applications:

• Multiple alignments allow detection of regions that are conserved over evolution. These regions very often correspond to domains associated witha key function of the molecule.

• Strictly conserved amino acids or nucleotides, like those that appear in grayin Figure 2.4, often play a direct role in the function. Using multiple align-ments, we are able to identify amino acids implicated in the catalysis orselective binding of a substrate by an enzyme.



Score = 68.7 bits (165), Expect = 5e-11 Identities = 29/77 (37%), Positives = 42/77 (53%), Gaps = 1/77 (1%)


Query: 618 KCYEAGMTLGARKLKKL 634 KC GM+ A + ++


Figure 2.15 Example of local alignment produced by BLAST. It concerns the hom*ology betweenthe androsterone nuclear receptor and an orphan nuclear receptor of unknown function.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (63)

• They sometimes permit a posteriori validation of an alignment found byBLAST or FASTA, whose low score is in the ‘twilight zone,’ where it isdifficult to ascertain its biological significance. If the regions and/or hom*ol-ogous amino acids are precisely the same as those conserved over evolu-tion, there is a strong likelihood that the alignment found reflects biologicalreality.

Multiple optimal alignment

Dynamic programming methods (Needleman and Wunsch and Smith andWaterman) may in principle be directly generalized to N sequences. The notionof alignment score as now defined must be extended to include use withBLOSUM and PAM comparison matrices. Given a multiple alignment column,the corresponding score is generally defined as the sum of the scores of all pair-wise combinations of sequences in the alignment:

For insertions/deletions, we extend matrix M as above, defining:

The same matrix column can contain an insertion in several sequences. Wecan thus also define a score for the alignment of two gaps, which we generallytake to be zero:

The fact of rendering this cost zero has the effect of favoring the grouping ofinsertions in the same columns in the multiple alignment. This is rather naturalbecause, as we have already discussed, for proteins, insertions/deletions areusually grouped in polypeptide loops located on the surface and not dispersedin the interior of the molecular sequence.

These extensions in the calculation mode and of the comparison matrixpermit calculation of an overall score for a multiple alignment containing inser-tions and deletions. This method of calculating the score in a multiple align-ment is called the sum of the pairs score. It is thus possible to search for anoptimal alignment, for which, by definition, the score of the sum of the pairswill be maximal.

To find it, we always calculate a table T of the partial scores. This table isnow N-dimensional, with T(i1, i2, . . . iN) = the best score for a partial alignmentwith the sub-sequences {i1, i2, . . . iN}. This table is calculated in the same way,taking the maximum of an alternative to 2N − 1 terms, and describes all pos-

M − −( ) =, 0

∀ −( ) = −( ) = <a M a M a, , , ,∆ 0 insertion penalty.

Score column i M S i S ik lk l N

( ) = ( ) ( )( )≤ < ≤∑ ,



Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (64)

sible combinations of insertions and ‘gaps’ in the N sequences. The followingis an example of the seven terms of the alternative for three sequences, S1, S2,and S3, where the score function is what was defined above for a column andwhere a dash (‘−’) designates a gap in the alignment.

While in principle this method is only slightly more complex than for twosequences, the computing time clearly becomes prohibitive as N increases. ForN sequences of length n, there are nN elements to be calculated in table T, foreach of which it is necessary to evaluate 2N − 1 scores, each requiring the sum-mation of N(N − 1)/2 terms, which engenders an overall algorithmic cost of

The required amount of memory space itself increases ‘only’ as nN. Whereasone second of computer time is needed to align two protein sequences of 100amino acids, ten minutes are required to align three sequences, and nearly threedays to align four. For nine or more sequences, the computation time requiredexceeds the age of the universe . . . Clearly, this method cannot be used inroutine practice. Today, families of hom*ologous proteins, such as those ofnuclear receptors, consisting of several dozen or even several hundred members,are common. The development of high-performance, good quality, multiplealignment algorithms is still an open problem and an area of active research.Numerous heuristic strategies exist that generally function in a satisfactorymanner when the set of proteins or DNA to be aligned consists of sequencesthat are rather similar to each other and contain few insertion and deletion sites.These algorithms function with more difficulty when the hom*ology is weaker.

‘Profile’ alignment: a simple and efficient heuristic method

Since it is difficult to undertake simultaneous alignment of N sequences directly,a simple approach is to progressively align them pairwise, starting with the mostsimilar. Other sequences are then aligned with the first alignments, and others

N NnN N−( ) −( )1

22 1

T i j k

T i j k score S i S j S k

T i j k score S j S k

T i j k score S i S k

T i j k, , max

, , , ,

, , , ,

, , , ,

, ,( ) =

− − −( ) + ( ) ( ) ( )( )− −( ) + − ( ) ( )( )

− −( ) + ( ) − ( )( )− −( ) +

1 1 1

1 1

1 1

1 1

1 2 3

2 3

1 3

scorescore S i S j

T i j k score S k

T i j k score S j

T i j k score S i

1 2







( ) ( ) −( )−( ) + − − ( )( )

−( ) + − ( ) −( )−( ) + ( ) − −( )

, ,

, , , ,

, , , ,

, , , ,


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (65)

are aligned with those alignments, until all the sequences are incorporated intoan overall alignment. These intermediate alignments are called profiles. Onlytwo elements, either sequence/sequence, sequence/profile, or profile/profile, arealigned at each step, using a method derived from the Smith & Watermanmethod, which considerably limits the complexity of the algorithm.

After the two sequences have been aligned, the gaps are filled with a neutralcharacter that does not correspond to any amino acid or nucleotide (forexample, − or ×). We thus obtain two sequences of the same length that con-stitute a profile. We can then align two profiles, or a profile and a sequence,with each other, still using the dynamic programming approach. It is enoughjust to modify calculation of the alignment score by means of the sum of thepairs method, as described above.

One of the properties of this method is that gap creation is definitive (‘oncea gap, always a gap’). If an insertion is created in a sequence during an earlypairwise alignment step, it cannot be abolished in a later step, even if that wouldbe favorable in terms of the overall score. This is certainly a weakness in themethod, and in order to minimize its harmful effects, it is crucial to avoid align-ing the least hom*ologous sequences in the early phases of multiple alignment.Indeed, not doing so would risk introducing insertions at the wrong places. Sincethis process is irreversible, the overall alignment quality would be stronglyaffected. All methods of profile alignment use a phylogenetic approach, whichfollows the tree of evolutionary ancestry among the various sequences (cf thenext paragraph: Construction of a phylogenetic tree).

The first step is usually to reconstruct this tree. The approach used consists infirst carrying out pairwise N(N − 1)/2 sequence alignments. Starting with thescores of these alignments, we construct a table of the evolutionary distancesbetween each pair of sequences. The lower the score, the greater the correspond-ing distance. The function employed to determine the distance between twosequences i and j starting from an alignment score is necessarily somewhat arbi-trary. Among those most frequently used are expressions of the following type:


2 x 2 alignment with insertion/deletion

filling of gaps and fusion to form a profile

Figure 2.16 Profile alignment. On the left side, two sequences are aligned using a classical dynamic programming method. The gaps are then filled by replacing them with a neutral character.The profile obtained may then be aligned to another sequence (right side), or to another profile.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (66)

where S is the alignment score, Srandom is the score obtained after random per-mutation of the amino acids of one of the two sequences, and Sid is the maximumscore obtained by aligning one of the two sequences with itself.

The table of distances Di,j is then used to construct a tree that serves as aguide used for iterative alignment. We start by aligning the sequences that cor-respond to the closest branches, following the branches to the roots of the tree.Figures 2.17 and 2.18 show an example of a phylogenetic tree and of multiplealignment constructed using the CLUSTAL program. CLUSTAL (for ClusterAlignment) is one of the most popular multiple alignment programs, availableon most computer platforms and online on various web servers.


S Si j


id random, log ,= − −


Figure 2.17 Example of a phylogenetic tree constructed from the ten sequences of nuclear recep-tors listed in Figure 2.14. The lengths of the tree branches are proportional to the evolutionarydistances (indicated above). The multiple alignment constructed from this tree is indicated inFigure 2.18.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (67)

Construction of a phylogenetic tree

Phylogenetics consists in reconstructing filiations and ancestral links amongvarious species in the context of the Darwinian Theory of Evolution. Such clas-sifications were long based on sometimes rather arbitrary external characteris-tics. Since the introduction of DNA sequencing, the criteria of similarity amongthe sequences of certain genes have been used to directly deduce weak evolu-tionary ancestry. These criteria permit the construction of genealogical specia-tion trees known as phylogenetic trees.

Two families of algorithmic methods are used for reconstructing phylogenetictrees:

• Progressive grouping methods, which use a distance matrix Di,j betweenpairs of sequences. These methods utilize the metrics determined by thesedistances to construct a binary tree in an iterative manner. First they groupthe sequences into branches, then the branches among themselves, start-ing with the nearest ones and ending with those most distant.

There are several variants of progressive grouping methods. The simplestgroups the two i and j vertices between which the distance is minimal ateach iteration. They are then replaced by a new vertex k, and the distancebetween it and the other vertices is determined by calculating the arith-metic mean of the distances to the fused vertices: Dk,m = (Di,m + Dj,m). Ateach iteration the number of vertices is reduced by one. When only oneremains, the tree has been constructed. This method, called UPGMA(Unweighted Pairgroup Method using the Arithmetic Mean), is very simpleand rapid, but in unfavorable cases does not yield an optimal tree in terms





Figure 2.18 Portion of the multiple alignment of 10 sequences of nuclear receptors indicated inFigure 2.17. Positions at which at least 7 out of 10 sequences are identical are highlighted. Strictconservations are framed. The region represented corresponds to the cofactor liaison domain.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (68)

of total branch length. Nevertheless, with a few modifications, it is pos-sible to overcome this shortcoming.

• Methods based on the individual nature of each amino acid or nucleotide:these take into account the number of mutations (changes in the sequence)necessary to go from one sequence to another. The sequence of thecommon ancestor of that branch is located at each vertex of the tree. Theterminal extremities (‘leaves’) are the sequences of the current species. Thelengths of the various branches that descend from a vertex correspond tothe number of mutations separating the common ancestor from its descen-dents. The so-called maximum parsimony method is used to find anoptimal tree the sum of whose branch lengths is minimal. The tree explainsthe biological diversity present starting from a minimum number of muta-tions. The search for this optimal tree is carried out using a systematicapproach, exploring all possible configurations. In view of the explosiveasymptotic character of the number of different N-leaved binary trees, par-simony methods resort to an elaborate branch-pruning strategy in theirsearch, which consists in the computer science principle known as branchand bound. We will not give a detailed description of this algorithm here.

Beyond phylogenetic analysis, which is certainly interesting with respect toevolutionary theory, several groups of bioinformaticians have become interestedin the systematic analysis of databases in their search for data concerning mul-tiple alignments among conserved protein regions. They realized that Natureoften reutilizes the same sequence element to perform analogous functions.



j k

Figure 2.19 Schematic principle of the construction of a tree by the UPGMA method. The twoclosest vertices i and j are replaced by a new vertex corresponding to their arithmetic mean k. Onthe right: the finished tree. In the center: the last summit added; that is, the root of the tree. Inthis simplified example, the distances Di,j are represented by the Euclidian distances in the plane.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (69)

These conserved elements are generally structured into independent domains inthe corresponding proteins and fulfill one of the ‘tasks’ of the molecule: DNAor RNA binding, ligand binding, enzymatic activity, association with other cellcomponents, etc . . . Nature is thus playing with a kind of molecular Lego,assembling various domains in constituting a new protein. A great many of thesedomains have been inventoried, and today specialized databases exist thatcompile multiple sequence alignments for each of them. For example, thePRODOM database developed by INRA (the French Institut National de laRecherche Agronomique) in Toulouse, France ( in May 2005 listed around 240,000 domains common to at leasttwo proteins. PRODOM is part of the InterPro project (, which integrates the data of most protein domain andfamily databases. The mean length of these conserved domains is slightly morethan one-hundred amino acids. For example, Figure 2.20 illustrates the break-down of nuclear androsterone and glucocorticoid receptors into domains.

The systematic alignment of these domains provides the biologist withanother very useful tool: the identification of characteristic motifs. When mul-tiple alignments like those in Figures 2.4 and 2.19 are carried out, there are positions in conserved regions (highlighted in Figure 2.4 and framed in Figure2.19) at which either hom*ologous amino acids, like those in the first column ofFigure 2.4, or only F or Y, two aromatic amino acids, are systematically found.A simple method for determining whether an unknown protein contains a conserved domain is to look not for a sequence alignment, but for the presenceof a pattern of a few amino acids. The definition of such a pattern is of the following type:

‘F or Y, followed by any amino acid, followed by C, followed by any fiveamino acids, followed by C.’

This can be written in a more synthetic manner as follows:

where the brackets represent the alternative, x indicates any amino acid, andthe parentheses surround a numerical value that is the number of repetitions of

FY C C[ ] × × ( )6




100 acides aminés

15754 186097


Figure 2.20 Breakdown of the nuclear receptors for androsterone and glucocorticoids by PRODOM.Each ‘box’ corresponds to a domain. The C-terminal third of the two receptors presents a certainnumber of common domains associated with DNA and ligand binding.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (70)


the preceding symbol. This type of pattern corresponds to what computer sci-entists call a ‘regular expression’ (cf, Chapter 4, Genetic Information and Bio-logical Sequences). There are numerous programs to search for these, especiallyusing UNIX (grep, lex). Several slightly different syntaxes exist for writing theseregular expressions; the one presented above is different from the one providedin UNIX tools. It corresponds to one used by biologists, who have compiled adatabase of functional patterns and regular expressions that can be associatedwith them. Called PROSITE (, it is maintained by the Swiss Institute of Bioinformatics in Geneva, which also runs the SWISSPROT database ( It contains 1,400 docu-mented patterns. The pattern that corresponds to nuclear receptors fits the following definition:

In Figure 2.2, we note that the three receptor sequences conform well to thesyntax of the PROSITE pattern, whereas the bacterial protein sequence whosehom*ology is random does not. This pattern correctly ‘recognizes’ 229 nuclearreceptors of the 233 identified to date, and no other proteins. We say that ‘thespecificity is 100 percent’ (0 false positives out of 229) and that its sensitivity98.3 percent (229 positives out of 233).




Regular expressions are powerful tools, since algorithms exist that allowsearching sequences in linear time (O(n)), using what are called finite stateautomata (cf. Chapter 4, Genetic Information and Biological Sequences). Owingto its great efficiency, the systematic search for known patterns is today a pre-cious complement to research methods for similarity by alignment.


Altschul S.F., et al. (1990). Basic local alignment search tool. J Mol Biol 215: 403–410.Altschul S.F., et al. (1997). Gapped BLAST and PSI-BLAST: a new generation of protein

database search programs. Nucleic Acids Res 25: 3389–3402.Bairoch A. (1991). PROSITE: a dictionary of sites and patterns in proteins. Nucleic

Acids Res 19 Suppl: 2241–2245.

C-x 2 -C-x- DE -x 5 - HN - FY -x 4 -C-x 2 -F-F-x-R( ) [ ] ( ) [ ] [ ] ( ) ( )

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (71)

Gotoh O. (1987). Pattern matching of biological sequences with limited storage. ComputAppl Biosci 3: 17–20.

Henikoff S., Henikoff J.G. (1992). Amino acid substitution matrices from protein blocks.Proc Natl Acad Sci USA 89: 10915–10919.

Henikoff S., Henikoff J.G. (1993). Performance evaluation of amino acid substitutionmatrices. Proteins 17: 49–61.

Higgins D.G., Sharp P.M. (1988). CLUSTAL: a package for performing multiplesequence alignment on a microcomputer. Gene 73: 237–244.

Karlin S., Altschul S.F. (1990). Methods for assessing the statistical significance of molec-ular sequence features by using general scoring schemes. Proc Natl Acad Sci USA 87:2264–2268.

Needleman S.B., Wunsch C. D. (1970). A general method applicable to the search forsimilarities in the amino acid sequence of two proteins. J Mol Biol 48: 443–453.

Pearson W.R., Lipman D.J. (1988). Improved tools for biological sequence comparison.Proc Natl Acad Sci USA 85: 2444–2448.

Smith T.F., et al. (1981). Comparative biosequence metrics. J Mol Evol 18: 38–46.


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (72)

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (73)

3Comparative genomics

Formerly, biologists used isolated genes to study molecular evolution, but sinceit has become possible to completely determine genomic sequences, evolution isinvestigated at a higher level: the whole genome. How can genomic informa-tion be used to understand the evolution of genomes?

Genomes can be analyzed and compared according to various criteria, suchas nucleotide and dinucleotide content, repetitions, coding-zone density,operons, gene family size distribution, etc. Genomes may also be studied on different levels, at the lowest of which they are regarded as a ‘sack of genes’,without considering interactions among their components. The next level takesthese interactions into account, as well as cross-correlations among the variousproperties of genomes.

3.1 General properties of genomes

One can only be struck by the extreme diversity of the genomes that havealready been studied, whether with respect to their general properties or to finercharacteristics. The first general property of genomes is their size, which, as maybe seen in Figure 3.1, ranges from 103 to 1011 basepairs if viruses are included,and from 105 to 1011 if only cellular life forms are considered. The sizes of bac-terial genomes overlap those of both viruses and eukaryotes.

3.1.1 Size and structure of eukaryote genomes

Table 3.1 lists the sizes and numbers of some unicellular and multicellulareukaryote chromosomes. However, comparisons of closely related species revealthat the great variability observed in the number of chromosomes evidently doesnot obey a golden rule. The human genome consists of 46 chromosomes,whereas that of the chimpanzee has 48 (certain monkeys have 70), althoughdivergence in DNA sequences between these two organisms does not exceed 1percent. Within a single vegetal genus, millet, for example (Table 3.1, bottom),

Bioinformatics: Genomics and post-genomics, Frédéric Dardel and François Képès© 2006 John Wiley & Sons, Ltd

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (74)

various species have haploid genomes whose sizes vary by a factor greater thanthree. The number of chromosomes ranges between 5 and 9 in the haploid milletgenome, and since millet ploidy can vary between two and six according tospecies, the number of chromosomes per cell in fact ranges between ten (2 × 5)and fifty-four (6 × 9).






103 104

Genome sizes (basepairs)

105 106 107 108 109 1010 1011

Figure 3.1 Order of magnitude of the genome sizes of viruses and of the three great realms ofcellular life forms: archaea, eubacteria, and eukaryota. Deviant values correspond to viroids amongthe viruses and to endosymbiotic green algae among the eukaryotes.

Table 3.1 Physical analysis of eukaryote genomes of the major phylogenetic groups by sizeand number of chromosomes (for a haploid cell).

Common name Size (Mbp) Number of chromosomes

Unicellular organismsBaker’s yeast 12 16

Multicellular organismsAnimals

Nematode 97 6Drosophila 137 5Mouse 3,000 20Human 3,100 23

PlantsArabidopsis 120 5Rice 450 12Wheat 15,000Mistletoe 75,000Millet 370–1,200 5–9

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (75)

The considerable diversity in genome size observed among organisms doesnot reflect differences in their complexity or capacity. The mistletoe genome (75,000Mbp) is around 25 times larger than that of the human (3,100Mbp).However, if we consider the basic criterion of the number of protein-codinggenes contained in a genome, it is probably true that the genomes of unicellu-lar organisms have fewer genes than those of multicellular organisms. Never-theless, the spread in the number of genes between unicellular and multicellularorganisms is low. Between the fly and the human, via Arabidopsis and nema-todes, the number of genes probably varies by no more than a factor of two,although this point remains controversial, because of the difficulty in locatinggenes in mammals and plants.

Comparative genomics becomes particularly efficient when it is based on thestudy of a large number of genomes belonging to related organisms. Indeed,knowledge of such cases facilitates disentanglement of the numerous and super-imposed evolutionary events that have resulted in genomes as we know themtoday. So far, the best case study in eukaryotes relates to yeasts. The genomes ofover a dozen species of yeasts and those of a few filamentous fungi have recentlybeen analyzed by complete or partial sequencing. These comparative genomicanalyses, which have no equivalent in other phyla, reveal that yeasts evolvedthrough interplay among gene duplication, formation of novel genes, and loss ofothers. The picture that emerges from these studies is that of a highly dynamicevolutionary process involving diverse mechanisms operating simultaneously.Longer-range duplications have also been demonstrated in yeasts, plants, andvertebrates. They consist of whole genome duplications, segmental duplications,and tandem gene array formation. Whole genome duplications are rare ineukaryotes, with the possible exception of polyploid plants, and are rapidly fol-lowed by extensive gene loss that leaves little trace, except for very recent events.

3.1.2 Diversity and plasticity of bacterial genome structure

Although bacteria display very limited variation in their dimensions and mor-phology, we should bear in mind that branches in their phylogenetic trees areseparated by immense temporal distances, at least as large as those that sepa-rate the longest branches of the eukaryote tree; for example, between fungi andvertebrates.

The following section explores the structural diversity and fluidity of the bac-terial genome, showing how individual bacteria within the same species differ,as well as how closely related species differ from each other. Eubacteria will beemphasized, since they are the most studied phylum in this respect. Large DNAreplicons are called chromosomes and small ones plasmids, although the defi-nitions of these terms have become blurred, since intermediate cases have beenrecognized.


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (76)

Replicon size

Bacterial genome sizes can vary by a factor greater than ten. The smallest knownbacterial genome is 0.58Mbp (Mycoplasma genitalium) and the largest is 9.2Mbp (Myxococcus xanthus). These values may be compared with the sizesof the largest viral genome, 0.67Mbp (bacteriophage G), and of the smallestknown eukaryote genome, 6.2Mbp (the microsporidium Spraguea lophii;Figure 3.1). The average gene size in bacterial genomes known today is ratheruniform, around 1kbp. Bacterial genes all seem to be compacted in a similarmanner, with around 90 percent of the DNA coding for macromolecules, pro-teins, and stable RNA. Large bacterial genomes therefore contain more genesthan small ones. The number of genes in a bacterial genome seems to reflect thelifestyle of the organism. Small-genome bacteria (~500 genes) are specialists; forexample, parasites that only thrive within living hosts or under other very specialconditions, whereas bacteria with large genomes (~10,000 genes) are general-ists from the metabolic point of view, or undergo certain forms of development,such as sporulation or mycelium formation.

As may be seen in Figure 3.1 and in Table 3.2, bacterial genome sizes varywidely, even within the same species. This variation can be explained by therapid loss of genes occurring after a species has invaded a highly specific niche,which would tend to reduce genome size. However, the discovery of high ratesof horizontal gene transfer among bacterial phyla (see details below) also makesincreases in genome size plausible.


Table 3.2 Physical analysis of bacterial chromosomes among the major bacterial divisions withrespect to size, shape (Circular or Linear), and number of chromosomes.

Bacterial division Genome size (Mbp) Genome shape Number of chromosomes

Aquificales 1.1 C 1Chlamydiae 1.0 C 1Cyanobacteria 2.7–6.4 C 1Fibrobacteriae 3.6 C 1Firmicutes 0.58–0.67 C 1Actinomycetes 1.6–8.2 C or L 1Fusobacteria 2.4 – 1Cytophagales 2.1–5.3 C 1Planctomycetes 5.2 C 1Proteobacteria 0.9–9.2 C, C and L 1–3Micrococcus 1.7–3.6 C 1Spirochetes 0.9–4.5 L or C 1–2

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (77)

Replicon geometry

In most bacteria, the genome is circular; however there are exceptions (Table3.2). Bacteria belonging to the genera Borrelia and Streptomyces have linearchromosomes, and often also have linear plasmids. The ends of BorreliaDNA are hairpin-shaped; that is, a Borrelia DNA strand forms a loop thatbecomes the second DNA strand. In one species of Borrelia, a plasmid that isordinarily linear has been found to be circular, indicating that linearity is notnecessary for replication. In contrast, Streptomyces DNA extremities are open,and specific proteins are attached to the free 5′-ends. If such a chromosome isartificially circularized, we again see that linearity is not a prerequisite for replication.

Number of replicons

Bacteria generally have a single large chromosome (Figure 3.1 and Table 3.2),and extra-chromosomal elements (plasmids) are found in many – if not all –species. However, members of several bacterial genera contain two or three largereplicons (>100kbp). The presence of multiple chromosomes is a stable char-acteristic of the Brucella and Burkholderia genera (Table 3.2), and ‘housekeep-ing’ genes are distributed among these multiple chromosomes. The possibleadaptive advantage conferred by chromosome multiplicity remains a mystery.If the circular chromosome of Bacillus subtilis is divided into two circular parts,the bacterium displays no particular phenotype.

Number of chromosome copies

The number of copies of plasmids found in natural bacterial isolates is nearlythe same as the number of principal chromosomes. It is an oversimplificationto consider bacteria to be haploid (one copy of each chromosome per cell), sincea few bacterial species have more than one chromosome in each cell. This is thecase for Deinococcus radiodurans, a bacterium that is highly resistant to radi-ation. It contains four copies of its chromosome, which, after severe radiationdamage, can recombine by hom*ology to regenerate an intact chromosome.

Also, even a haploid bacterium may contain more copies of certain genes thanothers. During rapid proliferation, bacteria have an average of four times morecopies of sequences that are close to the replication origin than of those closeto the replication terminus. This is because the time between two cell divisionsduring rapid proliferation is shorter than the total time required for completereplication (Figure 3.2), resulting in the coexistence of two or three levels of


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (78)

replication forks in the bacterium at a given time. Thus, double, and sometimesquadruple, copies of genes close to the origin are present in bacteria during rapidgrowth. In the absence of downstream regulation, this leads to two-to-four timesgreater expression than with a single copy. It is not absurd to imagine that overthe course of evolution, chromosome rearrangements have optimized the selec-tion of these genes as a function of their particularly intensive expression duringrapid growth.


Slow growth

Rapid growth

Figure 3.2 Replication of circular bacterial DNA. The circular chromosome of Escherichia coli repli-cates bi-directionally from the origin O toward a terminus situated opposite the origin. The periodof chromosome replication is relatively independent of growth conditions and of the time betweensuccessive cell divisions (generation time). Replication takes about 40 minutes for E. coli. Underconditions of slow growth (above), replication time is shorter than generation time. Re-initiationof replication does not occur until the preceding replication is complete. Replication is followedby a period in which there is no DNA synthesis. Under conditions of rapid growth (below), repli-cation time is longer than generation time. Re-initiation of replication nevertheless occurs at thesame rhythm as cell division. Thus, two or three levels of replication forks coexist following successive fork departures from the origin.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (79)

3.1.3 Bias, isochores, and CpG islands

Bacterial genomes display characteristic codon usage (some codons are preferredto others that are synonymous with them). These genomes have a characteris-tic G+C content (the molar proportions of guanine and cytosine in DNA), and a G and C bias for one DNA strand compared with the other, as well asparticular oligonucleotide frequencies (mono-nucleotides, di-nucleotides, etc).Although the mechanistic origin of these properties is not entirely clear, theyevolve slowly enough to be of use in inferring the evolutionary history of horizontal transfers.

The same biases, as well as other characteristics, are observed in eukaryotegenomes. Isochores are long DNA segments (>300kbp) of hom*ogeneous basecomposition in mammals. They may be divided into a small number of fami-lies covering a wide range of G+C content. Isochores that are poor in G+C tendto correlate with dark G bands (after staining with certain reagents, such asGiemsa stain, all human chromosomes display dark and light bands known asG bands). Isochores with a high G+C level have been observed to be richest in genes.

CpG islands (the ‘p’ is for phosphodiester bond), as their name indicates,contain numerous repetitions of the CG dinucleotide. This dinucleotide is gen-erally underrepresented in the human genome, but found in some islands of afew hundred basepairs, usually associated with the 5′-ends of genes.

3.2 Genome comparisons

3.2.1 Orthologous and paralogous genes

To compare genomes A and B in a meaningful manner, it is useful to determinewhich gene b in genome B corresponds to gene a in genome A. This corre-spondence relationship is primarily based on hom*ology. Functional hom*ologywithout sequence resemblance indicates evolutionary convergence (genes withno ancestral link but identical function.) Sequence hom*ology leads to a diag-nosis of evolutionary divergence, starting from a common ancestral sequence.

In the frequent case of evolutionary divergence, the genes evolved from acommon ancestor, but diverged after speciation or following a duplication event(Figure 3.3). When the hom*ology is the result of speciation, such as when thehistory of the gene reflects that of the species (for example, human and mousealpha-hemoglobins), the genes are called orthologs (ortho = ‘correct’). When thehom*ology is the result of gene duplication, in which the two copies are trans-mitted side-by-side over the history of a species (for example, mouse alpha- andbeta-hemoglobin), the genes are known as paralogs, (para = ‘in parallel’). Thus,


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (80)

orthology and paralogy are defined only with respect to the phylogeny of thegenes; not to their functions.

Identification of orthology using relative levels of sequence identity

Ideally, one would expect orthologous genes in the genomes of two species tohave the highest similarity, considering their relatively recent divergence. Themost direct approach to identifying orthologous genes therefore consists in com-paring all the genes in the two genomes with each other. The pairs of genes (a,b) that have the most similarity are selected and considered orthologs. Forexample, b in B is the gene most hom*ologous to a, which in A is the gene mosthom*ologous to b. This approach can be supported by auxiliary information,but may also encounter difficulties, described below.

Auxiliary information used for orthology detection

The first piece of auxiliary information useful in measuring orthology is synteny(see details below.) An example of synteny is the presence in genome B of con-tiguous genes b1 and b2, which are the orthologs of a1 and a2, themselves con-tiguous in genome A. When the evolutionary distance between two genomes issuch that the divergence between their orthologous genes is on average greaterthan 50 percent with respect to amino acids, gene order is no longer conservedbetween the genomes. Therefore, the use of synteny to identify orthologs ismainly limited to genomes in which divergence is relatively recent.


Speciation 1

Speciation 2

Geneduplication 1

Geneduplication 2

Figure 3.3 Diagram of divergent evolution illustrating orthology and paralogy. Speciation eventsyield species A, B, and C. Genes a1, b1, b2, c1, c2, and c3 are descended from the same ances-tral gene (above) via evolutionary speciation events and genetic duplication. Speciation 2, whichstarts from the same ancestral gene as b1 − c1, gives rise to two orthologous genes, b1 and c1,in species B and C, respectively. Gene duplication 2 in species C gives rise to two paralogous genes,c2 and c3.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (81)

A second kind of auxiliary information may be obtained by comparing a pairof presumably orthologous genes with a hom*ologous gene in a third genome.If two genes in different genomes have the highest level of identity both witheach other and with respect to the gene in the third genome, there is strong presumption that the two genes are in fact orthologs.

Sequence divergence

At great evolutionary distances, for example, between eubacteria and archae-bacteria, sequence similarities can be so eroded that the distances betweenorthologous genes are comparable to the distances between sequences fromsame gene family. The similarity may even become so low as to be undetectable.

Non-orthologous gene displacement

A second problem encountered in identifying orthologs is the displacement of a non-orthologous gene. This occurs rather frequently when two non-orthologous genes with no phylogenetic link carry the same function in differ-ent organisms. This is therefore a case of evolutionary convergence, which canbe confusing.

Gene duplication and loss

A third process that limits the identification of orthologous genes is gene lossin combination with gene duplication. If genomes A and B lose two paralogsa1 and b1 of an ancestral gene that was duplicated prior to speciation events,the remaining genes a2 and b2 in genomes A and B display the highest sequenceidentity. According to the above definition, they are considered orthologs,whereas in fact they are not. Such cases may be detected by verifying whetherthe percentage of similarity between their protein products lies within a certainrange of values.

Orthology of proteins consisting of several domains

Two levels of orthology can be identified in proteins that consist of severaldomains, one for each domain considered individually and one for the completeprotein. This can lead to situations in which non-orthologous genes code fororthologous protein domains.


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (82)

The concept of orthology is therefore an important refinement of the notionof hom*ology for use in describing phylogenetic relationships among genes, aslong as the problems outlined above are borne in mind and the methods usedto determine orthology are well-specified.

3.2.2 Synteny

As mentioned briefly above in the discussion of orthology, synteny concerns theconservation of the respective order of genes among genomes.

When species that are not closely related are considered, extensive rearrange-ment of gene order is usually observed. In contrast, any conservation of rela-tive gene order is remarkable. However, in a few cases, relative gene order isconserved over considerable evolutionary periods. In general, genes that ‘worktogether stay together.’ This is often the case for operons, suggesting that theadvantageous co-regulation of genes in an operon could justify synteny.However, this justification may be insufficient, since, for example, the trypto-phan operons of E. coli and B. subtilis genes display total gene order con-servation despite utilizing different regulatory mechanisms. In addition, genesthat are co-regulated within the same operon in a given species are sometimesdispersed in another, even though they are correctly co-regulated there.

Synteny within the same genus

It is possible to compare the genomes of two species of a given genus (Figure3.4(A)). The chromosome sizes of Mycoplasma pneumoniae and Mycoplasmagenitalium are, respectively, 816 and 580kbp. An ortholog of every M. geni-talium gene exists in M. pneumoniae. However, the structures of the twogenomes have greatly diverged, particularly as a result of deletions and inser-tions. The Mycoplasma chromosome may be considered to consist of six seg-ments whose order has been shuffled without affecting their orientations. It isinteresting to note that these six segments are flanked by repetitive sequenceswith the same orientation, known as ‘MgPa,’ which could permit suchrearrangements by hom*ologous recombination.

Synteny within the same species

More surprising than the differences found between closely related species arethe substantial ones observed among the genomes of natural isolates of the samespecies (Figure 3.4(B)). Among the 127 isolates of Salmonella typhi there are 17different arrangements of the seven chromosome segments, flanked by seven


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (83)

tRNA operons. Those most frequently found are presented in Figure 3.4(B).Many non-counter-selected rearrangements therefore seem to be produced byhom*ologous recombination among these tRNA operons. Consider enterobacte-ria of closely related genomic content, but from different genera (Escherichia andSalmonella) or species (paratyphi and typhimurium) (Figure 3.4(C)). Curiously,they display less difference in this particular case (Figure 3.4(C)) than do strainsof the same genus and species.

Synteny within the same strain

Even within the same strain, minor genomic differences can appear among indi-viduals, either as random events that accumulated over time, or in response to


A Mycoplasma genitalium

Mycoplasma pneumoniae

Salmonella typhi

Salmonella typhimurium

Escherichia coli

Salmonella paratyphi



Figure 3.4 Major differences in genomic structure among species within the same genus andamong strains of the same species. All the circular chromosomes in this representation have beenlinearized near the replication terminus. (A) Relations between Mycoplasma pneumonia andMycoplasma genitalium. Different shades of gray correspond to different genome segments in eachstrain. The same shade of gray indicates a segment that is hom*ologous in the two strains, eventhough each M. genitalium segment includes deletions with respect to its M. pneumoniae hom*olog.Arrows indicate the positions of repetitive sequences known as ‘MgPa’. (B) Rearrangements amongSalmonella typhi strains (127 natural isolates were analyzed). Arrows indicate the positions of theseven tRNA operons. (C) Rearrangements among three enterobacteria genera. Arrows indicate thepositions of the seven tRNA operons.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (84)

a particular situation. In the first case, semi-stable reversible phenotypic varia-tions occur in which members of the population spontaneously change proper-ties (gene expression states), with low probability at each generation. Theseproperties may later revert, with the same low probability. Such variations arecalled ‘epigenetic variations’ or ‘epimutations,’ and do not affect genomesequence itself, but rather gene expression status.

In the second case, programmed changes in the genome arise irreversibly inparticular cells not destined to produce descendants. Two such cases are known:the B. subtilis chromosome during spore formation and the heterocyst chro-mosome in certain cyanobacteria (heterocysts are described in Figure 8.4). In both these cases, the rearranged chromosome is no longer replicated and therearrangement causes major changes in gene structure and expression.

‘Cassette mechanisms’ also exist, in which a DNA sequence contained in anunexpressed pseudogene is physically placed in an expression site. These mech-anisms can be either distributive (pseudo-genes not grouped at a single site) ororganized (pseudo-genes in single-file at one or more privileged sites.) Somemembers of the genus Borrelia have a mechanism by which cassettes code exter-nal membrane proteins, allowing the bacterium to change its surface antigendeterminants, thereby escaping the response of the host immune system. Ineukaryotes, a cassette mechanism allows yeasts to switch mating type. The vertebrate immune system itself utilizes a basically similar strategy to createimmunoglobulin diversity.

3.2.3 Minimum gene set

Once complete genome sequences are available, it is possible to investigatewhether there exists a minimum set of genes necessary to maintain life. The M.genitalium bacterium includes only 468 genes that have been identified as codingfor proteins, which some biologists refer to as the minimum number necessaryto maintain life. While the M. genitalium genome is the smallest one knownamong cellular life forms, there is not the slightest experimental evidence to indi-cate that it is the smallest possible genome. It is also interesting to compare thisset of 468 genes with other small-genome bacteria, such as that of Haemophilusinfluenzae, which contains 1,703 protein-coding genes. Of the M. genitaliumgenes, 240 have orthologs in the H. influenzae genome. To these may be added22 genes whose functions are not assumed by any of the 240 orthologs due todisplacement of non-orthologous genes. Finally, six functionally redundantgenes may be eliminated from the count. Therefore 256 remain, and have beenpostulated to constitute the minimum set of genes necessary for cellular life(Figure 5.5).

This approach is obviously subject to criticism, since these small-genome bac-teria are parasites that have lost numerous genes necessary for autonomous life.


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (85)

Nevertheless, these 256 genes do provide basic life functions and a degree ofmetabolic autonomy. They have inspired the choice of the 127 basic genes thathave been modeled in the ‘Electronic-Cell (‘E-Cell’; see chapter 8) to simulatethe functioning of a minimal cell, including notions of energy cost, such as forthe synthesis of macromolecules.

The circles, drawn approximately to scale, represent the gene sets of H.influenzae, M. genitalium, and of the yeast Saccharomyces cerevisiae. The rec-tangles show the two successive stages in the construction of the theoreticalminimum set of 256 genes, in comparison with the two prokaryotes, and a puta-tive set of 96 ancestral genes derived by comparing this minimal set of geneswith the genes of the yeast, a eukaryote.

If this study is extended to a eukaryote such as the yeast, once the genes that code for mitochondrial proteins have been eliminated, 96 of these 256 genes are found to have an ortholog in the eukaryote (Figure 3.5). It is amusing


+22 displacements ofnon-orthologous genes

M. genitalium






468 genes

H. influenzae1,703 genes

262 common genes

256 non-redundant

common genes

6 redundant genesS. cerevisiae6,200 genes

96 ancestral genes

Figure 3.5 Construction of a minimum set of genes from two bacterial genomes and comparisonwith yeast genes.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (86)

to note that certain functional categories are totally absent from this reducedset of 96 genes, in particular, genes involved in DNA replication. This suggeststhat the last common ancestor of these organisms did not use DNA; perhaps itused another nucleic acid to replicate. While it is too early to tell, this study isinteresting in that it is the first to prefigure a theory of comparative genomics.

3.2.4 Pathogenicity islands

Pathogenicity islands are chromosome or plasmid regions in pathogenic bacte-ria that assemble genes that code for virulence factors, such as toxins, pili, andother host adsorption and invasion factors. These islands also have at least oneof the following characteristics: association with DNA mobility agents (inte-grase genes and insertion/deletion sites), non-universal distribution in naturalisolates, unusual codon use, and an abnormal percentage of G+C in compari-son with the other genes of the bacterium. Their size is highly variable; thelargest one known is 190kbp. By their sporadic presence and mobility, theyclearly contribute to genomic content variability in numerous pathogens. Thus,four different pathogenicity islands are present, alone or in pairs, in 10, 11, 28,and 28 of the 72 pathogenic isolates of E. coli that have been studied.

If the genomes of two closely related strains, one pathogenic and the otherbenign, are compared, it is often possible by studying the difference betweenthem to identify the genetic basis of the pathogenicity. Thus, 60 percent of thegenes of the pathogen Haemophilus influenzae that do not have hom*ologs in abenign strain of E. coli appear to be implicated in pathogenesis.

It seems likely that pathogenicity islands are just one manifestation of a more universal phenomenon that could be called ‘specialization islands,’ which would confer various specific capacities – metabolic, aggressive, or defensive –permitting a fraction of the members of a bacterial species to occupy very specialized niches.

3.2.5 Therapeutic targets

Comparing the genetic contents of the genomes of pathogens (bacteria andyeasts) and hosts (mainly human, but also animals and plants) can be extremelyrevealing. Indeed, any gene of a pathogen that has no equivalent in its host isa potential therapeutic target. The list of differences therefore amounts to a pre-liminary list of antibacterial or antifungal targets. The targets retained are thosethat are indispensable either for the survival of the pathogen, or for its invasiveor pathogenic capacity (for example, the gene for a pathogenicity island.) Ofcourse, before a drug can be considered a candidate therapeutic agent it is


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (87)

necessary to ascertain that it does not have an unexpected secondary effect ona host function.

Finally, the genome of each bacterial species may be considered to consist ofa universally present hard kernel of genes (the ‘endogenome’), plus a battery ofaccessory elements, which may be located on free replicons or well-integratedat various chromosome sites (the ‘exogenome’). The selective advantage pro-vided by the fact that certain genes are accessory elements could be related totheir genetic mobility. An accessory element may or may not be functional (forexample, it may have lost its function following evolutionary degeneration). Itmay appear to be selfish (a pro-virus integrated into the bacterial host’s chromosome), or of great adaptive value under certain circ*mstances (a patho-genicity island or integrative element that confers antibiotic resistance), or bothat the same time (a provirus bearing a bacterial virulence gene). It is difficult todraw similar conclusions regarding eukaryotes, since few complete genomeshave so far been sequenced. However, it appears that eukaryote genomes harborsome ‘orphan’ genes with no known hom*olog in other genomes.

3.3 Gene evolution and phylogeny: applications to annotation

3.3.1 Gene evolution

Various criteria may be applied in evaluating the rate of genome evolution. Inan initial approach, we will consider the genome as a ‘sack of genes’ and counthow many orthologous genes the compared genomes have in common. Corre-lations between genes in these genomes will then be taken into account, andgene synteny, regulatory modifications, and co-occurrence will be evaluated.

Sharing of orthologous genes

In order to evaluate the rate of genome evolution, we can examine how thenumber of orthologs shared by two genomes decreases over the time since theirdivergence. The number of amino acid substitutions per protein per positionwill be used to estimate divergence time. Figure 3.6 shows that the fraction oforthologous sequences shared by two sequences rapidly decreases over evolu-tion, faster than the decrease in the level of sequence identity between thesesame sequences. The most rapid decrease occurs over short time scales, whenprotein identity between a given pair of genomes is still greater than 50 percent.

Time since divergence is measured by the number of amino acid substitutionsper protein per position in a set of 34 orthologs (abscissa). If eubacteria and archaebacteria diverged 3.5 billion years ago, each unit of the abscissa


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (88)

represents 875 million years. Each point on the graph indicates the fraction ofsequences in genome A that have an ortholog in genome B (left ordinate). Thecurve indicates the percentage of identity of the protein sequences (right ordinate) calculated as a function of the abscissa, according to the equation of N.V. Grishin [J Mol Evol (1995), 41:675–679].

Horizontal transfer

Sometimes a species shares a few orthologous genes with a phylogeneticallydistant one (for example, eubacteria and archaebacteria) that it does not sharewith closely related species. This reveals an aspect of the evolution of the geneticcontent of genomes that does not derive from the ‘vertical’ notion of the ‘phy-logenetic tree’: ‘horizontal’ transfer. Based on nucleotide and dinucleotide fre-quencies, it appears that at least 10 to 15 percent of the genome of a bacteriumsuch as E. coli consists of sequences that have been transferred horizontally;that is, bacterium-to-bacterium. This phenomenon is rarer among archaebacte-ria, although some particular mechanisms have recently been discovered bywhich they integrate mobile genetic elements. It is also believed that in thedistant past, metabolic genes were horizontally transferred from eubacteria toarchaebacteria, whose genomes would therefore be chimeric.


Gene order conservation may be analyzed in terms of genome divergence time,as above (Figure 3.7). A drastic rearrangement of genomes is observed over short


Genome divergence









n of



s w

ith a

n or



















1.0 2.0 3.0 4.0 5.0 6.0

Figure 3.6 Relationship between genome similarity and evolutionary time.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (89)

periods, while protein pair identity between the genomes is still over 50 percent,below which saturation is reached. Compared with the preceding figure, theorder of orthologous genes is seen to be somewhat less well-preserved than theiractual presence. Genes that are still paired by the time the saturation level isattained are generally those that code for proteins that physically interact. Gene order conserved in the absence of physical interaction indicates recent horizontal transfer.

Horizontal transfer

The above argument may be exploited to detect horizontal gene transfer bysynteny. Since gene order is poorly conserved over evolution, the ordered pres-ence of a few genes in two evolutionarily distant branches raises the suspicionthat they have undergone horizontal transfer. This suspicion is reinforced if thegene order is not conserved in closer evolutionary branches.

Selfish operons and Fisher’s hypothesis

Bacterial operons are DNA segments in which genes coding for several proteinsare contiguous and co-regulated. Considering the universality of operons andtheir structuring effects, the weak conservation of bacterial gene order is surprising (Figure 3.7). Except for cases of horizontal transfer, conservation of bacterial gene order appears to be limited to genes whose products interact.


Genome divergence


e or











1.0 2.0 3.0 4.0 5.0 6.0

Figure 3.7 Relationship between the conservation of gene order in a genome and evolutionarytime.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (90)

According to Fisher, this is justified because the physical proximity – geneticlinkage – between genes whose products function well together tends to increasein order to prevent the separation of co-adapted allele pairs by recombinationevents.

The time since divergence is measured by the number of amino acid substi-tutions per protein per position in a set of 34 orthologs (abscissa). If eubacte-ria and archaebacteria diverged 3.5 billion years ago, each unit of the abscissacorresponds to 875 million years. The ordinate displays the number of geneswith orthologs in the two genomes that also have at least one neighboring genethat is orthologous in the two genomes. This number is compared with the totalnumber of orthologs shared by the pair of genomes. The rapid differentiationin gene order during evolution may then be readily observed.

These interactions should impose selective constraints, thereby slowing theevolution of genes whose products are co-adapted. As a result, sequence con-servation in genes that are conserved in pairs would, on average, be greater thanin other genes. In bacteria, the co-folding of two proteins (destined to belongto the same complex) during co-translation has been proposed as a possibleexplanation of why their genes are maintained side-by-side, producing their messenger RNAs at the same physical site.

The ‘selfish operon’ theory proposes that operons simply increase the prob-ability that genes that function in concert stay together during horizontal trans-fer. Of course, the transferred DNA is more likely to be conserved over evolutionif it contains genes necessary for a biochemically advantageous function. Thismodel applies only to non-essential genes; i.e., those that can be lost and thenreintroduced by horizontal operon transfer (pathogenicity islands, etc). Forexample, it does not apply to ribosome genes, which are essential and closelygrouped, and for which there is no proof of horizontal transfer.

While many operons include genes whose products do not physically inter-act, those operons that have been conserved over long evolutionary periods doseem to conform to Fisher’s hypothesis; i.e., they contain genes whose productsinteract.

Regulatory modifications

Once orthologous genes and synteny have been determined, it is possible toexamine the conservation of the regulatory sequences in which the links thatcontrol gene expression are located may be considered. This conservation hasbeen found to be remarkably low and to diminish much more rapidly than theconservation of gene order. However, there are exceptions to this rule. A second-ary RNA structure at the 5′-end of E. coli ribosome genes rpl1 and rpl11 has beenfound to be involved in the regulation of operon expression. The same secondaryRNA structure has been identified in all bacterial genomes studied to date.


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (91)

Gene co-occurrence

Some genomes are more organized than others

If neighboring genes in a genome express products that cooperate in a givenfunction, as is the case for operons, these same genes should also coexist in othergenomes, even in those in which they are not neighbors or located in the sameoperon. The absence of even one such cooperative gene should render the entireset non-functional. This phenomenon has indeed been observed. If gene a1 ingenome A has gene a2 as a neighbor, and if ortholog b2 of a2 is found in genomeB, the probability that ortholog b1 of a1 is also present in B is increased. Revers-ing this proposition, orthologs shared by genomes tend to be physically close insome of these genomes. Of course, b1 and b2 are often also neighbors; exceptfor such cases, this tendency remains. Genomes therefore are often organized,some more so than others, in the sense that grouping is more frequent in them(Figure 3.8).





e or


og g



B. s



E. c


M. t






M. p




H. i








M. j




M. g



H. p







Figure 3.8 Bacterial genomes are more or less well organizedThe ordinate gives the ratio of the number of genes in genome A that have an ortholog and atleast one neighboring gene that also has an ortholog in genome B to the number obtained afterrandomizing gene order. This analysis includes only genes that are neighbors in A but not in B (orthe inverse), so that the results will not be influenced by the phenomenon of synteny, discussedpreviously. The relative grouping of orthologs, given on the ordinate, is the result of the meancomparison of one genome with the other eight. The species studied are given on the abscissa.The M. genitalium genome (right) is the only species in the study in which there is not moregene grouping than predicted by a random model. The explanation for this could be that the verysmall M. genitalium genome makes neighbors of all of its genes that display orthology with thegenes of another species. Genome size therefore does not systematically justify relative grouping,since, for example, Synechocystis has a large genome but displays little organization. E. coli andB. subtilis have the most organized genomes.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (92)

Gene co-occurrence and metabolic pathway conservation

Besides the spatial association of orthologs, it is also possible to determinewhether orthologs reveal ‘genomic association’. Are both orthologs either coex-istent or absent in a genome; i.e., is one not found without the other? Suchanalysis makes it possible, in principle, to infer which genes are functionallylinked. As we have just seen, orthologs found in two genomes have an increasedprobability of being grouped together (and possibly of being functionally linked)in one of the genomes, even if not in the other. This information may be usedto help reconstruct metabolic and signaling pathways. This works well only ifthe structure of the pathway concerned is well conserved over evolution. Thedisplacement of a non-orthologous gene, by which one gene takes over the roleof another in a given pathway, suggests that orthologous gene pathways arebetter conserved than their presence.

These observations have consequences in selecting strategies for genomicanalysis. In particular, they indicate that, according to the phenomenon beingstudied, the evolutionary distance between organisms included in a study mustbe carefully chosen. Therefore, when studying the evolution of gene regulation,it is better to compare species that are close to each other than when studyingthe evolution of gene order. When studying the evolution of coding sequences,it is better to compare more distant species. Finally, the most distant speciesshould be used when studying the evolution of metabolism.

3.3.2 Using genomic context to predict functions

This section appears here because it makes use of some of the concepts thathave just been discussed. However, the perspective is different now, involvingthe annotation of unknown genes that derive from systematic sequencing. More specifically, the aim here is to predict functional interactions among proteins, based on their gene context. In bacteria, this covers three different possibilities:

Type 1: Gene fusion, forming a single hybrid gene that codes for a singleprotein in another genome;

Type 2: Gene order conservation or gene co-occurrence in putative operons;

Type 3: Co-occurrence of genes in genomes (phylogenetic profile).

In the first type, interaction between the two proteins of an organism is demonstrated by their covalent bonding in another species. In the second type, ithas been established experimentally that proteins coded by 75 percent of con-served gene pairs interact physically, as seen above. To these 75 percent may be


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (93)

added 20 percent of conserved gene pairs for which gene product interaction hasbeen predicted but not proven. For the third type, as discussed above in ‘Gene co-occurrence and metabolic pathway conservation’, a positive result suggeststhat the genes concerned contribute to the same major function or pathway.

Conservation of gene order (Type 2) is the most useful of these three con-cepts, since it provides contextual information for 37 percent of the genesstudied in the M. genitalium standard case. If there is gene co-occurrence inoperons without order conservation, Type 2 provides information in 8 percentof the cases. Type 1 provides contextual information in 6 percent of the cases,and Type 3 in 11 percent. Combining these criteria, significant information concerning the genomic context may be obtained for 50 percent of genes (thecategories overlap).

This approach based on genomic context therefore holds promise and is com-plementary to the more classical approach based on hom*ology search. In prin-ciple, hom*ology search allows prediction of a molecular function, while thegenomic context approach addresses a higher functional level by predicting towhich process or pathway a given protein belongs, or with which other proteinit interacts.

3.3.3 The genomic tree of life

An important consequence of the availability of numerous complete genomicsequences is that it becomes possible in principle to construct a universal treeof life based on genomic rather than genetic phylogeny. The existence of theuniversal tree, which until recently was based on 16S rRNA genes, led during1987–90 to a proposal of the existence of three major biological reigns: eukary-otes (Eukarya), eubacteria (Bacteria), and archebacteria (Archaea). For variousreasons, this classification was criticized over the following ten years. Certainother genetic phylogenies have yielded the same result, whereas still others haveproposed different topologies, and others ended up with only two realms. Thesedifficulties in part derive from the fact that archebacteria are close to eukary-otes with respect to transcription/translation machinery and close to bacteriawith respect to metabolism. These problems, as briefly mentioned above, aredue to horizontal transfer, which may have been particularly intense during earlyevolution, and from unequal nucleotide substitution rates, according to thelineage. More generally, these problems reflect the fact that the trees representthe evolutionary distances between genes, rather than whole organisms orgenomes.

The first attempts to analyze genome macrostructure for the purpose of phy-logenetic reconstruction utilized DNA hybridization or restriction fragmentanalysis. As is the case for genetic phylogenies, these techniques ultimatelydepend on the degree of divergence of the sequences compared. In contrast, oncethe orthologs have been identified, the comparative analysis of gene order


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (94)

discussed above is independent of the degree of divergence. A different but com-plementary approach is described here, based not on evolutionary lineage, buton the hierarchical classification of genomes according to their gene content andoverall similarity.

This approach consists in comparing the full set of gene products predictedfrom the complete genome of an organism, either with itself or with sets deriv-ing from each of the other organisms studied. First, using a classical method,the product of each open reading frame (ORF) is determined and comparedwith all the others. Then the proportion of ORFs in genome A with at least onesimilar ORF in genome B is determined. Overall comparison of n organismsgenerates an n × n matrix of these proportions. Multifactorial correspondanceanalysis is then used to reveal the axes of greater data variability (by rotatingthe axes while keeping them orthogonal). Each organism is represented by apoint in this n-dimensional space. The distances between pairs of organisms are then calculated. Organisms are classified according to their neighborhood,yielding a hierarchy known as a ‘genomic tree.’ A genomic tree is the graphicrepresentation of the relationship between sets of organisms, which indirectlydepends on genome size, internal redundancy due to ancestral duplication, andoverall gene loss/acquisition events. Nevertheless, this tree is independent of thefunctional identity of the genes. It is also possible to render it independent ofduplication events by eliminating redundant genes within the same species fromthe set of initial data.

Applying this method to the first 20 sequenced genomes generates the graphin Figure 3.9, in which only the two axes with the greatest data dispersion arerepresented. Based on this graph, the distances calculated then permit con-


Main component













0.50 0.75

Figure 3.9 Representation of the proportion of ORFs of one organism for which there is at leastone similar ORF in another organism. After multifactorial correspondence analysis, the first andsecond axes represented here contain 48 percent and 26 percent of the total data dispersion,respectively. Each point in this space corresponds to an organism (see their names in Figure 3.10).Groups of points correspond to major phylogenetic divisions, represented here in different shadesof gray.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (95)


Escherichia coli

Synechocystis SD.Bacillus subtilisAquifex aeolicus

Mycobacterium tuberculosisCampylobacter jejuniHaemophilus influenzaeHelicobacter pyloriRickettsis prowazekiiChlamydia trachomatis

Treponema pallidumBorrella burgdorferiMycoplasma pneumoniaeMycoplasma genitaliumMethanococcus jannaschiiArchaeoglobus fulgidusMethanobacterium thermoantetrophicumPyrococcus horikoshiiComorhabditis elegansMus musculushom*o supiensSchieosaocharomyces pombeSaocharomycos cerevisiae

Figure 3.10 Genomic tree. This tree is obtained by the pair-wise hierarchical classification oforganisms, based on their neighboring distances (see Figure 3.9). Gray-level codes are the same.Lengths of the horizontal lines between nodes are proportional to their degree of similarity.

struction of the genomic tree (Figure 3.10). Four well-defined groups of organ-isms appear in this tree: a) an archebacterial group; b) a eubacterial group; c)a group of mycoplasms close to the eubacterial group, and d) a eukaryote group.This result is perfectly compatible with the ‘three kingdoms’ perspective.


Casjens S. (1998). The diverse and dynamic structure of bacterial genomes. Annu RevGenet 32: 339–377.

Dujon B., et al. (2004). Genome evolution in yeasts. Nature 430: 35–44.Jaillon O., et al. (2004). Genome duplication in the teleost fish Tetraodon negroviridis

reveals the early vertebrate proto-karyotype. Nature 431: 946–957.Casjens S. (1998). The diverse and dynamic structure of bacterial genomes. Annu Rev

Genet 32: 339–377.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (96)

Dujon B., et al. (2004). Genome evolution in yeasts. Nature 430: 35–44.Fisher R. A. (1930). In The genetical theory of natural selection. Oxford University Press,

Oxford, UK.Huynen M.A., Bork P. (1998). Measuring genome evolution. Proc Natl Acad Sci USA

95: 5849–5856.Jaillon O., et al. (2004). Genome duplication in the teleost fish Tetraodon negroviridis

reveals the early vertebrate proto-karyotype. Nature 431: 946–957.Kellis M., et al. (2004). Proof and evolutionary analysis of ancient genome duplication

in the yeast Saccharomyces cerevisiae. Nature 428: 617–624.Koonin E.V., Mushegian A.R. (1996). Complete genome sequences of cellular life forms:

glimpses of theoretical evolutionary genomics. Curr Opinion Genet Develop 6:757–762.

Tekaia F., et al. (1999). The genomic tree as revealed from whole proteome com-parisons. Genome Res 9: 550–557.


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (97)

4Genetic information andbiological sequences

4.1 Introduction: Coding levels

The information contained in a genome is stored at several levels, the most basicof which associates each amino acid of each protein coded by a gene to a singletriplet of DNA bases (codon). Besides this elementary code, simple punctuationsignals identify the beginnings and ends of genes. In addition to this ‘raw’ data,the genome contains expression, regulation, and alternative splicing signals (ineukaryote cells) that govern how cells implement the information it contains.The genome also contains specific signals unrelated to expression of the geneticmessage, which concern the metabolism of the DNA molecule itself, includingreplication, recombination, methylation, and restriction sites.

These data are all coded in the DNA sequence, and often mutually overlap.Genes thus contain methylation and recombination sites; certain genes partiallyoverlap; the expression signals of one gene are sometimes located withinanother. . . . The unraveling of these various coding levels is of primary impor-tance to the biologist seeking access to information contained in the genome inorder to understand the functions of living matter, as well as to devise experi-mental strategies and to analyze results.

Information technology may be used to efficiently extract the informationcoded in DNA. The remainder of this chapter recalls and describes various typesof signals coded in DNA, as well as specific patterns and sequences with whichthey are associated.

4.2 Genes and the genetic code

The principal information contained in the genome consists in the genes them-selves. Confronted with the raw sequence data of a genome, the biologist firstseeks to identify the various genes they contain, in order to study the proteins

Bioinformatics: Genomics and post-genomics, Frédéric Dardel and François Képès© 2006 John Wiley & Sons, Ltd

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (98)

for which they code. Translation of a gene into a protein associates onenucleotide triplet (codon) with each of the 20 natural amino acids that con-stitute proteins. Since there are four different nucleotides, there exist 43 = 64possible different codon triplets. The meaning of each of these 64 codons is uni-versally conserved throughout all forms of life, and is known as the genetic code(Figure 4.1). Sixty-one codons specify the 20 amino acids, and three, TAG,TAA, and TGA, are translation stop signals, also called termination or non-sense codons. In addition, protein translation can only be initiated by certaincodons known as start signals or codons. The main start codon is ATG, whichspecifies the amino acid methionine; however, translation can also begin withGTG. Whatever the start codon, the first amino acid incorporated into a proteinis always methionine, even though the GTG codon normally specifies valine inpeptide chains. A gene that codes for a given protein of length n therefore hasthe following structure, known as an open reading frame (ORF):

When examining a sequence of cDNA or of a genome that has no introns(bacteria), analyzing the distribution of its nonsense codons furnishes valuableinformation concerning the potential positions of genes. In a purely randomsequence in which the four nucleotides, A, T, G, and C, would be distributedequally, the statistical frequency of occurrence of these stop codons would be3/64, or around one stop for every 21 codons. However, a protein chain gen-erally consists of between 100 and 1,000 amino acids, which is very significantlylonger. The presence of an ORF of length equal to or greater than 100 codonsin a DNA sequence is thus a very strong indication of the presence of a gene,thereby providing a predictive method. This simply requires examining stop

start codon n codon specifying an amino acid termination codon− −( ) × { } −1


TTT: Phe TCT: Ser TAT: Tyr TGT: Cys TTC: Phe TCC: Ser TAC: Tyr TGC: Cys TTA: Leu TCA: Ser TAA: STOP TGA: STOPTTG: Leu TCG: Ser TAG: STOP TGG: Trp CTT: Leu CCT: Pro CAT: His CGT: Arg CTC: Leu CCC: Pro CAC: His CGC: Arg CTA: Leu CCA: Pro CAA: Gln CGA: Arg CTG: Leu CCG: Pro CAG: Gln CGG: Arg ATT: Ile ACT: Thr AAT: Asn AGT: Ser ATC: Ile ACC: Thr AAC: Asn AGC: Ser ATA: Ile ACA: Thr AAA: Lys AGA: Arg ATG: Met ACG: Thr AAG: Lys AGG: Arg GTT: Val GCT: Ala GAT: Asp GGT: Gly GTC: Val GCC: Ala GAC: Asp GGC: Gly GTA: Val GCA: Ala GAA: Glu GGA: Gly GTG: Val GCG: Ala GAG: Glu GGG: Gly

Figure 4.1 The genetic code.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (99)

codon distribution in the six possible reading frames (three in each of the twoDNA strands).

4.3 Expression signals


The expression of the genes coded in DNA begins with transcription of thegenetic message into messenger RNA molecules (mRNA). In eukaryotes, eachmRNA molecule codes for only one gene, whereas in bacteria, several genesmay be transcribed by the same RNA molecule, constituting what is known asa transcription unit. In all cases, the DNA sequence situated around the begin-ning of the transcription sequence contains a specific pattern known as a pro-moter sequence, which enables the cell to begin polymerizing mRNA.

Cells contain one or several enzyme complexes known as RNA polymerases,which carry out transcription. When transcription begins, certain protein factorsor supplementary subunits whose function is to recognize the transcription pro-moter sequence bind to the RNA polymerase. Once transcription has begun,these factors have fulfilled their roles and detach from the RNA polymerase.For example, in Escherichia coli, RNA polymerase is a pentamer, a2bb′w, towhich a supplementary subunit called the s factor, which recognizes the pro-moter, is added at the start of transcription. The normal s factor, known as s70,recognizes the ‘classical’ sequence, located a dozen nucleotides upstream fromthe beginning of transcription:

5′TTGACA- ~17 bases- TATAAT

E. coli cells can also contain s factors specific for other DNA sequences thatdetermine other promoter families. Various s factors are expressed under par-ticular physiological conditions, such as heat shock and low nitrogen levels, inresponse to which they permit selective activation of the transcription of a wholegroup of genes.

A comparable, but much more complex situation is observed in eukaryotes,in which three distinct RNA polymerases exist, each interacting with a largenumber of transcription factors. RNA polymerase II, which transcribes most ofthe genes for proteins, binds in the vicinity of a conserved pattern known as aTATA box, situated around 30 nucleotides upstream from the transcriptionstarting site:

5′TATA (A or T) A (A or T)

This sequence is recognized by a specific protein known as TFIID, by whichthe polymerase binds. Upstream eukaryote transcription promoters usually


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (100)

include other large patterns that enhance transcription efficiency, such as a 5′-GG(T or C)CAA(T or C)CT around 70 basepairs upstream from the transcrip-tion starting site, and sometimes a GC-rich motif of the 5′-GGGCGG or5′-CCGCCC-type one-hundred basepairs upstream.

Since promoters are found upstream from transcription units, specific down-stream sites exist that determine the end of transcription. In prokaryotes, thesegenerally contain an inverted repeat (see Chapter 6 on structure prediction) ableto fold into RNA, forming a hairpin structure with a stem and a single-strandloop. This sequence is followed by a series of T (U in RNA).

Transcription termination is more complex in eukaryotes. Many eukaryotemessenger RNA molecules are polyadenylated at the 3′-end, and the enzymeresponsible for this, poly(A) polymerase, recognizes a 5′AAUAAA3′ segmentdownstream from the gene. This site also appears to be involved in the termi-nation of transcription by RNA polymerase.

Alternative splicing

The product of DNA transcription in eukaryotes is precursor mRNA, whichmay contain non-coding segments (introns). The precursor RNA then under-goes a process called alternative splicing, which excises the introns, yieldingmature messenger RNA that can then be translated into protein. This matura-tion is carried out by complex assemblies of proteins and nucleic acids calledSmall NUclear Ribonucleo-Proteins (snRNP or SNURPs), which recognize specific sequences at intron-exon junctions:

Exon Intron Exon5′ . . . (C ou A)AG GURAGU . . . YYUUYYYYYYNCAG G . . . 3′

Y indicates a pyrimidine (C or U), R a purine (A or G), and N any nucleotide(A, U, G, or C).




A — UG — C C — GC — GC — GG — CC — G





5' 3'



Inverted repeat

Figure 4.2 Prokaryote transcription terminators.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (101)


In eukaryotes, the ribosome sweeps along messenger RNA from the 5′-end,starting translation with the first ATG codon it encounters. In bacteria, wheremessenger RNA may be polycistronic, the translation start codon is selected bythe ribosome, which recognizes the Shine-Dalgarno sequence, AGGAGGU,located 5 to 10 nucleotides upstream from it:

5′AGGAGGU-{5 to 10 nucleotides} (A or G)TG . . .

Regulation of gene expression

A wide variety of mechanisms regulate genetic expression. One or more spe-cialized proteins exist in both eukaryotes and prokaryotes that recognize spe-cific sites on DNA around gene promoter regions. These regulatory proteins caneither activate or inhibit transcription, depending on the physiological condi-tions and external stimuli.

Modification signals

DNA is not only the medium upon which genetic information is based and thatthe cell uses to synthesize proteins, it is also a macromolecule that the cell mustsynthesize, maintain, and repair. DNA therefore also contains signals involvedin its own metabolism that are recognized by specialized enzymes of which it isthe substrate.

Methylation: Conservation of the genetic heritage is a concern of vital impor-tance for cells, which have developed mechanisms for protecting the quality oftheir DNA. Polymerization errors may occur during the replication process,resulting in mispairing between the template strand and the complementarystrand that is being synthesized. However, cells contain enzymes that can repairsuch defects. In the repair process, one of the two mispaired bases is excisedand replaced by the base that restores correct pairing. The repaired DNAsequence differs according to the stand from which the base is excised:

• If the excision is on the newly synthesized strand, the repaired double-stranded DNA retains the native sequence;

• If excision is on the replication template strand, the repaired double-stranded DNA bears a mutation.

The cell has developed a strategy for distinguishing the template strand fromthe newly synthesized strand. Specialized enzymes introduce chemical modifi-


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (102)

cations (methylation of certain bases) at specific sites on the template strand,but not on the just-synthesized strand, enabling the repair enzyme to distinguishone strand from the other. For example, in E. coli, there are two modifica-tion systems, dam methylase, which methylates adenines in the symmetricalmotif



and dcm methylase, which methylates cytosines in the quasi-symmetrical pattern

5′CC(A or T)GG3′

3′GG(T or A)CC5′

Similar specific cytokine methylations exist in animal cells for 5′-CG-3′sequences and in plants for 5′-CNG-3′ sequences (N can be any nucleotide; A,T, C, or G).

Restriction: An analogous mechanism is used by prokaryotes for defenseagainst bacteriophage viruses. The bacterial cell produces an endonuclease thatrecognizes a specific DNA sequence and cleaves the two strands of the doublehelix. This action of this enzyme permits degradation of the viral genome, pre-venting infection. In order to avoid cutting its own chromosome, the bacteriaalso contains a methylase that recognizes the same site as the endonuclease andmodifies some of the site’s bases, inhibiting the nucleolytic activity of theendonuclease. At least one strand of the bacterial chromosome is always methy-lated at these sites, protecting it from the endonuclease.

Several varieties of restriction enzymes exist, all of which recognize very local,well-defined motifs of between four and a dozen bases. The strands are usuallycut at the recognized site, or in the immediate vicinity. Some endonucleases/methylases, such as BamHI, have perfectly symmetrical (palindromic) sites:



Others, such as HindII, have quasi-palindromic sites, with several possiblebases at certain positions:

5′GT(C or T) (A or G) AC3′

3′CA(A or G) (C or T) TG5′

Finally, a small number, such as HphI, have non-palindromic sites:




Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (103)

Replication: DNA contains specific patterns involved in replication. Replica-tion usually starts at sites known as replication origins, according to variousmechanisms that vary as a function of the organism and type of DNA (chro-mosome, plasmid, or virus). Replication generally begins with the binding of aspecialized protein to a nucleotide pattern located at the replication origin. Thisprotein promotes fusion of the double helix, required to initiate replication. Forexample, in E. coli, the chromosome replication origin bears four tandem repetitions of the following nine-base sequence:

5′TT(A or T)T(A or C)CA(A or C)A

which is recognized by the initiation protein known as DnaA.In general, a great number of proteins involved in DNA metabolism ‘selec-

tive’ for certain nucleotide sequences. Therefore it is possible to identify specificpatterns at which certain events occur, for example, recombination, integrationof a viral genome, and the action of certain gyrases.

4.4 Specific sites

Therefore, as just discussed, identification of the DNA sequence of a site thatcorresponds to a biological function consists in searching for a pattern. Theessential characteristics of the principal types of sites encountered in biologicalsequences are listed below. Studying these signals reveals the existence of twodistinct families, according to whether recognition of the specific motif occurson template or transcribed RNA. The mechanisms involved are quite different,leading to the identification of two major classes.

4.5 Sites located on DNA

Proteins are able to ‘read’ double-stranded DNA bases

As seen in the above examples of signals directly recognized on double-strandedDNA, there is almost always a protein that ‘reads’ the pattern. To understandhow proteins are able to decode a sequence of bases requires a review of theDNA double helical structure.

The two deoxyribose phosphate chains wind around each other on theoutside of the molecule, inside which the basepair plates are stacked perpen-dicular to the axis of the helix. This arrangement forms two lateral grooves of unequal size in which the basepairs are accessible edgewise. Proteins can slide into either of these grooves (most often the larger one), thereby enteringinto specific interactions with particular basepairs. These basepairs expose different chemical functions, which play either acceptor or donor roles in


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (104)

hydrogen bonding with the lateral chains of the amino acids of proteins lodgedin the grooves.

Dimeric proteins bind to palindromic sites

In analyzing the various types of signals that are recognized by proteins, itbecomes clear that a large number of them correspond to palindromic or quasi-palindromic sites. For the DNA molecule, this consists of two-fold symmetryaround an axis centered on the middle of the motif. In the majority of cases,these sites are recognized by proteins that consist of an even number of sub-units that also have two-fold symmetry, dimers, tetramers, etc. . . . Each half ofthe DNA site is recognized by one of the halves of the protein, which has the same axis of symmetry as the DNA in the complex.

This symmetrical subunit arrangement can accommodate a large number ofregulatory proteins, methylation enzymes, and restriction enzymes, and allowsthe cell to increase interaction specificity by doubling the size of the pattern recognized without increasing the size of the protein subunit involved in therecognition. Figure 4.3 illustrates such an interaction between a dimeric proteinand a palindromic site, the protein cro of bacteriophage l. This regulatoryprotein binds to viral DNA on sequences of the following type:


Figure 4.3 Structure of a complex consisting of the bacteriophage lambda cro regulatory proteinand its specific DNA site.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (105)



and stimulates transcription of genes implicated in the lytic cycle of the virus.

Some complex sites consist of several subsites

Certain factors involved in the recognition of specific DNA sites are complexstructures, such as nucleoprotein assemblies (for example, ribosomes) and verylarge proteins consisting of multiple subunits (for example, RNA polymerase).They may therefore be in contact with several distinct regions on the DNA. Eachregion constitutes an independently recognized sub-site, the spacing betweenwhich varies over a more or less elastic range of values. Since the contact zonesare discontinuous, therefore distributed over a greater distance, this variablespacing may possibly be due to the flexibility of both the DNA double helix andthe molecule with which it is in contact.

Transcription promoters are examples of the most well known complex sites,both in prokaryotes and eukaryotes. They consist of several distinct ‘boxes’ separated by zones of various sizes. Translation start sites in prokaryotes mayalso be considered complex sites, since they consist of both the Shine-Dalgarnosequence and the start codon, situated about a dozen bases downstream.

Expression signals are often fuzzy patterns

Until now, only patterns strictly defined by a given nucleotide sequence havebeen covered in this chapter, although in some cases, one or several ambiguitieshave been introduced in the form of alternatives of the type ‘(A or T)’, whichhas the simple effect of multiplying the acceptable sequences; for example, GT(Cor T)(A or G)AC, defines one of the following sequences: GTCAAC, GTCGAC,GTTAAC, and GTTGAC. However, the corresponding sites remain clearlydefined and easily recognizable.

Some proteins, such as methylases and restriction enzymes, recognize suchstrict patterns and tolerate no variation in the admissible sequences. The situa-tion is different with respect to regulation signal sites. In general, it is imposs-ible to define a precise sequence that exactly corresponds to a regulation signalsite. However, based on experimental observations, it appears that sets ofsequences corresponding to regulation signal expression sites more or lessalways tend toward an ideal sequence, which is known as a consensus sequence.

Such fuzzy patterns allow cells to modulate the efficacy of various expressionsignals according to their needs. The closer a signal expression sequenceapproaches to the consensus sequence, the more effective it is. The protein(s)


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (106)

involved can effectively enter into a certain number of molecular interactions(hydrogen and ionic bonding, hydrophobic interactions, etc. . . .) with the basesof the consensus sequence. At positions where the sequence of a site differs fromthe canonic sequence, interactions with the corresponding bases will be broken.The interaction energy will therefore be lower, which destabilizes formation ofthe DNA/protein complex to the extent that the site sequence differs from theconsensus sequence.

Most of the expression and regulation signal sequences mentioned above arecanonic consensus sequence patterns that correspond to maximum effectiveness.For example, the sequences of transcription promoters corresponding to genesthat are very strongly expressed in prokaryotes are often practically identical tothe following canonical sequence:

5′TTGACA = ~17 bases –TATAAT

Less strongly expressed gene promoters may differ more or less significantlyfrom this ideal sequence, either in the spacing between the two patterns or inthe nature of the bases in each conserved cell. In addition, the variation observedamong the bases of the consensus sequence is not constant. Constraints oncertain positions seem to be greater, probably because the interactions they makewith RNA polymerase are stronger and/or more numerous.

This naturally leads to the weighting of various bases recognized at a givensite. Certain bases may be strictly conserved, and are therefore probably essen-tial, whereas others are less strictly conserved. Figure 4.4 illustrates the resultsof statistical analysis of the conservation of the consensus sequence bases in acompilation of 112 experimentally characterized E. coli promoters.

Patterns that induce structural changes in the double helix

Until now, the signals discussed have generally constituted sites recognized byother macromolecules, essentially proteins. In addition, specific sequences existthat induce deformations or other structural changes in the DNA molecule, inthe absence of an external protein factor. These regions can nevertheless also be the targets of interaction with proteins that then recognize the perturbationin the structure, without necessarily specifically ‘reading’ its sequence.


T T G A C A — T A T A A T82% 84% 79% 64% 54% 61% — 79% 95% 44% 59% 51% 96%

Figure 4.4 Average frequency of various bases in the consensus sequence of E. coli transcriptionpromoters.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (107)

The most basic structural change is the appearance of a curvature in the axisof the double helix. The simplest interpretation is that the curvature is causedby an accumulation of deformations at the level of basepair stacking. The geom-etry of this stacking is determined by three angles, as indicated in Figure 4.5.

The twist angle determines the step of the double helix, which on the averageis 34.7° for type B DNA. The roll and tilt angles vary by a few degrees, as afunction of the nature of the basepairs, i and i + 1, which can introduce a slightbend. In a given sequence, these slight deformations cancel each other out,resulting in a nearly straight double helix axis. Nevertheless, in particular cases,the small deformations distribute in phase with the step of the helix, and theircumulative effect yields a non-zero curvature.

In particular, the dinucleotide AA (or TT on the complementary strand)causes significant variations in the roll and tilt angles. Some sequences in whichseries of 3 to 5 A (or T) are regularly spaced every 10 or 11 nucleotides havebeen described. Since the double helix contains around 10.4 basepairs per turn,all the A in such a region are located on the same side of the double helix; there-fore the effects of the induced curve are cumulative.

In general, regularly repeating nucleotide sequence patterns whose periodic-ity corresponds to that of the double helix indicate the presence of a curvedregion and/or the binding site for a protein that interacts with one face of theDNA. Figure 4.6 illustrates a region located upstream from a bacterial gene thatbinds a regulatory protein called Lrp. This sequence is probably curved.




Basepair i

Basepair i+1

Figure 4.5 Angles that determine the local geometry of the stacking of two basepairs.









Figure 4.6 Binding site of the Lrp protein. A T-rich segment repeats approximately every 10 to11 bases. The structure of the double helix is indicated below on the same scale.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (108)

4.6 Sites present on RNA

Certain sites are recognized by other RNA molecules with which they can pair

RNA present in the cell is usually single-stranded, which allows it to locally pairwith another RNA strand. The sites recognized are usually fuzzy, and thestrength of the interaction depends on the stability of the duplex RNA formed.Using the empirical rules described in chapter 6 to calculate the sites, it is generally possible to estimate their stability. To do this, the sequence of the complementary RNA strand involved in the recognition must be known, which requires studying the underlying biological mechanism.

Among the recognition mechanisms described that utilize RNA:RNA pairing are translation initiation in prokaryotes, in which the 3′-end of the 16S ribosomal RNA fragment pairs with the Shine-Dalgarno sequence of messenger RNA, and alternative splicing in eukaryotes, in which intron-extronjunctions interact with small RNA fragments in the nucleus (snRNP orSNURPs).

Certain sites correspond to local stem and loop RNA folding. Single-strandedRNA can also pair with itself (internal pairing) and fold locally, forming stemand loop structures. Such secondary structures can play multiple roles; forexample, they can serve as specific sites for proteins that recognize the shaperather than the sequence of an RNA segment. Inversely, RNA regions involvedin internal pairing cannot interact with other RNA molecules. Some sequencesmay be ‘masked’ by the formation of a stem and loop structure. Finally, as inthe case of DNA transcription terminators, the presence of such a secondarystructure can hinder the progress of a polymerization enzyme, leading to a pauseor block.

The existence of an RNA segment able to fold into stems and loops is revealed by the presence of extended Palindromic regions (repeated inverseregions), such as the DNA binding sites of dimeric proteins (Section 4.5). When such palindromes are detected in a sequence, the biological context must be analyzed in order to distinguish between these two alternatives: a symmetric site on the DNA molecule, and a stem-and-loop structure on the transcribed RNA.

4.7 Pattern detection methods

The sequencing programs used today generate huge quantities of raw data fromwhich biologically pertinent information must be extracted, such as the loca-tions of genes, expression signals, etc. This analysis essentially involves two com-plementary approaches:


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (109)

• Pattern detection, which consists in locating defined sequence segmentsthat correspond to documented biological functions, such as thosedescribed in the preceding paragraphs;

• Content search, which analyzes the properties of the statistical distribu-tion of nucleotides or amino acids in a sequence. This approach will bedescribed in Chapter 5, ‘Statistics and sequences’.

The following is an intentionally simplified description of the methods usedin pattern detection. Known as pattern matching in basic computer science, thisheavily studied domain employs high-performance algorithms that employ veryformal methods. In what follows, these algorithms will be approached usingonly practical ‘biological’ examples. For a fully formal treatment of this subject,the reader is advised to refer to specialized works (see bibliography.)

Simple searches

Numerous patterns associated with biological functions are relatively simple innature. For example, the eukaryotic messenger RNA polyadenylation signal isa non-degenerate hexanucleotide whose sequence is AAUAAA (AATAAA in the corresponding DNA sequence). The search for occurrences of such a non-degenerate pattern in a long DNA sequence may be conducted using the following naïve algorithm:

i ←1k ← 1Do

If Pattern[k] = Sequence[1+k] thenk ←k+1

Elsek ← 1; i ← i+1

If k > Length(pattern) thenpattern found at position ik ← 1; i ← i+1

While i ≤ Length (Sequence)

This simple algorithm consists in nucleotide-by-nucleotide comparison of thepattern with the target sequence at a given position. If the comparison fails, thepattern is advanced one position and the operation repeated. This algorithm hasthe advantage of being simple; its cost is simply a function of the target sequencelength. Nevertheless, it is not optimal, since it may require a number of com-


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (110)

parisons (line shaded in gray) that are greater than the length of the sequenceitself. To demonstrate this, it suffices to observe what happens when the targetsequence is very rich in A and U1: For example,


At the beginning of the search, the situation is the following:


The naïve algorithm will produce two positive comparisons (+) for the twofirst pairs of A, before failing in the third (−). It will then shift the pattern oneposition and repeat the analysis. This time, it will fail in the fifth comparison,after four positive tests:


The algorithm compared a total of eight positions just for the two positionstested. This example was particularly unfavorable, but for strongly biased targetsequences, it is possible to conduct twice as many comparisons as the numberof nucleotides they include2.

This method is not optimal, since when the naïve algorithm yields severalcoincidences between the pattern and the sequence before failing it returns tothe target sequence, in which case it reads some target sequence nucleotidesseveral times. It is possible to improve the naïve algorithm by making it ‘remem-ber’ what has already been read, thus avoiding going back.

Searching with a finite state automaton

To keep track of information concerning nucleotides that have already beenread, a virtual machine known as a finite state automaton is devised. Themachine is characterized by a finite number of certain states. The automaton


1 While not a textbook case, this problem is frequently encountered with the genomes of some organisms that are particularly rich in A and T (see chapter 5, ‘Statistics and sequences’).2 Even in the best case, that of a target sequence in which the four nucleotides are distributed equally,it may be shown that the cost of the naïve algorithm is 35 percent greater than that of an optimal algorithm.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (111)

begins to function in a particular state, called the initial state. The machine readsone nucleotide in the sequence during each cycle, then evolves toward a newstate, according to the nucleotide and its present state. If the automaton reachesanother specific state, known as the final state, the pattern has been found andthe automaton stops.

The following three elements constitute a finite automaton:

• An alphabet Σ of characters for the target sequence (for biological appli-cations, four nucleotides or twenty amino acids);

• A list E of n states e0, e1 . . . en. This list includes an initial state, e0, andone or more final states.

• A transition function T which is an application of E × Σ → E. This func-tion determines the i + 1 state, starting from the current state i and thecharacter read. This transition function is stored in the form of a matrixof dimensions (n, k) where n is the number of states and k is the size ofthe alphabet Σ (four nucleotides or twenty amino acids).

The automaton can then be simulated by applying a very simple algorithm:

i ← 1state ← e0Do

State ← T[state, sequence[i]]; i ← i+1If state = final_state then

Pattern found at position i; QuitWhile i ≤ Length(sequence)

This algorithm must read each nucleotide or amino acid in the sequence onlyonce and not go back over it. The problem is obviously to specify the list ofstates and transition table, so that the automaton will search for the pattern.Taking the example of the eukaryote polyadenylation pattern, AAUAAA, it ispossible to achieve this with the automaton, a graphic representation of whichappears in Figure 4.7.

This helpful graphic representation permits better understanding of the operating principle of the automaton. In particular, it is possible to see howreading the pattern sought, AAUAAA, leads directly to the final state (state 6).The transition table, T, that corresponds to this graph is indicated below (there is no entry for state 6, since it is the final state).

Using the above automaton to analyze the target sequence, the reader mayverify whether it goes through the states indicated below and identifies thepattern (underlined):


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (112)

target: AAAUAUUAAAUAUUAGACGAAUAAAAGUAUAUUUAGstate: 01223400122340010100123456

The state of the automaton at a given instant constitutes its ‘memory’. Itpermits knowing which left sub-part of the pattern has already been identified.If the automaton is in state 2, this indicates that it has just read two A (adenine).If it reads a third one, which does not correspond to the sequence sought,AAUAAA, the two last A of the three read still correspond to the beginning ofthe pattern.

Degenerate patterns

The Knuth-Morris-Pratt (KMP) method functions well for simple, non-degenerate patterns and may readily be adapted to other, partially degeneratetypes of patterns, such as consensus sequence sites that give rise to alternativesplicing in eukaryotes (exon-intron junction), for example, (C or A) AGGU(AAor C)AGU. Generalizing the KMP approach, it is possible to devise automatacapable of seeking such patterns. For example, Figure 4.8 is a graphical repre-sentation of a finite automaton that searches for a pattern corresponding to pre-


nucleotide readA G C U

0 1 0 0 01 2 0 0 02 2 3 0 03 4 0 0 04 5 0 0 05 6 0 0 3





10 62 3 4 5A A





Figure 4.7 Graphic representation of a finite state automaton seeking a polyadenylation signal,AAUAAA. States are symbolized by numbered circles. The initial state is state 0 and the final stateis state 6. Transition rules are represented by arrows beneath the nucleotides associated with eachtransition. Transitions not represented return to the initial state (for example, if the automatonreads something other than A in state 0, it remains in state 0).

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (113)

sumptive alternative splicing sites. It is more complex than the one in the pre-ceding paragraph, consisting of 13 states and presenting a branched structurethat allows managing the degenerated positions of the pattern. Some arrowscorresponding to the return to states 0, 1, and 2 have been omitted in order tosimplify the figure.

Regular expressions

Finite state automata are extremely powerful tools that can recognize a widevariety of different patterns, not only in the context of biological sequences, butin other domains as well. They are often utilized by text and program editingsoftware that includes sophisticated functions of the ‘search/replace’ type. Ingeneral, these search functions employ a language for identifying somewhatcomplex patterns in the files analyzed. In general, this rather extensive languagemay be reduced to what are called ‘regular expressions’.

Regular expressions are defined with respect to three basic operations:

1. Concatenation, the consecutive linking of two patterns: pattern__1pattern__2

2. Alternation between two patterns, notated ‘or’: pattern__1 or pattern__2

3. Repetition a given number of times (0 or more) of a pattern, notated ‘*’:pattern*


0 14







1 3A



8 10

















Figure 4.8 Simplified representation of a finite state automaton that searches for the pattern (C or A)AGGU(A or G)AGU, which corresponds to sites that give rise to alternative splicing ineukaryote RNA. Not all transitions are represented. Missing transitions, which correspond to reading a C, return to state ➀; those corresponding to reading an A return to state ➁, and theothers return to the initial state.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (114)

According to the application concerned – word-processing, biologicalsequences, program editors, etc – the basic language can adopt slightly differ-ent syntaxes and take on operators that simplify notation, but which can alwaysbe reduced to the above three fundamental operations. The formal frameworkof regular expressions allows defining a very wide variety of interesting biolog-ical patterns, such as codons, open reading frames, elementary signals, and consensus signals, of which some examples follow. In order to reduce the complexity of the expressions, simple notations are introduced, some of whichderive from the syntax of the Unix-based grep program.

For the simple alternatives using alphabetic symbols, the following notationwithin brackets is used:

[ATG] ⇔ (A or T or G)[AG] ⇔ (A or G)

Sometimes ‘N’ or ‘X’ are used as wild cards to indicate any nucleotide oramino acid, respectively:


Among these simple constructions, it is possible to define ‘punctuation’codons, which identify the beginning and end of an open reading frame:

start codon: [AG] TGstop codon: [T([GA]) or A[AG])

A coding codon, that is, a non-stop codon, is longer, but still accessible:

coding codon: [ACG]NN or T([CT]N or G[CGT] or A[CT])

Starting from these constructions, it is possible to define an open readingframe as being a start codon assembly followed by some number of codingcodons (identified by a ‘*’ symbol) and ending with a stop codon.

open reading frame: [AG]TG([ACG]NN or T([CT]N or G[CGT] orA[CT])* T([GA] or A[AG])

It is also possible to define variable spacing between patterns, like the oneseparating the prokaryote ribosome binding site from its start codon (between6 and 12 nucleotides).


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (115)

Ribosome binding site: AGGA or GGAG or GAGGTranslation start site: (AGGA or GGAG or GAGG)(N or NN or NNNor NNNN or NNNNN or NNNNNN or NNNNNNN) [AG]TG

This type of pattern definition by regular expression can be very readily gen-eralized to protein sequences, using the 20-character amino acid alphabetinstead of the four nucleotides.

The PROSITE pattern library (, see Chapter 2)employs regular expression syntax in this way to define more than a thousandpatterns and protein signatures.

Finite automata and regular expressions

The fundamental property of patterns defined by regular expressions is that theyexactly correspond to patterns that may be sought using finite state automata.In other terms, if it is possible to represent a biological pattern of interest usinga regular expression, then it is possible to construct a finite state automatonwhich seeks it in a sequence, and reciprocally.

The following is an example of an automaton that seeks a pattern which isa simplified version of the expression that defines an open reading frame:

[AG]TG (NNN)* T([GA] or A[AG])

Figure 4.9 displays an example of an automaton that recognizes patternsdefined by the above regular expression:

This automaton is different from those described above in that its behavioris not deterministic. For example, if a T is read in state 6 of the sequence ana-lyzed, there exist two possible actions: either going on to state 7, which amountsto trying to identify a stop codon, or returning to state 4; that is, continuingthe open reading frame. Therefore the automaton has several different possibleroutes and a priori it is impossible to say which one leads to the final state, state








0 1 2 3 4 5 6 7




Figure 4.9 A non-deterministic finite state automaton that recognizes open reading frames. Thearrows indicating the return to the initial state are not represented.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (116)

10, which is why this construction is known as ‘a non-deterministic finite stateautomaton’. If in reading a DNA sequence, one of the possible paths in thegraph in Figure 4.9 leads to the final state (state 10), then one occurrence of thepattern sought has been found; that is, an open reading frame.

Non-deterministic automata may be used to conduct searches for complexpatterns contained in biological sequences. In order to simulate the multiplicityof possible paths in graphs for a non-deterministic automaton, a set of statesreplaces the single state used in deterministic automata. The transition functionT must then be modified. Deterministic automata use the application E × Σ →E, in which E is the list of states and Σ the alphabet utilized. For non-deter-minist automata, starting from a given state ei, for a character read in Σ, it ispossible to end up at several different states. In the example used in Figure 4.9,starting from state 6, when a T is read, the automaton ends up in states 7 and4. The transition function T is therefore an application of E × Σ → P(E), whereP(E) represents the set of all subsets of E.


nucleotide readA G C T

0 {0.1} {0.1} {0} {0}1 {0} {0} {0} {0.2}2 {0} {0.3} {0} {0}3 {0.4} {0.4} {0.4} {0.4}4 {0.5} {0.5} {0.5} {0.5}5 {0.6} {0.6} {0.6} {0.6}6 {0.4} {0.4} {0.4} {0,4,7}7 {0.8) {0.9} {0} {0}8 {0.10} {0.10} {0} {0}9 {0.10} {0} {0} {0}





The above transition table corresponds to the automaton represented inFigure 4.9. The lists of states obtained after each transition are indicated inparentheses.

The functioning of such non-deterministic automata may then be simulatedby applying the following algorithm, which is a generalization of the one givenabove for deterministic automata:

i ← 1state_list ← {e0}Do

list ← {}For ei in state_listlist ← union (T[ei, sequence[I]], list)

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (117)

state_list ← listi ← i+1If final_state ∈ state_list thenpattern found; quit

While i ≤ Length (sequence)

The following is an example of an analysis obtained using this automaton.The sequence read is indicated above the list of states through which theautomaton passes upon analyzing each nucleotide. Start and stop codons in the miniature open reading frame detected by the automaton are underlined inthe sequence.


1112 12112115123134564 645479610








A*A or B

Figure 4.10 Principles of automaton assembly. The initial and final states are represented byboldface or double circles, respectively. Concatenation is achieved by merging the final state ofthe first automaton with the initial state of the second. Alternation is obtained by merging theinitial and final states of the two automata. Repetition is obtained by merging the initial and finalstates of the automaton which recognizes the pattern that must be repeated.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (118)

Non-deterministic automata may be constructed directly from regular expres-sions by applying the three basic rules defined above, concatenation, alterna-tion, and repetition. Figure 4.10 is a schematic representation of how toconstruct automata that recognize the expressions (AB), (A or B), and (A*),starting from automata that recognize the expressions A and B, respectively. Theprinciple of these constructions consists in appropriately merging the initialand/or final states of elementary automata. By iterating these assemblies, it ispossible to progressively construct a non-deterministic finite state automatonthat corresponds to any regular expression.


Aho A.V. (1990). Algorithms for finding patterns in strings. In Handbook of Theoreti-cal Computer Science, Volume A: Algorithms (van Leeuwen, J., ed.), pp. 255–300.Elsevier, Amsterdam.

Attwood T.K. (2000). The role of pattern databases in sequence analysis. Brief Bioinform 1: 45–59.

Bairoch A. (1991). PROSITE: a dictionary of sites and patterns in proteins. NucleicAcids Res 19 Suppl: 2241–2245.

Bairoch A., Bucher P. (1994). PROSITE: recent developments. Nucleic Acids Res 22:3583–3589.

Gattiker A., et al. (2002). ScanProsite: a reference implementation of a PROSITE scanning tool. Appl Bioinformatics 1: 107–108.

Gusfield D. (1997). Algorithms on strings, trees, and sequences: In Computer Scienceand Computational Biology. Cambridge University Press, Cambridge, UK.

Knuth D.E., et al. (1977). Fast pattern matching in strings. SIAM J Comput 6: 323–350.Sigrist C.J., et al. (2002). PROSITE: a documented database using patterns and profiles

as motif descriptors. Brief Bioinform 3: 265–274.


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (119)

5Statistics and sequences

5.1 Introduction

Nucleotide sequences extracted from a database usually seem quite unreadable,and at first sight may be difficult to distinguish from a random series of lettersA, T, G, and C. However, this impression is erroneous, since ‘real’ nucleotidesequences are replete with redundancies and statistical biases resulting from theprocesses of evolution and natural selection to which they have been subjected.Studying these biases is extremely informative for the biologist, since theyprovide information concerning the origins of the phenomena responsible forthem, leading to a better understanding of how living cells exploit their geneticinformation. Once the mechanisms involved in such biases have been charac-terized, their analysis and systematic investigation become valuable tools for usein predicting the properties of other biological sequences.

5.2 Nucleotide base and amino acid distribution

This chapter covers the distribution of nucleotide bases in genomes and of amino acids in proteins. Analysis will proceed in the direction of increasing complexity, beginning with monomer frequency, followed by more complex patterns of length n (n-tuples), including their frequencies and correlations withmonomer frequency, as well as with various types of sequences (coding and non-coding strands, introns, exons, etc).

Genome composition

Since the sequence of bases constituting the genome of living species consists ofthe four nucleotide bases, A, T, G, and C, the first point to consider with respectto genome composition is the relative frequencies of the bases. The principle ofbase complementarity in the DNA double-helical structure imposes overall

Bioinformatics: Genomics and post-genomics, Frédéric Dardel and François Képès© 2006 John Wiley & Sons, Ltd

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (120)

equality in the numbers of A and T, and G, and C. What is generally studied isthe (G+C)/(A+T) ratio or the GC percent in a given genome or part thereof.

Well before sequencing methods had been developed, physicochemical tech-niques, such as the measurement of DNA density and fusion temperature, wereused to demonstrate the wide disparities in the genomic GC content of variousspecies. These differences were later confirmed by statistical analysis ofsequences extracted from databases.

GC/AT distribution in the E. coli genome is more or less equal; around 51to 49 percent. However, equality is not the general rule, and the GC contentcan range between 15 and 70 percent, according to the species. Thus, onlyaround 18 percent GC is found in the genome of the malaria protozoan Plas-modium falciparum, versus 68 percent in Thermus thermophilus, a bacteriumthat lives in hot springs able to thrive in temperatures exceeding 80°C (seeFigure 5.1). The overall GC level in vertebrate genomes is intermediate, rangingbetween 40 and 45 percent. However, the situation is somewhat more complex,since the vertebrate genome is segmented into large regions called ‘isochores’,in which the GC level is locally constant, but different from that in neighbor-ing regions.

The origin of the variation in the GC rate among various living species is notentirely clear. In thermophilic organisms, the genome generally includes a highproportion of GC, which permits DNA to better resist thermal denaturation,since G-C pairing is more stable than that of A-T.

Comparison of proteins

In contrast to DNA, the average protein composition remains relatively con-stant throughout all living organisms. However, there is some variation, accord-


Plasmodium falciparum (18 % GC)





Thermus thermophilus (68 % GC)





Figure 5.1 Sequence of two gene fragments derived from organisms whose genomes contain verydifferent proportions of GC. Initiation codons are underlined.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (121)

ing to the type of protein considered. For example, membrane proteins are richin hydrophobic residues, and proteins that interact with nucleic acids are oftenmore basic, but these are general tendencies that depend little on the species. Inthe ‘standard’ protein composition indicated in Figure 5.2 it may be seen thatthe distribution of the 20 amino acids is unequal: some amino acids are rela-tively abundant, while others are less common. The average amino acid distri-bution is a kind of universal signature that may be used to distinguish a realprotein sequence from a random one.

N-tuple frequency

Analysis of the incidence of simple nucleotides may be extended to dinu-cleotides, trinucleotides . . . and more generally, to n-tuples. Triplets will bespecifically treated below, in connection with the genetic code.

In this type of analysis, the frequency of occurrence of an n-tuple of the basesB1B2 . . . Bn, fB1B2 . . . Bn, is compared with the product of the frequencies of occurrence of the individual bases fB1.fB2 . . . fBn. If it is lower, the n-tuple isunderrepresented, if higher, it is overrepresented.

This kind of analysis may be carried out on an entire genome, or on two fam-ilies of disjoined sequences in parallel, in attempting to discriminate betweenthem (e.g., coding or non-coding sequence; intron or exon). Among the mostspectacular results of such analysis are:

• Marked underrepresentation of the CG dinucleotide throughout all verte-brate genomes. The CG sequence is a cytosine methylation signal. The 5-methylcytosines located at these sites can be transformed into T (thymine)by deamination, a common type of chemical damage. It is clear then that,little by little, owing to the effect of successive mutations, CG sequencescan become TG sequences, thus becoming increasingly rare in the genome.


Alanine A: 8.3% Methionine M: 2.4% Cysteine C: 1.7% Asparagine N: 4.4% Aspartate D: 5.3% Proline P: 5.1% Glutamate E: 6.2% Glutamine Q: 4.0% Phenylalanine F: 3.9% Arginine R: 5.7% Glycine G: 7.2% Serine S: 6.9% Histidine H: 2.2% Threonine T: 5.8% Isoleucine I: 5.2% Valine V: 6.6% Lysine K: 5.2% Tryptophan W: 1.3% Leucine L: 9.0% Tyrosine Y: 3.2%

Figure 5.2 Average protein composition throughout all living organisms.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (122)

• Underrepresentation of the CTAG palindromic quadruplet in the E. coligenome:

In general, palindromes appear to be underrepresented in bacterial genomes.

Frequency of base triplets; use of the genetic code

The fact that DNA includes genes that code for proteins imposes additional con-straints on the nucleotide sequence that constitutes a gene. Since translation ofDNA into protein uses elementary words called codons, which consist of threebases, it is reasonable to examine the frequencies of the 64 possible triplets. Thepresence of long open reading frames without stop codons is already a first indi-cation of the non-random character of their distribution.

Respecting the reading frames, it is possible to analyze the frequency of the61 base triplets corresponding to the 20 amino acids. Codon distribution mustnecessarily follow that of the corresponding amino acids in the genetic code.For example, on average, tryptophan residues constitute 1.3 percent of proteins;therefore genes should consist of 1.3% TCG, the only codon that correspondsto that amino acid.

It is more interesting to look at what happens in the case of amino acids thatare coded by several synonymous codons. For example, lysine, which represents5.7 percent of the amino acids present in proteins, can be specified by twocodons, AAA and AAG. Are these two triplets represented in an equal manner?For 100 lysine codons in the human genome, there are only around 38 AAAfor every 62 AAG. In the E. coli genome, the proportion is reversed, with 60AAA for every 40 AAG. Cells therefore express a preference for certain syn-onymous codons. This preference is also species-specific.

It is possible to compute the statistics of the occurrence of various codons inthe genes of a species, compiling what is known as a codon usage table. Figure5.3 gives the codon usage tables for E. coli and for the human. The variouscodon frequencies observed are the result of two superimposed effects: theamino acid composition of the proteins (which is not uniform) and the system-atic preference for certain codons among the various possible synonymouscodons. Certain triplets, such as the arginine codon AGG in E. coli and thealanine codon GCG in the human, which appear to be systematically avoided,are known as ‘rare codons’.

When compiling codon usage tables for different genes derived from the sameorganism, it is clear that they are all very similar, in general, faithfully reflect-ing the overall table. Thus, every gene generally conforms to these preferencerules, which are a kind of ‘signature’ of the genome in which it occurs.

f f f f fCTAG C T A G= × << ≈ ×− −3 6 10 3 9 104 3. .


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (123)

In addition, notice that the preferential use of certain synonymous codons isconserved over the course of evolution. The tables of neighboring species oftenreveal very similar codon usage, and to some extent, their comparison may beused to evaluate their degree of evolutionary relationship.

It has been observed that extreme (very high or very low) GC levels in thegenomes of organisms have a significant effect on codon selection. In genomesthat are GC-rich, codons ending in C or G are very strongly preferred, whereasthe opposite is true for genomes that are AT-rich. Generally, even in specieswhose genomes have a nearly equal GC/AT ratio, the constraints imposed bythe codon usage table result in a more accentuated bias for bases located in thethird codon position.


Phe TTT: 1.9 Ser TCT: 1.0 Tyr TAT: 1.5 Cys TGT: 0.6 Phe TTC: 1.8 Ser TCC: 1.0 Tyr TAC: 1.4 Cys TGC: 0.5 Leu TTA: 1.0 Ser TCA: 0.6 TAA: STOP TGA: STOP Leu TTG: 1.1 Ser TCG: 0.8 TAG: STOP Trp TGG: 1.3 Leu CTT: 1.0 Pro CCT: 0.6 His CAT: 1.1 Arg CGT: 2.5 Leu CTC: 1.0 Pro CCC: 0.4 His CAC: 1.1 Arg CGC: 2.2 Leu CTA: 0.3 Pro CCA: 0.8 Gln CAA: 1.3 Arg CGA: 0.3 Leu CTG: 5.5 Pro CCG: 2.4 Gln CAG: 3.0 Arg CGG: 0.4 Ile ATT: 2.7 Thr ACT: 1.1 Asn AAT: 1.6 Ser AGT: 0.7 Ile ATC: 2.8 Thr ACC: 2.4 Asn AAC: 2.5 Ser AGC: 1.5 Ile ATA: 0.4 Thr ACA: 0.6 Lys AAA: 3.7 Arg AGA: 0.2 Met ATG: 2.7 Thr ACG: 1.2 Lys AAG: 1.2 Arg AGG: 0.1 Val GTT: 2.1 Ala GCT: 1.8 Asp GAT: 3.2 Gly GGT: 2.9 Val GTC: 1.4 Ala GCC: 2.3 Asp GAC: 2.3 Gly GGC: 3.1 Val GTA: 1.2 Ala GCA: 2.0 Glu GAA: 4.4 Gly GGA: 0.7 Val GTG: 2.5 Ala GCG: 3.3 Glu GAG: 2.0 Gly GGG: 0.9

Phe TTT: 1.6 Ser TCT: 1.3 Tyr TAT: 1.3 Cys TGT: 1.0 Phe TTC: 2.3 Ser TCC: 1.8 Tyr TAC: 1.9 Cys TGC: 1.5 Leu TTA: 0.5 Ser TCA: 0.9 TAA: STOP TGA: STOP Leu TTG: 1.1 Ser TCG: 0.4 TAG: STOP Trp TGG: 1.4 Leu CTT: 0.1 Pro CCT: 1.6 His CAT: 0.9 Arg CGT: 0.5 Leu CTC: 2.0 Pro CCC: 2.0 His CAC: 1.4 Arg CGC: 1.1 Leu CTA: 0.6 Pro CCA: 1.4 Gln CAA: 1.1 Arg CGA: 0.5 Leu CTG: 4.3 Pro CCG: 0.6 Gln CAG: 3.4 Arg CGG: 0.4 Ile ATT: 1.5 Thr ACT: 1.3 Asn AAT: 1.7 Ser AGT: 1.0 Ile ATC: 2.4 Thr ACC: 2.3 Asn AAC: 2.3 Ser AGC: 1.9 Ile ATA: 0.6 Thr ACA: 1.4 Lys AAA: 2.2 Arg AGA: 1.0 Met ATG: 2.3 Thr ACG: 0.7 Lys AAG: 3.5 Arg AGG: 1.1 Val GTT: 1.0 Ala GCT: 2.0 Asp GAT: 2.2 Gly GGT: 1.1 Val GTC: 1.6 Ala GCC: 2.9 Asp GAC: 2.9 Gly GGC: 2.5 Val GTA: 0.6 Ala GCA: 1.4 Glu GAA: 2.7 Gly GGA: 1.7 Val GTG: 3.1 Ala GCG: 0.7 Glu GAG: 4.1 Gly GGG: 1.7

Figure 5.3 Escherichia coli (top) and human (bottom) codon usage tables. The frequencies ofvarious codons are indicated in percentage. Highly preferred codons are shaded gray.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (124)

Context-dependent codon biases

In cases in which several synonymous codons are able to code for the sameamino acid, each organism expresses preferences that may be expressed as fre-quencies and listed in a codon usage table. Other effects may add to this schemeof preferences. When several codons of nearly equal probability are possible,the choice among them may be influenced by the neighboring codon; that is, bynucleotides located immediately upstream and/or downstream. In E. coli, it hasbeen noticed that for the two lysine codons AAA and AAG, the former is morefrequently encountered when the next codon starts with a G, and reciprocally,that AAG is preferred when there is a C immediately downstream.

These biases are collectively known as ‘context effects,’ some of the mostmarked of which are indicated below for the Escherichia coli genome:

AAA-G > AAG-G AAG-C > AAA-C LysineGAA-G > GAG-G GAG-C > GAA-C GlutamateGGC-G > (GGG-G, GGT-A > (GGT-C, GGT-T, GGT-G) Glycine

GGA-G, GGT-G)TTT-G > TTC-G Phénylalanine

5.3 The biological basis of codon bias

Codon utilization adapts to transfer RNA concentrations

It has been shown that the frequency of utilization of each codon in the yeastSaccharomyces cerevisiae and in the bacterium E. coli is directly proportionalto the intracellular concentration of transfer RNA (tRNA) that decodes it. Forexample, E. coli has two transfer RNA isoacceptors for isoleucine, tRNAIle

1,which decodes the ATT and ATC codons (respective frequencies: 2.7 and 2.8percent), and tRNAIle

2, which decodes the ATA codon (frequency: 0.4 percent).The intracellular tRNAIle1/tRNAIle2 ratio is 20/1, which is close to the ratio ofthe decoded codons (27 + 28)/4.

This adaptation optimizes the protein translation process by exactly adjust-ing the tRNA demand of the translation machinery to the amount available inthe cell.

The most highly expressed genes are the fittest

Comparing the codon usage table of individual genes with the ‘canonic’ tablebased on the full genome reveals that the genes which most closely follow thecanonic table are those most strongly expressed in the cell (ribosomal proteins,


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (125)

translation factors, structural proteins, etc). The selection pressure on thesegenes is the greatest, resulting in stronger bias than for weakly expressed genes.

At each stage of the translation process, tRNA molecules in the cytoplasmdiffuse into the active site of the ribosome, which ‘tests’ whether the codon-antocodon interaction is correct and ‘rejects’ inappropriate tRNA molecules.The ribosome ‘waits’ for the tRNA molecule that possesses the anticodon cor-responding to the codon. This wait is proportional to the number of unfruitfultests that the ribosome must carry out before the appropriate tRNA moleculearrives, which depends on the relative abundance of the isoacceptor soughtamong the set of tRNA molecules:

The most abundant glycine codons in E. coli are GGT and GGC. They aretranslated by tRNAGly3, which accounts for 6.5% of all tRNA in the cell. Ribo-somes must therefore carry out around 15 tRNA trials (1/0.065) when trans-lating one of those codons. It takes the ribosome more than 300 attempts totranslate a rare codon such as ATA, which codes for isoleucine, since the cor-responding isoacceptor represents only 0.3% of the total tRNA. When ribo-somes encounter a rare codon, there is a pause in translation. By utilizing onlycodons that correspond to the most abundant isoacceptors, translation of themost highly expressed genes is able to avoid delays that can slow cell growth.

The rarity of the CG dinucleotide in vertebrates affects the codon usage table

As described earlier (5.2), due to the effect of mutations, the CG dinucleotide,which is a methylation site, tends to disappear in vertebrates. As a result, codonsthat include the CG doublet are underrepresented in vertebrate codon usagetables. As may be seen in Figure 5.3, this phenomenon is particularly evidentfor proline (CCN), threonine, (CAN), and alanine (GCN), in all of which a Gin the third position is strongly disfavored. This is also reflected in contexteffects; codons ending in C upstream from a codon that begins with G are systematically avoided.

5.4 Using statistical bias for prediction

The statistical biases described in the preceding sections may be used for pur-poses of prediction in answering questions of biological importance, such as the following: ‘Does this DNA region code for a protein?’ ‘Which is the coding

< >= [ ] [ ]

number of tRNA molecules tried

total tRNA molecules tRNA molecules sought


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (126)

strand in this DNA sequence?’ ‘Which is the coding phase?’ ‘Does this sequencecontain an error?’ ‘What are the boundaries between introns and exons?’ ‘Isthis gene strongly expressed?’

In chapter 4, sequence characteristics were investigated according to signalsdetermined by specific patterns. In this section, they will be pursued by identi-fying their contents; i.e., by determining the statistical properties of their dis-tribution as a function of the bases that constitute them.

The search for coding sequences

The most powerful method for determining whether a stretch of DNA is acoding region, and which relies on the fewest a priori hypotheses, consists inseeking period 3 irregularities in nucleotide distribution. Non-uniform codonusage within a gene or exon should be revealed by period 3 bias in the fre-quencies of occurrence of individual nucleotides. This method requires no priorknowledge of the codon usage table for the organism, or even of the geneticcode. The frequency of occurrence of each base at positions 3i, 3i + 1, and 3i + 2 are simply calculated and compared with the average frequency of occur-rence in the sequence.

The value of ∆ is calculated within a window of around 100 nucleotidesmoved progressively along the sequence being studied. If the value of ∆ is tracedas a function of the position of the window, a profile is obtained whose peaksfairly well represent the coding regions.

This method is generally well adapted to seeking introns and exons in eukary-otic genomic sequences. Splice site consensus sequences are usually insufficientto accurately determine intron/exon boundaries, for which the content researchmethod can be an effective complement.

Coding phase analysis

The above technique may be considerably refined by utilizing the codon usagetable for the species from which a sequence derives, thereby allowing predic-tion of which of the three phases is coding. The principle is to compare the frequencies of triplets appearing in the three phases with the canonical tripletfrequencies, then to determine which of the former are most similar to the latter.

∆ = −∑∑=

f fN phase i NphasesN A T G C 3, , ,


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (127)

Practically speaking, the frequency of occurrence f1 of N codons in a windowon phase 1 is compared with the individual frequencies of codons in the stan-dard genetic code usage table:

By shifting first one and then two nucleotides, the calculation is repeated forphases 2 and 3, which gives two other frequencies, f2 and f3. The probability ofeach phase being the coding phase is given by Bayes’ formula:

By displacing the length N window along the sequence, it is possible to tracethe profiles of probabilities p1, p2, and p3, whose peaks indicate the positions ofgenes that code for proteins with remarkable precision (see Figure 5.5).

This method is very sensitive and can be used with shorter windows (around12 codons). It is extremely useful for detecting insertion and deletion sequenc-ing errors, which can result in a reading frame shift and thus easily be detectedby using the probability graph, p1, p2, and p3 (Figure 5.4).

p f f f f3 3 1 2 3= + +( )

p f f f f2 2 1 2 3= + +( )

p f f f f1 1 1 2 3= + +( )

f fcodon ii










Figure 5.4 Coding probability profile obtained by analyzing genetic code usage. Thick lines cor-respond to actual coding regions. Short vertical lines in the phase graphs indicate the positionsof stop codons.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (128)

Gene expression level

The expression level of a gene may be estimated by comparing its codon usagefrequency with the standard codon frequencies given in the codon usage tablefor the species. The gene expression level may be quantified using a quantityknown as the codon adaptation index (CAI), which is calculated as follows:

Each codon i contained in the gene is assigned a score wi equal to the ratioof its frequency to the frequency of the most frequent codon that codes for thesame amino acid. If codon i is the most frequent, then wi equals 1. For a codonthat is systematically avoided, wi is close to 0. The codon adaptation index isthe geometric mean of the wi scores of the set L of the gene’s codons.

The CAI score obtained ranges between 0 and 1, increasing as the gene con-forms to the standard utilization frequency of the species’ genetic code. Forexample, protein genes that are strongly expressed in the yeast, such as riboso-mal proteins and histones, have scores of between 0.52 and 0.92, whereas regulatory protein genes, of which there are only a few copies per cell, have ascore of 0.1.

The CAI may be used to estimate the expression level of a gene whose func-tion is unknown. It is also useful when expressing a recombinant protein in a het-erologous host, for example, a human protein in a bacterium. The CAI for ahuman gene, in combination with the bacterial codon usage table, allows us topredict whether a gene will be efficiently expressed, and may be used to guide themodification of certain codons in order to better adapt the gene to its new host.

5.5 Modeling DNA sequences

In his novel, ‘Jurassic Park’, the American author, Michael Crichton, tells thestory of a molecular biologist, a certain Dr. Wu, who sequences the DNA ofdinosaurs preserved in yellow amber in order to resuscitate the extinct animals.

Index wii







3’ nucleotide A C G T A 0.102 0.054 0.071 0.074 C 0.077 0.057 0.010 0.069G 0.059 0.046 0.054 0.048

editoelcun’5 T 0.062 0.057 0.072 0.087

Figure 5.5 Dinucleotide frequencies in the human genome.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (129)

To support his fiction, the author even went to the extent of devising a figurethat supposedly represented a ~1.5 kbase fragment of dinosaur genome!

Is this dinosaur DNA sequencing credible? For the connoisseur of biomath-ematics, it does not withstand scientific scrutiny, hence it is just pseudoscience.Subjecting the fictional sequence to the statistical tests described above revealsthat it does not present the biases expected for vertebrate DNA, such as the40% G + C level and low CG dinucleotide frequency (see Section 5.2). But whatwould a novelist or biological faker have to do to concoct a pseudosequencecapable of mystifying the shrewdest specialist?

Clearly, the solution would be to find a method that generated sequences dis-playing exactly the same statistical biases as real biological ones.

Markov chains

A very simple first approach is to attempt to reproduce both the nucleotide com-position of the genome and its CG nucleotide content. Elementary statisticalanalysis of a body of vertebrate genomic sequences yields the following resultsfor the relative frequencies of nucleotides:

The nucleotide frequencies are indicated in Figure 5.5, in which the frequencyof the CG dinucleotide (in boldface) is observed to be lower than that of theothers:

A very simple approach for generating a sequence that reproduces these fre-quencies involves iteration, by adding one nucleotide at a time. Suppose that aC has just been added at position i. Using the above table, it is easy to calcu-late the probabilities of encountering A, C, G, or T after that C. For example,the probability for an A to follow a C is:

Equivalent probabilities may be symmetrically calculated for the other threenucleotides, C, G, and T. They are indicated in Figure 5.6 for the example ofvertebrate sequences, expressed in percentages. Within rounding-off errors, linetotals should equal 1 (100 percent).

Utilizing these probabilities in conjunction with a random number generator,it is possible to progressively generate a pseudosequence that mimics the biasesobserved in vertebrate sequences. This method of random fabrication using aprobability table to generate state i + 1 from state i is called a Markov chain


ff f f f





( ) = =+ + +

f f f fA T C G= = = =0 30 0 29 0 21 0 20. , . , . , .


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (130)

(or Markov process), well known in the field of applied mathematics. If state iis called ei, the probability p(ei+1|ei) that i will go to state ei+1 is given by theMarkov process transition table, as indicated above. A simple Markov chain isa discrete random process that takes only the immediately preceding state intoaccount in generating the current state.

Higher-order biases

The simple Markov chain just described is a perfect tool for reproducing DNAcomposition in terms of mono- and dinucleotides. Nevertheless, as seen above,higher-order irregularities exist in genomes, in particular, period 3 biases relatedto the translation of the genetic message by codons (i.e., nucleotide triplets). Tosimulate these biases and produce ‘plausible’ sequences, we can use an exten-sion of simple Markov chains. All that is required to produce state i is to utilizemore complex probability tables that not only take state i − 1 into account, butalso states i − 2, i − 3 . . . i − k. In this case, we speak of an ‘order k Markovprocess’. State ei is then produced with probability p(ei|ei−1 . . . ei−k). With aMarkov process of order n − 1, it is possible to produce pseudosequences thatreproduce the frequencies of the k-tuples of nucleotides of a family of biologi-cal sequences for all k between 1 and n.

Using Markov processes to study DNA sequences

Up to this point, the objective has been to produce random pseudosequencesthat would fool a biologist, which might seem to be of quite trivial interest.Indeed, the utility of Markov chains lies elsewhere, since they are powerful prob-abilistic tools for testing hypotheses concerning the nature of DNA sequences.Using a Markov transition table makes possible a posteriori computation of theprobability that a given sequence e1,e2 . . . en could have been produced by thecorresponding Markov process. This equals the product of the probabilities ofthe individual transitions:


3’ nucleotide A C G T

A 34% 18% 24% 25% C 36% 27% 5% 32%

G 28% 22% 26% 23%


T 22% 21% 26% 31%

Figure 5.6 Markov chain transition probabilities for human DNA sequences.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (131)

For example, in the above model, it is possible to calculate the probability ofsequence CGCG and to compare it with probability of AATG. The firstsequence, which is rich in CG, is thus shown to be much less probable than thesecond:

prob (CGCG) = 0.05 × 0.22 × 0.05 = 5.5 × 10−4

prob (AATG) = 0.34 × 0.25 × 0.26 = 2.2 × 10−2

This kind of calculation is useful when comparing various hypotheses for thesequence of a given DNA segment, as long as it is possible to associate a Markovprocess with each hypothesis envisaged. The following is a concrete exampleillustrating this type of application.

A biologist has just determined some cDNA sequences obtained from mes-senger RNA derived from a mammalian cell culture. Even when all the requiredprecautions are taken, cell cultures are sometimes contaminated by intracellu-lar bacterial parasites known as mycoplasms. The RNA extraction process doesnot permit separating the nucleic acids of the mammalian cells from those ofthe mycoplasms. At the end of the experiment, some of the sequences obtainedtherefore could have come from the mycoplasms rather than from the mam-malian cells. How can the intrusive sequences be distinguished from the onesof interest?

Several mycoplasm genomes have already been completely sequenced, there-fore it is possible to study their dinucleotide composition and to deduce aMarkov process transition probability table for them, as was done above forthe vertebrate sequences. The resulting values, compiled for Mycoplasma geni-talium, are indicated in Figure 5.7.

Given a fragment of a sequence obtained by our biologist, it is possible tocalculate the probability that it was produced by Markov processes, utilizingthe probabilities derived from either the vertebrate or the mycoplasm genomes.Comparing these two probabilities, it is often also possible to decide whichhypothesis is the most likely. Consider the following sequence as an example:

prob e e e p e en k kk n

1 2 11

. . .( ) = ( )−< ≤∏


3’ nucleotide A C G T A 42% 15% 17% 26% C 40% 18% 6% 36% G 31% 19% 18% 32%


T 26% 14% 19% 42%

Figure 5.7 Markov chain transition probabilities for mycoplasm DNA sequences.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (132)


A posteriori calculation utilizing the two transition tables (vertebrate andmycoplasm) given above yields the following results:

Despite the short length of this sequence (only 25 nucleotides), there is a ratherclear difference, since the Markov process based on mycoplasm DNA producesa result that is four times more probable than the Markov process based on vertebrate DNA. This sequence may therefore be attributed to contaminationwith a mycoplasm bacterium. The differences are even clearer (by several ordersof magnitude) in longer sequences, usually eliminating any ambiguity.

5.6 Complex models

When examining protein-coding regions in a given organism, it is preferable toutilize the codon usage table for the species, such as those shown in Figure 5.3for the human and for E. coli. However, the simple Markov processes describedabove do not allow this to be done directly, since codon usage tables only con-sider triplets that are in phase with reading frames that code for proteins.Triplets that are shifted +1 or +2 nucleotides with respect to the reading frameare not taken into account, unlike in a simple Markov process.

For sequences that encode proteins, it is possible to compile triplet frequencytables for both the correct phase (phase 0), which corresponds to the classicalusage table of the genetic code, and for the two other phases (phase +1 andphase +2). If this statistical approach is used for E. coli, the three tables shownin Figure 5.8 are obtained.

These three tables are rather complicated, but carefully examining themreveals certain interesting features: While some triplets, such as TAG and GGG,are systematically lacking in all three phases, others have very different fre-quencies, according to the phase considered. For example, whereas GAA is relatively abundant in the coding phase (4.05%), it is rare in phase +1 (only0.36%). This observation calls for two important remarks:

• The simple Markov processes discussed above cannot capture the full com-plexity of biases imposed by using the genetic code within coding regions.

• To predict the coding phase, it must be possible to use biases that are notthe same among the three phases.

pmycoplasm S( ) = × −18 7 10 15.

pvertebrate S( ) = × −4 3 10 15.


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (133)


TTT 188 CTT 104 ATT 268 GTT 196TTC 194 CTC 108 ATC 276 GTC 141TTA 110 CTA 29 ATA 29 GTA 116TTG 107 CTG 553 ATG 257 GTG 250TCT 92 CCT 55 ACT 106 GCT 207 TCC 88 CCC 50 ACC 278 GCC 257 TCA 54 CCA 88 ACA 59 GCA 219

phase 0 TCG 78 CCG 262 ACG 133 GCG 337 TAT 157 CAT 119 AAT 142 GAT 309TAC 131 CAC 111 AAC 220 GAC 226TAA 19 CAA 141 AAA 353 GAA 405TAG 0 CAG 309 AAG 87 GAG 185TGT 42 CGT 222 AGT 62 GGT 280 TGC 58 CGC 223 AGC 153 GGC 323 TGA 4 CGA 29 AGA 16 GGA 58 TGG 145 CGG 42 AGG 11 GGG 99

TTT 111 CTT 79 ATT 94 GTT 115TTC 151 CTC 87 ATC 188 GTC 145TTA 185 CTA 86 ATA 183 GTA 157TTG 310 CTG 210 ATG 261 GTG 188TCT 120 CCT 105 ACT 110 GCT 120 TCC 139 CCC 111 ACC 148 GCC 146 TCA 205 CCA 186 ACA 169 GCA 166

phase 1 TCG 253 CCG 271 ACG 262 GCG 325 TAT 39 CAT 61 AAT 102 GAT 18TAC 90 CAC 111 AAC 228 GAC 28TAA 83 CAA 93 AAA 242 GAA 36TAG 71 CAG 157 AAG 346 GAG 24TGT 156 CGT 99 AGT 99 GGT 38 TGC 303 CGC 278 AGC 183 GGC 111 TGA 302 CGA 167 AGA 126 GGA 65 TGG 406 CGG 266 AGG 174 GGG 83

TTT 149 CTT 150 ATT 97 GTT 203TTC 102 CTC 90 ATC 48 GTC 71TTA 108 CTA 98 ATA 36 GTA 65TTG 40 CTG 118 ATG 38 GTG 52TCT 184 CCT 150 ACT 150 GCT 311 TCC 105 CCC 100 ACC 93 GCC 158 TCA 160 CCA 156 ACA 131 GCA 234

phase 2 TCG 121 CCG 138 ACG 83 GCG 173 TAT 219 CAT 201 AAT 156 GAT 252TAC 141 CAC 173 AAC 120 GAC 143TAA 217 CAA 227 AAA 134 GAA 224TAG 32 CAG 125 AAG 44 GAG 41TGT 153 CGT 208 AGT 108 GGT 233 TGC 289 CGC 281 AGC 173 GGC 277 TGA 329 CGA 347 AGA 197 GGA 252 TGG 406 CGG 266 AGG 174 GGG 83

Figure 5.8 Relative frequencies of various triplets in the three reading phases within E. coli genes(occurrences per 10,000 genes).

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (134)

It is nevertheless possible to devise a sequence-coding model that includes thenotion of phase. This requires construction of a kind of Markov process thatutilizes three probability tables deduced from the frequencies listed in Figure 5.8in rotation. During each cycle, this process generates a nucleotide that shifts itsreading phase one position, as per the diagram in Figure 5.9.

This is a modified order 2 Markov process, because knowledge of the twopreceding nucleotides, ei−1 and ei−2, is used to complete the nucleotide triplet byreferring to probability tables: p0(ei|ei−1,ei−2), p+1(ei|ei−1,ei−2) and p+2(ei|ei−1,ei−2) foreach of the three phases, 0, +1, and +2. This Markov process could be consid-ered an automaton, permitting fabrication of very convincing E. coli codingpseudosequences. These pseudosequences also reproduce the preferential codonusage of this bacterium by utilizing the phase 0 probability table, as well as thecontext effects described above, which are captured by +1 phase probabilitiesfor the 3′ context and +2 phase probabilities for the 5′ context. An example ofa generated sequence is:


As for classical Markov processes, the main utility of this type of model obvi-ously is not to produce pseudosequences, but rather to test a posteriori whethera biological sequence does indeed correspond to a coding phase. To do this, itis possible, as shown above, to calculate the probability that a given sequencehad been produced by the modified Markov process:

This probability evaluation alone does not provide much information, but itis interesting to compare it with other probabilities. As discussed above, a directapplication is prediction of the species from which a sequence derives; for

prob e e e p e e en k k k kk n

1 2 3 1 22

. . . ,( ) = ( )− −< ≤∏ mod


Figure 5.9 Phase shifting in a standard coding sequence.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (135)

example, determining whether a sequence has been contaminated by exogenousgenetic material. To do this requires calculating and comparing the probabilities,using Markov processes adapted to various frequencies in the organisms studied.

Another interesting application concerns searching for the coding phase.Given a cDNA sequence presumed to code for a protein, is it possible to deter-mine which is the coding phase? The following three probabilities may be cal-culated, each of which is shifted by one nucleotide with respect to the precedingone, and which correspond to the three phases, 0, +1, and +2:

Their comparison usually allows unambiguous prediction of which of thethree phases is coding. In the following example, the 25 first codons of a naturalE. coli gene are:


In spite of the short length of this sequence, the results obtained using themodified Markov model described above are very clear: the correct phase (phase0) has a probability between ten and one-hundred million times greater thanthe other two phases:

This result is all the more remarkable in that it uses only statistical com-position information and does not directly depend on the presence of start orstop codons, which, as seen in the preceding chapter, is necessary for patternrecognition. This approach is therefore very robust and may be applied withina fragment of any coding sequence. It can also be applied to very short openreading frames (a few dozen codons) to predict whether they are in fact realgenes coding for small proteins or peptides.

5.7 Sequencing errors and hidden Markov models

Recognition of coding phases is a fundamental problem faced by biologistsattempting to annotate genomes. Despite the technical progress that has beenmade in sequencing methods (see Chapter 1), errors inevitably slip into rawdata. When this concerns the insertion or deletion of a nucleotide, the result isa shift in the DNA reading frame. The pattern recognition methods mentionedin the preceding chapter cannot be used to solve this type of problem, whichresults in possible failure to detect a gene in a genome.

prob prob probphase phase phase0 10 1 10 2 1097 105 104( ) ≈ +( ) ≈ +( ) ≈− − −; ;

prob e e e prob e e e prob e e en n n1 2 2 3 1 3 4 2. . . ; . . . ; . . .( ) ( ) ( )+ +


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (136)

The sensitivity of the statistical approach provided by Markov processes inlocating coding phases makes it possible to detect this type of error. In explain-ing how to proceed, the following section will again hypothesize using aMarkov-type process to fabricate false genomic sequences containing errors.

Sequencing errors are relatively rare, with an incidence of around 1 insertionor deletion per 1,000 nucleotides in the raw data. The model employed in Figure5.9 may be used, in which the phases are incrementally cycled, utilizing the asso-ciated probability table each time. The only modification that will be introducedis that the system will occasionally be allowed to ‘slip’; i.e., jump to a phaseother than the one expected. The new diagram of the process (Figure 5.10) presents frequent transitions (solid lines), which correspond to those of thecorrect model, as well as rare transitions, (dotted lines), which correspond toinsertions/deletions.

This system is a probabilistic automaton consisting of 48 possible states. Infact, each transition depends on the two preceding nucleotides e−1 and ei−2, whichrepresent 16 different dinucleotides. These 16 possibilities combined with thethree phases amount to 48 states. The automaton goes from one state to another,as a function of its probability tables. At each transition, it also produces anucleotide that is added to the end of the sequence being fabricated.

The objective now is to utilize this model for a posteriori calculation of theprobability of a given sequence, as was done earlier for classical Markovianprocesses. However, this cannot be done, since if only the final sequence is avail-able, it is impossible to reconstitute the route taken in Figure 5.10. In fact, unlikeall the automata constructed earlier in the chapter, in this one, several differenttransitions can yield the same nucleotide at each step, according to whether anormal transition or an insertion/deletion (dotted arrow shafts) occurs. For thisautomaton, it all depends on whether the current state is in phase 0, +1, or +2.Such a model, in which the sequence produced does not allow reconstruction


Figure 5.10 Schematic representation of the modified model, allowing phase jumps. Solid arrowshafts correspond to normal transitions and dashed arrow shafts to transitions associated withinsertions or deletions. The diagram is simplified, each arrow in reality representing four transi-tions, each of which is associated with one of the four nucleotides, A, T, G, and C.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (137)

of the progression of events of the automaton, is called a hidden Markovprocess, precisely because its states are unknown.

For hidden Markov processes, there is a very large number of possible routes,therefore a series of possible states yielding the same sequence. In the aboveexample, given the three phases, 3n−2 routes exist for a sequence of length n.However, the probabilities associated with them may be very different. In par-ticular, since transitions associated with phase jumps (insertions/deletions) havevery low probabilities, the routes that take them are usually highly unlikely.

The problem at hand is to identify the coding phase and any shifts in it causedby sequencing errors. This information corresponds to the sequence of the statesof the automaton. The aim therefore is to reconstitute the pathway shown inFigure 5.10.

According to which criteria should the route be selected among the 3n−2 pos-sible? The overall probability of each route may be calculated by computing theindividual probabilities associated with each transition as it proceeds. The routewith the highest probability can then be selected in order to reconstruct the onewith the most plausible sequence of states.

The Viterbi algorithm

There are obviously too many probabilities associated with each route to cal-culate them exhaustively. The Viterbi algorithm avoids this difficulty by usinga dynamic programming approach. For a sequence e1e2 . . . en of length n, itbegins by calculating the probability of the best route from e1 to ek for k pro-gressively varying between 1 and n and terminating in each of the possible statesof the automaton. In the example of the search for coding phase, this amountsto seeking the probability of the best route over the k first mucleotides and terminating in either phase 0, phase +1, or phase +2. A table P of dimension 3 × (n − 1) containing all these possibilities is drafted. The table is initialized at1 in the three phases for the second nucleotide e2. By recurrence, it is then pos-sible to calculate the kth column of table P:

Variable i and j represent the index of the phase and pj,i(ei|ek−1, ek−2) repre-sents the probability of transition of the dinucleotide ek−2ek−1 toward thenucleotide ek passing through phase j toward phase i. If i = j + 1 modulo 3, thenit is a standard transition, corresponding to the solid-stem arrow in Figure 5.10,and the probability can be calculated from the tables in Figure 5.8. In the con-trary case, the transition corresponds to an insertion or deletion (dotted-shaftarrow in Figure 5.10) and the probability is lower, on the order of the errorlevel, estimated to be around 10−3.

P i k P j k p e e ej

j i k k k, max , ,, ,

,( ) = −( ) ( )( )=

− −0 1 2

1 21


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (138)

When the table is complete, the last (nth) column indicates which of the threephases has the highest probability. The inverse route is then taken in the tablein order to determine the path it followed. For each column k, this amounts tofinding the value of j in the above equation corresponding to the maximum ofthe three terms. This value of j indicates the most probable state of the hiddenMarkov process at the level of the k − 1th nucleotide. The list of the most prob-able states as a function of k yields a prediction of the correct coding phase (Figure 5.12).

In practice, it is not these probabilities that are calculated directly, but theirlogarithms, which has two advantages: (i) calculation of a product is replacedby calculation of a sum, which is faster, since all that is necessary is to add thelogs of the transition probabilities log(pj,i(ei|ek−1,ek−2)); (ii) using logarithmsavoids the risk of exceeding the capacity of the machine (underflow), since theoverall probability of long sequences rapidly becomes ridiculously small.

Predictions carried out using the Viterbi algorithm are often very accurateand of much higher quality than those produced using profile score methods,such as those presented in Figure 5.4. In the following example, which uses thenucleotide sequence of an E. coli gene, two errors have been introduced inten-tionally: one deletion and one insertion. This kind of double error can be espe-


phase 0

phase 1

phase 2




k-1 k n

Figure 5.11 Completion of table P. For each cell, the probabilities of the three possible transi-tions are evaluated, then the maximal one is selected. The table is completed from left to right.

phase 0

phase 1

phase 2


phase shifts


Figure 5.12 Reverse reconstitution of the path across table P, starting from the final maximumprobability score. In this diagram only the route of the first nucleotide of each codon is drawn;this eliminates the normal cycling across the three phases that would make the figure illegible.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (139)

cially difficult to detect, since the effects on the phase mutually compensate foreach other such that the length of the reading frame is not affected. Neverthe-less, the coding sequence between the two errors is profoundly affected.However, the Viterbi algorithm nearly perfectly detects the two ill-timed changesin the sequence (Figure 5.13).

5.8 Hidden Markov processes: a general sequence analysis tool

The phase shift search discussed above is just one relatively simple exampledeveloped to illustrate the material treated in this chapter. However, the rangeof Markov model applications in the domain of biological sequence analysis ismuch greater. Refining this type of model and having it take into account fre-quencies in non-coding regions and in start and stop codons can convert it intoan automatic genome gene-seeking tool. In eukaryote genomes, in which codingregions are interrupted by introns, statistical analyses have been conducted inboth intron and exon regions and the frequencies obtained used to constructhidden Markov processes. The most probable route reconstituted using theViterbi algorithm may then be utilized to predict spliced regions in messengerRNA. Finally, hidden Markov processes have also been used with proteinsequences to predict secondary structures (a-coils and b-sheets; see Chapter 6).

5.9 The search for genes – a difficult art

The exhaustive search for all the genes contained in a large genome is a complextask. In higher eukaryotes, it is complicated by the ‘dilution’ of relevant infor-mation in non-coding sequences of repetitive DNA and in intergene regions. The







Figure 5.13 Sequence of a fragment of the Escherichia coli infC gene with two phase shifts; firsta deletion, then an insertion. The gray-highlighted area indicates the zone between the two errors.The Viterbi algorithm phase shift prediction is underlined. Only the first nucleotide of the shiftedzone is incorrectly predicted.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (140)

presence of lengthy introns separating short exons renders precise assembly ofcoding zones rather delicate. In such case, none of the methods discussed untilnow is by itself sufficient to identify genes. In general, several differentapproaches are necessary to achieve a reliable result. Currently used approachesare based on the following strategy:

1. Predict exon and intron regions, using a statistical method, most often ahidden Markov process.

2. Search for consensus patterns of donor and acceptor splicing sites, usinga finite automaton method.

3. Combine the two types of information in order to precisely predict intronand exon borders.

4. Assemble the predicted exons and compare the sequence obtained withcDNA and EST sequence databases of the same organism (see Chapter 1).All or part of the predicted sequence may be found in one of thesesequences, which derive from messenger RNA.

5. If this search fails, it is still possible to compare the protein sequences pre-dicted from the genomic DNA (translated into all possible phases) withprotein sequence databases. If the genomic DNA region being consideredeffectively codes for a protein that possesses a hom*olog identified inanother species, a BLAST-type search based on translated sequence frag-ments may be carried out to identify these hom*ologies. This would permitidentifying or confirming some coding parts, thus some intronic parts ofthe gene.

Several very elaborate tools that combine several of these approaches areaccessible on the web, for example, GenScan (, which has been utilized to analyze and document the entire humangenome (accessible on


Burge C., et al. (1992). Over- and under-representation of short oligonucleotides in DNAsequences. Proc Natl Acad Sci USA 89: 1358–1362.

Burge C., Karlin S. (1997). Prediction of complete gene structures in human genomicDNA. J Mol Biol 268: 78–94.

Burge C.B., Karlin S. (1998). Finding the genes in genomic DNA. Curr Opin Struct Biol8: 346–354.

Durbin R., et al. (1998). Biological sequence analysis. In Probabilistic models of pro-teins and nucleic acids. Cambridge University Press, Cambridge, UK.


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (141)

Eddy S.R. (2004). What is a hidden Markov model? Nat Biotechnol 22: 1315–1316.Gautier C. (2000). Compositional bias in DNA. Curr Opin Genet Dev 10: 656–661.Lukashin A.V., Borodovsky M. (1998). GeneMark.hmm: new solutions for gene finding.

Nucleic Acids Res 26: 1107–1115.Zhang M.Q. (2002). Computational prediction of eukaryotic protein-coding genes. Nat

Rev Genet 3: 698–709.


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (142)

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (143)

6Structure prediction

6.1 The structure of RNA

As the material basis of genes and the vector of heredity, DNA has for manyyears occupied a preeminent place in the family of nucleic acids. In 1953,Watson and Crick predicted its double-helical structure, now recognized as practically universal, one of the essential advances in modern molecular biology.

Compared with DNA, RNA seemed to be a poor relative, and was relegatedto secondary roles. Three main families of RNA molecules could be identified:

• Messenger RNA (mRNA) molecules, considered to be ephemeral copies ofDNA used in protein translation;

• Ribosomal RNA molecules, long reduced to the role of internal ribosomescaffolding, with no real function, since ribosomal functions were attrib-uted to their protein components;

• Transfer RNA (tRNA) molecules, considered to be little more than molecular adapters for ferrying amino acids to the ribosome.

The rehabilitation of RNA began in the early 1980s, when the introns of somemRNA molecules were found to be capable of excision in the absence of pro-teins. Other RNA molecules were later found to be endowed with catalytic prop-erties; for example, ribonuclease P, an enzyme responsible for the maturation oftRNA. This amounted to a conceptual revolution leading to the recognition ofthese RNA enzymes, which came to be known as ribozymes.

There is now good reason to believe that RNA is responsible for polymeraseactivity in the ribosome, and that the proteins that adorn RNA molecules arenot there just to stabilize and/or enhance that activity. This has recently beenconfirmed by resolution of the three-dimensional structure of the ribosome,revealing the active site of the ribonucleoproteic assembly to consist exclusivelyof RNA.

Bioinformatics: Genomics and post-genomics, Frédéric Dardel and François Képès© 2006 John Wiley & Sons, Ltd

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (144)

Thus, one of the essential reactions in all living cells, the translation of geneticmessage into protein, is carried out mainly by RNA molecules, not proteins.RNA not only carries genetic information and is replicated like DNA, but alsocatalyzes the reactions necessary for life. In 1986, the versatility of RNA ledFrancis Crick, co-discoverer of the double-helical structure of DNA, to proposethat life itself developed via an ancestral organism that used only RNA to carryout its functions. This hypothesis, known as the ‘RNA world,’ obviates thechicken-or-egg dilemma of whether proteins or DNA came first during evolution.

Our knowledge of the family of RNA molecules has also considerablyexpanded. We now know of RNA molecules that play important, if not essen-tial roles in the control of plasmid replication, in the regulation of geneticexpression, in eukaryote intron splicing machinery (the ‘splicosome’), in themanufacture of telomeres (special structures found at chromosome ends), andin numerous processes in RNA viruses, retroviruses, for example.

This chapter covers how RNA acquires the complex and varied structuresthat allow it to exercise these diverse functions. It may even be stated that‘unstructured linear RNA does not exist in vivo,’ since under physiological con-ditions, practically all RNA, including mRNA, is more or less elaborately folded.

Predictive methods exist today that permit a priori calculation of the sec-ondary structure of RNA, starting from its primary sequence. It is possible tovalidate and clarify these theoretical structures by phylogenetic comparison andbiochemical experimentation, which in certain cases has led to the developmentof three-dimensional models of the folding of some very complex RNA molecules.

6.2 Properties of the RNA molecule

In the cell, RNA may be distinguished from DNA by the presence of a hydroxylgroup at the 2′-position of the ribose molecule (Figure 6.1) and by the fact thatRNA is mostly present as a single strand. The thymine residues in DNA arereplaced by uracil in RNA (Figure 6.2).








5 '

3 ' 2 ' 2 '3 '

5 '






Figure 6.1 RNA has an additional hydroxyl group on the ribose 2′-carbon atom.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (145)

Like DNA, RNA can form Watson–Crick-type nucleotide pairs (A:U andG:C) and locally fold into a double-helix, either with another RNA strand orby forming a heteroduplex with a DNA strand. However, the presence of theribose 2′-OH group imposes additional steric constraints and usually preventsformation of a B-type DNA, the major DNA conformation. RNA thus formsan A-type helix, which is more open and includes 11 nucleotides per turn insteadof 10. Figure 6.3 displays the differences between the A and B helical forms. InA-form RNA, the large groove is appreciably deeper and very narrow, whereas














Thymine Uracil

Figure 6.2 Uracil (U) is the thymine analog in RNA. It has the same pairing capacity as thymine,but lacks the methyl group at position 5.

A Form

B Form


Figure 6.3 Structures of the A and B types of double helices formed by polynucleotides.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (146)

the small groove is much shallower. Finally, the basepair stacks are inclined anddisplaced with respect to the axis of the helix.

Finally, RNA forms non-canonical pairs that are different from standardWatson–Crick pairs (A:T and G:C), among which the most frequently observedare G:U, known as ‘wobble pairs.’ Compared with Watson–Crick pairs, wobblepairs require displacement of the pyrimidine nucleotide (uracil) toward themajor groove, which deforms the ribose-phosphate backbone (Figure 6.4).

6.3 Secondary RNA structures

Classical structures

Given a segment of an RNA sequence, for example, 5′-AGCGGUU-3′, it issimple to deduce the complementary sequence 5′-AACCGCU-3′ by reversing theorder of the nucleotides and replacing each with its complement. When an RNAsequence has both a sequence and its complement on the same strand, it is saidto contain an inverted repeat.













A-U pair
















G-C pair















G-U wobble pair

Figure 6.4 Watson–Crick A-U, G-C, and ‘wobble’ pairings in RNA structures.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (147)

——————> <——————5′——————AGCGGUU———————————AACCGCU———————–


When a single-strand RNA molecule contains an inverted repeat, it can foldonto itself locally, forming small double-helical segments called ‘stems’, sepa-rated by single-stranded regions known as ‘loops.’ Four types of loops may beidentified, according to the local folding topology: bulges, terminal loops or‘hairpins’, internal loops, and branched or multiple loops (Figure 6.5).

This planar representation of pairing topology and helix formation in anRNA molecule is called its secondary structure. Determining this topology is thefirst step in the three-dimensional modeling of helix layout, known as tertiarystructure.

The curved lines connecting the tandem reverse sequences in the diagram atthe top of Figure 6.5 do not cross each other. Such pairings are said to be clas-sical, since there is no entanglement of the strands of two helical regions. Thisnon-entanglement constraint is analogous to that used for writing parenthetical



Terminal loop

Terminal loop

Internal loop


Inverted repeats

Figure 6.5 Example of secondary structures formed by the pairing of complementary sequencesin an RNA molecule, showing the various types of loops encountered. ‘Classical’ structures do nothave the long-distance interactions among loops that correspond to the complementarity indicatedin dashed lines in the upper part of the figure.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (148)

expressions in mathematics. If an opening parenthesis is associated with the firststrand of a paired region in a classical secondary structure and a closing paren-thesis is associated with the second, the syntax of the expression obtained iscorrect. For example, in the structure represented in Figure 6.5, the parenthet-ical expression associated with the diagram is

( [ ( ) [ ( ) ] ] )

We arbitrarily chose to alternate parentheses and brackets in order to make theexpression more understandable, but whatever the symbol (parentheses, brack-ets, braces, etc), the expression remains correct as long as two of the samesymbols are attributed to each strand of a paired region. If non-classical pairingoccurs between two terminal loops (indicated by the dashed line in Figure 6.5)the associated parenthetical expression becomes

( [ ( [ ) [ ( ] ) ] ] )

which clearly is incorrect.The reason for making this rapprochement between parenthetical expressions

and secondary RNA structure (which at first sight may appear somewhat arti-ficial) will become apparent later. The methods used to analyze and predict theclassical secondary structure of RNA derive from tools developed by computerscientists for analyzing the syntax of mathematical expressions and the struc-tures of computer programs.

True knots and pseudoknots

The preceding paragraph described classical secondary RNA structures and thelimitation they impose on long-distance interactions. However, there is no bio-logical reason for RNA folding to be limited to classical folds. In fact, thereeven exist known and well-described secondary RNA structures called pseudo-knots that do not obey this rule. Pseudoknots can occur in a simple structureconsisting of a stem and a terminal loop when part of the loop sequence is com-plementary to a region situated outside the stem. A second helical region thenforms, often as a prolongation of the first region, in which the ribose-phosphatebackbone takes on an S-shape (Figures 6.6 and 6.7).

They are called pseudoknots because the RNA strand does not really form aknot in the topological sense, which would be physically unrealistic and bio-logically catastrophic. This imposes a major topological constraint on thesepseudoknot structures:

Thenon-canonicalpairedregionof apseudoknotcannotexceed9or10basepairs


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (149)

This constraint is the result of the helical character of the stem that is formed.The helices formed by the RNA contain 10 or 11 basepairs per turn. Startingwith this length, the full-turn formed by the rolling-up of the two strands trans-forms the pseudoknot into a real knot (cf. Figure 6.6).

Strategy for analyzing and predicting RNA secondary structures

This constraint on the length of pseudoknots is an advantage for the bioinfor-matician, since it considerably limits the number and extent of pseudoknots inreal structures, greatly simplifying analysis. In fact, the only efficient algorith-mic tools that exist are those used for seeking classical secondary structureswithout pseudoknots.

After identifying the most probable classical secondary structure heuristically,the search continues for possible pseudoknots that the RNA molecule might


Pseudoknot “True” knot

Figure 6.6 Formation of a pseudoknot by a loop pairing with a region outside its stem.

Figure 6.7 Three-dimensional structure of a pseudoknot present at the 3′-end of the genomicRNA of the tobacco mosaic virus.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (150)

contain. This approach yields satisfactory results; since pseudoknots are short,they usually do not significantly contribute in a definitive manner to overall sec-ondary structure stability.

6.4 Thermodynamic stability of RNA structures

A priori, there are numerous ways an RNA molecule of known sequence andlength can fold into stem and loop structures. One would expect these in vivostructures to correspond to the lowest energy conformations. A method isneeded that can be used to develop approaches for predicting the relative sta-bility of these various structures for quantitative evaluation.

The ‘nearest neighbors’ hypothesis

The factors that play a role in stabilizing nucleic acids in duplex form areknown. There are three principal contributions: First, hydrogen bond forma-tion between the two bases of a pair; then Van der Waals interactions betweenconsecutive layers of helically stacked basepairs; and finally, the ‘hydrophobiceffect,’ which also promotes the stacking of basepairs by protecting theirhydrophobic faces from the aqueous environment.

In a classical helix structure, these three interactions take only the nature ofthe basepair and its immediate neighbors (the basepairs before and after it) intoconsideration. The absence of long-distance effects suggests that it could be pos-sible to calculate the energy associated with the formation of a helical region(pairing ∆G0), by ‘cutting elementary slices’ of a basepair and summing theenergy contributions of each slice:

Each ∆G0 depends on the nature of the basepair (A-U or G-C), as well as onthat of its nearest neighbors, the basepairs situated immediately before and afterit in the helix. Empirically, it turns out that this simplification, while ratherdrastic, allows accurate prediction (≤10%) of the stability of RNA helices(without loops).

Empirical measurement of parameters

The thermodynamic parameters associated with basepair formation in a helicalstructure must be available in order to implement this simplified method. Theseparameters are based on sets of experimental measurements carried out on short

∆ ∆G Ghelix0

basepairs0≈ ∑


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (151)

synthetic oligoribonucleotides. For example, by placing a solution of theoctameric palindrome sequence 5′-AGAUAUCU-3′ in the cuvette of a spec-trophotometer, it is possible to follow the UV absorption of the RNA bases asa function of temperature.

At low temperature, the oligonucleotide self-pairs, forming a duplex:


Heating melts the Watson–Crick pairs, separating the two stands. In the duplexform, the bases are stacked on top of each other, which places them in ahydrophobic environment in which their UV absorption is lower than for thesingle-strand form, in which the bases are exposed to an aqueous environment1.

Analysis of the melting curves obtained with various concentrations of RNAstrands provides access to the thermodynamic equilibrium parameters:

In particular, it is possible to calculate variation in the values of the enthalpyparameter ∆G0, ∆H0, and ∆S0 associated with duplex formation2. For example,for the octamer AGAUAUCU mentioned above, ∆G0 was found to be −6.58kcal/mol at 37°C by experimentation.

Systematic analysis of a large number of oligonucleotides permits determi-nation of all the incremental ∆G0 corresponding to various stacking combina-

strand strand duplex[ ] + [ ] ↔ [ ]


1 This property of paired nucleic acids is called hypochromicity.2 The following relations are utilized:

Utilizing the fact that at the melting temperature Tm, 2[duplex] = [single strand], we obtain

Finally, the conservation of the total quantity of RNA strands is written

from which the following relation is extracted:

Studying the variation in the melting temperature Tm (in °Kelvin) as a function of the RNA concentra-tion in the spectrophotometer cuvette, it is easy to extract the values of ∆H0 and ∆S0. These values allowcalculation of the free enthalpy associated with pair formation at any desired temperature. For RNAfolding, the value at the physiological temperature of 37°C is used: ∆G0





=+ [ ]


0 log

2 duplex single strand RNA total[ ] + [ ] = [ ]

RT T Sm mlog .2 0 0single-strand H[ ]( ) = −∆ ∆

∆ ∆ ∆G RT K T S K0 0 0 2= − = − ⋅ = [ ] [ ]log H where duplex single-strand

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (152)

tions of a new basepair onto an already formed basepair. Taking into accountG-U pairs, which are rather frequent in RNA, there are three types of pairs: G-C, A-U, and G-U, each with two orientations. The set of incremental ∆G0 at37°C is presented in Figure 6.10.

Systematically repeating this operation with a great number of oligonu-cleotides makes it possible to calculate the individual contributions of each typeof basepair stacking. For example, by comparing the ∆G0 associated with theformation of a duplex of n nucleotides and the ∆G0 associated with the forma-tion of a duplex of n + 1 nucleotides containing the n first basepairs of theformer (see Figure 6.9), it is possible to deduce the incremental ∆G0 obtainedby adding the supplementary basepair.

This is known as the Freier–Turner Rules table, after the names of its authors,and merits comment:

Several other authors have identified sets of thermodynamic parameters foruse in evaluating the stability of RNA structures, finding the same qualitativetendencies but obtaining slightly different values. The Freier–Turner rules aremore recent, and were calculated for the folding of RNA at 37°C, whereas thoseof other authors were calculated at different temperatures (e.g., 25°C). TheFreier–Turner rules are also more accurate, since the determinations are basedon a more complete set of experimental data.

Note that the table is asymmetric. For example, the contributions of G-C fol-lowed by C-G, and of C-G followed by G-C (from 5′ toward 3′) are different



n 062( ecna







Figure 6.8 Fusion of a duplex RNA molecule as a function of its UV absorption. The fusion temperature Tfus corresponds to the temperature at which half of the RNA strands are in duplexform.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (153)

(−3.4 and −2.0kcal/mol, respectively). This may be explained by the differencein stacking interactions between the two types of duplex, which are muchstronger for the former than for the latter. Indeed, coverage is better when a 5′-pyrimidine is stacked onto a 3′-purine than for the reverse, as may be verifiedin Figure 6.11.

To obtain the total ∆G0 of the pairing of a duplex RNA molecule, the con-tributions of all the nucleotides it contains plus the nucleation ∆G0 (correspon-ding to the formation of the first pair) are summed. This nucleation term hasalso been determined by thermodynamic study of short duplexes:






∆G0n ∆G0


∆G0 = ∆G0n+1 - ∆G0


Figure 6.9 Incremental determination of ∆G0 obtained by stacking a G-C basepair at the 3′-endof an A-U basepair.

5' basepair GU UG AU UA CG GC GU -0.5 -0.6 -0.5 -0.7 -1.5 -1.3 UG -0.5 -0.5 -0.7 -0.5 -1.5 -1.9

AU -0.5 -0.7 -0.9 -1.1 -1.8 -2.33' basepair UA -0.7 -0.5 -0.9 -0.9 -1.7 -2.1

CG -1.9 -1.3 -2.1 -2.3 -2.9 -3.4 GC -1.5 -1.5 -1.7 -1.8 -2.0 -2.9

Figure 6.10 Helices: table of incremental ∆G037°C in kcal/mol (Freier–Turner rules).









Figure 6.11 Pyrimidine-purine (left) and purine-pyrimidine (right) nucleotide stacking interactions.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (154)

which is positive, since it corresponds mainly to entropic loss during associa-tion of the two RNA strands.

Example of calculation

For the formation of the following duplex:


contributions are summed:


Freier–Turner rules are relatively dependable for helical regions, allowing pre-diction of the ∆G0 with around 5 to 7 percent accuracy.


The same kind of thermodynamic analysis of thermal denaturation curves maybe carried out on oligonucleotides that include loops in order to try to quantifythe destabilizing contributions they make to the overall structure. The valuesobtained by Turner and Freier at 37°C for ∆G0 associated with the loops areindicated in Figure 6.12.

+ − − − − − − − − = −3 4 3 4 2 0 2 3 0 9 0 9 0 5 1 3 2 0 9 9. . . . . . . . . . kcal mol

∆ ∆ ∆ ∆ ∆

∆ ∆ ∆ ∆












+ + +

+ + + +

∆G kcal mol0 3 4= + .


Loop size 1 2 3 4 5 6 7 8 9 10 12 14 16 18 20 25 30



5.23.3 6.0 6.7 7.4 8.2 9.1 10.0 10.5 11.0 11.8 12.5 13.0 13.6 14.0 15.0 15.8Terminal loops • • 7.4 5.9 4.4 4.3 4.1 4.1 4.2 4.3 4.9 5.6 6.1 6.7 7.1 8.1 8.9Internal loops — 0.8 1.3 1.7 2.1 2.5 2.6 2.8 3.1 3.6 4.4 5.1 5.6 6.2 6.6 7.6 8.4

Figure 6.12 Loops: table of incremental ∆G037°C in kcal/mol.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (155)

Bulges are considered not to interrupt the continuity of the helix. The basepair situated immediately before a bulge is stacked onto the one immedi-ately after it. Both its stabilizing contribution and that of the bulge are takeninto account. Since there are not enough experimental data for multiple loops,they are usually treated as though they were internal loops.

Freier–Turner rules also take into account certain terminal mispairings inloops, as well as the fact that unpaired 5′ and 3′ nucleotides in helices can bestacked onto the last basepair, stabilizing the structure. (These are special casesthat will not be developed here.)

Hyperstable tetraloops

Contributions to the empirical free energy associated with loops are in generalmuch less accurate than those associated with the formation of paired regions.This low accuracy results mainly from a lack of data on loop conformation.Certain terminal loop sequences appear to be able to assume considerably morestable conformations than predicted by Freier–Turner or other empirical rules.Four-base terminal loops (tetraloops) that include the sequences GNRA, UNCG,and CUYG (Y = [C or U], R = [A or G], and N = A, U, G, or C]) have beenfound to be exceptionally stable. Structural studies using NMR and crystallog-raphy have revealed that the stability of GNRA and UNCG tetraloops is dueto a particular nucleotide conformation that allows numerous base-base andbase-phosphate interactions (Figure 6.13). Such loops are found in the riboso-mal RNA of numerous species, in the genomic and mRNA of some viruses, andin certain prokaryote transcription terminators.






Figure 6.13 Structure of a hyperstable GNRA tetraloop. The G and A form a non-canonical paircalled a ‘sheared G-A’. The two other bases are stacked onto the A.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (156)

The systematic search for motifs corresponding to these hyperstable loopsbetween two inverse repeat sequences is a very efficient method of detecting ofsuch stem and loop structures.

6.5 Finding the most stable structure

These empirical thermodynamic rules may be used to calculate the ∆G0 associ-ated with the formation of any RNA secondary structure. Finding the moststable structure; that is, the one with the lowest ∆G0, simply requires calculat-ing the ∆G0 of every possible structure. The obvious problem is that the numberof different ways to fold an RNA molecule of length N becomes very great asN increases.

As discussed above, there is a parallel between classical secondary nucleotidestructures and parenthetical expressions in mathematics, and this parallel isstrict for the topology of secondary structures that alternate between helices andloops. It is thus possible to ascertain that there are 14 ways to write an expres-sion consisting of four pairs of parentheses, and that 14 different topologiesexist for the secondary RNA structures of an RNA molecule that includes fourhelical paired regions (cf. Figure 6.14).


Figure 6.14 The 14 different possible topologies for an RNA molecule containing four helices.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (157)

If this approach is generalized to the level of individual pairings, associatinga pair of parentheses to each basepair formed within the structure, it becomespossible to estimate the number of ways to form classical secondary structuresin an RNA molecule of length N. This number will be of the same order as thenumber of ways to form expressions containing p = N/2 pairs of parentheses3.That number is known and related to the central coefficient of the binomial thatwe have already mentioned in chapter 2 (Sequence Comparisons). The numberS(p) of ways to form a syntactically correct expression containing p pairs ofparentheses is expressed as:

For example, whereas for p = 4 there are only 14 possibilities (cf. Figure 6.14),for p = 20 the number of possibilities exceeds 6.5 billion.

It is clearly impossible to test all the possible combinations and to calculatethe associated ∆G0, even for very low values of N (or p). As in the case for theoptimal alignment of two sequences, a more astute strategy is required. Someelements or substructures are shared by a great number of different possiblestructures. For example, the five structures illustrated in the top line of Figure6.14 share the bottom helix, which forms the ‘trunk’ of these various tree-liketopologies. The problem may be considerably simplified by calculating theenergy of a given subsequence of length N associated only once with the for-mation of an RNA sequence structure. The recurrent calculation of the ∆G0


associated with the optimal local foldings for longer and longer subsequencesfrom i to j is used to calculate the optimal ∆G0 for the whole sequence, ∆G0

1,N.This constitutes a dynamic programming method similar to that used for theoptimal alignment of two sequences.

The Nussinov algorithm

Nussinov proposed the first application of dynamic programming for the pre-diction of secondary RNA structure in 1978. Based on a drastic simplification ofthe rules governing energetics, it takes into account only the contributions ofpaired regions (helices), attributing a score to each basepair and disregarding thedestabilizing contributions of bulges, loops, and unpaired extremities. However,

S pp


ppp( ) =



222 which increases as for large ,


3 In fact, the number of ‘realistic’ secondary structures is a bit smaller than the number of parentheti-cal expressions, since certain structures thus formed contain very short loops or non-canonical pairs.However, this approximation provides an idea of the asymptotic behavior of the number of structures.A probabilist asymptotic estimation of the number of ‘biologically reasonable’ structures S(N) that anRNA sequence of length N can form in which there is equal distribution of the four nucleotides A, G,C, and U was given by Zuker and Sankoff: S(N) ≈ 1.8N.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (158)

its functioning principle is developed in the next sections, since it remains thebasis of most of the more sophisticated algorithms, which merely improve on it.As for optimal sequence alignment, the Nussinov method consists of two steps:first completing a table, then identifying the optimal path in it.

For a given RNA sequence of length N, we will call E(i,j) the energy of themost stable structure in a sequence extending from nucleotide i to nucleotide j(1 ≤ i ≤ j ≤ N). With the Nussinov algorithm, the set of these ‘partial’ energiesis progressively tabulated in an N × N table, using the following relation:

where e(i, j) is the individual pairing score of base i with base j.While this expression may appear complex at first glance, it corresponds to

the example schematized in Figure 6.15.In the best structure for the i,j subsequence, either base j is paired to base i

(the left-hand example in Figure 6.15) and the stem that had nucleotides i + 1and j − 1 as its extremities is prolonged, or the structure is divided into twoparts, from i to k − 1 and from k to j, and the best combination of two suchpieces sought.

This recurrence relation over E(i,j) makes it very easy to complete the energytable progressively, proceeding by successive diagonals that correspond to a con-stant sub-segment length L = j + 1 − i. Since it is impossible for a structure toconsist of a single nucleotide, diagonal values in table E(i,i) are initialized atzero. In order to prohibit loops that are too small, E(i,i + 1) and E(i,i + 2) mayalso be initialized at zero. The diagonal terms corresponding to segments of

E i jE i j e i j

E i k E k ji k j

, min, ,

min , ,( ) =+ −( ) + ( )

−( ) + ( )( ) < ≤

1 1





ji+1 j-1 k-1 k

Figure 6.15 Seeking the minimal energy conformation for the [i,j] subsequence. The i,j pair iseither formed (left) or not formed (right). Substructures are represented in simple stem-and-loopform, but may be more complex, or even unpaired (single strand).

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (159)

increasing length are then progressively calculated. The energy correspondingto the complete sequence E(1,N) is obtained upon arriving at a corner of thetable.

In order to reconstitute the structure associated with this global minimumenergy, it is necessary to go backwards through the table, seeking the path thatled to the minimum energy value. The term in the above alternative that wasused to calculate the calculation of E(i,j) must be identified each time. The limiting step in the Nussinov algorithm is the completion of the table, which is O(N3) for computation time and O(N2) for memory space.

The Zuker algorithm

In the early 1980s, Michael Zuker perfected Nussinov’s method, adoptingenergy rules that were more realistic, such as those of Freier and Turner, includ-ing stacking interactions between nearest neighbors in helices and the contri-butions of unpaired regions. An exhaustive description of the Zuker algorithmis complex, since it takes into account a great number of special cases. Only itsmain principles are outlined in the following sections.

The principal modification with respect to the Nussinov method is the intro-duction of a second table. In addition to the E(i,j) table that still gives the energyof the best structure of the [i,j] segment, Zuker introduced the V(i,j) table, which


1 2 3 k

E(1, N)

Figure 6.16 Completing the energy table.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (160)

contains the energy of the best structure in which the i − j pairing is formed.Since the i − j pair is forced, it might be a sub-optimal structure. Thus V(i,j) ≥E(i,j).

The V(i,j) and E(i,j) tables are then progressively calculated, using the fol-lowing recurrence relations:

Calculation of the E(i,j) table is the same as for the Nussinov algorithm,except that two terms are explicitly added in order to take into account casesin which either base i or base j is unpaired. The table and the path search arecompleted in the same way. Calculating V(i,j) is more complex: The ee(i,j) func-tion represents the gain in energy obtained by stacking the i,j basepair onto thei + 1, j − 1 basepair. Among other reasons, it is for the purpose of calculatingthis term that E and V must be tabulated separately, since it must be possibleto use V(i + 1, j − 1) to calculate V(i,j).

The various Loop (i,j) functions represent the energy of the best structure inwhich the i,j basepair close an internal loop, a terminal loop, or a multiple loop.Calculation of the terminal loop function is rather simple using empirical ther-modynamic rules, but the two other loop types are much more complex, oftenrequiring the use of ad hoc hypotheses to limit calculation time (loop size upperlimit, simplification of free energy rules). The more complex of the two, mul-tiple loop, is at best O(N3) in calculation time for each V(i,j) term, which makesthe Zuker method an algorithm of complexity O(N5). With currently usedmachines, the Zuker algorithm allows handling of sequences of between severalhundred and several thousand nucleotides, which is generally adequate for mostcommon biological problems.

Limits of in silico predictive methods

The computer methods for predicting secondary RNA structure like the onespresented above have two types of limitation:

V i j

Terminal loop i j

V i j ee i j

Internal loop i j

Multiple loop i j

, min


, ,



( ) =

( )+ −( ) + ( )

( )( )

1 1

E i j

E i j unpaired i

E i j unpaired j

V i j i j pairing

E i k E k j broken down into substructuresi k j

, min




min , ,

( ) =

+( )−( )

( ) −( ) + +( )( )

< <



1 2


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (161)

(i) The inaccuracy of empirical rules combined with certain approximationsnecessary for the efficient calculation of energy associated with loops pro-duces a relatively high degree of error in the energy values obtained;

(ii) The algorithms presented yield only a single solution, which correspondsto the minimal energy.

The combination of these two limitations may result in failure to detect thecorrect structure, especially if there are several alternative solutions whose freeenergies are very similar. One necessary improvement in predicting secondaryRNA structure is therefore to generate suboptimal solutions whose energy scoresfall a few percent below the highest score. The biologist must then sort the struc-tures within the family of candidates, based on additional criteria.

6.6 Validation of predicted secondary structures

One of the main problems encountered in trying to predict RNA folding, forexample, ribosomal RNA folding, is that there are usually many alternative conformations with rather similar calculated energies. However, biologists havetwo complementary approaches for solving this problem: i) phylogenetic analy-sis; ii) enzyme or chemical probes for use in characterizing effectively pairedregions.

Phylogenetic analysis

When the RNA whose folding is being determined belongs to a family for whichseveral sequences from related organisms are available, comparisons may beused to confirm the existence of a paired region. The various secondary struc-ture components are usually conserved, and the appearance of a mutation onone strand in a helical region must be accompanied by the appearance of a compensatory mutation on the other strand, in order to restore a basepair.Observation of this type of covariation between two positions in several RNAsequences of the same family is an extremely strong argument for the existenceof an interaction between the two corresponding bases in the two- or three-dimensional structure.

Analyses of this type are systematically carried out on ribosomal RNA andon several families of autocatalytic introns, for which dozens, even hundreds ofhom*ologous sequences have been compiled in databases. This work has allowedreconstitution of the secondary structures of RNAs with sequences exceedingseveral kilobases with a high degree of certainty.

If a reliable multiple alignment in a family of RNA hom*ologs is available, itis possible to describe a quantitative way of measuring the correlation between


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (162)

variations in the bases at positions i and j. This relation is called mutual infor-mation M(i,j), and is written:

where bi and bj are the various bases possible at positions i and j, fbi, fbj, theirindividual frequencies of appearance at positions i and j, and fbibj the frequencyof simultaneous appearance of bi, and bj at positions i and j.

This quantity, assigned a value between 0 and 1, increases as the correlationbetween positions i and j increases. If there is no correlation, or if the positionsdo not vary, it is 0. Since mutual information analysis makes no a priorihypothesis regarding the nature of the pairs (Watson–Crick, wobble, or other),it allows detection of standard interactions among helical segments, as well asmore exotic, non-canonical interactions. Systematic study of M(i,j) can alsodetect more complex interaction zones of the pseudoknot type, as well as amongsecondary structural components, permitting construction of a primitive three-dimensional folding model.

6.7 Using chemical and enzymatic probing to analyze folding

If purified RNA is available, probes may be used that react differently withsingle- and double-stranded regions to provide information concerning its structure in solution.

RNA can be subjected to limited enzyme attack. Ribonucleases (RNAses)exist that preferentially cleave either unstructured regions (e.g., RNAseT1,nuclease S1) or helical regions (e.g., RNAse V1). Comparison of the experi-mental distribution of phosphodiester bonds cleaved by these two classes of

M i j ff

f fb b

b b

b b

b bi j

i j

i j

i j

, log,

( ) = ∑












Figure 6.17 Carbodiimides (highlighed) react with the G and U imino groups. This reaction ispossible only if the base is not involved in pairing. Cartography of the reactive positions identi-fies single-strand regions of the RNA sequence.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (163)

enzymes with paired regions in various computer-predicted models allows elimination of those that are inconsistent with experimental data.

The use of enzyme probes is not always sufficient, since ribonucleases aremacromolecules and therefore occupy a non-negligible volume. For stericreasons, they sometimes do not reach all the theoretically cleavable sites of theRNA sequence being studied, which results in ambiguities and difficulties ininterpretation for certain regions. In order to complete the analysis, smaller chemical probes, such as alkylating reagents and metal ions, are often used.Combining these probing strategies usually allows precise identification ofpaired RNA regions in the structure.


chemical probe

Modifications are revealed by reverse transcriptionusing a labeled primer

Polymerization is blocked by the modified base

Modified positions are identified byelectrophoretic separation of the labeled DNA fragments

Figure 6.18 Principle of secondary RNA structure analysis using chemical probes, such as carbodiimides.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (164)

The ‘signature’ thus obtained directly from the molecule may then be usedto screen the various secondary structures predicted by computer.

6.8 Long-distance interactions and three-dimensionalstructure prediction

The three-dimensional structure of highly structured RNA, such as ribosomalRNA, is well defined. As in the folding of peptide chains of proteins, it is sta-bilized by long-distance interactions between secondary structural elements.Three types of long-distance RNA structural interactions are known today:pseudoknots, tetraloop–receptor interactions, and triple helices. Systematic phy-logenetic analysis permits location of some of these interactions. If enough ofthese long-distance contacts are known, in some cases it is possible to constructa more or less detailed model of the three-dimensional structure of the RNAsequence. The formation of pseudoknots was mentioned previously, and othertypes of interactions will be described in this section.

Triple helices

Watson–Crick-type pairing is not the only form of basepairing possible betweena purine (A and G) and a pyrimidine (C and U). Purines have a second face,known as the Hoogsteen face, which can form hydrogen bonds with a pyrimi-dine. Both Hoogsteen and reverse Hoogsteen pairing are possible, according tothe orientation of the pyrimidine (see Figure 6.19). Hoogsteen pairings may be
















Watson–Crick Hoogsteen






























Figure 6.19 Hoogsteen-type pairings. In direct Hoogsteen pairing, the two RNA strands are par-allel, whereas in reverse Hoogsteen pairing, the two strands are antiparallel, as in Watson–Crickpairing.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (165)

formed between A and U (Figure 6.19) and between G and C, provided thatthe N3 carbon atom is protonated.

Examining the geometry of these various basepairings reveals nothing thatcan simultaneously form Watson–Crick and Hoogsteen pairings. A triplet ofbases therefore forms. The stacking of several such triplets leads to the forma-tion of a triple helix that consists of a standard double helix plus a third strandinserted into the major groove of the duplex.

However, a major criterion must be fulfilled for the stacking of several basetriplets to occur: the purines must succeed each other on the same strand (Figure6.20). Therefore, a triple helix can form only in regions that contain consecu-tive purine sequences, known as ‘hom*opurine sequences.’ These triple helicalstructures can interact at long distances in an RNA sequence. An unpairedregion, such as a loop of appropriate sequence can in effect insert into the majorgroove of a stem containing a hom*opurine strand and a hom*opyridine strand(Figure 6.21).

Tetraloop–receptor interactions

Tetraloops of the GNRA family (see Hyperstable tetraloops, above) are able tointeract over long distances with more or less regular helical structures. Thesestructures, known as tetraloop receptors, are specific for the associatedtetraloop. As an example, Figure 6.22 shows the structure of tetraloop GAAA,which was found in the structure of the core of a self-splicing intron.


















































Figure 6.20 When one of the two classical Watson–Crick strands contains a hom*opurine tract,Hoogsteen or reverse Hoogsteen pairing can form a triple helix with a third hom*opyrimidine strand,as shown here.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (166)

Prediction of three-dimensional structure

When a sufficiently full set of related sequences is available, systematic analy-sis of the covariation between tetraloops and receptors or between sequencesand loops able to pair and form pseudoknots, and the search for hom*opurineand hom*opyridine sequences able to form triplexes, can generate enough con-


Triple helices

Figure 6.21 Triple helix formation.







Tetraloop–receptor interaction

Figure 6.22 Tetraloop–receptor interaction. The sequence of the receptor indicated is specificfor the GAAA tetraloop. Other GNRA loops have different receptors associated with them.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (167)

straints to unequivocally model the three-dimensional structure of some RNAmolecules. These models are often constructed manually or semiautomatically,with fixed helical regions in the form of rigid cylinders. These models can betested and validated by incorporating chemical and enzymatic reactivity data.In some cases, covalent bridging between nucleotides may be obtained, eitherby ultraviolet radiation or the action of a chemical reagent. The incorporationof supplementary data can refine and confirm the model. Several RNA com-plexes have been modeled in this way, permitting remarkable functional predictions.

6.9 Protein structure

The stakes and the difficulties

The rapid high-throughput genomic sequencing programs operating today are filling protein sequence databases, whose sizes double every 15 to 18months. At the same time, although progress has been made, resolution of three-dimensional protein structure remains a relatively difficult task. At present, such structures are available for only a small fraction of proteins. The hiatusbetween protein sequences and three-dimensional structures will persist in the


Figure 6.23 Model of the structure of a self-splicing intron. The essential characteristics of this model were subsequently validated by determination of the crystallographic structure of thecatalytic heart of the intron.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (168)

foreseeable future, resulting in an obvious interest in predictive methods thatallow avoiding the experimental step, at least for some applications.

Regrettably, the problem of predicting protein folding remains very difficultand much more complex to formalize than that of RNA folding. Four types ofdifficulty exist:

• There are no simple interaction rules, such as Watson–Crick pairing forRNA.

• Energy-prediction methods are inadequate. The free energy associated with folding (∆G0

folding) is relatively small, on the order of a few dozenkcal/mol, or less. The imprecision of the methods used for energy calcula-tion makes any attempt at prediction impractical.

• The more complex alphabet used for proteins (20 amino acids versus 4nucleotides) makes using phylogenetic methods based on comparativesequence analysis more difficult. For the same reason, the wide chemicaldiversity of amino acids complicates the table by increasing the numberand nature of interactions encountered in three-dimensional protein struc-tures (e.g. hydrophilic, ionic, polar).

• Proteins adopt a very wide variety of three-dimensional structures. A ‘topo-logical’ classification based on the nature of the secondary structural elements (a-helices and b-sheets) they contain has been established (seeFigure 6.24). Some contain only a-helices or b-sheets, while most containa combination of the two.

• Finally, proteins longer than 300 amino acids are almost always organizedinto two more or less separate domains. Starting from the sequence, thismorphological richness renders ab initio prediction of three-dimensionalprotein structure more difficult.

However, these obstacles are not totally insurmountable, and it has been possi-ble to make some progress, most often of limited scope, but sometimes leadingto spectacular success.

Given a protein for which only the primary amino acid sequence is known,it is at present reasonable to envisage three approaches to predicting three-dimensional protein structure:

• In the absence of all other information, it may be possible to predict thepositions of the secondary structural elements in the protein (a-helices, b-sheets);


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (169)

• If the protein contains sequence hom*ologies with a protein of known struc-ture, it is possible to construct a more or less accurate three-dimensionalfolding model based on alignment of the two sequences;

• It is possible to determine whether the three-dimensional folding of theprotein corresponds to folding that has already been determined and ref-erenced in databases. To do so, ad hoc methods based on the distributionof secondary structural elements predicted for the sequence may be used,as well as more systematic methods that consist in attempting to ‘thread’







Figure 6.24 Various types of three-dimensional protein structures: A) parallel a-helices (ferritin);B) a-helices of mixed orientations (myoglobin); C) a-helices and antiparallel b-strands (amidino-transferase); D) b-sheets (collagenase); E) b-sheet sandwich (CD4 antigen); F) barrel formed by b-strands alternating with a-helices (triose phosphate isomerase); G) multi-domain protein(aminoacyl-tRNA synthetase).

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (170)

the protein sequence onto a three-dimensional fold and evaluating thedegree of correspondence, using more or less elaborate cost functions.

These three strategies are briefly developed below.

6.10 Secondary structure prediction

There exist a great number of methods for predicting elements of secondaryprotein structure, which it would be impossible to present here. However, theyall more or less directly depend on a combination of three physicochemical andgeometric properties of proteins:

• In soluble proteins, hydrophilic amino acids are usually exposed on thesurface, where they are in contact with the aqueous environment, whilehydrophobic amino acids face the interior of the protein.

• b-sheets are structures of periodicity 2 (along a strand, the lateral chainsare alternatively above and below the plane of the sheet) and the a-helicesare structures of periodicity 3.6 (the lateral chains point toward the outsideof the cylinder of the helix, whose pitch is 3.6 residues per turn.)

• A protein domain contains an average of 100 to 300 amino acids. Theaverage length of secondary structural elements therefore corresponds tothe diameter of such a domain, approximately 30 to 40Å. This length cor-responds to between 15 and 20 amino acids for an a-helix and to betweenfive and 12 amino acids for a b-sheet whose peptide chain is completelyextended. Between these elements of regular structure, the protein back-bone forms loops, most often exposed on the surface.

There are two main classes of predictive methods, both based on amino aciddistribution in the sequence: pattern recognition methods, which try to use theabove observations directly and statistical methods, which use them in a moreindirect way, calculating a probabilistic score.

Most methods used at present utilize a three-state secondary structure model:(1) a-helix; (2) b-sheet; (3) loop or irregular structure. The model aims to predictin which of these three states each amino acid in the sequence is located.

Pattern recognition

The principle of pattern recognition methods is to try to recognize the specificdistribution of residues in each type of secondary structure. For example, if a


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (171)

strand of a b-sheet is located on the surface of a protein, one would expect halfthe residues to be hydrophilic and half to be hydrophobic. The same would bethe case for an a-helix located on the surface, which would be expected to havea hydrophobic side facing the inside of the protein and a hydrophilic side facingthe solvent.

By examining hydrophilic and hydrophobic residue distribution in thesequence of a protein of unknown structure, one can discern the specific alter-nation of a-helices and b-sheets. Surface loops are usually continuous regionscomposed of highly hydrophilic residues.

One of the first methods used to detect helices was a graphic representationknown as the ‘helical wheel’, in which the sequence of the backbone is woundonto an a-helix viewed perpendicularly along its axis (Figure 6.25). It is poss-ible to directly visualize the presence or absence of hydrophilic and hydropho-bic faces on the cylinder. Other graphic methods that use the surface of thecylinder have also been proposed.

Two approaches may be used to identify these distributions, one of which ispurely formal, while the other is more quantitative. The first approach, pro-posed by the Russian biologist Valery Lim, consists in searching for certain patterns of hydrophobic residues that are characteristic of various types of sec-ondary structures. It defines a set of 22 complex rules that allow prediction ofthe whole sequence. Some examples of the Lim rules are indicated below (Φdesignates a hydrophobic amino acid, and – a hydrophilic one.

a −− −

− −

helixΦΦ ΦΦ ΦΦ

b −− − −

sheetΦΦΦΦΦ Φ Φ Φ


I23 G22









Figure 6.25 Representation of residues 12 through 23 of flavodoxin in the form of a helicalwheel. Highly hydrophilic residues are in squares and hydrophobic ones in circles. The asymmetricdistribution characteristic of helical regions is clearly demonstrated.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (172)

The quantitative approach consists in attributing a score to each amino acid.Several empirical hydrophobicity scales exist, based on either physicochemicalparameters (e.g. solubility, partition coefficient) or protein-surface distributionstatistics. For a protein sequence of length N, it is possible to use this score tocalculate a vector quantity known as the hydrophobicity moment:

where Hi is the hydrophobicity score of the ith amino acid of the segment andd is a periodicity parameter taken to be equal to 180° in order to detect b-sheetsand to 100° to detect a-helices. The vector (cos di, sin di) therefore correspondsto the orientation of the lateral chain of the ith residue in the corresponding periodic structure. If the sequence of N residues studied does fold according to a structure of periodicity d, the hydrophobicities Hi will sum in a construc-tive manner and the moment module ⟨ ⟩ will be large; if not, the hydropho-bicities will on the average cancel out and the ⟨ ⟩ module will be very small.

It is also possible to sweep a window of N residues through a sequence beingstudied (N typically being on the scale of the average length of secondary struc-tural elements). The positions in which ⟨ helix⟩ or ⟨ sheet⟩ are higher than a thresh-old value respectively indicate the probable presence of a helix and a sheet atthat place.

Statistical methods

The statistical methods are based on systematic study of amino acid distribu-tion in three-dimensional protein structures available in databases. The prin-ciple underlying these methods is determination of whether there is a bias in thecomposition of helices, sheets, and loops that could be used for prediction. Itwas noticed very early that the amino acids glutamate, methionine, and alaninewere over-represented in a-helices. This simple frequency analysis led to an earlysimple predictive method, known as the method of Chou and Fasman.

The principle of this method was later modified and extended by Garnier andRobson to take into account effects at a distance. The principle is to observethe influence of the residue i + k(−8 ≤ k ≤ +8) on the conformation(helix/sheet/loop) of residue i by carrying out detailed frequency analysis. Thisinfluence is quantified in terms of information, I, in the Shannon sense. Giventwo events X and Y, this is:

Starting with frequency analysis in the structure database, these authors tabu-lated the following information regarding the conformation of residue i, Si:

I X; Y P X Y P X( ) = ( ) ( )[ ]log














N cos



Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (173)

Starting with this information, Garnier and Robson were able to reconstitutethe information associated with subsequence ai-8 . . . ai . . . ai+8 for the three typesof Si structures.

These information profiles can attribute a probable conformation to eachresidue i.

Efficiency and limits

At present, and despite numerous improvements, when applied to proteins ofunknown structure, the accuracy of these predictive methods does not exceed70% correctly predicted residues. Two kinds of difficulty are encountered; onone hand, there are boundary effects, since it is always difficult to preciselydetermine the borders of secondary structure elements. On the other hand, thesemethods often miss small b-sheets, since they are too short to give a detectablesignal.

Among the efficient improvements made to this method, one approach con-sists in using not just a single protein sequence, but a set of hom*ologoussequences previously subjected to multiple alignment analysis. At position i inthe sequence, there is not just a single amino acid, but a set of possible aminoacids corresponding to all the residues found in the ith column of the multiplealignment. This information is precious, since it provides data concerning thevarious amino acids that are compatible with the structure of the protein.

Several methods using multiple alignment of hom*ologous sequence profileshave been implemented. To date, one of the most efficient remains PredictProtein,a neural network system that is ‘trained’, using the profiles of known structuralproteins ( Thissystem is related to statistical methods, since it learns by effectively taking intoaccount the frequencies of the various types of amino acids. However, it is a ‘blackbox’; therefore it is difficult to figure out how it functions.

6.11 Three-dimensional modeling based on hom*ologousprotein structure

The Protein Data Bank protein structure database library ( a great number of hom*ologous protein structures. Its use makes it pos-sible to see how the observed similarities in protein amino acid sequences are

I S a a I S ai i i i i kk

; . . . ;− + +=−

( ) = ( )∑8 88


I -helix; a I -sheet; a and I loop aS S Si i k i i k i i k=( ) =( ) =( )+ + +a b, , ;


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (174)

reflected in their three-dimensional structure. This systematic study was carriedout by Chothia and Lesk around fifteen years ago. After eliminating proteinsurface loops, the conformation of which is often variable, from the study andretaining only the protein core, they found a very strong correlation betweensequence conservation and structural similarity.

The similarity of two structures may be quantitatively evaluated by calculat-ing the standard deviation of the positions of atoms in the peptide backbone.When the sequence identity level (percentage of identical residues) is 50%, thestandard deviation of the atom positions is of the order of 1Å or less. Even with25% identity, the standard deviation generally remains under 2Å. This is illus-trated in Figure 6.27, which shows the structural similarity between human col-lagenase, an enzyme that digests collagen, and serralysin, a bacterial protease.However, alignment of the two sequences indicated below reveals the similarity to be tenuous, with only 27% identity and major insertions/deletions.

Reciprocally, this correlation between sequence hom*ology and structural similarity can be used to construct a three-dimensional model of a protein of




Figure 6.26 Alignment of the sequence of human collagenase with the catalytic domain of serralysin, a protease of the bacterium Serratia marcescens.

Serralysin (catalytic domain) Collagenase

Figure 6.27 Three-dimensional structures of serralysin and collagenase.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (175)

unknown structure, starting from the known structure of a hom*ologous protein.The protocol consists of four successive steps:

Core construction: Conserved regions corresponding to secondary structural ele-ments in the core of the known structure are identified. Corresponding proteinbackbone regions are then placed onto the unknown structure.

Loop construction: Protein surface loops are extracted from the Protein Data-base Library. The loop of the unknown protein sequence that best correspondsin terms of length, sequence, and end-orientation is sought. The backbone ofthe selected loop is then grafted onto the backbone of the core constructed inthe preceding step.

Sidechain budding: Following the first two steps, the polypeptide backbone is constructed, after which the two lateral chains are added. The conformationof the lateral chain of the reference protein is generally used directly at posi-tions where there is sequence identity between the reference protein and theprotein under construction. For positions at which the sequences diverge, asmany as possible of the atoms the two lateral chains have in common are used.

Model refinement: Following the sidechain construction step, the model usuallycontains a number of defects, such as steric violations due to atoms being tooclose to each other, or unfavorable electrostatic interactions. It is possible toquantify these defects by using a classical formal energy calculation to sum thevarious potential energy terms associated with the molecular conformation. Thisincludes the terms associated with deformation of the covalent geometry andthe terms of non-covalent interaction: the electrostatic potential and Van derWaals interaction.

Covalent geometry

The first term represents the energy associated with modifications in the lengthsof the covalent bonds; the second corresponds to the modifications in the valence angles; and the third to deformations in the dihedral angles (torsion ofplanar bonds and aromatic cycles). Parameter n describes the periodicity of thedihedral angle (for example, n = 2 for a planar liaison, corresponding to cis- and trans-isomers.) The ai, fj, and ck parameters are related to spring stiff-ness and have been empirically determined from the vibration frequencies ofsmall molecules.

E a l l b c nibonds

i i jangles

j j kdihedral angle

k kcov cos= −( ) + −( ) + + −( )( )∑ ∑ ∑0 2 0 2 01J J j j


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (176)

Non-covalent interactions

The first term corresponds to classical electrostatic interaction, in which param-eter D describes the dielectric constant of the medium. The second term corre-sponds to the Van der Waals interaction, with a strongly repulsive 1/r12 part,which prevents interpenetration of the atoms and an attractive 1/r6 part. TheAij and Bij parameters are adjusted to yield a minimum that corresponds to thesum of the Van der Waals radii of atoms i and j (Figure 6.28). The two sumsare carried out on all the pairs of atoms of the molecule.

Minimizing the energy of the reconstructed molecule

Starting from a conformation of the molecule defined by the Cartesian coordi-nates xi, yi, zi of its N atoms (1 ≤ i ≤ N), it is possible to calculate the associ-ated potential energy E(xi, yi, zi), using the above terms. This function of 3Nvariables can then be minimized, using classical numerical methods (e.g., steep-est descent, conjugated gradient, allowing descent into a potential energy well).However, this approach is usually far from satisfactory, since the potential func-tion is very rough and contains a great number of local minima. Direct mini-mization usually leads to one of these local minima, which often are far froman acceptable solution (Figure 6.29). To avoid these potential traps, a dynamicmethod called ‘simulated annealing’ is used.

Movements of atoms in the molecule can be simulated by the action of theabove potential. To do this, initial random velocities are assigned to each atom

Eq qDr



noni j

iji j




iji j−

≠ ≠

= + −

∑ ∑cov 12 6


Sum of vdW radii



r = interatomic distance


r12A —

Figure 6.28 Potential of the Van der Waals interaction.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (177)

of the molecule. The distribution of these velocities is selected such that itfollows a Boltzmann distribution at a given temperature (e.g., 300°K) The equa-tion for the dynamics of the movement is then numerically introduced for eachatom i:

This gives the movement trajectories of the atoms in the molecule. Follow-ing the few dozen picoseconds it takes for the system to equilibrate, the veloc-ities of the atoms4 fall progressively during digital simulation of cooling. Themolecule then drops into a much lower potential well, approximating a rea-sonable conformation (Figure 6.30). This method remains heuristic, and it isusually necessary to conduct several simulations in order to test convergence. Italso still consumes much calculation time. An ‘average’ protein contains severalthousand atoms, which makes calculation of the potential and trajectories rathercumbersome. In addition, it is necessary to resort to slow cooling, requiringnumerous (104 to 106) integration steps, which takes several, or even dozens, ofhours of CPU time. Happily, this kind of calculation lends itself to parallel pro-cessing, and most existing molecular dynamics programs are able to work onmulti-processor platforms.

The final conformation obtained using this approach is usually a goodapproximation of the real structure of the molecule, especially in the conservedcore.

r r

rF md rdt


i ii


= = −2





nElai tnet


Starting state

Final state


Figure 6.29 Because of the presence of numerous local minima in the potential function, thedescent minimum does not give satisfactory results.

4 For a given particle, mv2 = 3kT, where k is the Boltzmann constant.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (178)

6.12 Predicting folding

As seen in the preceding paragraph, two proteins that present amino acidsequence hom*ologies usually have closely related three-dimensional structures.However, the reciprocal of this observation is not true, and it has been shownthat two proteins can have very closely related topologies without necessarilyhaving detectable sequence hom*ology.

It is therefore possible for a protein of unknown structure to take on a knownthree-dimensional topology in the absence of detectable sequence hom*ologywith other proteins that adopt the same folding topology. There is even a widelyadmitted postulate according to which there are a finite number of foldingtopologies for individual protein domains, which would explain these structuralcoincidences. From geometric or structural arguments, some authors even con-jecture this number to be of the order of 1,000 to 2,000.

The protein structure database already includes several hundred differenttopologies. Therefore, when a new protein or family of proteins of unknownstructure is analyzed there is a reasonable probability that one or severaldomains within it correspond to an already known topology. On the basis ofthis property David Jones and Janet Thornton several years ago proposed a new method for predicting global folding called threading. Their idea was toselect a set of representative proteins from among all the known topologiesincluded in protein structure databases and place the sequence of the unknownprotein over the backbone of one of the selected proteins. A cost function isthen applied to evaluate the quality of the interactions in the protein core andof the solvation of the surface polar groups. To optimize this structural thread-ing, it is usually necessary to introduce insertions and deletions, especially inloops. Structural alignment by threading is therefore similar to sequence align-ment, mentioned earlier, which dynamic programming methods may also be








Figure 6.30 Annealing principle: Adding kinetic energy to the system first exceeds the poten-tial barriers, then descends ‘below’ them in the potential function during cooling.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (179)

used to achieve. During recent ‘competitions’, in which structural biologistschallenge bioinformaticians to predict the folding of newly determined struc-tures prior to making them public, this method has been able to correctly predictthe global topology of 60 to 70 percent of the proteins submitted. This amountsto remarkable progress in a domain that continues to undergo major expansionand advancement.


Chou, P.Y., Fasman, G.D. (1974). Prediction of protein conformation. Biochemistry 13:222–245.

Freier, S.M., et al. (1996). Improved free-energy parameters of RNA duplex stability.Proc Natl Acad Sci USA 83: 9373–9377.

Garnier, J., et al. (1996). GOR method for predicting protein secondary structure fromamino acid sequence. Methods Enzymol 266: 540–553.

Jones, D., Thornton, J. (1993). Protein fold recognition. J Comput Aided Mol Des 7:439–456.

Lim, V.I. (1974). Structural principles of the globular organization of protein chains. A stereochemical theory of globular protein secondary structure. J Mol Biol 88:857–872.

Michel, F., Westhof, E. (1990). Modelling of the three-dimensional architecture of groupI catalytic introns based on comparative sequence analysis. J Mol Biol 216: 585–610.

Moult, J. (2005). A decade of CASP: progress, bottlenecks and prognosis in proteinstructure prediction. Curr Opin Struct Biol 15: 285–289.

Nussinov, R., Jacobsen, A.B. (1980). Fast algorithm for predicting the secondary struc-ture of single-stranded RNA. Proc Natl Acad Sci USA 77: 6903–6913.

Pley, H.W., et al. (1994). Model for an RNA tertiary interaction from the structure ofan intermolecular complex between a GAAA tetraloop and an RNA helix. Nature372: 111–113.

Rost, B., Sander, C. (1993). Prediction of protein secondary structure at better than 70%accuracy. J Mol Biol 232: 584–599.

Sternberg, M.E. (1997). Protein Structure Prediction: A practical approach. IRL Press.Zuker, M. (1989). On finding all suboptimal foldings of an RNA molecule. Science 244:

48–52.Zuker, M. (1989). Computer prediction of RNA structure. Methods Enzymol 180:



Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (180)

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (181)

7Transcriptome and proteome:macromolecular networks

7.1 Introduction

The traditional approach in molecular biology research is local; it consists inexamining and collecting data on one gene, one protein, or one reaction at atime. We may recognize the classical reductionist approach, understand the partsin order to understand the whole, which has permitted remarkable advancesover the years and resulted in the development of extremely precise biochemi-cal models.

However, the advent of genomics has led to the emergence of an entirely newclass of abundant data, which until now have principally been exploited instudying previously unknown genes, genes that are over- or under-expressed incertain circ*mstances, etc. While these data do constitute an important resourcefor researchers working on individual genes, is it reasonable to try to charac-terize all relevant molecular interactions one at a time when devising a pre-dictive model for a human disease, given that the objective is to identifypharmacotherapeutic targets?

The changes introduced by (post-)genomics may also be described in termsof information theory: Knowledge of a biological system may be defined as theratio of extracted information to the relevant information potentially present inthe system. Until 1990, this ratio (which was very low) determined the overallstrategy of biological research, which may be summarized as follows: Sinceaccess is available to only a limited quantity of data, experimental strategiesmust be found that focus on those parameters (genes, for example) whose effectspredominate. The extraction of information from biological systems is lessskewed today. Indeed, by furnishing data on thousands of genes, genomics hasincreased the ratio of extracted to potentially relevant information present in agiven biological system by several orders of magnitude.

New methods are clearly required in order to comprehend these data on aglobal basis and to analyze these large-scale systems at an intermediate level

Bioinformatics: Genomics and post-genomics, Frédéric Dardel and François Képès© 2006 John Wiley & Sons, Ltd

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (182)

without descending to the level of precise biochemical reactions. At the veryleast, such analysis would be useful in guiding traditional pharmacological andbiochemical approaches toward the genes that deserve the most attention,among the thousands recently discovered. Ideally, a sufficiently predictive andexplicative model at the intermediate level would obviate the need for an exactunderstanding of the system at the biochemical level.

In this chapter, after examining the methods and data made available by post-genomics, we will undertake global analysis of the data.

7.2 Post-genomic methods

The use of (post-)genomic tools allows the acquisition of massively parallelmolecular data, such as those that concern genomic DNA sequences (genomes),as well as the concentration, activity, localization, and interaction of messengerRNA (transcriptomes1) and proteins (proteomes). This section describes theprincipal techniques used to acquire transcriptome and proteome data.

7.2.1 Proteomics

Separation, identification, and quantification of protein

The method of choice for fractionating a large number of proteins contained ina natural extract is two-dimensional gel electrophoresis (Figure 7.1). This tech-nique separates proteins in a plane, first in one direction, as a function of theirisoelectric point2, then in the orthogonal direction, according to their molecu-lar weight. After staining, the result is a two-dimensional image consisting of alarge number of spots corresponding to the constituent proteins. The intensityof spot coloring with certain stains is approximately proportional to the quan-tity of protein present. However, spot resolution may not be sufficient to sepa-rate all the proteins; therefore the results obtained using two-dimensionalelectrophoresis gels are subject to problems of reproducibility and artifacts.Recently, these problems have been partially resolved by using highly stan-dardized protocols and high-precision techniques. Today, it is possible to sepa-rate 2000 spots on a single gel. Proteins that have extreme isolectric points areunder-represented (supplementary gels with an extreme pH can partly remedy


1 Although the notions of activity, localization, and interaction also apply to mRNA, they are at presentinaccessible on a large scale.2 pH at which the overall charge on the molecule is neutral; otherwise stated, the pH at which the mol-ecule does not migrate in the applied electric field. The first dimension of the gel therefore correspondsto a stable pH gradient.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (183)

this), as are particularly hydrophobic proteins, which are not solubilized by theweak detergent used in the first separation.

It is impossible to know at the outset which proteins are present in a spot.Identification may be achieved in part by referring to a database containing two-dimensional gel electrophoresis results. Such databases exist, notably for bacte-ria, yeast, fruitfly, mouse, rat, and human proteins. When such information isnot available in a database, and if the sequences of the proteins are known, theirpositions may sometimes be estimated by calculating their isoelectric points andmolecular weights, providing they have not undergone post-translational mod-ification. Otherwise, the spots must be removed and the proteins they containeluted from them. If the sequence of a protein is known, it may be identifiedafter mild hydrolysis, either by microsequencing or by mass spectrometry of thepolypeptides obtained. In rare instances, an eluted protein can be renatured andits biological activity tested.

Intracellular localization

It is more difficult to evaluate the cellular localization of proteins on a largescale; however, some attempts have been made to do so for yeast proteins. Allcoding sequences have been fused to a fluorescent reporter protein, such as greenfluorescent protein (GFP), each strain containing only one such fusion sequence.The strain collection includes at least one representative of each coding sequencefused to GFP. The fluorescence is then located in the cells of each strain, usinga light microscope equipped with a fluorescence device. This method is limited


Isoelectric point



ar w



Figure 7.1 Two-dimensional protein electrophoresis gel. Proteins contained in a natural extractare fractionated in a polyacrylamide gel subjected to an electric field. They are separated firstaccording to their isoelectric point in a stabilized pH gradient and a weak detergent (horizontalaxis), then by their molecular weight, in the presence of a strong ionic detergent (vertical axis).Finally, they are stained in the gel to render them visible.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (184)

by the low resolution of the light microscope (0.3µm) compared with the sizeof the cell, as well as by localization artifacts sometimes caused by the fusedreporter protein, or by over-expression.

Protein–protein interactions

The vast majority of proteins interact with other proteins, either in a stablemanner, in which case they are known as a complex of polypeptide subunits,or in a more or less transitory manner. All degrees of stability are possible, andmay be expressed in terms of the half-dissociation time, or of the dissociationconstant, to which the half-dissociation time is inversely proportional. The variation in free energy is proportional to the logarithm of the association/dissociation equilibrium constant. Like the tertiary structure of individualpolypeptides, association between polypeptides involves weak chemical bondsor covalent disulfide bridges between two cysteine residues. These interactionsmay be demonstrated in various ways.

Traditional molecular biological methods

Classically, various molecular biological methods are used to demonstrate thatproteins co-purify, thus interact in a non-transitory manner. Biochemists gener-ally rely on a succession of chromatographic and biophysical methods whentrying to purify a molecular entity responsible for a biological activity that theyknow how to assay. They evaluate the nature of such a molecular agent by frac-tionating it on a dissociation or so-called ‘denaturing’ gel; i.e., under conditionsthat dissociate molecules from their neighbors but do not cleave them. If themost active purified sample corresponds to several protein bands in the dis-sociation gel, there is reason to suspect that the biological activity is due to a polypeptide complex. The biochemist can then try to identify the genes thatcode for the various polypeptides, using well-tested techniques inappropriatelyknown as ‘reverse genetics’.

If available, antibodies against individual polypeptides presumed to be con-tained in the complex can be used to test co-purification directly. A polypeptideis immunoprecipitated with its antibody under controlled conditions, and dif-ferent antibodies are used to verify whether other polypeptides are also presentin the immunoprecipitate (for example, by immunoblotting, or by successiveimmunoprecipitations with various antibodies). Determining the most dissoci-ating conditions that maintain co-immunoprecipitation (for example, by varyingthe salt concentration) permits qualitative evaluation of polypeptide interactionstability. But even this relatively direct approach is difficult to implement on the


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (185)

scale of the complete set of proteins in an organism. One limitation of theimmunoprecipitation technique is its poor ability to detect transitory interac-tions, unlike the approach described in the next section.

Two-hybrid approaches

An alternative to the classical approaches described above is to test the inter-action between polypeptides A and B by their ability to bring artificial elementsadded to them by genetic engineering close enough to each other. In a versionof this approach (Figure 7.2), one of the manipulated genes encodes a chimericprotein containing polypeptide A fused to the DNA-binding domain of a regu-latory protein. The other manipulated gene encodes a chimeric protein consist-ing of polypeptide B fused to the transcription-activating domain of a regulatoryprotein (the same or another). Both genes are then expressed in the same cell(usually yeast), which also includes a reporter gene whose transcription is acti-vated by the regulatory protein. The transcription-activating domain alone hasno effect, since it lacks the capacity to bind DNA; the DNA-binding domainalone has no effect because it cannot activate transcription; hence, the reportergene is not expressed. However, if polypeptides A and B form a complex, thetwo domains are brought close enough to each other to restore the full activityof the regulatory protein. The complex then binds to the DNA, thereby acti-vating transcription of the reporter gene, which codes for a protein that is readilyassayed, usually by a colorimetric method, providing semi-quantitative data onthe binding of the two polypeptides. As a precaution, it is wise to test the inter-






Reporter geneRegulatory region



Figure 7.2 Two-hybrid assay. A functional activation complex is reconstituted by interactionbetween two tested polypeptides, A and B. A is fused to the DNA-binding domain of a regulatoryprotein. B is fused to the transcription-activating domain of a regulatory protein. Once the DNAbinding domain is associated with DNA upstream from the coding sequence, interaction betweenA and B causes the cell transcription machinery to activate expression of the gene. The gene codesfor a protein whose biological activity serves as a reporter for the transcription level.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (186)

action by reversing the roles of A and B, this time fusing A to the transcription-activating domain of the regulatory protein and B to its DNA-binding domain.Strong interaction in both configurations indicates true association of the twopolypeptides. In certain real cases, it may be possible to demonstrate interac-tion in one direction but not in the other, justifying the use of a one-way arrowto symbolize the interaction. However, this reaction between proteins is funda-mentally symmetrical: A binds to B ⇔ B binds to A. In other real cases, thereporter gene is highly expressed whatever B may be, since A is capable ofbinding to DNA by itself, or of activating transcription, or both; for example,if A is a regulatory protein.

In another version, A and B are both fused to fluorescent polypeptides, oneof which emits blue light under ultraviolet excitation. If the other is very near,it absorbs the blue light and emits green. A fluorimeter set to detect the greenlight records a signal only when A is very close to B, since the energy transferis proportional to the distance between the two fluorophores to the sixth power.

Various other possibilities exist; for example, testing a polypeptide fused toone regulatory protein fragment against a ‘two-hybrid bank’ of random fusionsto the other regulatory protein fragment. The result of such an assay is synthe-sis of a dye that visibly stains the yeast colony, or of a protein essential for cellsurvival, thereby permitting direct selection of cells that bear the winning com-bination. In the ‘reverse two-hybrid’ technique, the reporter gene is a ‘killer,’allowing only cells bearing combinations that do not associate to survive. Thetwo-hybrid technique is binary, which also constitutes a major limitation: if athird protein is required, two polypeptides being tested will never be found toassociate. A so-called three-hybrid variant technique therefore also exists, inwhich the interaction between modified polypeptides A and B is measured inthe presence of a required third protein sandwiched between them. Note that ifthe two-hybrid approach is carried out in the organism from which the testedpolypeptides come, a sort of ‘accidental’ three-hybrid may occur, since the thirdcomponent is already present in the organism.

It is equally possible to test portions or domains of proteins rather than com-plete polypeptides. This has several advantages: The domain that interacts withthe other polypeptide is clearly specified. The validity of the result may be esti-mated by the degree of redundancy and connectivity of the observed interac-tions of domains from the same natural polypeptide. Additionally, it becomespossible to test a polypeptide (for example, a membrane protein) that in itsnatural state would be distant from the nucleus, where the reporter gene islocated.

The two-hybrid technique is sometimes used on the full set of proteins of anorganism. If the number of proteins to be tested in both configurations (baitand test) is N, there are obviously 2N genetic constructions, and around N2 testswould have to be carried out. This is an operation that calls for major automa-


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (187)

tion. One limitation of this technique is that it is possible to test only proteins,which is not the case for the technique described below.

Protein chips

The affinities of a given protein for other molecules may be tested by purifyingit and attaching it to a microarray. This may be systematically carried out byestablishing as many strains as there are genes in the organism being studied.Each strain over-expresses a different gene, which encodes its natural productfollowed by a constant stretch of a few amino acids with a strong affinity fora highly specific reagent (for example, the stretch may be an epitope againstwhich specific antibodies are available). A microarray is coated with the spe-cific reagent and a crude protein extract of each culture is deposited at eachspot. The microarray is washed under conditions that keep the modified proteinfixed and eliminate the others. The microarray is then incubated in the presenceof the substance to be tested (generally a protein, nucleic acid, lipid, etc.) thatbinds at spots where its binding partner is located, which can then be assayed.

Unlike the two-hybrid technique, this approach has the disadvantage of beingable to test interactions only under completely artificial conditions, and islimited to proteins that maintain their functional conformation when purifiedand deposited onto the solid surface of a microarray. An older and less fre-quently used alternative consists in attaching a different specific antibody at eachspot on the microarray and testing for the presence of proteins that are recog-nized by the antibodies in a cell-free extract. Antibody preparation is obviouslya time-consuming task with this technique. The protein chip described herewould require constructing N strains in order to prepare an array consisting ofN protein probes.

A future type of protein chip might be based on attaching a different DNAfragment at each spot and programming in situ synthesis of each protein ofinterest. Such a protein would simply have to remain bound to its mRNA, whichis itself bound to the attached DNA. This could easily be achieved by remov-ing the translation and transcription termination signals from the DNA.

Systematic identification of protein complexes

New techniques feature relatively mild purification of complexes containingseveral polypeptides. The complexes obtained are resolved by ultrasensitivemass spectrometry, which, as discussed in section 2.1.1, often permits identifi-cation of polypeptides on the basis of their molecular weight, predicted withthe aid of a sequence databank. This method overcomes the binary nature ofthe two-hybrid technique.


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (188)


As discussed in the chapter on comparative genomics (Chapter 3), conservationof gene proximity over long evolutionary periods, at least in bacteria, is a strongindication that such genes code for proteins that physically interact.

7.2.2 Transcriptomics

Large-scale monitoring of genetic expression is inspired by the premise that thefunctional state of an organism is largely determined by the expression statusof its genes. The latter may be described as the quantity of each mRNA mole-cule present in the cell at a given instant. These quantities evolve over time asa function of the interactions between regulatory proteins and regulated genes.Methods for measuring the quantity of mRNA, as well as for investigating therelationship between regulatory proteins and their gene targets, are briefly exam-ined below. These methods are somewhat similar.

Complementary DNA microarrays

Used throughout the world today, microarrays are plates (usually glass) ontowhich complementary DNA (cDNA) probes are deposited by high-speed robo-tized printing methods. They are very well adapted to analyzing the expressionof up to 10,000 genes derived from sequencing projects (Figure 7.3).


Samplescontrol experimental


cDNAGene G

DNA microarray

Figure 7.3 A microarray. Complementary DNAs (cDNAs) prepared from two messenger RNA (mRNA)sources (for example, the undisturbed sample on the left and the stimulated sample on the right),are marked by two different fluorophores, one red and one green. The cDNA is then probed in amicroarray bearing gene probes representative of the entire genome, for example, gene G. Theprobes themselves consist of cDNA that has been attached to the microarray.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (189)

Measurement is carried out by differential hybridization, in order to mini-mize errors due to variations in cDNA printing. mRNA from two differentsources (for example, a drug-treated assay and an untreated control) is usuallyreverse-transcribed into cDNA and marked with either of two fluorophores, onefor the assay and one for the control. Differential hybridization of the micro-array probes is carried out with a mixture of both liquid phase cDNA prepa-rations, thereby minimizing systematic errors (Figure 7.3). After hybridizationand washing, each fluorescent signal is independently evaluated and used to calculate the ratio of the concentrations of the experimental and control nucleicacids. This ratio is generally used as the starting point for interpreting the results.

Such microarrays have already been used to measure the genetic expressionlevels of the complete S. cerevisiae genome (around 6,400 distinct cDNAsequences) during various kinds of treatment, such as transition from sugar- toethanol-based metabolism, or sporulation, or throughout the entire cell cycle.Such data sets are available in the public domain. Guides explaining how to con-struct devices that deposit cDNA onto the surfaces of microarrays, as well as howto analyze fluorescence levels, can also be found on the Internet. In addition, pre-prepared microarrays are commercially available for an increasing number oforganisms, including human, rat, mouse, plant (Arabidopsis), and some bacte-ria. However, as already mentioned, they are readily adaptable to a wide rangeof available cDNA probes corresponding to individual laboratory requirements.

Oligonucleotide chips

These chips are produced mainly by Affymetrix®, and consist of small siliconplates to which thousands of short oligonucleotides (20 nucleotides or more)are attached. The oligonucleotides are synthesized directly on the surface of thechip by photolithography and light-controlled chemical synthesis. Due to thecombinatory nature of the process, it is possible to probe a very large numberof mRNA molecules simultaneously. Today, chips consist of as many as 200,000different probes, usually including several for each mRNA molecule (Figure 7.4).Chips may contain different exons of an intronic gene or some perfect-pairingprobes as well as a few mispaired probes with a single nucleotide mismatch.

However, the preparation and reading of oligonucleotide chips requiresexpensive equipment, and at present only commercially produced standard chipsare available at an accessible price. This does not provide laboratories the oppor-tunity to pose specific questions.

The future of the microarray/oligonucleotide chip industry would appear tolie in the development of synthetic probes that are longer than the earlier ver-sions (but shorter than cDNA sequences) and that are attached after synthesisand verification. A disadvantage of the methods described above is their lack ofsensitivity, which means that tens of thousands of cells have to be mixed in order


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (190)

to obtain enough material for a microarray experiment. This can be a problemif only a small amount of tissue is available, e.g., from an isolated embryonictissue. In addition, the response dynamics are diminished by auto-absorption offluorescence, steric hindrance reducing access to the probe, etc. These problemsare exacerbated by the cost of microarrays, which limits kinetic experiments,since each time point corresponds to the consumption of one chip. A factor oftwo is usually estimated to be the practical limit of resolution for massively par-allel quantification; therefore two values must be separated by a factor greaterthan two in order to be sure they are different. Of course, this factor is only anorder of magnitude, since it is a function, among others, of the absolute valuesinvolved. In any case, it imposes limits on the quantity of useful informationthat can be obtained from massively parallel measurements.

Reverse-transcription–polymerase chain reaction

In order to evaluate gene expression using the reverse-transcription–polymerasechain reaction (RT–PCR), mRNA is first reverse-transcribed into cDNA, thenamplified by PCR until detectable levels are reached. Using internal calibrationtechniques, it is possible to obtain exceptionally high levels of sensitivity (on the order of 1 molecule per microliter of sample volume) and dynamic range (6to 8 orders of magnitude). This method requires using primers for all genes of


Figure 7.4 Portion of an oligonucleotide chip. Each fluorescent patch corresponds to a specificgene probe.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (191)

interest and, unlike the techniques described above, cannot be run in parallel.It is therefore crucial that the procedure be automated in order for it to func-tion on a large scale. In practice, it is used to obtain more precise data on asmall number of genes that have previously been identified as especially interesting.

Serial analysis of gene expression (SAGE)

Serial analysis of gene expression (SAGE) utilizes a very different technique tomeasure mRNA levels. First, cDNA is synthesized from mRNA; then a DNAtag long enough (~10–20 basepairs) to unambiguously identify a gene is cutfrom each cDNA fragment at a precise site. The tags are then concatenated intoa long double-stranded DNA sequence, and the concatenated DNA is amplifiedand sequenced. If tag T1 is 10 times more frequent than tag T2 in the con-catenated DNA sequence, it indicates that the mRNA containing T1 is 10 timesmore abundant than the mRNA containing T2.

The SAGE method has two advantages: (i) it is not necessary to know themRNA sequence in advance, therefore allowing detection of unknown genes,and (ii) it utilizes a sequencing technology already commonly used in numerouslaboratories. However, SAGE involves a somewhat complicated procedure andnecessitates massive sequencing. SAGE has already been used to analyze theentire set of S. cerevisiae genes expressed during various phases of the cell cycle,as well as the expression of tens of thousands of human genes, comparinghealthy and cancer cells.

Chromatin immunoprecipitation

A recently developed approach has permitted direct investigation on a genomicscale of the gene targets of proteins that regulate transcription by binding toDNA3 (Figure 7.5). During the first phase, a chemical fixative, such as formalde-hyde, is added to a culture of living cells. The bivalent reagent rapidly entersthe cells, where it forms covalent bonds between two chemical groups in closespatial proximity. In particular, it can form a solid bridge between a regulatoryprotein and its DNA binding site, if one happens to be bound at that instant.The cells are then ruptured and their DNA. The extracted and broken mechani-cally into random fragments of about 1kb. DNA fragments associated with theprotein of interest are then immunoprecipitated, using an antibody specificallydirected against the protein. The second phase consists in deproteinization, afterwhich PCR may be used to amplify the DNA; alternatively, the DNA can be


3 This method may also be used to study other DNA-binding proteins.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (192)

identified by hybridization to a microarray or biochip. Ideally, the microarrayshould include probes that are representative of both gene and intergene regions.Since regulatory regions are usually intergenic, in principle, one would expectthe precipitated DNA to hybridize with intergene rather than gene probes. The first phase is known as Chromatin ImmunoPrecipitation (ChIP). Since thesecond phase requires the use of a microchip, the whole method is called theChIP–Chip technique.

However, this approach has its problems, especially those involving back-ground noise due to low antibody specificity. These problems require the use of


DNA-protein bonding

DNA cleavage

Protein immunoprecipitation

DNA microarrayprobing

DNA deproteinizationand marking

Marked referenceDNA


Figure 7.5 The ChIP–Chip technique. Proteins associated with DNA are covalently bonded by abivalent reagent such as formaldehyde, which penetrates and kills living cells. The DNA is mechan-ically broken into random fragments of average length on the order of the size of a gene. A DNA-binding protein is then immunoprecipitated with specific antibodies. Fragments of theco-precipitated DNA are deproteinized and marked with a red fluorophore (for example). In paral-lel, DNA fragments representative of the entire genome are marked with a green fluorophore (forexample). The red and green DNA fragments are then mixed, and the mixture used to probe amicroarray or biochip that contains both gene and intergene regions of the organism being studied.After washing, the intensity of the two fluorescences is measured. For each microarray probe, theintensity ratio indicates the level of enrichment in protein-associated DNA.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (193)

relatively arbitrary thresholds to analyze the results. Nevertheless, as expected,the precipitated DNA preferentially hybridizes with the intergene probes, ratherthan with the gene probes. It also hybridizes more often with DNA regulatoryregions that contain the motif recognized by the regulatory protein than withzones that do not. The results indicate that the number of regulated targets persequence-specific regulatory protein is higher than previously believed based onclassical genetics and biochemistry experiments. Thus, the yeast (S. cerevisiae)regulatory protein Rap1p has around 300 targets, which is around 5 percent ofthat organism’s genes.


A protein that regulates transcription preferentially binds to DNA sites identi-fied by a certain sequence or by a small subset of neighboring sequences (Figure7.6). According to principles discussed in the preceding chapters, these poten-tial sites may be detected using textual analysis of chromosome sequences. Ifthe predicted position of such a site relative to the coding part of a gene matchesthe actual site, it is reasonable to assume that transcription of the gene is regu-lated by the DNA-binding protein. However, in practice, this approach encoun-ters various difficulties, the most serious of which is that such sequences areusually very short and degenerate. The number of potential protein binding sitesis therefore disproportionately large and identification of real sites is not basedon solid criteria, except for eliminating sites that are located in coding regions,which are often as numerous as those located in regulatory zones. It is some-times possible to obtain better results by noting when these potential sites appearthree or more times in rapid succession in the regulatory region of the same


Figure 7.6 Visual representation in the form of a logo of a consensus DNA site at which a reg-ulatory protein is likely to bind. The motif presented in this example is 12 nucleotides long. Onlyone DNA strand is shown. The complete data could be represented in the form of a 4-row (ACGT),12-column (positions) array giving the percentage of each nucleotide at each position calculatedfrom a compilation of sequences shown to bind to the regulatory protein. The sum of the heightsof the letters at all the alignment positions indicates the information content in bits. The rela-tive sizes of the letters correspond to the frequency of the nucleotide at each position. For example,only A can be found at position 11; A is nearly always found at position 10; both G and A arefound at position 6.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (194)

gene. Empirically, such repetition often corresponds to effective regulation,which may be demonstrated using a direct approach. In higher eukaryotes, the respective positioning of binding sites for different regulatory proteins may be used in a similar manner as a criterion for improving the quality of predictions.

The emerging practice is to use three approaches simultaneously to discrim-inate among significant interactions between genes. Genes are considered to bethe probable targets of a regulatory protein if they simultaneously satisfy thefollowing three conditions: i) the regulatory region contains at least one site thatis recognized by the regulatory protein (bioinformatics); ii) the DNA co-precipitates with the regulatory protein (ChIP–Chip); iii) the transcription levelis modified following a stimulus known to trigger a response involving the regulatory protein (kinetic experiment using a microarray or biochip).

7.3 Macromolecular networks

Most cellular processes are rooted in dynamic interactions among a greatnumber of biological molecules that both implement and undergo regulation.This section covers such interactions involving macromolecules. Distinctionswill be made among interactions between proteins, between an enzyme proteinand its substrate, and between a regulatory protein and DNA. Other types ofinteractions, which are potentially as relevant to the physiology of the organ-ism but are not currently subject to massive genome-wide investigation, will notbe covered here.

Before analyzing these networks separately, it is important to recall that theydo not act independently in the cell. Biological information is usually describedas flowing from DNA to RNA to protein to function (Figure 7.7, left). However,this view overlooks the fact that function emerges from a network of interac-tions among active macromolecules and small molecules, and only exception-ally from a single macromolecule. The network constitutes a filter between theisolated macromolecule and the function. This view also fails to take intoaccount that the state of the protein interaction network feeds back onto theRNA state (affecting alternative splicing, for example). Likewise, the states ofthe protein and RNA networks feed back onto the DNA state (for example, thepattern of active and inactive genes) (Figure 7.7, right side).

7.3.1 Protein-protein interactions

When a large number of binary interactions between proteins are detected, aninteraction map may be drafted. This map summarizes knowledge of the proteininteractome, which is the set of all interactions among proteins in a cell. The


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (195)

interactome is one component of what is known as the proteome. This mapmay be abstracted into a graph in which each node represents a protein andeach edge indicates a non-directional interaction between two proteins4. As anexample, Figure 7.8 presents a small portion of such a graph. When the tech-nique employed permits, a number between 0 and 1 may be assigned to eachedge in order to quantify the stability of the link. These quantitative and semi-quantitative aspects can strengthen purely topological considerations intendedto distinguish subsets of proteins in the graph. The idea is that a subset (suchas the one shaded in dark gray in Figure 7.8) has dense internal connectivitybut few connections to other subsets (such as the one shaded in light gray).

7.3.2 Enzyme–substrate interaction

Interactions between enzymes and their substrates roughly correspond to whatis known as metabolism, although some cases are borderline. For example, whatif the substrate is a protein that the enzyme modifies irreversibly? The subject



Information and its flow in








Epigenetic networks



Figure 7.7 Genetic and epigenetic concepts of information flow in biology. In the all-geneticview, represented on the left, information flows from DNA to RNA to protein (except for reverse-transcription of RNA into DNA), and the bioactive macromolecule directly determines a function.In the epigenetic view, represented on the right, the network of interactions among bioactive molecules determines the biological functions. These networks feed back onto the state of themolecules represented higher up; for example, protein bioactivity, RNA alternative splicing, andtranscriptional activity.

4 It is tempting to orient links involving an asymmetric transitory interaction between two proteins, oneof which is the substrate of the other (for example, when one protein is phosphorylated by another).This point is discussed below, precisely in the context of enzymatic reactions.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (196)

here is the metabolome, the set of metabolic pathways in a given organism.Metabolisms were essentially described several decades ago, using what are nowconsidered classical biochemical and genetic approaches. These approaches arenot the subject of this section, which will focus on first the local, then the globalcharacteristics of metabolism.

Metabolic pathways

A metabolic pathway consists of a series of successive chemical reactions thattransform a substrate into a product. Each reaction is catalyzed by an enzymethat is characterized by its catalytic properties (Figure 7.9). In classical cases,the effective catalytic rate is a hyperbolic function of the substrate concentra-tion, involving an affinity parameter, Km, and a maximum rate, Vmax (see Insert7.1, below).


Protein 11Protein 12

Protein 10

Protein 9Protein 8

Protein 1

Protein 6Protein 5

Protein 4Protein 2

Protein 7

Protein 3
















Figure 7.8 Protein–protein interaction map. In principle, links between proteins are non-directional; if protein A is associated with protein B, the reciprocal is also true. In certain cases,it is possible to assign weights to edges, represented here as a number between 0 (no link) and1 (strong link). Such interaction charts may be obtained by means of either molecular biologicalor bioinformatics methods. The molecular methods are traditionally based on the co-purificationof two polypeptides, indicating that they probably belong to the same complex. More recently, thetwo-hybrid approach, co-precipitation of protein complexes, and the use of protein chips have per-mitted more systematic experimentation. On the bioinformatics side, conservation of the proxim-ity of two genes over major phylogenetic distances often indicates physical interaction betweentheir products.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (197)

Insert 7.1The Michaelis–Menten Law (1913)

Given a chemical reaction that transforms substrate S into product P whencatalyzed by enzyme E, it is possible to account for experimental observa-tions by postulating the formation of a transitional complex ES between Eand S. This complex then dissociates, recycling E and releasing P. Theforward rate constants are k1 and k2, and the reverse constants are k−1 andk−2. The latter constant (k−2) is neglected (see below).

Michaelis and Menten found a simple relationship between the catalysisrate V and the substrate concentration [S], given three hypotheses. This rela-tionship involves only two parameters, the Michaelis–Menten constant Km and the maximum catalysis rate Vmax. These hypotheses in turn imposecertain experimental constraints.

1) Initial rate conditions:

In practice, enzyme kinetics is measured only during the initial reactionperiod, when substrate consumption is a linear function of the reaction time.During this initial period, so little P is generated that the reverse reactionmay be neglected. Concentrations appear in brackets.

v P− ( )2 disappearance of negligible:

v P k ESsynthesis of ( ) = [ ]2

E S ES E Pk k

k+ ← → → +

1 2



A > B


Km, Vmax

B > C


Km, Vmax

C > D


Km, Vmax

Figure 7.9 In this metabolic pathway, three successive chemical reactions transform one smallmolecule A into another D. Each reaction is catalyzed by a different enzyme. The first one, notatedA > B, accelerates the transformation of substrate A into product B. Product B in turn serves asthe substrate for enzyme B > C, which accelerates its transformation into product C, etc. Biochemists measure two parameters for each enzyme, purified separately: the Michaelis–Mentenconstant, related to the reciprocal of the affinity for its substrate (Km), and the maximum rate of catalysis (Vmax).

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (198)

2) Stationary conditions:

The first fractions of a second of the initial reaction period, during which the ES complex has not reached a constant concentration, are ignored. Whenthe reaction becomes stationary; that is, when ES synthesis equals ESconsumption:

3) Enzyme dose and mass conservation:

The hypothesis is that the enzyme-catalyst is added in much lower concen-tration than the substrate. In addition, the masses of the enzyme and the sub-strate are conserved.

Let [E]T be the total concentration of the enzyme and [S]T that of the sub-strate (experimentally controlled quantities):

⇒ [ ] + [ ]( )= [ ] [ ]





m m

⇒ [ ] = [ ] − [ ]( )[ ]ES




S S T[ ] ≈ [ ]

S S ES E ST T[ ] = [ ] + [ ] [ ] << [ ]and

E E EST[ ] = [ ] + [ ]

Letting mKk k

K= +−1 2


Michaelis–Menten constant


[ ] = [ ][ ]m

k E S k k ES1 1 2[ ][ ] = +( )[ ]−

v v1 1= −

v E k k ES− −( ) = +( )[ ]1 1 2consumption of S

v E k E S1 synthesis of S 1( ) = [ ][ ]


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (199)

The special case of enzyme saturation:

If the substrate concentration is ‘infinite’ (in practice, very high), all theenzyme is complexed to the substrate and the reaction rate is maximum.

The relation sought may then be written as:

This hyperbolic relation can be linearized by double inversion:

It suffices to represent the reciprocal of the initial slope of the formation ofP as a function of the reciprocal of the concentration of substrate added,varying the latter, then drawing a straight line through all the experimentalpoints to calculate the values of the two parameters, Km and Vmax (Figure 7.10).

1 1 1V V K V S T= + ⋅ [ ]max maxm Lineweaver–Burk equation




= [ ][ ] +max


Michaelis–Menten equation


V k ET


[ ] ≈ [ ]= [ ]max 2

maximum reaction rate

⇒ = [ ] = [ ] [ ][ ] +

V k ES k ES



2 2m

⇒ [ ] = [ ] [ ][ ] +



T m


Figure 7.10 Linear relationship between the reciprocals of the measured reaction rate andthe total substrate concentration, which may be experimentally varied. The two parameters ofthe Michaelis–Menten equation, Km and Vmax, are, respectively, read directly on the abscissa at the ordinate at the origin. The slope is Km/Vmax.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (200)

Remember that Vmax is the maximum reaction rate obtained for a given doseof enzyme acting on a substrate of infinite concentration. The Km is the concentration of substrate that gives a catalysis rate one-half the Vmax. It represents the reciprocal of enzyme–substrate affinity.

Regulation of cell metabolite concentrations

The stable metabolite at the end of a metabolic pathway (an unbranched linearchain) often has an inhibitory effect on the enzyme that catalyzes the first stepof the pathway; this is called negative feedback. The concentration of the finalmetabolite is regulated by the sensitivity of the initial enzyme to its inhibition.Other, more specific, regulatory mechanisms exist, operating either by positivefeedback or by connecting to two different pathways in competition with eachother.

Ignoring these specific regulations, the metabolic network may be regardedas a set of interconnected pathways in which the first step of each pathway isregulated by feedback inhibition. Flux across this pathway would therefore beexpected to depend strongly on the concentration of the rate-limiting enzyme.However, this is not the case. Indeed, comparison of two diploid organisms thatpossess respectively one and two copies of a gene that codes for an enzymewhose concentration is the limiting factor in a given metabolic pathway, (sur-prisingly) reveals their metabolic fluxes to be rather comparable. Artificiallyincreasing the quantity of enzyme also has little effect on metabolic flux. A meta-bolic pathway is significantly affected only by total loss of enzymatic activity.Thus, negative feedback is not the principal way that cells regulate their meta-bolic pathways. Control is exercised in a more distributive fashion.

Distributive control of metabolic pathways

In 1973, Kacser and Burns studied the integrated kinetics of a linear sequenceof reactions catalyzed by enzymes that convert one stable metabolite intoanother. They described how genes influence flow through a metabolic pathwayas:

According to Haldane’s modification of the Michaelis–Menten equation, theconversion rate of Si into Sj may be expressed as:

(7.1)v V K S S K S K S Ki i mi i j i i mi j mj= − ( )[ ]{ } + ( ) + ( ){ }1

. . .E E E

S S S S1 2

2 3 1


1 n⇔ ⇔ ⇔ +


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (201)

in which vi is the effective conversion rate, Vi is the maximum conversion rate(which depends on the enzyme concentration Ei and the rate constants); Kmi isthe Michaelis constant for Si; Kmj is the Michaelis constant for Sj; and Ki is theequilibrium constant for the Si ⇔ Sj reaction.

Most enzymes function under conditions in which their substrates are notsaturating, such that Si << Kmi and Sj << Kmj. Thus, the value of the denomina-tor in Equation 7.1 is approximately 1.

At equilibrium, all vi rates become equal to the overall rate of the sequenceof reactions. Let us call that overall rate ‘Flux F’, and write the following seriesof corresponding equations:

. . .

In summing all these elements, note that the terms on the right cancel out two-by-two, except (of course) for the first and last. This gives:

Metabolites S1 and Sn+1 are the molecules at the beginning and end of the reac-tion, whose concentrations are relatively constant. Thus, the term on the rightside of the above equation is a constant, which we will call Cs.

Each Vi is the product of a rate constant Ki and the enzyme concentration Ei.These two parameters are genetically determined. Each term on the left side ofthe equation may be replaced by 1/ei, a composite term for the genetically deter-mined parameters that contribute to the overall flux. The following simplifiedequation is obtained:


The essential point expressed by Equation 7.2 is that control of flux across ametabolic pathway does not occur in a single key step – the initial one, forexample – but is distributed among all the enzymes in the pathway. It thusbecomes possible to examine the dependence of F on the concentration of a singleenzyme Ei, which could change due to inactivation of one of two cell copies ofthe gene that codes for Ei, for example. The sum of the terms in the denomina-tor of equation (7.2) is a constant Ce for all the enzymes except Ei, thus:

F Cs e e e= +( )1 1 11 2 . . . n


S S K K K* * . . . * * . . . *

* . . . *

m m 2 mn n n

n n

1 1 2 1 1 2 1

1 1 1 2

+ ( ) + (( )[= − ( )[ ]


K F V K K K S K K K S K K Kmn n n n n n n* * * . . . * * . . .* * . . . *[ ] [ ] = [ ] − ( )[ ]− − +1 2 1 1 2 1 1 1 2

K F V K S K S K Km2 2 1 2 1 3 1 2* * *[ ] [ ] = [ ] − ( )[ ]

K F V S S Km1 1 1 2 1*[ ] = − [ ]


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (202)

or even


This hyperbolic relationship between F and ei yields an initial slope Cs, thenasymptotically attains the value Cs/Ce. Consider the case in which F is near theoptimum, when there are two active genes coding Ei in a diploid organism. Wesee that with only one gene, and a halved E1 concentration, F is little altered.The overall flux through the metabolic pathway weakly depends on each indi-vidual enzyme. This conclusion is supported by the experimental observationthat enzymes function physiologically in the region of the hyperbolic curve nearthe asymptote, where the slope is shallow.

The metabolome

The entire set of interactions among enzymes and small molecules contained ina cell is known as the metabolome. It may be represented as a graph in whicheach node represents a small molecule and each edge corresponds to a chemi-cal reaction catalyzed by an enzyme (Figure 7.11). Linear and unbranched meta-bolic pathways may be detected with the naked eye. Although going from a mapto a graph is straightforward, it raises the issue of directionality. Since eachchemical reaction is reversible, the graph must be non-directional. Whereas thisis certainly true under conditions in which the reactions are likely to attain equi-librium, it is rarely the case in living cells, which function far from equilibrium.In practice, most cell reactions are drawn in one direction by coupling to astrongly exergonic reaction and by continuous consumption of the final product.For example, during protein synthesis, an amino acid is consumed as soon asit is produced. This is what the directed representation in Figure 7.9 is meantto convey. It is therefore justified to consider the network of metabolic interac-tions to be a directed graph. However, there are some exceptions, in which apathway functions in one direction or the other, according to the circ*mstances.But even in this case, one direction is privileged at any given time.

7.3.3 Interactions between regulatory proteins and DNA regulatory regions

Regulatory proteins and DNA regulatory regions merit an introduction prior todiscussing their interactions.

F Cs e Ce e= × + ×( )i i1

F Cs Ce e= +( )1 i


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (203)

Genes and DNA regulatory regions

According to one definition, a gene is an abstract entity bearing two kinds ofinformation: the first is the code for the sequential assembly of a macromole-cule; the second specifies the quantity of that macromolecule that is to be syn-thesized. The physical support of genes is always a nucleic acid polymer, usuallyDNA. The coding region of DNA bears the first piece of information and theregulatory region bears the second. If the coded macromolecule is a protein, theDNA coding zone is the open reading frame (ORF), and the DNA regulatoryregion is usually (often only, in microorganisms) located upstream from thecoding region (see Gene A, on the left of Figure 7.12).

Regulatory proteins

When a regulatory protein binds to a DNA regulatory region, it modulates theexpression of one or more genes. These genes are either grouped together andshare the same DNA regulatory region, to which the regulatory protein binds


Figure 7.11 Metabolic network of a bacterial cell. Strictly speaking, this simplified map repre-sents interactions between small substrate molecules (round nodes), each linked by catalyzedchemical reactions (light edges). Reciprocally, it would also be possible to represent the enzymesas vertices and the substrates/products of the reactions they catalyze as edges. The darker pathwaycorresponds to a major metabolic pathway, glycolysis (glucose breakdown), ending in the Krebscycle, on the right.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (204)

(bacterial operons, for example), or are dispersed along various points of theDNA molecule, where copies of the regulatory protein can bind.

What are the limits of the definition of a regulatory protein? We are inter-ested here in proteins that regulate a subset of the genes of an organism. Otherproteins also called regulatory play a more general role in initiating transcrip-tion (for example, most eukaryote transcription factors of type II). In principle,these other proteins, like RNA polymerase itself, are essential for the tran-scription of all genes that code for proteins. However, since their action is non-specific, they are not covered in this section. Nevertheless, the boundariesbetween ‘generalist’ and ‘dedicated’ regulatory proteins have recently becomeblurred. Indeed, the new post-genomic ChIP–Chip technique (see section7.2.2.5) may be used to locate regulatory proteins at their DNA binding sites.This technique is sometimes used in combination with specific inactivation ofeach regulatory gene. Studies with yeasts have shown that certain dedicated reg-ulatory proteins affect up to 5% of the genes of an organism, whereas certaingeneralists affect only 3%. Therefore, there is a quasi-continuum of regulatoryproteins that covering the whole spectrum of genes in the genome. Accordingto current knowledge, the threshold between generalist and dedicated regula-tory proteins may be empirically established at around 3 to 5% of all genes. Itis the dedicated regulatory proteins that are of interest here.


Figure 7.12 Interactions among some fictional genes. DNA, the material support of all genes, issymbolized in two parts: the regulatory region (RR) on the left and the coding region (ORF, oropen reading frame, on the right). Gene A is neither regulatory nor regulated. Gene B is not regulated, but its product is a regulatory protein, which, when it binds to the regulatory regionsof genes C and D, activates transcription of the former and inhibits transcription of the latter. Theproduct of gene C inhibits transcription of gene E as well as of itself. In so doing, gene C consti-tutes a feedback circuit, or unary self-regulating loop. Gene C also participates in a genetic reg-ulation pathway that links genes B and E, since it is both regulatory and regulated. In this figure,only B and C are regulatory genes, and only C, D, and E are regulated genes.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (205)

Genetic interactions

Regulatory proteins are coded by genes called ‘regulatory.’ ‘Regulated’ genesbear regulatory regions. Regulatory genes may themselves be regulated. It is cus-tomary to speak of interactions between regulatory and regulated genes, or‘genetic interactions.’ However, this terminology poses a problem, since there isa fundamental asymmetry between regulated and regulatory genes. Regulatorygenes exert their effects via their products, which are regulatory proteins. Thisimplies that a graph of genetic interactions is basically of the directed type.Bearing that in mind, we will continue to use the customary terminology.

Genetic interaction map

At the qualitative level, a regulatory gene can activate (positive effect) or inhibit(negative effect) a regulated gene target. A dual effect is also sometimesobserved, which may be either positive or negative, according to the circum-stances. The target gene may itself be regulated, in which case it participates ina genetic regulatory pathway. If such a regulatory pathway is closed onto itself,it forms a feedback circuit. Since the incoming and outgoing connections of agene can be multiple, the circuits and pathways are sometimes linked in a fullyconnected component of variable size. Some of these situations are representedin Figure 7.12, including the case of a gene that self-regulates, which is other-wise said to form a unary feedback circuit.

7.4 Topology of macromolecular networks

The set of molecular networks in a cell no doubt constitutes a heterogeneousweb with respect to both its nodes and edges. If we emphasize the geneticnetwork, then information rapidly flows from the patterns of genetic activityand through a cascade of intercellular and intracellular signaling functions,before slowly returning toward the regulation of gene expression. DNA regu-latory and coding sequences thus unfold into spatiotemporal structures thatdefine the organism. Other perspectives are possible, starting from and return-ing to protein (and possibly membrane) activity patterns (Figure 7.7). The challenge is therefore to identify significant connections in these regulatory net-works and to determine the abstract principles underlying the architecture anddynamics of the network that allow it to function in a reliable and flexiblemanner.

The accumulation of data has made network architecture accessible.Expressed in the symbolism of graphs, network architecture consists in describ-ing links (edges) that connect nodes (vertices), and eventually in the rules,


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (206)


Figure 7.13 Partial map of genetic interactions in Saccharomyces cerevisiae. Genes are identi-fied by the systematic names that were assigned to them during sequencing projects. For clarity,arrows are not used, but transcriptional influences go from left to right, as indicated by the thick arrow at the bottom. One-quarter of the 500 genes studied are regulatory, 52 of which arelinked to at least one other regulatory gene. These 52 genes are represented here in a graph indi-cating causal relationships. In order to reflect the causal flow, genes that lack a known regulator(except for self-regulation) are listed in the left-hand column. The other genes are then placed inthe left-most column, such that all their regulators are to their left. Respecting this rule, thenumbers of non-regulatory genes regulated by the regulatory genes in each column are indicatedon the bottom of the figure. For example (right column), YBR083W and YJR094C together regu-late a total of 10 genes. Self-activation is indicated in bold, self-inhibition in bold italics, acti-vation in solid lines, inhibition in dashed lines, dual regulation (one case) in dotted lines, andessential genes (whose knock-out is lethal) in frames. Genes belonging to the recurrent kernel areunderlined. (A recurrent kernel is a set of genes with at least one direct or indirect target in thesame group. According to this definition, a recurrent kernel includes feedback circuit genes andall genes that regulate them.)

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (207)

functions, and weights that may be assigned to links. Often the only informa-tion available is whether or not a link exists, which is insufficient for modelingnetwork dynamics. In addition, such dynamics would require introducing thenotion of temporal delay; thus it often continues to escape us.

The following section presents some elements of global and local networktopology analysis, which will then be separately applied to each hom*ogenousmacromolecular network. Only one, the genetic network, will be examined indetail. In doing so, we will bear in mind the reductionism to which this per-spective subjects cell functioning, both by artificially separating networks andby disregarding the spatial organization of biological objects.

7.4.1 Topological analysis

Global topology

Empirical and theoretical results indicate that networks may be divided into twomajor categories, according to the connectivity distribution pk, which indicatesthe probability that a node is connected to k other nodes. The first category ofnetworks is characterized by a pk that reaches a maximum at an average valuekaverage and that exponentially diminishes for higher values of k: pk ~ C e−bk,where b and C are constants. In such exponential networks, each node hasapproximately the same number kaverage of links. In the second category of networks, pk decreases according to a power law: pk ~ C k−g, where g and Care constants. Thus, the distribution tail for high k is thicker, making the nodepopulation much less hom*ogeneous than for the case of exponential distribu-tion. There would be many nodes with few links and a small number of nodeswith many links.

One property of a network is its average diameter, which is the minimumnumber of edges connecting any two nodes in the network, averaged over theset of all possible pairs5. Intuitively, the diameter of a network must have someimpact on its dynamics. For example, information takes longer to flow througha large-diameter network. One of the expected properties of an inhom*ogeneousnetwork that follows a power-law distribution that has been demonstrated bynumerical simulation is that destruction of a randomly chosen vertex has littlechance of modifying the diameter of the network, since most vertices are nothighly interconnected. Such networks are said to be ‘robust’ with respect to acci-dent. On the other hand, destruction of a well-chosen, strongly connected vertex– a hub – will significantly increase the diameter of the network. Such networks


5 It is also possible to use the maximum diameter, which is the minimum number of edges connectingthe most distant pair.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (208)

are said to be ‘vulnerable’ to sabotage attack. A hom*ogeneously distributedexponential network does not display these two characteristics.

Another general network property is the presence of a giant component, i.e.,a large sub-network in which a path connects any pair of vertices. The thresh-old size that the word ‘giant’ applies to is of little importance, since the giantcomponent phenomenon corresponds to a sudden change of the network phasefrom fluid to solid. This frank jump in the number of vertices in the largest connected component marks the phase transition into a network that includesa giant component. In fact, the existence of a giant component, a global feature,if there is one, also depends on local topology.

Local topology

For a given average connectivity (the total number of edges divided by the totalnumber of vertices), edges may be apportioned in extremely diverse ways. Inde-pendent of the type of overall distribution, edges may be distributed uniformlyamong vertices or display various degrees of local clustering. Figure 7.14 pres-ents the case of increasingly stronger clustering, from left to right. While theedge distribution is rather uniform in the left panel, it is locally clustered in thecenter panel, as would be typical of ‘small worlds’. In the right panel, the clus-tering is so extreme that the network has broken down into small fragmentsnot connected with each other, resulting in the loss of the connected giant com-












Figure 7.14 For the same number of vertices and edges, edge-per-vertex apportionment can varyfrom uniform to highly clustered. The example shown here, consisting of 4 vertices and 5 directededges, is easy to visualize, but too small to be realistic. Distribution is rather uniform on the leftand unequal in the center, with strong clustering around B and A. However, the graph is not frag-mented. In a much larger network, this would correspond to a ‘small world’ (see text for a moredetailed definition). On the right, excessive local clustering has fragmented the graph into threedisconnected parts. The largest connected component is A–B.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (209)

ponent. The formal criterion determining the presence of a ‘small world’ isreduced diameter, as in random networks, combined with strong local cluster-ing, as in regular networks. In spite of such strong local clustering, a ‘smallworld’ still includes a connected giant component.

A clique is a set of vertices that are all interconnected. Local clustering is eval-uated by the clustering coefficient, that is, by a number between 0 and 1, where1 corresponds to a true clique. The clustering coefficient of a vertex taken asthe reference is the ratio of the number of its actual connections to the numberof possible connections (Figure 7.15).

7.4.2 The interactome

Maps of interactions among yeast proteins, which for the moment are quiteincomplete and contain errors, generate undirected graphs whose connectivitydistribution obeys a power law. The most heavily connected vertices usually cor-respond to products of essential genes whose inactivation is lethal. Much morecomplete data must be obtained before any definite conclusions can be drawn.

7.4.3 The metabolome

The data concerning metabolic pathways and their some 800 metabolites arerelatively mature; it is therefore possible to draw some conclusions. If themetabolites are the vertices in an undirected graph and if the reactions con-necting these molecules – each catalyzed by an enzyme – are the edges, the con-


Figure 7.15 The cliquishness or clustering coefficient is a measure of local clustering. Takingnode R as the reference, the clustering coefficient is the number of edges connecting its immedi-ate neighbors a–f, yielding the maximum number of such edges. For N nodes, the maximum numberis [N(N-1)/2] for an undirected graph such as in this example, and [N(N-1)] for a directed graphsuch as shown in Figure 7.14. Here N = 5; therefore the maximum number of undirected edgesamong neighboring nodes is 10. In the left panel, there are no edges connecting the neighborsof R to each other, so the clustering coefficient is 0. All 10 possible edges are present in the rightpanel, so the coefficient is 1; therefore it is a clique. The center panel has 5 edges, so the coef-ficient is 5/10, and it is not a clique.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (210)

nectivity distribution of the metabolic network obeys a power law for which(according to the author) the exponent varies between −1.5 and −3. The resultsvary, particularly depending on whether or not very common small molecules,such as water and ATP, are included.

7.4.4 The genetic network

Data concerning the genetic network are obtained from classical genetic exper-iments, as well as from using the more exhaustive ChIP–Chip technique (Figure7.5). For the moment, the genetic networks of S. cerevisiae and E. coli havebeen sufficiently investigated to allow some observations.

The graph of genetic interactions is signed; that is, each edge bears an inter-action sign, positive for activation and negative for inhibition. It is also directed,for the molecular reasons discussed above. As a consequence, incoming and out-going connectivities will be considered separately.

The distribution of incoming interactions obeys an exponential law whoseexponent is −0.45 for S. cerevisiae and −1.2 for E. coli (Figure 7.16A). In prac-tice, the shallower slope observed for the yeast indicates that the maximumnumber of different regulatory proteins that may regulate the same gene ishigher than in bacteria. The average and maximum connectivities are, respec-tively, 2.3 and 13 for S. cerevisiae, and 2 and 6 for E. coli. This no doubt indicates the greater sophistication of the machinery regulating eukaryotic transcription.

Outgoing connectivity has no such limits; the total number of targets and theprotein concentration are its only molecular limits. Outgoing connectivity doesnot obey an exponential law, but approaches a power law (Figure 7.16B). Theexponent is around −1 for both organisms, indicating that the number of out-going connections kpk is distributed equally over k. This −1 value also corre-sponds to the phase transition of a generalized random graph. Averageconnectivity is 8.3 for S. cerevisiae and 3 for E. coli. It is among the essentialgenes (framed in Figure 7.13), therefore the most sensitive to disruption, thatthose richest in direct and indirect targets are found.

In any case, in these networks we observe a small maximum diameter (6 steps;see Figure 7.13), general fragmentation, and very strong local clustering (caseshown in Figure 7.14, right). The number of feedback circuits seems small, evenif it is greatly superior to what would be predicted for a random graph con-strained by the same empirical connectivity distributions. In E. coli, self-inhibition (unary negative circuits) widely predominates. In S. cerevisiae thereis a slight excess of positive circuits, nearly all of which are self-activations (Figure 7.13). Positive circuits are implicated in various differentiation pro-grams, of which there are more examples in the eukaryotic yeast than in E. coli,which is a non-sporulating bacterium. The overwhelming predominance of the


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (211)

shortest possible circuits could reflect the savings in both stimulus response timeand biosynthetic energy required at each step.

Finally, overall fragmentation of the genetic network could limit informationcrosstalk at the transcriptional level. It is also important to recall here that thenumber of feedback circuits would greatly increase and fragmentation woulddiminish, if the genetic network were thrust into the heart of the cell; that is, ifit were in contact with other molecular networks.

7.5 Modularity and dynamics of macromolecular networks

In the preceding section, we saw that the remarkable progress of post-genomicsleads to a map – now much closer to completeness – of the elements that con-stitute Life. This map should be used to understand the regulatory logic of living





ed g


Regulatory proteins






Incoming connectivityA


10 2 4 6 8 10 12 14


Regulated genes

Outgoing connectivityB



Figure 7.16 Connectivity of the genetic regulatory network of yeast. This network includes 500genes and 900 interactions, a small number of which are represented in Figure 7.13. (A) Incom-ing connectivity (semilog plot): the number of regulatory proteins per regulated gene follows anexponential distribution. (B) Outgoing connectivity (log–log plot): the number of regulated genesper regulatory protein approximates a power law distribution. Nil values have been discarded.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (212)

organisms. However, that will be possible only if we know how to read the map,a task beyond the grasp of the unequipped human mind. Two complementaryapproaches are of great importance in trying to interpret this map. First, it seemsthat the implications of regulatory logic will be revealed only when new ex-periments are undertaken that combine laboratory bench and computerapproaches. Second, it is important to shed light on the inherent modularity ofthe map in attempting to reduce the otherwise insoluble overall problem. Therelevance of this modularity is discussed below.

7.5.1 Modularity – why and how

The modularity gamble

Partitioning a molecular network into sub-networks is of interest only underthree conditions: First, such modules must be biologically relevant; for example,by underlining functionality. For instance, a module could include all the actorsinvolved in the response to a hormone. If this condition is not satisfied, parti-tioning is just a mathematical game of no biological interest.

The second condition is that it must be possible to attribute a characteristicdynamics to a module. For example, a negative circuit (see below) could generate homeostatic behavior. If this condition is not satisfied, partitioning will not help achieve the objective of understanding the functioning of thewhole.

The third condition is that it be possible for the modules to be self contained(or that when present as part of a larger network they retain their principalproperties, especially their dynamics). For example, the negative circuit, whichin isolation engenders homeostasis, must retain that property when placed in awider context. When this condition is satisfied, modularity permits the mergingof certain representations and facilitates comparison among organisms. Modularity loses much of its interest if this condition is not met.

Implementing modularity

It is possible to imagine two ways of partitioning. In both cases, the three con-ditions described above must be satisfied. Constructively, partitioning consistsin selecting and assembling a small number of vertices into a sub-network.Reductively, a mathematical criterion, based for example on local connectivity,is applied to the whole network, so as to fragment it into sub-networks.

Statistical analysis can demonstrate the interest of the modules detected, usingan analytic approach or numerical simulation to demonstrate that they are significantly over- or under-represented in a natural macromolecular network.


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (213)

Note, however, that there is not necessarily a correlation between the naturalrepresentation of a module and its biological relevance.

Until now, the constructive approach, in conjunction with generally rudi-mentary statistical analysis, has been used more frequently.

Properties of a module

Let us again use the example of the negative circuit in Figure 7.17. This smallmodule is characterized by its topology, e.g., a self-regulating vertex; by itsdynamic property, e.g., homeostasis; and by its biological properties, e.g., stableregulation or oscillation. Its behavior may be investigated by numerical simu-lation or by in vivo simulation (see Chapter 8). In the next section, these prop-erties will be examined for some examples of modules.


Figure 7.17 Main properties of two families of feedback regulatory circuits. The two families aredistinguished by the number of inhibitory interactions (negative interactions are represented bya square arrowhead) connecting a vertex to itself. The number of activating (positive) interactionspresent along the regulatory path is no more relevant than the total number of steps in the path;only the number of inhibitory steps counts. The positive circuit on the left consists of twoinhibitory interactions: A self-activates via B. If A is high, it will remain so, and B will remainlow, and vice-versa. The negative circuit in the center includes an inhibitory interaction: A self-inhibits. If A is high, it further self-inhibits, thus will diminish. If A is low, it self-inhibits less,thus will increase. The negative circuit on the right consists of three inhibitory interactions. Iftime delays are introduced for each step, and the products of A, B, and C are short-lived, thesystem may behave as an oscillator.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (214)

7.5.2 A sketch of module taxonomy

A few modules obtained constructively from directed graphs, which have beensubjected to various degrees of investigation, are gathered here, along with anexample of topology and some other properties, in particular, a probabledynamics.

Feedback circuits

A directed interaction pathway that is closed on itself constitutes a loop or feed-back circuit. Interactions can be activating or inhibiting. Two types of feedbackcircuits may be identified (Figure 7.17): ‘positive’ circuits, which include an evennumber of inhibitory interactions, and ‘negative’ circuits, which include an oddnumber of inhibitory interactions. This terminology is justified by the fact thata vertex has an activating (positive) effect on itself in a positive circuit and aninhibitory (negative) effect on itself in a negative circuit. These two types of circuits have very different dynamic and biological properties.

A positive circuit can contribute to differentiation (‘multi-stationarity’; i.e.,several possible stationary states), since if the concentration of the molecule thatcorresponds to node N increases, its formation will be further activated, whereasif its concentration diminishes, its formation will be less activated. Thus, as a first approximation, the dynamics tends toward either the maximum orminimum value, and any intermediate equilibrium is metastable. A number ofpositive circuits are found in natural genetic networks and their frequency seemsto increase in proportion to the complexity of the developmental program ofthe organism involved.

A negative circuit contributes to homeostasis (parameter stability), since ifthe concentration of a molecule that corresponds to node N increases, its for-mation will be further inhibited, whereas if the concentration of the moleculedecreases, its formation will be less inhibited; it therefore tends towards a stableequilibrium value. A negative circuit can also lead to more or less dampenedoscillation, according to how the parameters for the interactions are set and tothe half-lives of the molecules that correspond to the nodes. Intuitively, suchhalf-lives must be short in order to produce oscillation. Unary negative circuits(auto-inhibition) are considerably over-represented in E. coli.

Regulatory triangles (‘feedforward loops’)

A ‘triangle’ in directed graphs is also known as a feedforward loop, and con-sists of an input vertex (‘In’) that influences a second vertex. These two verticesjointly influence an output vertex (‘Out’). The feedforward loop is said to be‘coherent’ if the direct effect of the input vertex on the output vertex has the


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (215)

same sign (activating or inhibiting) as its net effect through the indirect path. Ifnot, the loop is said to be ‘incoherent’ (Figure 7.18). Each of these two circuitfamilies includes four possible topologies, with different dynamic and biologi-cal properties.

Coherent triangles of the type represented at the left of Figure 7.18 are foundto be over-represented in genetic networks. If activation of C requires simulta-neous activation of A and B (A and B), B must progressively accumulate underthe effect of A in order to cross its threshold, finally allowing activation of C.Thus, this triangle filters the transients (which do not leave time for B to accu-mulate), responds only to persistent stimulation (which does allow B to accu-mulate), and quickly shuts down when A ceases to activate B and C. Moregenerally, numerical simulations indicate that coherent triangles introduce adelay into the response when the signal goes either up or down.

The same approach suggests that incoherent triangles introduce accelerationinto the response when the signal goes either up or down. While the incoher-ent triangle represented on the right in Figure 7.18 has been observed, it is infrequent. It also includes an additional interaction (B self-inhibits). These char-acteristics make any prediction of dynamic behavior difficult, unless the modelis constrained by some prior biological knowledge.


Figure 7.18 Main properties of two families of regulatory triangles (‘feedforward loops’). Severaltopologies exist for each family. These two families are distinguished by the coherence of theaction of A (Input) on C (Output), whether this action is via B or direct. For example, on the left,A activates C directly (bottom arrow) and indirectly via B (upper arrows); the action of A on C iscoherent, and the resulting dynamics is rather easy to predict qualitatively. On the right, A inhibitsC directly (bottom arrow), and activates C indirectly (upper arrows); the action of A on C is inco-herent. To predict the resulting dynamics, some biological knowledge must be introduced (basedhere on some real cases, in which B also self-inhibits).

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (216)


A cascade is a chain of vertices that influence one another sequentially. For thecase in which each interaction introduces a non-negligible temporal delay (forexample, the time necessary for biosynthesis in a genetic network), the delayintroduced by the cascade is a function of its length, as well as of individualdelays (Figure 7.19). It is also interesting to note that cascades are often shortin microorganisms, for which a quick reaction to an external stimulus is essen-tial. In contrast, cascades in multicellular organisms are often long, therebyintroducing delays that the organism can use for its developmental program.This phenomenon is accentuated by the larger quantity of introns in multicel-lular organisms, which increase the time required for mRNA synthesis.

If several steps each introduce an amplification factor (for example, kinasecascades), the cascade permits strong amplification globally. If several interac-tions are cooperative, the bottom of the cascade responds in a quasi all-or-nonemanner.

Combinations of cascades and positive circuits

A developmental program consists of a series of irreversible steps, each of whichtakes a relatively precise length of time. One way to satisfy these constraints


Figure 7.19 Main properties of regulatory cascades. The number of steps in a cascade, as wellas the duration of each step, obviously determines the delay of the response to the initial stim-ulus. Note the predominance of short cascades in microbes and of long ones in multicellular organ-isms. This probably corresponds to the physiological requirements for rapid reaction in microbesand to long delays between successive developmental events in multicellular organisms.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (217)

would be to introduce a cascade that inserts a delay, followed by a positivecircuit that irreversibly locks the mechanism. A linear suite of such mechanismswould implement the series of steps that constitute the entire program (Figure7.20). Control points could be added to this simplified diagram, each of whichwould represent a prerequisite for the next key step. For example, attaining aminimum mass would be a condition for the next cell division.


A fan consists of a few upstream vertices (‘In’) that influence some downstreamvertices (‘Out’) under closure conditions. All the incoming influences of thedownstream vertices are in the fan as are, reciprocally, all the outgoing influ-ences of the upstream vertices. At present, we can say nothing about the generalcase of fan dynamics; we are discussing only single-output and single-inputmodules.

A single-output module is a set of vertices that jointly and exclusively regu-late a single output vertex (‘multigene regulation’), allowing, for example, fineregulation of genes by combining numerous inputs, each of which representsone aspect of the state of the cell. In known cases, the single-output module


Figure 7.20 Principal properties of some long regulatory cascades combined with positive-feed-back circuits. Each cascade, for example, A–C, introduces a long time delay, corresponding to prepa-ration of the next developmental step. This developmental event is then locked on by a positivecircuit, with no possibility of return (‘ratchet mechanism’), for example, self-activation of C. Successive developmental events in multicellular organisms therefore correspond to the successivetriggering of C, then F, then Z.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (218)

implements an ‘AND’ logical gate; that is, all the regulators must be presentand activating in order for the output to be activated (Figure 7.21).

A single-input module includes a vertex that regulates a set of exiting ver-tices with no other input (‘pleiotropic regulation’). Single-input modules havebeen found to be under-represented in genetic networks. In some cases it hasbeen possible to demonstrate that a single-input module permits time-stagedtriggering of exiting genes. It is only necessary that the dose of the single regu-lator increase over time and that the regulated genes respond to different dosesof the regulator. The most sensitive gene then triggers the early events in theglobal response to the common regulator, whereas the least sensitive gene trig-gers the late events (Figure 7.21).

7.6 Inference of regulatory networks

Inference consists in learning from a model of the relations among variables,starting from observations of those variables. The variables here are genes andproteins, and the objective is to infer maps of their molecular interactions, start-ing from post-genomic measurements.


Figure 7.21 Main properties of two sub-families of fans. Only the single-input (SIM) and single-output modules (SOM) are represented. The SIM may provide a temporal expression program, accord-ing to an activation threshold hierarchy for regulated genes. When the single regulator A rises orfalls, gene B, which is sensitive to the lowest threshold, will be activated first and deactivatedlast. Gene C, which is sensitive to the highest threshold, is the last activated and first deacti-vated. The SOM can provide an ‘AND’ logical gate. The single target, D, is triggered only after allA–C regulators are active.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (219)

7.6.1 The data

Post-genomic data obtained from kinetic experiments may be considered tocontain the information required for inferring the underlying network. However,two kinds of difficulties lie in the path of this inference. On one hand, experi-mentally measured parameters reveal intrinsic variability and dispersion, sincethe observations almost always concern a population, not an individual. On theother hand, experimental options, which are often constrained by practical con-siderations, engender extrinsic variability (due to measurement), as well as datawhose structure may not be well adapted to inference.

More generally, a typical set of post-genomic data concerns thousands of vari-ables, either genes or proteins. However, even in the best of circ*mstances, thisset includes only a few hundred experimental situations, thus a few hundrednumerical values for each variable. Under these conditions, in which the modelis under-determined by the experimental facts, one would expect the inferencemethod to propose a model which, in ‘attempting’ to account for the facts, con-tains some false correlations.

These data can be analyzed using a variety of methods capable of inferenceat various levels, such as clustering analysis, correlation analysis, and mutualinformation content. Abstract computational models serve as the basis for devel-oping these inference techniques. Such models are required in order to link the dynamic behavior of variables (trajectory, attractor) to a specific networktopology.

7.6.2 Models

In this reverse-engineering work, the choice of a modeling formalism is, ofcourse, crucial. Since modeling will be described in Chapter 8, only a few modelsused in inference will be mentioned here, without details.

The first models used were Boolean; that is, based on variables that couldtake on only one of two values. To infer a Boolean network of N potentiallycompletely connected genes theoretically requires measuring 2N pairs ofinputs/outputs, which is obviously inconceivable. Assuming that the genes donot have more than k inputs from other genes, the quantity of independentexperimental data required becomes proportional to 2k logN. In practice, algo-rithms proposed that are based on a Boolean model function correctly only withartificial data.

Bayesian models have sometimes been successful in real situations. In themore general framework of machine learning, they allow deciphering the prob-abilistic structure of dependence among observed variables. Variants have beenproposed whose training is more or less robust with respect to the paucity ofdata. For instance, it is possible to group variables with the same statistical


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (220)

behavior (e.g., genes having the same regulatory inputs) in order to reduce thespace of the models and parameters. This recent approach led to some predic-tions that have subsequently been validated at the laboratory bench. However,until now, nearly all attempts have taken a temporal series of N experimentalpoints to be N independent experimental situations. Temporal information istherefore lost in current approaches, except in the dynamic Bayesian framework.

7.6.3 Prior knowledge

The above-mentioned under-determination suggests that it would be useful tohave supplementary data available to constrain the model. Such data couldderive from prior knowledge of molecular interactions, or of the consensusbinding sequences. It is also sometimes possible to convert prior knowledgeobtained from higher organizational levels into constraints that are expressedin the same language as the model.

Constraining the model by using prior information concerning what is knownor plausible from the biological point of view probably remains the best toolfor tackling the curse of dimensionality! How to include this information in theinference process is the real art of the modeler.


Atlan H. (1999). La fin du ‘tout génétique’. INRA Éditions.Bonnet G., et al. (1999). Thermodynamic basis of the enhanced specificity of structured

DNA probes. Proc Natl Acad Sci USA 96: 6171–6176.Fields S., Song O. (1989). A novel genetic system to detect protein–protein interactions.

Nature 340: 240–246.Gavin A.-C., et al. (2002). Functional organization of the yeast proteome by systematic

analysis of protein complexes. Nature 415: 141–147.Guelzim N., et al. (2002). Topological and casual structure of the yeast genetic network.

Nature Genetics 31: 60–63.Ho Y., et al. (2002). Systematic identification of protein complexes in Saccharomyces

cerevisiae by mass spectrometry. Nature 415: 180–183.Kacser H., Burns J.A. (1973). The control of flux. Symp Soc Exp Biol 27: 65–104.Lee T.I., et al. (2002). Transcriptional regulatory networks in Saccharomyces cerevisiae.

Science 298: 799–804.Legrain P., et al. (2001). Protein–protein interaction maps: a lead towards cellular func-

tions. Trends Genet. 17: 346–352.Milo R., et al. (2002). Network motifs: simple building blocks of complex networks.

Science 298: 824–827.Perrin B.E., et al. (2003). Gene networks inference using dynamic Bayesian networks.

Bioinformatics 19: Suppl 2: II138–II148.


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (221)

Segal E., et al. (2003). Module networks: identifying regulatory modules and their con-dition-specific regulators from gene expression data. Nature Genetics 34: 166–176.

Thomas R., d’Ari R. (1990). Biological Feedback CRC Press, Boca Raton, FL.Zhu H., et al. (2001). Global analysis of protein activities using proteome chips. Science

293: 2101–2106.


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (222)

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (223)

8Simulation of biological processesin the genome context

The advent of the genomics era will strongly influence the field of modeling andsimulation in biology. By massively producing data at the molecular level,genomics effectively pulls the center of gravity of experimental biology towardsthe ground level of the molecule. It is tempting to try to build cells – and evenorganisms – from their genomes and related molecular data, in order to rapidlyand cheaply simulate their functions and dysfunctions. As we shall see, suchattempts are doomed to failure if they omit knowledge that originates from theanalysis of upper levels in living systems. One reason for the failure of the purelybottom-up approach is that it is based on the common misconception of thegenome as by far the major cellular information store. Another reason is that,despite all the advertized abundance of postgenomic data, models and theoriesin biology are still largely under-determined by the available facts, except incarefully defined and small domains. In the end, this cornucopia of genomicdata may be a boost to the field of modeling/simulation, not because it has filledthe gaps in our knowledge, but rather because, compared with the previous situation, the gaps are so much smaller that it has become reasonable to hopethat models will legitimately fill in some of these gaps, making simulations morefruitful than ever before.

One central problem in post-genomic biomedical research is to forge newtools and improve our capacity to anticipate a cellular or organismal pheno-type, starting from the data generated by high-throughput biology: genomicDNA sequences (genome), RNA (transcriptome) and protein (proteome) con-centrations, activities, localizations, and interactions. The typical approachtowards this goal has been to try to establish statistical correlations between agiven molecular polymorphism and an individual feature (Figure 8.1). However,these correlations have no validity outside the feature under scrutiny, and theydo not entail any causal link. In contrast, simulation demands that causal linksof general validity be established. The famous example of sickle cell anemia canserve to illustrate this point (Figure 8.2). The phenotype at the organism level

Bioinformatics: Genomics and post-genomics, Frédéric Dardel and François Képès© 2006 John Wiley & Sons, Ltd

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (224)

(symptoms) has been reduced to a genotypic cause in several steps. In the oppo-site direction to this triumph of reductionism, and on largely unknown ground,the goal could be to re-establish a causal tree strongly rooted in the moleculardata. In networks, straightforward causality is replaced by a ‘diluted’ causalitythat obviously constitutes a major difficulty in achieving this goal. It is, however,a highly desirable goal, since, in the long run, present users of statistical corre-lations (biotechnologists, clinicians, etc) would benefit more than anyone elsefrom this causal and more generic approach that cuts the tremendous costs ofbenchwork.


Polymorphismat molecular level

Susceptibility todisease or to medicine

Low propensity– A T G C C A –*

– A T G C G A – High propensity

Figure 8.1 Statistical correlation between individual variation at the molecular level and an indi-vidual phenotype. As an example, a single nucleotide polymorphism (SNP) on the genome has beencorrelated with the propensity of the individual for diabetes. This correlation does not entail anycausal link. A causality tree may later be built, either by costly and lengthy clinical or by labora-tory bench work, or sometimes by simulation.





Patient has anemia

Red blood cells are fragile

Hemoglobin forms filaments

Genomic mutation

Figure 8.2 Causal analysis. In this instance, the phenotype includes anemia (organism level)caused by the brittleness of the red blood cells (cell level), due to abnormal formation of fibersby the hemoglobin (protein level), a consequence of a nucleotide substitution in the gene thatencodes hemoglobin (genotype level). The genomicist receives abundant information, mainly atthe molecular level, and must therefore learn how to connect the causal links in the oppositedirection, on largely unknown ground.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (225)

8.1 Types of simulations

In biology, simulations have been employed in three distinct ways: in vitro, invivo, and in silicio. It is important to bear in mind the necessity of validatingthe results of any of these simulations by observations conducted in vivo.

The oldest way consists in cell-free assays, which reproduce a certain func-tion in the test tube (in vitro) with a mixture of chemicals and components pre-pared from live material. Although such assays are quite useful in determiningthe minimal set of components that supports the desired function, they will notbe discussed further, since they fall outside the scope of this book.

The most recent way consists in implementing small genetic circuits com-prising one, two, or three foreign genes in live cells (in vivo). The reporter ofthe network state is in all cases a fluorescent protein. Among the early attempts,the following functionalities have been tested: Homeostatic gene regulation bya self-inhibitory gene; a toggle switch made from two mutually inhibitory genes;an oscillator comprising three genes that inhibit each other in a circular per-mutation. Typically, this approach is preceded or accompanied by computa-tional simulation in order to predict the behavior and tune the parameters ofthe biological circuit, yielding the desired features.

Computational (in silicio) simulations are the main subject of this chapterand will be developed further below.

8.2 Prediction and explanation

Before discussing some simulation approaches, let us illustrate them with a fic-titious example that will take us from live cells all the way to simulation, whileemphasizing predictivity and control (Figure 8.3).

Differential gene expression has been measured on a microarray in tumorversus normal cells. Let us assume that it is possible to infer the underlyinggenetic network from such experimental data (see Chapter 7). A portion of theresulting network can be used as a model for running a simulation. This simu-lation may, for instance, predict that the network can exist in either of twostates, depending on the level of mitogenic stimulation of the main regulator‘REG’. The network state in a normal cell corresponds to its survival, while thatin a tumor cell leads to a ‘CDK’-mediated cell division. As this example implies,simulation can be useful to orient a costly laboratory bench or clinical experi-ment, or to attribute a probable function to a gene (annotation), and more gen-erally, to generate a prediction that can be tested on live material. For instance,how would the genetic network re-equilibrate once its REG′ node was inacti-vated by a drug (Figure 3, lower right)? Running a simulation might indicatethat in the presence of a drug that inactivates REG′ (but not REG″), division


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (226)

would be less favored than in an untreated tumor cell. Furthermore, apoptoticcell death (via APO), being no longer inhibited via REG′ (or via REG″), wouldbe directly triggered by REG, which is abundant in tumor cells. This predictioncan then be tested on the bench.

Besides its predictive capacities, simulation may have the explanatory powerto falsify or validate the coherence of a model. For instance, the growth of cellchains of the Anabaena microbe (Figure 8.4) has been simulated, using a rewrit-



normal APO






































Bio-chip Network Simulation

Figure 8.3 From biological sampling to simulation to control. Messenger RNAs (mRNAs) areextracted from normal or tumor cells. The ratio of their concentrations is measured on a biochip.The mRNAs transcribed from the ‘REG’ or ‘CDK’ genes are more abundant in the tumor cells thanin normal cells, giving a red color (not observable on this B&W picture) at the corresponding spot.The opposite holds true for the ‘APO’ gene, giving a green spot. The other spots are yellow, indi-cating that the concentrations of the corresponding genes are similar in both samples. These partialresults are compatible with the idea that ‘REGulator’ encodes a protein that activates the expres-sion of the ‘Cell Division Kinase’ gene and inhibits that of the ‘APOptosis’ gene. By inference, theresults obtained with the whole biochip may allow us to postulate a portion of a genetic network,such as the one shown on top, with REG′ and REG″ encoding intermediate regulators (→, activa-tion; �, inhibition). In this model, activation of REG by a mitogenic agent leads to hyperactiva-tion of CDK, directly and via REG′. This activation results in cell division and tumor proliferation(bottom). In the absence of a mitogenic agent, division and apoptosis remain balanced and thecell survives without dividing (top). This equilibrium can be studied through simulation, e.g., byproviding measured or calculated coefficients ‘c’ to the arrows that link genes (top). Similarly, inthe presence of a drug that inactivates REG′, the simulation may predict in which direction thenetwork will re-equilibrate. In this simple example, it can readily be seen that REG′ inactivationopposes division and relieves apoptosis inhibition (bottom). The prediction is that the tumor cellwill therefore be killed by this drug. Please note that the simulation could have been based on anetwork model provided by a classical approach or by theoretical considerations, just as well asby inference from transcriptomic data (like here).

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (227)

ing approach called Lindenmayer’s systems. The first modelling attemptincluded only an inhibitor of the differentiation into cells called heterocysts. Itaccounted well for the average distance of 10 undifferentiated cells between twoheterocysts, but was hypersensitive to the actual parameter values. Addition ofan activator coupled to the inhibitor by a specific regulatory circuit (Figure 8.4)made the model more robust (i.e., no longer critically dependent on the param-eter values). Independently, molecular biological studies later proved this modelessentially correct. In essence, simulation was used as an investigative tool thatdemonstrated the incoherence of the first model. The expert consequently pro-posed the minimal organization that would re-establish the explanatory coher-ence of the model, as verified by running a simulation. The molecular structurethat implemented the proposed organization was later discovered at the labo-ratory bench.

8.3 Simulation of molecular networks

The dynamic implications of the underlying logic of regulatory and metabolicnetworks cannot be deduced solely from laboratory experiments, in particular,because the molecular components are entangled in a complex web of interac-tions. Increasingly, formal models and computational simulations are requiredin conjunction with laboratory bench studies. This section reviews the for-malisms that are commonly used to describe and study regulatory and meta-bolic networks. Although it is very interesting to simulate both types ofnetworks in an integrated manner, which often requires the use of hybrid for-malisms, here we will address one formalism at a time.



Activator Inhibitor

Figure 8.4 Model that explains the relative constancy of the distance between two successiveheterocysts in a chain of Anabaena cells. This microorganism forms chains of cells (bright, wider)and differentiates into heterocysts (dark, narrower), present about every 10 cells. The simulationindicates that through the action of a couple of genes, it is possible to account for the relativeconstancy of the distance separating two successive heterocysts, with no critical dependence onthe parameter values. It is both necessary and sufficient to assume that the activator induces theproduction of the inhibitor and of itself, while the inhibitor represses the production of the activator. Simulation demonstrates the coherence of the model.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (228)

8.3.1 Graphs and their derivatives

A graph is a set of vertices and edges connecting some pairs of vertices. It iscustomary at all scales of biology to use graphs to represent interaction maps.At the supramolecular level, they have been used, for instance, to representprotein interaction maps, metabolic charts and genetic networks. At the supra-cellular level, they have traditionally been used to represent neural and immuno-logical networks. One potential problem with pure graphs is that they describea static relational topology, while data are sometimes more informative thanjust the presence or absence of a relation between vertices. Starting with graphs,it is possible to add various types of conditional, directional, and spatiotempo-ral information to the relations between vertices. For example:

– Biological networks may be directed, when the relation connects twoobjects that play different roles (e.g., transcriptional regulator/target) orundirected, in other cases (e.g., protein interactions).

– Vertices may be of the stoichiometric or catalytic type. Stoichiometric ver-tices describe resources that are consumed (e.g., a metabolic graph in whichvertices denote metabolites). Catalytic vertices are recycled to the identi-cal form (e.g., a metabolic chart in which vertices denote enzymes, or atranscriptional regulator/target map, in which the gene is still there afterits product has activated – or has been regulated by – another gene).

– A function may be assigned to each edge that formally describes the inter-relation between the vertices it connects, depending on the availability ofspecific knowledge about the interaction they represent. The minimal infor-mation is whether or not there is an edge. When available, more accurateinformation may be supplied in the form of a sign (activation or inhibi-tion), or an amplification factor (how much more or less of this mRNA isproduced per hour from the target gene when the regulator is present).Notions of space and time may also be introduced in the edge function,allowing more dynamic representation of the phenomena captured by thegraph (e.g., when enough of the transcription factor has accumulated inthe head of the embryo, that gene will be turned on, after a certain delay).Predicates may be embedded within the edge function, allowing somerefinements (e.g., if the A gene is on and galactose is present, then the Bgene is turned off). Such refined edge-assigned functions effectively provideinformation relevant to the dynamics of the graph, thereby allowing it toescape its inherent limitation of depicting a static topology.

– Graphs are not suitable for expressing interactions that involve more thantwo partners in an obligate fashion (e.g., three proteins assemble into a


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (229)

stable complex, provided that they are all present simultaneously; twoalone do not form a stable complex). The formalism of hypergraphs allowsexpression of non-binary interactions, and could be used in the abovecontext. If computational efficiency requires using graphs, at the cost of alimited loss of information, a n-ary complex can be broken down into aset of all possible binary interactions between subcomponents (a ‘clique’,see Chapter 7).

8.3.2 Boolean modeling

An idealized model based on elementary mechanisms may sometimes capturethe essence of a complex behavior. In a Boolean model, each gene may receiveone or several inputs from other genes or from itself. Assuming a sigmoidal(highly cooperative) relation between input and output, a gene may be con-sidered as a first approximation to be either on (1, transcribed) or off (0, untranscribed) (Figure 8.5). Time takes discrete values and all gene states are simultaneously computed at each time point. The output at time t + 1 is calculated from the input at time t according to Boolean functions.

Consider a simple Boolean network with three genes, each receiving inputsfrom the other two (Figure 8.6(a)). Gene A is an AND gate, meaning that genesB and C must both be active before A is activated at the next time point. GenesB and C are OR gates, meaning that they will be activated if one or the otherof their inputs is on. With three genes that may take two values, the networkcan assume eight possible states, from (000) to (111). Reading from left to right,the table in Figure 8.6(b) shows, for each current state at time t, the next stateof the genes, at time t + 1. For instance, starting from state (001) at t = 0 (secondline), the system will move to state (010) at the next tick of the clock, at t = 1.




















Figure 8.5 From continuous to discrete. Assume that the output is some sigmoidal function ofthe input (triangles), due to cooperative intermolecular interactions. The Boolean simplificationreplaces this curve with a step function (squares). Only two output states remain, 0 and 7.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (230)

This table therefore provides state transitions. Starting in one state, over time,the system will flow through some sequence of states. Such a sequence consti-tutes a trajectory. Given a finite number of states (eight here), the system willnecessarily reach a state it has previously encountered. From that moment on,it will loop forever through the same cycle of states, called for this reason anattractor. This cycle of states may be limited to one single state, in which caseit is called a point attractor. The collection of trajectories flowing into an attrac-tor constitutes its basin of attraction. For example, we have seen that (001)moves to (010), and the third line indicates that (010) moves to (001). There-fore, the system oscillates between these two states which, together constitute astate cycle (Figure 8.6(c)). Its basin of attraction is limited to the state cycle,since no other state flows into either of those two. This is in contrast to state(111), a point attractor with a basin of attraction consisting of five states, includ-ing itself. For instance, state (101) moves to (011), which in turn moves to thesteady state (111).

Boolean modeling has been used for network inference (see Chapter 7). It hasalso been used to study the global dynamic properties of large-scale regulatorysystems, in particular genetic networks, given local rules that bear, for instance,




Timet t+1

0 0 0 0 0 00 0 1 0 1 00 1 0 0 0 10 1 1 1 1 11 0 0 0 1 11 0 1 0 1 11 1 0 0 1 11 1 1 1 1 1

A B C A B CGene




111 011 101




(a) (b)

Figure 8.6 A small Boolean network. (a) The wiring diagram in a Boolean network with threegenes (A, B and C), each an input to the other two. (b) The Boolean rules for the diagram shownin (a), assuming that gene A represents an AND gate, while genes B and C each represent an ORgate. Given three binary elements, there are eight (23) possible states at any given time t. Thistable shows the successor state at time t + 1 for each state at time t. (c) The state transitiongraph of the Boolean network depicted in (a) and (b). Each triplet of digits corresponds to a statefor genes A–C, from left to right. State transition to a successor state is shown by an arrow. Threestate cycles are observed, a point attractor (000), a state cycle (001 and 010), and a basin ofattraction centered on (111).

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (231)

on the average degree of connection between genes. It is efficient, even for largegenetic networks, at the expense of greatly simplifying assumptions regardingthe absence of intermediate gene expression levels.

8.3.3 Generalized logical modeling

Transcriptomic data generally do not show extreme gene expression values, butrather intermediate ones. Although this observation may often reflect a mixturein varying proportions of cells that are each in different extreme states, carefulstudies conducted on small scales have suggested that at least some genes areexpressed at more than two levels. More importantly, transcription factors likelyhave different thresholds for their different target genes. The generalized logicalmethod in part corrects this problem by allowing logical variables to assumeseveral discrete values. Here, a variable is an abstraction for the cellular con-centration of one transcription factor. If a transcription factor encoded by geneA influences k genes, each with a different threshold, then the logical variablefor A can take at most k + 1 values, one for each threshold and 0 if no thresh-old is crossed. In practice, A will have only a limited number of significantthresholds, and consequently the number of values A can take will often besmaller than k. State transitions are not necessarily synchronous for this for-malism. Indeed, a synchronous step would sometimes entail jumping severalthresholds at once, which cannot occur in vivo because the processes are con-tinuous (e.g., protein accumulation). In logical networks, transitions are mademore realistic by being desynchronized, i.e., by jumping one threshold at a time.Furthermore, time delays, such as those arising from biosynthetic steps, can betaken into account.

Consider a regulatory graph consisting of two genes A and B that encodetranscription factors, where A activates B and itself and B inhibits A (Figure8.7(a)). Thus, A has two output links, corresponding to two thresholds in themost general case, so that A can take a value among 0, 1, and 2, while B iseither 0 or 1. The asynchronous state graph in Figure 8.7(b) indicates possibletrajectories in the space defined by all possible states for A and B. This allowscomputation of the next possible states. Note that this asynchronous state graphis only one among all possible state graphs and parameter sets that are com-patible with the regulatory graph and associated knowledge, viewed in a staticmanner. Finally, by applying temporal logic, models that are consistent with thetemporal properties of the system can be automatically extracted from theremaining set of all graphs that were compatible with the static view.

Generalized logical networks have been used to model various small regula-tory systems, including developmental networks and bacterial genetic switches.In the case of developmental networks, it has been possible to manually intro-duce notions of compartmentalization. This approach seems to be an efficient


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (232)

compromise between the wild simplifications of the Boolean model and theexcessive parameter-dependence of the differential approaches. Moreover, itoffers the possibility to exhaustively verify the temporal properties of a system,by taking advantage of the whole body of formal methods from computerscience.

8.3.4 Petri nets

A Petri net typically allows the description and modeling of concurrent systems.Although so far they have mostly been used to model technological systems (seatreservations, communication protocols, etc.), they have also been proposed todescribe and model biological networks, such as metabolic pathways or geneticnetworks. A Petri net consists of places, transitions, and arcs. A place containstokens that may flow through arcs according to some general rules. An arc con-nects a place to a transition, and vice versa. A transition comprises incomingand outgoing arcs that connect to places (Figure 8.8). When a transition is trig-


(a) (b)


Mucus production



2+1 - 1

0 1 2

01 11 21

00 10 20



Figure 8.7 A small logical network. (a) The regulatory interactions for mucus production in theopportunistic pathogen Pseudomonas aeruginosa. Two genes, encoding an activator (A) and aninhibitor (B) of mucus production, are considered. Each edge in the graph is labeled with the ranknumber of the threshold, followed by the sign of its regulatory influence (−, inhibition; +, acti-vation). Given parameters (not shown here), the dynamics may be deduced. (b) The asynchronousstate graph. This graph is one among several graphs that would fulfill the constraints based onbiological knowledge or hypotheses. A can take any value among {0, 1, 2}, and B among {0, 1}.Thresholds are represented by dashed lines, and transitions by arrows. The graph shows two steadystates, one for A = 0, and one cycle: 11 → 10 → 20 → 21 → 11.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (233)

gered, a token is taken from each input place and one is added to each outputplace through the corresponding arcs. This is used to represent reactions thatconsume their substrates and produce new elements. A ‘test’ arc checks for thepresence of a token in its source place but does not consume it. Therefore, itmay be used to represent enzyme activity or gene regulation, since the enzymeor the gene is not consumed by the reaction.

Some extensions to the classical Petri net have become popular for biologi-cal applications, including the hybrid Petri net, which has two kinds of placesand two kinds of transitions, discrete and continuous. The discrete places andtransitions are defined above. A continuous place holds a real positive number.A continuous transition continuously fires at a rate determined by parametersassigned to the transitions in hybrid Petri nets, or to the places in hybrid dynamicnets, or to both, in hybrid functional Petri nets.

Petri nets have been employed to model a variety of small biological systems,including signaling and metabolic pathways and regulatory gene networks. It isa very natural formalism for stoichiometric networks, such as metabolic path-ways, and has been adapted and extended to handle catalytic networks, suchas regulatory systems. It presents the visual inconvenience of requiring a largenumber of symbols and links to represent a small network (Figure 8.8).

8.3.5 Bayesian networks

In Bayesian formalism, a chart of regulatory interactions is represented by adirected acyclic graph, i.e., a graph with oriented edges deprived of circuits. In this graph, a vertex corresponds to a molecular entity, such as a gene or aprotein, bearing a random variable representing the gene expression level orprotein concentration (Figure 8.9(a)). A conditional probability distribution isdefined for the variable of each vertex, given the variables of its direct inputs inthe directed graph (Figure 8.9(b)). A joint probability distribution is finallydefined from all conditional distributions (Figure 8.9(c)).




Figure 8.8 A Petri net representation of a set of regulatory interactions. Circles denote placesidentified by a letter, black rectangles are transitions and arrows are arcs. Only discrete elementsare shown.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (234)

As such, this formalism allows propagation of information within the model.Moreover, when data are available, it is possible to apply algorithms of statis-tical inference to estimate parameters of the conditional probability distribu-tions, and to identify plausible structures. In this vein, Bayesian modeling has been successfully used for small network inference of microbial transcrip-tional interactions (see Chapter 7). This formalism is interesting because it isstrongly anchored in statistics, and appears to be well-suited for handling noisydata. Furthermore, it can be used even under conditions of incomplete knowl-edge, and prior knowledge can be introduced. Bayesian formalism can beextended to time-dependent variables, thus allowing the inclusion of regulatorydynamics.

8.3.6 Ordinary differential equations

Using the widespread formalism of ordinary differential equations, molecularconcentrations are represented by time-dependent variables. Regulation ismodeled by expressing the rate of synthesis of a molecule as a function of theconcentrations of all molecules, using rate equations of the form:

or of the more complete form:

where i is an integer that denotes the molecule under consideration (between 1and the number of molecules in the system); fi is a function (non-linear in thegeneral case); x is the vector storing all the molecular concentrations as real

dx dt f x P xi i i i= ( ) −; ,g

dx dt f xi i= ( ),




p(XA)p(XB)p(XC XA ,XB)p(XD XC)


(a) (b)

p(X) = p(XA) p(XB) p(XC XA ,XB) p(XD XC)

Figure 8.9 A Bayesian network. (a) In this directed acyclic graph, a vertex corresponds to amolecular entity such as a gene or a protein, and holds a random variable representing the geneexpression level or the protein concentration. (b) A conditional probability distribution is definedfor the variable of each vertex, given the variables of its direct inputs in the directed graph. (c)A joint probability distribution is finally defined from all conditional distributions.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (235)

positive numbers; P is a parameter. A degradation term of the form −gi xi, wheregi is a degradation constant, can be added to account for the exponential decayof the ith molecule, which may result from true degradation, but also from dilu-tion due to growth or diffusion. In this case, the equation represents a balancebetween synthesis and decay. Time delays can also be easily introduced into thefunction.

Analytical solutions are often impossible to reach for non-linear functions,and practitioners usually resort to numerical simulations that calculate approx-imate values for the variables at each successive time point. However, it is some-times possible to analyze specific features of the dynamic system known assteady states and limit cycles. The robustness of these steady states or limit cyclesto alteration of parameter (P) values may additionally be assessed using bifurcation analysis. Numerical simulations of non-linear ordinary differentialequations have been used to study systems such as the regulatory switch of bacteriophage l between host cell lysis and lysogenic growth, where the kineticparameters are few and have been measured very carefully. A general difficultywith ordinary differential equations is that they rely on accurate knowledge ofthe numerical parameters, and this knowledge is seldom available at present,although this situation can only improve. In the absence of proper experimen-tal measurements, e.g., in cell cycle models, parameter values can be chosen,using a manual or semi-automatic procedure, to fit the experimentally observedbehavior. However, mere fitting does not guarantee that the parameters are rightor that the numerical model is relevant to the biological situation. Furthermore,predictions following fitting are unsafe.

One popular non-linear function that accounts for real cases of sigmoidalresponse curves is the Hill function:

for activation, and

for inhibition. The function H+ ranges from 0 when xi is null, to 1 when xi tendsto infinity. It equals 0.5 when xi = qi, hence qi is called the threshold of the reg-ulatory influence of the ith molecule on its targets (Figure 8.10). The Hill func-tion admits a parameter S that reflects the steepness of the curve at H = 0.5. AsS increases, so does the sigmoidicity of the curve, corresponding to increasedcooperativity between interacting molecules. Extreme sigmoidicity brings usback to a step function, as discussed in the paragraph on Boolean networks (Figure 8.5).

One way to circumvent the analytical difficulties encountered with non-lineardifferential equations is to approximate them with a series of linear differential

H H− += −1

H x S x xi i i i i+ ( ) = +( ); ;q q


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (236)

equations, yielding ‘piecewise-linear differential equation’ models. For instance,the sigmoidal relation depicted in Figure 10 could be approximated by a stepfunction. Often, this approximation does not change the qualitative propertiesof the solutions. Thus, piecewise-linear differential equation models standbetween non-linear ones and logical models and have the advantage of stronglyconstraining local behavior in the phase space. They can be drawn even closerto logical models through qualitative analysis, at the expense of scalability.Qualitative analysis spots transitions that join different qualitative statesthrough trajectories: a state transition graph can thus be elaborated, akin to astate graph in the generalized logical formalism. This approach is adapted tothe usual lack of quantitative knowledge about regulatory mechanisms, but itcannot be readily scaled up, because qualitative constraints are often not avail-able in sufficient number or strength.

8.3.7 Partial differential equations

So far, spatial aspects have not been considered, except in a rather superficialway, in a few applications. In all other cases, spatial hom*ogeneity was assumed,or the formalism could not handle spatial aspects. However, simple considera-tions of how a cell functions, even a prokaryotic cell lacking any internal mem-branes, tell us that this assumption is wrong within a cell, not to mention thecase of multicellular organisms.

Consider a set of n cells arranged in a row. Each cell is identical to all theothers, and within each cell c (1 < c < n), gene expression is ruled by an iden-tical rate equation such as that shown in the previous section, using some func-



00 θi


H+ 0.5

Figure 8.10 Hill activation function H+. qi is the threshold of the regulatory influence of the ith

molecule on its targets, for which H = 0.5. S > 0 is the steepness parameter. For S > 1, Hill curvesshow a sigmoidal shape, as shown on this graph. As S increases, so does the sigmoidicity of thecurve.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (237)

tion f. Now consider a vector xc(t) holding the time-dependent concentration ofmolecules in cell c. Between two adjacent cells c and c + 1, diffusion of mole-cule i occurs proportionally to a diffusion constant di and to the concentrationdifference xi

c+1 − xic. When n is large enough, the integer c is replaced by a real

number c between 0 and C. Hence the partial differential equation:

This type of equation, and its close relative, the reaction–diffusion equation,have been extremely popular in studies of morphogenesis and pattern forma-tion, especially in their two-molecule version, one activator and one inhibitor.Such approaches may be extended to cope with higher spatial dimensionality.However, the predictions made on the basis of partial differential and reac-tion–diffusion equations are generally sensitive to parameter values, boundaryconditions, and domain shape. This is in stark contrast to the relative robust-ness of the developmental processes observed in vivo.

8.3.8 Stochastic equations

Bacteria may contain as few as 10 molecules of a given transcription factor, orone molecule of a given mRNA. It is thus questionable to assume, as has beendone so far with differential approaches, that molecular concentrations varycontinuously. It is equally questionable to neglect fluctuations (internal noise)in the timing of molecular processes and assume a perfect determinism, i.e., thattwo identical systems starting from the same initial states will follow an iden-tical trajectory. Accordingly, regulatory systems have also been modeled in astochastic fashion, to account for the imperfect determinism, and in a discretefashion, to account for the small number of molecules. One possibility withrespect to the lack of determinism is to add a term to a rate equation thataccounts for the noise in the system (Langevin’s equation). Another possibilityis to simulate step-by-step the temporal evolution of the system. In this lattercase, stochasticity is introduced at the level of two variables, which representthe time interval between two successive steps, and the next reaction to occur(Gillespie’s algorithm). Care is taken that the value distributions of these vari-ables allow a large set of stochastic simulations to approximate, on average, thebehavior of a so-called master equation, not discussed here. In this way, themaster equation provides the average number of molecules and associated vari-ances, while each stochastic simulation represents one individual trajectory.

Stochastic simulations have been notably applied to the developmental choicemade by bacteriophage l between host cell lysis and lysogenic growth. An inter-esting outcome of the observed fluctuations is that stochasticity may be onegood way to account for phenotypic diversity, i.e., the fact that different indi-

∂ ∂ ∂ ∂x t f x d x ci i i i i= ( ) + ( )⋅ ⋅2 .


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (238)

viduals in an apparently hom*ogeneous population have different behaviors. Inpractice, the stochastic approach may yield realistic simulations, provided thatthe reaction mechanisms are known in great detail, at the cost of heavy com-putations, given the number of complex simulations to be run. When it is pos-sible to expand either the time or the space scale, the phenomenon under studymay be approximated by less costly deterministic models.

In conclusion, the choice among the formalisms discussed above must bebased on careful consideration of their shortcomings and their strong points,given the problem and data, and the available computer power. Finally, it maybe anticipated that the most successful approaches will involve the union ofcomputational and bench experiments.

8.4 Generic post-genomic simulators

8.4.1 State of the art

A small number of highly funded, highly publicized, projects in the U.S.A. andin Japan propose simulation frameworks or platforms. To save the cost of devel-oping new specific software for each model, these frameworks or platforms mayin principle be generically applied to a variety of problems. Although theyrevolve around metabolism, they each have specific properties. Some, like Bio-Drive, provide for extracellular signaling and signal transduction through thecell membrane. Electronic-Cell (E-Cell) starts from a small set of genes encod-ing the minimal house-keeping metabolic functions, and implements the notionsof energetic cost and protein degradation. Virtual-Cell (V-Cell) provides a computational framework that accommodates several formalisms, takes intoaccount cell geometry and includes the notion of transmembrane flux, using it,for example, to simulate a calcium wave in the neuron. However, all these com-putational simulators are deeply rooted in a similar philosophy. Their startingpoint is a list of a few molecular components and their initial concentrations,and a set of reactions among these components. Generally, at each time point,a few differential equations are integrated and the list of concentrations isupdated.

To illustrate one of the criticisms that can be made regarding all the existinggeneric simulators, let us consider the simple case of a cellular metabolon, i.e.,a complex of several enzymes that each catalyze one of the successive reactionsof the same metabolic chain (Figure 8.11). Each enzyme of this complex hasbeen purified separately by biochemists over the past decades, and their param-eters measured in the test tube. These parameters are now fed into the simula-tors (Figure 8.11(a)). However, even the aqueous cellular compartments arehighly organized. As has been proven in an increasing number of cases, cellsexploit local concentration effects to allow for a sufficiently rapid metabolic


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (239)

pace. In particular, a metabolon maintains an elevated local concentration ofthe successive products of the reactions, or even funnels them in a channel in afew proven cases (‘solid-state metabolism’; Figure 8.11(b)). At such high localconcentrations, the enzymes turn over at full speed and diffusion times become negligible. Taking these facts into account would allow us to reduce the numberof parameters, while obtaining fewer erroneous simulations. More generally, forthe purpose of a realistic simulation, it is often an inacceptable first approxi-mation to consider the cell as a stirred mini-reactor and to ignore the ubiqui-tous local concentration effects. Since the existing generic simulators useparameters obtained outside the organizing context of the cell in a purelybottom-up (from molecules up) approach, they have to introduce ad hoc para-


A > B(a)



Km, Vmax, D

A > B


B > C


Km, Vmax, D

C > D


Km, Vmax, D

B > C


C > D

C Dc3

Figure 8.11 Metabolic chain and metabolon. In this short metabolic chain (a), three successivechemical reactions transform molecule A into molecule D. These three reactions are catalyzed bythree enzymes. The first enzyme, named ‘A > B’, accelerates the transformation of substrate A intoproduct B. Product B serves as a substrate for enzyme ‘B > C’ which accelerates its transformationinto product C, etc., hence the notion of a ‘chain’. This chain is part of the whole metabolism ofthe organism; it is a metabolic chain. The parameters of each enzyme have been measured sepa-rately following purification (affinity Km, kinetic Vmax, and diffusion D). Proper fulfilment of thisset of chemical reactions could rely on either a), the chemical specificity of the interaction betweenan enzyme and its substrate (order based on thermodynamics), or (b) spatial isolation that wouldprevent unwanted interactions (order based on localization). It appears that these two mecha-nisms are simultaneously active to varying degrees. In the absence of a membrane border, howcan a metabolic chain be spatially isolated in an aqueous compartment? It suffices that the enzymesof this chain have a tendency to associate, either among themselves – dependent or not on thepresence of their substrate – or to form fibers (a certain class of the cell inner skeleton). Numer-ous cases of multi-enzymatic complexes that process their substrates/products efficiently havealready been described, including the glycolytic enzyme complex, a major and central metabolicpath. For a metabolon, the most relevant parameter may be the coefficients (c) that relate to thefluxes traversing it.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (240)

meters to fit the observations made on living cells. Hence, they lose any pre-dictive value, retaining little explanatory value.

One way to overcome these problems would be to rely preferentially on datagenerated from live cells rather than in vitro. Along these lines, the potential ofsome in vivo approaches has not yet been fully evaluated. This is typically thecase for metabolic flux measurements using either liquid chromatography or iso-topic labeling and nuclear magnetic resonance, complemented by calculationsbased on, for example, flux balance analysis.

8.4.2 Problems and outlook

In a back-and-forth loop between laboratory bench and computer experimen-tation, metabolic simulations require three steps: 1) set-up of stoichiometric andkinetic equations; 2) experimental determination of the distribution of fluxes ina stationary state; 3) simulation of flux dynamics. It is clear that the present generation of post-genomic and generic simulators fall short of allowing thisback-and-forth loop. We shall now briefly examine the ontogenic and episte-mologic obstacles to fruitful simulations.

One of the major challenges ahead of us is to make the best use of the massiveamount of molecular data generated by the tools of (post-)genomics. It is clearthat the abundance of such data has increased tremendously in the recent past.Yet a closer look reveals that, with the exception of genome sequencing in mostcases, the exhaustivity of such data is a ‘sales pitch’, and their quality is gener-ally poor, although slow improvements can be foreseen. To take just one simpleexample, the lack of resolution of transcriptomic data is such that roughlyspeaking, one cannot distinguish values such as 1.0 and 1.9, or 1.9 and 3.8. Incontrast, one can tell 1.0 from 3.8. Thus, the curve in Figure 8.12 should be





0 1 2



Figure 8.12 Low resolution of the numerical data generated using biochips. As a general rule,the resolution of these data is close to a factor of 2, i.e., it is not possible to contend here thatthe first value is different from the second, or the second from the third. However, the first valueis definitely different from the third one. This is just one example of the general problem thathigh-throughput biology produces non-exhaustive, non-quantitative data.

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (241)

honestly described as: ‘This concentration increases in two hours’, which is aqualitative statement. For a long time to come, the lack of exhaustivity and thepoor quality of the data will be a serious ontogenic obstacle to quantitative pre-dictions. Besides, it would be wasteful to convert the few available quantitativedata into qualitative results for the sake of format hom*ogeneity. It will there-fore be crucial to succeed in combining, in the same simulation, the use of quan-titative and qualitative results.

In terms of prediction, the difficulty is to attain sufficient accuracy, so thatthe prediction is useful or testable. The number of variables in a biologicalsystem of interest is such that even small errors in their values may prevent thegeneration of a useful prediction. If, however, the purely quantitative approachturns out to be deceptive, other approaches described briefly in the previous sections may sometimes permit making useful predictions from qualitative orsemi-quantitative results. Importantly, it often suffices to be able to give the evolutive trend of the final parameter, in order to determine the outcome.

Coming back to the example of the drug effect on the network (Figure 8.3,lower right), if apoptosis outweighs division in the tumor cell, cancer willregress, which is the single most important fact. This qualitative approach is common among biologists, as evidenced by the widespread use of ‘models’,little symbolic drawings, at the end of numerous primary publications. In otherterms, qualitative reasoning is fundamental for the elaboration of knowledge inbiology. More generally, and to cut down on costly laboratory benchwork, itwould probably be a good idea to organize and improve the synergy betweenconventional and computational experimentations. The Anabaena case (Figure8.4) pleads for such a synergy.

The temptation expressed in the introduction to this chapter; to build a celland an organism from their genomes and related molecular data, implicitly relieson the concept that all the cellular information resides in the genome. This viewpoint is not supported by facts, yet is widespread, thus creating an episte-mologic obstacle. An obvious counterexample is that two human cells endowedwith an identical genome may exhibit extremely different phenotypes (compare,for instance, a stomach cell and a neuron). It is thus clear that part of the cellular information is held in its genetic material, while another part is dis-tributed elsewhere in the cell. The genetic information is easier to identify, sinceit is contained mostly in one DNA molecule and is relatively accessible (in theform of the genome sequence). This may explain why the genetic part is sostrongly emphasized. Another explanation may be traced back to the wronglynamed ‘central dogma’ of molecular biology (Figure 7.7, left), according towhich a gene encodes a one-dimensional protein sequence and the encodedprotein folds into a unique three-dimensional structure to fulfill one function.Yet it has been known for a long time that several genes may contribute to theexpression of one phenotypic character, and that a gene may participate in theexpression of several characters. In addition, most proteins have the potential


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (242)

to fold into more than one structure. A famous case is that of the prion protein,whose ‘abnormal’ conformation is transmitted to the descendants of its cellu-lar host. This is a case of epigenetic heredity. The very fact that a stomach cell,by division, produces two stomach cells, is really another case, albeit non-pathological.

What is lacking in the scheme of the unidirectional flow of information toget closer to the epigenetic point of view (Figure 7.7, right)? Firstly, the notionof a network between the macromolecule and its function, which now resultsfrom the dynamics of several interacting molecules in the network, not fromjust one molecule. Secondly, the notion of feedback loops linking moleculartypes positioned at different levels in the flow of information, i.e., regulatoryproteins and DNA. The existence of such feedback loops tells us that a cell cantransmit not only the sequence of a gene to its offspring, but also its activitylevel. These observations suggest that we should discard the naïvety of a purelygenetic determinism, but certainly not replace it with an undeterminism that isnot warranted by macroscopic observations (think of the robust developmentalprocess of an animal). Instead, they suggest that we face up to an epigeneticdeterminism, more difficult to comprehend, which connects the constituents ofthe biological object under study in a complex web. Importantly, this scientificapproach must integrate knowledge that originates from the analysis of upperlevels in living material. Perhaps we should use the term ‘Epigenomics’, which,by analogy to ‘Epigenetics’, alludes to the bottom-up construction of biologicalobjects at increasing levels of integration, while refering to the genome insteadof the gene.


Arkin A., et al. (1998). Stochastic kinetic analysis of developmental pathway bifurca-tion in phage Lambda-infected Escherichia coli cells. Genetics 149: 1633–1648.

Becsksei A., Serrano L. (2000). Engineering stability in gene networks by autoregula-tion. Nature 405: 590–593.

Bernot G., et al. (2004). Application of formal methods to biological regulatory net-works: extending Thomas’ asynchronous logical approach with temporal logic. J Theor Biol 229: 339–347.

De Jong H. (2002). Modeling and simulation of genetic regulatory systems: a literaturereview. J Comput Biol 9: 67–103.

Doi A., et al. (2004). Constructing biological pathway models with hybrid functionalPetri nets. In Silico Biol 4: 271–291.

Elowitz M.B., Leibler S. (2000). A synthetic oscillatory network of transcriptional reg-ulators. Nature 403: 335–338.

Friedman N., et al. (2000). Using Bayesian networks to analyze expression data. J Comput Biol 7: 601–620.


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (243)

Gardner T.S., et al. (2000). Construction of a genetic toggle switch in Escherichia coli.Nature 403: 339–342.

Gillespie D.T. (1977). Exact stochastic simulation of coupled chemical reactions. J PhysChem 81: 2340–2361.

Glass L. (1975). Classification of biological networks by their qualitative dynamics. J Theor Biol 54: 85–107.

Hammel M., Prusinkiewicz P. (1996). Visualization of developmental processes by extru-sion in space-time. Proceedings of Graphics Interface ’96: 246–258.

Kauffman S.A. (1993). The origins of order: self-organization and selection in evolu-tion. Oxford University Press, New York.

Kyoda K.M., et al. (2000). Construction of a generalized simulator for multi-cellularorganisms and its application to SMAD signal transduction. Pacific Symposium onBiocomputing 4: 317–328.

Meinhardt H. (1982). Models of biological pattern formation. Academic Press, London.Murray J.D. (2003). Mathematical Biology I & II, 3rd edition. Springer, New York.Schaff J., et al. (1997). A general computational framework for modeling cellular struc-

ture and function. Biophysical J 73: 1135–1146.Segal E., et al. (2003). Module networks: identifying regulatory modules and their con-

dition-specific regulators from gene expression data. Nature Genetics 34: 166–176.Thomas R., D’Ari R. (1990). Biological Feedback. CRC Press, Boca Raton, FL.Tomita M., et al. (1999). E-Cell: software environment for whole-cell simulation.

Bioinformatics 15: 72–84.Turing A.M. (1951) The chemical basis of morphogenesis. Philos Transact Royal Soc

B237: 37–72.Tyson J.J. (1999). Models of cell cycle control in eukaryotes. J Biotechnol 71: 239–244.


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (244)

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (245)

a-helices 156–7, 158–61alternating sequences 101, 105alternative splicing 88amino acid distribution 107–12androsterone receptors

BLAST method 44, 47–50dynamic programming 35–6hom*ologies 27–8structure 26

annealing principle 165–6APO genes 214automatic sequencing 1–4

b-sheets 156–7, 158–61BACs see bacterial artificial chromosomesbacteria

chromosome copies 65–6diversity and plasticity 63–6gene evolution 75–81genomic phylogeny 81–3Haemophilus influenzae genome

19–20minimum gene set 72–4number of replicons 65pathogenicity islands 74replicon geometry 65replicon size 64synteny 70–2, 76–8therapeutic targets 74–5

bacterial artificial chromosomes (BACs)10–12, 18

Bayesian models 207–8, 221–2

bias 67BioDrive 226bioinformatics 176, 181–2BLAST method 43–50, 128block substitution matrices (BLOSUM)

30–3BLAST method 44, 46–7dynamic programming 38, 51insertions and deletions 33multiple alignments 51

Boolean models 207, 217–19bulges (RNA structures) 142–3

CAI see codon adaptation indexcarbodiimides 150cascades (macromolecular networks)

204–5cassette mechanisms 72causal analysis 211–12CDK genes 213–14cDNA see complementary DNACentre d’étude du polymorphisme

humain (CEPH) 10chemical probing 150–2chromatin immunoprecipitation

(ChIP–Chip) 179–81, 182, 192chromosomes 61–6classical structures 134–6clone cell lines 6–8CLUSTAL (cluster alignment) program

54co-occurrence of genes 79–81


Bioinformatics: Genomics and post-genomics, Frédéric Dardel and François Képès© 2006 John Wiley & Sons, Ltd

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (246)


coding levels 85coding phase analysis 114–16codon adaptation index (CAI) 116codons

biases 112–13usage tables 110–11, 113

coherent triangles 203collagenase 162comparative genomics 61–84

bacteria 63–6, 70–8, 81–3bias 67chromosomes 61–6co-occurrence of genes 79–81conservation of gene order 80–1CpG islands 67eukaryote genomes 61–3, 67, 73–4,

82–3Fisher’s hypothesis 77–8function prediction 80–1gene evolution 75–80genome properties 61–7genomic phylogeny 81–3hom*ologies 67–72, 81horizontal transfer 76, 77isochores 67metabolic pathway conservation

79–80minimum gene sets 72–4non-orthologous gene displacement 69orthologous genes 67–72, 75–8paralogous genes 67–8pathogenicity islands 74plasmids 63–5regulatory modifications 78replication 65–6selfish operons 77–8synteny 68, 70–2, 76–8therapeutic targets 74–5

comparisons 25–59BLAST method 43–50comparison matrices 28–38, 44, 46–7,

51confidence levels 46–50deletions and insertions 33, 38, 51–2diagonal strip method 39–41

dynamic programming 34–8, 53fast heuristic methods 38–46hom*ologies 27–8, 36–7, 43–50human androsterone receptors 26–8,

35–6, 44, 47–50k-tuples 40–5maximum parsimony method 56multiple alignments 50–8phylogenetic trees 53, 54–8profile alignment 52–5progressive grouping 55–6reverse genetics 50sensitivity/specificity 46–50

compensatory mutations 149complementary DNA (cDNA)

expressed sequence tags 20–2, 25, 128

microarrays 176–7concatenation 101, 105consensus sequences 13, 93conservation of gene order 80–1content searches 97context-dependent codon biases 112contigs 12–13, 18convergence (evolutionary) 67–8, 69core construction 163cosmids 10–12, 18CpG islands 67

ddNTP see dideoxyribonucleotidedegenerate patterns 100–1deletions 33

dynamic programming 38, 51–2multiple alignments 51–2

diagonal strip method 39–41dideoxyribonucleotide (ddNTP) method

1–4differential equations 222–5dinucleotides

phylogenetic analysis 150RNA structure 133–4, 139–42thermodynamic stability 139–42triple helices 152–3

divergence (evolutionary) 67–8, 69DNA-DNA hybridization 12

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (247)


duplication of genes 67–8, 69dynamic programming 34–8

multiple alignments 51–2, 53predicting folding 166–7protein structure 166–7

Electronic Cell (E-Cell) 226electrophoresis 170–1endogenomes 75enzymatic probing 150–2enzyme saturation 187enzyme–substrate interactions 183–90epigenetic variations 72epigenomics 230ESTs see expressed sequence tagseukaryotes

genome size and structure 61–3genomic phylogeny 82–3isochores 67minimum gene set 73–4transcription 87–8

exogenomes 75expressed sequence tags (ESTs) 20–2, 25,

128expression signals 87–91

codon biases 112–13, 116fuzzy patterns 93–4

fans (macromolecular networks)205–6

fast heuristic methods 38–46FASTA/FASTP 39–41

BLAST method 43–4sensitivity/specificity 46

feedback circuits 202feedforward loops 202–4finite state automata

pattern detection 98–100, 103–6phylogenetic trees 58regular expressions 103–6

Fisher’s hypothesis 77–8flavodoxin 159four-fluorophore technique 2–4fractionation 170–1fragmentation strategies 8–12

Freier–Turner rules 140–3function prediction 80–1functional hom*ologies 27–8, 67–8, 69

gel electrophoresis 170–1gene evolution 75–80generalist bacteria 64generalized logical models 219–20generic post-genomic simulators

226–30Généthon 10–11genetic

code 85–7, 110–11interactions 193–4maps 15marker identification 12networks 198–9

genome sequencing see sequencinggenomic phylogeny 81–3GenScan 128Gillespie’s algorithm 225

Haemophilus influenzae genome 19–20haploid bacteria 65–6hidden Markov processes 123–7hom*ologies

comparative genomics 67–72, 81comparisons 27–8, 36–7, 43–50functional 27–8, 67–8, 69protein structures 161–6sequence 67–8, 69

hom*opurine sequences 153hom*opyrimidine sequences 153Hoogsteen dinucleotides 152–3horizontal transfer 76, 77human androsterone receptors

BLAST method 44, 47–50dynamic programming 35–6hom*ologies 27–8structure 26

human genome libraries 10–12hybridization 12hydrophobicity moment 160hyperstable tetraloops 143–4, 153–4hypochromicity 139

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (248)


immunoprecipitation 172–3, 179–81in vitro/in vivo simulations 213incoherent triangles 203inference of regulatory networks 206–8insertions 33

dynamic programming 38, 51–2multiple alignments 51–2

Institute of Genome Research, The(TIGR) 19–20

integration of genetic maps 15interactomes 182–3, 197intermediate mapping 11internal loops 148International Human Genome sequencing

consortium 10–11intracellular localization 171–2inverted repeats 134–5isochores 67, 108iterative contig assembly 13

k-tuples 40–5Knuth-Morris-Pratt (KMP) method


Langevin’s equation 225large clones 18–19large-genome bacteria 64Lineweaver–Burk equation 187local hom*ologies 36–7logical models 219–20loops

protein structures 163RNA structures 142–3, 146, 148

macromolecular networks 182–206cascades 204–5distributive control 188–90dynamics 200–6enzyme saturation 187enzyme–substrate interactions 183–90fans 205–6feedback circuits 202genetic interactions 193–4genetic networks 198–9global topologies 195–6

inference of regulatory networks206–8

interactomes 197Lineweaver–Burk equation 187local topologies 196–7metabolic pathways 183–90metabolomes 184, 190, 197–8Michaelis–Menten equation 185–7,

188–9modularity 199–206protein–protein interactions 182–3regulation of cell metabolites 187–8regulatory protein–DNA interactions

190–3regulatory triangles 202–4topologies 193–9

Markov chains 117–20, 123–7maximal hom*ology segments 45maximum parsimony method 56maximum scoring pair (MSP) 45messenger RNA (mRNA)

hyperstable tetraloops 143proteomics 175simulations 213–14structure prediction 131–2, 143transcriptomics 176–9

metabolicchains 226–8pathways 79–80, 183–90

metabolomesenzyme–substrate interactions 184,

190topology of macromolecular networks

197–8metabolons 226–8methylation 89–90, 93Michaelis–Menten equation 185–7,

188–9minimal cost pathway 14minimum gene sets 72–4modeling DNA sequences 116–27

complex models 120–3hidden Markov processes 123–7higher order biases 118Markov chains 117–20, 123–7

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (249)


sequencing errors 123–7Viterbi algorithm 125–7see also simulations

modification signals 89–91molecular networks 215–26

Bayesian models 221–2Boolean models 217–19generalized logical models 219–20graphical methods 216–17ordinary differential equations 222–4partial differential equations 224–5Petri nets 220–1stochastic equations 225–6

mRNA see messenger RNAMSP see maximum scoring pairmulticapillary sequencers 5multiple alignments 50–8multiple loops 148mutual information 150

n-tuples 109–10nearest neighbor hypothesis 138Needleman & Wunsch algorithm 35–8non-deterministic automata 103–6non-hom*ologies 37non-orthologous gene displacement 69nonsense codons 86nucleotide base distribution 107–12Nussinov algorithm 145–7

oligonucleotide chips 177–8oligonucleotide primers

automatic sequencing 1, 3sequencing strategies 4, 6–8

open reading-frames (ORFs)comparative genomics 82–3genetic code 86–7macromolecular networks 191pattern detection 102sequencing 22, 25

optimal alignment 34–8, 51–2ordinary differential equations 222–4ORFs see open reading-framesorthologous genes 67–72, 75–8overlap identification 14

palindromic sitesDNA sites 92–3restriction enzymes 90RNA sites 96

PAM see probability of acceptablemutation

paralogous genes 67–8partial differential equations 224–5pathogenicity islands 74pattern detection 96–106

degenerate patterns 100–1finite state automata 98–100, 103–6protein structures 158–60regular expressions 101–6simple searches 97–8

PCR see polymerase chain reactionPetri nets 220–1phase analysis 114–16phase shifting 122–7phylogeny

genomic 81–3phylogenetic trees 53, 54–8profile alignment 53–5secondary structure validation 149–50

plasmids 63–5polymerase chain reaction (PCR) 15–16,

18–19post-genomics 169–82

bioinformatics 176, 181–2chromatin immunoprecipitation

179–81, 182, 192complementary DNA microarrays

176–7inference of regulatory networks

206–8information theory 169–70intracellular localization 171–2oligonucleotide chips 177–8protein–protein interactions 172–6proteomics 170–6reverse-transcription–polymerase chain

reaction 178–9serial analysis of gene expression 179simulations 226–30transcriptomics 176–82

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (250)


PredictProtein neural network 161probability of acceptable mutation

(PAM) matrices 30–3BLAST method 46dynamic programming 38, 51multiple alignments 51

PRODOM database 57profile alignment 52–5progressive grouping 55–6prokaryotes 87–8promoter sequences 87–8PROSITE 58, 103Protein Data Bank 161proteins

a-helices 156–7, 158–61amino acid distribution 107–12b-sheets 156–7, 158–61bioinformatics 176core construction 163covalent geometry 163difficulties with predicting structures

155–8dynamic programming 166–7efficiency and limits 161genetic interactions 193–4hom*ologies 161–6interactomes 182–3intracellular localization 171–2loop construction 163model refinement 163non-covalent interactions 164pattern recognition 158–60preparative methods 170–1protein chips 175protein–protein interactions 172–6,

182–3proteomics 170–6, 182–3regulatory protein–DNA interactions

190–3secondary structures 158–61sidechain budding 163statistical methods 160–1structure prediction 155–67systematic identification 175

thermodynamic stability 164–6two-hybrid assays 173–5

proteomics 170–6, 182–3pseudoknots 136–8, 150

random fragmentation 8–10REG genes 213–14regular expressions 101–6regulation of gene expression 89, 92–3regulatory

modifications 78protein–DNA interactions 190–3triangles 202–4

repetitive sequencesfinite state automata 105reconstruction 16–17regular expressions 101, 105

replication 65–6, 91replicons 61–6restriction enzymes

endonuclease sites 6–8fuzzy patterns 93palindromic sites 90segmentation after mapping 12

reverse genetics 50, 172reverse transcription 20–1reverse-Hoogsteen dinucleotides 152–3reverse-transcription–polymerase chain

reaction (RT–PCR) 178–9ribonucleases 150–2ribonucleic acid (RNA)

bioinformatics 181–2chemical/enzymatic probing 150–2chromatin immunoprecipitation

179–81, 182, 192classical structures 134–6complementary DNA microarrays

176–7dinucleotides 133–4, 139–42, 150,

152–3empirical measurements 138–42Freier–Turner rules 140–3hyperstable tetraloops 143–4,


Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (251)


limitations of predictive methods148–9

loops 142–3, 146, 148messenger RNA 131–2, 143, 176–82,

213–14molecular properties 132–4nearest neighbor hypothesis 138Nussinov algorithm 145–7oligonucleotide chips 177–8phylogenetic analysis 149–50pseudoknots 136–8, 150reverse-transcription–polymerase chain

reaction 178–9ribosomal RNA 131, 143secondary structures 134–52serial analysis of gene expression 179structure prediction 131–55tertiary structures 135, 152–5tetraloop-receptor interactions 153–4thermodynamic stability 138–49topologies 144–5transcriptomics 176–82transfer RNA 112–13, 131triple helices 152–3true knots 136–7validation of predicted structures

149–50Zuker algorithm 147–8

ribosomal RNA 131, 143RT–PCR see reverse-transcription–

polymerase chain reaction

SAGE see serial analysis of geneexpression

Sanger method 1–4secondary protein structures 158–61secondary RNA structures 134–52segmentation after mapping 10–12, 18self-splicing introns 155selfish operons 77–8sequence hom*ologies 67–8, 69sequenced fragment assembly method

12–13sequencing 1–23

amino acid distribution 107–12automatic 1–4base triplets 120–1coding levels 85codon biases 112–13comparison matrices 28–38, 44, 46–7,

51comparisons 25–59complementary DNA 20–2complex models 120–3complex sites 93degenerate patterns 100–1deletions and insertions 33, 38, 51–2DNA sites 91–5dynamic programming 34–8, 53errors 123–7expressed sequence tags 22expression signals 87–91, 93–4fast heuristic methods 38–46filling gaps 9–10, 14–16, 18–19finite state automata 98–100, 103–6fragmentation strategies 8–12genetic code 85–7, 110–11genetic information 85–106Haemophilus influenzae genome

19–20hom*ologies 27–8, 36–7, 43–50integration of genetic maps 15large clones 18–19modeling DNA sequences 116–20multiple alignments 50–8nucleotide base distribution 107–12overlap identification 14pattern detection methods 96–106phase shifting 122–7polymerase chain reaction 15–16,

18–19prediction using biases 113–16regular expressions 101–6repetitive sequences 16–17RNA sites 96search procedures 127–8segmentation after mapping 10–12, 18sequence assembly 12–18

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (252)


statistical methods 107–29strategies 4–8unclonables 17–18

serial analysis of gene expression (SAGE)179

serralysine 162shotgun method see random

fragmentationsickle cell anemia 211–12sidechain budding 163simulations 211–31

Bayesian models 221–2Boolean models 217–19causal analysis 211–12explanation 213–15generalized logical models 219–20graphical methods 216–17in vitro/in vivo 213molecular networks 215–26ordinary differential equations 222–4partial differential equations 224–5Petri nets 220–1post-genomics 226–30prediction 213–15problems and outlook 228–30stochastic equations 225–6transcriptomics 219types 213

single nucleotide polymorphisms (SNPs)211–12

small nuclear ribonucleo-proteins(SNURPs) 88

small-genome bacteria 64Smith & Waterman method 37–8SNPs see single nucleotide

polymorphismsSNURPs see small nuclear ribonucleo-

proteinsspecialist bacteria 64specialization islands 74speciation 67–8start codons 86statistical methods

amino acid distribution 107–12base triplets 110–11, 120–1

coding phase analysis 114–16codon biases 112–13complex models 120–3gene expression level 116hidden Markov processes 123–7higher order biases 118Markov chains 117–20, 123–7modeling DNA sequences 116–27nucleotide base distribution

107–12phase shifting 122–7prediction using biases 113–16search procedures 127–8sequencing 107–29sequencing errors 123–7Viterbi algorithm 125–7

stochastic equations 225–6structure prediction 131–67

chemical/enzymatic probing 150–2dinucleotides 133–4, 139–42proteins 155–67ribonucleic acid 131–55secondary protein structures 158–61secondary RNA structures 134–8tertiary RNA structures 135, 152–5thermodynamic stability 138–49validation 149–50

sum of the pairs scores 51–2, 53SWISSPROT database 48, 58synteny 68, 70–2, 76–8

terminal loops 148termination codons 86tertiary RNA structures 135, 152–5tetraloop-receptor interactions 153–4tetraloops 143–4therapeutic targets 74–5threading 166thymine 132–3TIGR see Institute of Genome Research,

The tiling pathways 11transcription 87–8, 93transcriptomics 176–82

bioinformatics 181–2cDNA microarrays 176–7

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (253)


chromatin immunoprecipitation179–81

expressed sequence tags 21generalized logical modeling 219oligonucleotide chips 177–8reverse-transcription–PCR 178–9serial analysis of gene expression

179transfer RNA (tRNA) 112–13, 131translation 89, 113triple helices 152–3tRNA see transfer RNAtrue knots 136–7two-hybrid assays 173–5

ultrasound 8unclonables 17–18universal primers 6, 8unweighted pairgroup method using the

arithmetic mean (UPGMA) 55–6uracil 132–3

V-Cell see Virtual Cellvector clone sites 6–8Virtual Cell (V-Cell) 226viruses 61–2Viterbi algorithm 125–7

Watson–Crick dinucleotidesphylogenetic analysis 150RNA structure 133–4,

139–42thermodynamic stability 139–42triple helices 152–3

wobble pairings 134, 150

yeast artificial chromosomes (YACs)10–12

yeasts 62–3, 73

Zuker algorithm 147–8

Index compiled by Neil Manley

Bioinformatics - Genomics and Post-Genomics - F. Dardel, F. Kepes (Wiley, 2006) WW - [PDF Document] (2024)
Top Articles
Latest Posts
Article information

Author: Zonia Mosciski DO

Last Updated:

Views: 5917

Rating: 4 / 5 (71 voted)

Reviews: 86% of readers found this page helpful

Author information

Name: Zonia Mosciski DO

Birthday: 1996-05-16

Address: Suite 228 919 Deana Ford, Lake Meridithberg, NE 60017-4257

Phone: +2613987384138

Job: Chief Retail Officer

Hobby: Tai chi, Dowsing, Poi, Letterboxing, Watching movies, Video gaming, Singing

Introduction: My name is Zonia Mosciski DO, I am a enchanting, joyous, lovely, successful, hilarious, tender, outstanding person who loves writing and wants to share my knowledge and understanding with you.