Bioinformatics:

Sequence, structure, and

databanks
'Atomics is a very intricate theorem and can be worked out with algebra
but you would want to take it by degrees because you might spend the
whole night proving a bit of it with rulers and cosines and similar other
instruments and then at the wind-up not believe what you had proved at
all.

'Now take a sheep', the Sergeant said. 'What is a sheep only millions of
little bits of sheepness whirling around and doing intricate convolutions
inside the sheep? What else is it but that?'
(from The Third Policeman, Flann O'Brien)
The Practical Approach Series

Related Practical Approach Series Titles

Protein-Ligand Interactions: structure and spectroscopy*
Protein-Ligand Interactions: hydrodynamic and calorimetry*
DNA Protein Interactions
RNA Protein Interactions
Protein Structure 2/e
DNA and Protein Sequence Analysis
Protein Structure Prediction
Antibody Engineering
Protein Engineering

* indicates a forthcoming title

Please see the Practical Approach series website at

http://www.oup.co.uk/pas
for full contents lists of all Practical Approach titles.
 Bioinformatics:
Sequence, structure,
and databanks

A Practical Approach

Edited by
D. Higgins
Department of Biochemistry,
University College, Cork,
Ireland
and
W. Taylor
Division of Mathematical Biology,
National Institute for Medical Research,
The Ridgeway, Mill Hill,
London NW7 1AA, UK

OXFORD
UNIVERSITY PRESS
OXPORD
UNIVERSITY PRESS
Great Clarendon Street, Oxford 0x2 6DP
Oxford University Press is a department of the University of Oxford.
It furthers the University's objective of excellence in research, scholarship,
and education by publishing worldwide in
Oxford New York
Athens Auckland Bangkok Bogota Buenos Aires Cape Town
Chennai Dar es Salaam Delhi Florence Hong Kong Istanbul Karachi
Kolkata Kuala Lumpur Madrid Melbourne Mexico City Mumbai Nairobi
Paris Sao Paulo Shanghai Singapore Taipei Tokyo Toronto Warsaw
with associated companies in Berlin Ibadan
Oxford is a registered trade mark of Oxford University Press
in the UK and in certain other countries
Published in the United States
by Oxford University Press Inc., New York
© Oxford University Press, 2000
The moral rights of the author have been asserted
Database right Oxford University Press (maker)
First published 2000
Reprinted 2001
All rights reserved. No part of this publication may be reproduced,
stored in a retrieval system, or transmitted, in any form or by any means,
without the prior permission in writing of Oxford University Press,
or as expressly permitted by law, or under terms agreed with the appropriate
reprographics rights organization. Enquiries concerning reproduction
outside the scope of the above should be sent to the Rights Department,
Oxford University Press, at the address above
You must not circulate this book in any other binding or cover
and you must impose this same condition on any acquirer
A catalogue record for this book is available from the British Library
Library of Congress Cataloging in Publication Data
(Data available)
ISBN 0 19 963791 1 (Hbk.)
ISBNO 19 963790 3 (Pbk.)
10 9 8 7 6 5 4 3 2 1
Typeset in Swift by Footnote Graphics, Warminster, Wilts
Printed in Great Britain on acid-free paper by
The Bath Press, Avon
 Preface

Bioinformatics an emerging field

In the early eighties, the word 'bioinformatics' was not widely used and what we
now know as bioinformatics, was carried out as something of a cottage industry.
Groups of researchers who otherwise worked on protein structures or molecular
evolution or who were heavily involved in DNA sequencing were forced,
through necessity, to devote some effort to computational aspects of their
subject. In some cases this effort was applied in a haphazard manner, but in
others people realized the immense potential of using computers to model and
analyse their data. This small band of biologists along with a handful of
interested computer scientists, mathematicians, crystallographers, and physical
scientists (in no particular order of priority or importance), formed the fledgling
bioinformatics community. It has been a unique feature of the field that the
most useful and exciting work has been carried out as collaborations between
researchers from these different disciplines.
By 1985, there was the first journal devoted (largely or partly) to the subject:
Computer Applications in the Biosciences. Bioinformatics articles tended to
dominate it and the name was changed to reflect this, a few years ago when it
was re-christened, simply, as Bioinformatics. By then, the EMBL sequence data
library in Heidelberg had been running for four years, followed closely by the US
based, GenBank. The first releases of the DNA sequence databases were sent out
as printed booklets as well as on computer tapes. It was routine to simply dump
the tape contents to a printer anyway as computer disk space in those days was
expensive. This practice became pointless and impossible by 1985, due to the
speed with which DNA sequence data were accumulating.
During the 1990s, the entire field of bioinformatics was transformed, almost
beyond recognition by a series of developments. Firstly, the internet became the
standard computer network world wide. Now, all new analyses, services, data
sets, etc. could be made available to researchers across the world by a simple
annoucement to a bulletin board/newsgroup and the setting up of a few pages
on the World Wide Web (WWW). Secondly, advances in sequencing technology
have made it almost routine to think in terms of sequencing the entire genome
of organisms of interest. The generation of genome data is a completely

v
PREFACE

computer-dependent task; the interpretation is impossible without computers

and to access the data you need to use a computer. Bioinformatics has come of
age.

Sequence analysis and searching

Since the first efforts of Gilbert and Sanger, the DNA sequence databases have
been doubling in size (numbers of nucleotides or sequences) every 18 months or
so. This trend continues unabated. This forced the development of systems of
software and mathematical techniques for managing and searching these
collections. Earlier, the main labs generating the sequence data in the first place
had been forced to develop software to help assemble and manage their own
data. The famous Staden package came from work by Roger Staden in the LMB in
Cambridge (UK) to assemble and analyse data from the early DNA sequencing
work in the laboratory of Fred Sanger.
The sheer volume of data made it hard to find sequences of interest in each
release of the sequence databases. The data were distributed as collections of
flat files, each of which contained some textual information (the annotation)
such as organism name and keywords, as well as the DNA sequence. The main
way of searching for sequences of interest was to use a string-matching program
or to browse a printout of some annotation by hand. This forced the develop-
ment of relational database management systems in the main database centres
but the databases continued to be delivered as flat files. One important early
system, that is still in use, for browsing and searching the databases, was ACNUC,
from Manolo Gouy and colleagues in Lyon, France. This was developed in the
mid-eighties and allowed fully relational searching and browsing of the data
base annotation. SRS is a more recent development and is described fully in
Chapter 10 of this volume.
A second problem with data base size was the time and computational effort
required to search the sequences themselves for similarity with a search
sequence. The mathematical background to this problem had been worked on
over the 1970s by a small group of mathematicians and the gold standard
method was the well-known Smith and Waterman algorithm, developed by
Michael Waterman (a mathematician) and Temple Smith (a physicist). The snag
was that computer time was scarce and expensive and it could take hours on a
large mainframe to carry out a typical search. In 1985, the situation changed
dramatically with the advent of the FASTA program. FASTA was developed by
David Lipman and Bill Pearson (both biologists in the US). It was based on an
earlier method by John Wilbur and Lipman which was in turn based on an
earlier paper by two Frenchmen (Dumas and Ninio) who showed how to use
standard techniques from computer science (linked lists and hashing) to quickly
compare chunks of sequences. FASTA caused a revolution. It was cheap (basically
free), fast (typical searches took just a few minutes), and ran on the newly
available PCs (personal computers). Now, biologists everywhere could do their
own searches and do them as often as they liked. It became standard practice, in

vi
 PREFACE

laboratories all over the word, to discover the function of newly sequenced
genes by carrying out FASTA searches of databases of characterized proteins.
Fortunately, by this time the databases were just big enough to give some chance
of finding a similar sequence in a search with a randomly chose gene. Sadly, the
chances were small initially, but by the early nineties they had risen to 1 in 3 and
now are well over 50%.
By 1990, even FASTA was too slow for some types of search to be carried out
routinely, but this was alleviated by the development of faster and faster
workstations. A parallel development was the use of specialist hardware such as
super-computers or massively parallel computers. These allowed Smith and
Waterman searches to be carried out in seconds and one very successful service
was provided by John Collins and Andrew Coulson in Edinburgh, UK. The snag
with these developments was the sheet cost of these specialist computers and
the great skill required to write the computer code so networks were important.
If you could not afford a big fast box of specialized chips, you might know
someone who would allow you to use theirs and you could log on to it using a
computer network.
In 1990, a new program called BLAST appeared. It was written by a collection
of biologists, mathematicians and computer scientists, mainly at the new NCBI,
in Washington DC, USA. It filled a similar niche to the FASTA program but was
an order of magnitude faster for many types of search. It also featured the use of
a probability calculation in order to help rank the importance of the sequences
that were hit in the search (see Chapter 8 for some details). Probability calcula-
tions are now very important in many areas of bioinformatics (such as hidden
Markov models; see chapter 4).

Protein structure analysis and prediction

Protein structure plays a central role in our understanding and use of sequence
data. A knowledge of the protein structure behind the sequences often makes
clear what mutational constraints are imposed on each position in the sequence
and can therefore aid in the multiple alignment of sequences (Chapters 1, 3, and
6) and the interpretaion of sequence patterns (Chapter 7). While computational
methods have been developed for comparing sequences with sequences (which,
as we have already seen, are critical in databank searching), methods have also
been developed for comparing sequences with structures (something called
'threading') and structures with structures (Covered in Chapters 1 and 2,
respectively). All these methods support each other and roughly following the
progression: (1) DB-search -» (2) multiple alignment —> (3) threading -> (4)
modelling. However, this is often far from a linear progression: the alignment
can reveal new constraints that can be imposed on the databank search, while at
the same time also helping the threading application. Similarly, the threading
can cast new (structural) light on the alignment and all are carried out under
(and also affect) the prediction of secondary structure.

vii
PREFACE

Before the advent of multiple genome data, this favoured route often came to
a halt before it started: when no similar sequence could be found even to make an
alignment. However, with the genomes of phylogentically widespread organisms
either completed or promised soon (bacteria, yeast, plasmodium, worm, fly, fish,
man) there is now a good chance of finding proteins from each that can compile
a useful multiple-sequencing alignment. At the threading stage (2) in the above
progression, the current problem and worry is that there may not be a protein
structure on which the alignment can be fitted. Failiure at this stage generally
compromises any success in the final modelling stage (unless sufficient struc-
tural constraints are available from other experimental sources). This problem
will be eased by structural genomics programmes (often associated with a
genome program) for the large-scale determination of protein structures. As
with the genome, these data will greatly increase the chance of finding at least
one structure onto which the protein can be modelled.

The future of a mature field

With several complete genomes and a reasonably complete set of protein struc-
tures, the problems facing Bioinformatics shifts from its past challenge of
finding weak similarities among sparse data, to one of finding closer similarities
in a wealth of data. However, concentrating on protein sequence data (as distinct
from the rawgenomic DNA) eases the data processing problem considerably and
the increased computation demands can be met by the equally rapid increase in
the power of computers. In this new situation, perhaps all that will be needed is
a good multiple sequence alignment program (such as CLUSTAL or MULTAL)
with which to reveal all necessary functional and structural information on any
particular gene.
The most fundamental impact of the 'New Data' is the realization that the
biological world is finite and, at least in the world of sequences, that we have the
end in sight. We have already, in the many bacterial genomes and in yeast, seen
the minimal complement of proteins required to maintain independent life—
and at only several thousand proteins, it does not seem unworkably large. This
will expand by an order-or-magnitude in the higher organisms but it is already
clear that much of this expansion can be accounted for by the proliferation of
sequences within tissue or functionally specific families (such as the G-protein
coupled receptors). Removing this 'redundancy' might still result in a set of
proteins that, if not by eye, can be easily analysed by computer.
The end-of-the-line in protein structures may take a little longer to arrive, but,
by implication from the sequences, it too is finite—and indeed, may be much
more finite than the sequence world. This can be inferred from current data by
the number of protein families that have the same overall structure (or fold), but
otherwise exhibit no signs of functional or sequence similarity. Besides com-
paring and classifying the different structures, an interesting aspect is to develop
models of protein structure evolution, perhaps allowing very distant relation-
ships between these different folds to be inferred. It might be hoped that this

viii
 PREFACE

will shed light on the most ancient origins of protein structure and on the distant
relationships between biological systems.
The ultimate aim of Bioinformatics must surely be the complete understand-
ing of an organism—given its genome. This will require the characterization and
modelling of extremely complex systems: not only within the cell but also
including the fantastic network of cell-cell interactions that go to make-up an
organism (and how the whole system boot-straps itself). However, as Sergeant
Pluck has told us: what is an organism but only millions of little bits of itself
whirling around and doing intricate convolutions. If a genome can tell us all
these bits (and sure it will be no time till we have the genome for a sheep) then
all we have to do is figure out how it all whirls around. For this, without a doubt,
the Sergeant would have recommended the careful application of algebra—and,
had he known about them, I'm sure he would have used a computer.
D.H. and W.T., 2000

ix
This page intentionally left blank
 Contents

Preface page v
List of protocols xvii
Abbreviations xix

1 Threading methods for protein structure prediction 1

David Jones and Caroline Hadley
1 Introduction 1
2 Threading methods 1
1-D-3-D profiles: Bowie a al. (1991) 5
Threading: Jones et al. (1992) 5
Protein fold recognition using secondary structure predictions: Rost (1997) 7
Combining sequence similarity and threading: Jones (1999) 7
3 Assessing the reliability of threading methods 8
Alignment accuracy 9
Post-processing threading results 10
Why does threading work? 30
4 Limitations: strong and weak fold-recognition 11
The domain problem in threading 11
5 The future 12
References 12

2 Comparison of protein three-dimensional structures 15

Mark S. Johnson and Jukka V. Lehtmen
1 Introduction 15
2 The comparison of protein structures 16
General considerations 16
What atoms/features of protein structure to compare? 17
Standard methods for finding the translation vector and rotation matrix 20
Standard methods to determine equivalent matched atoms between structures 25
Quality and extent of structural matches 29
3 The comparison of identical proteins 31
Why compare identical proteins? 31
Comparisons 31

xi
CONTENTS

4 The comparison of homologous structures: example methods 32

Background 32
Methods that require the assignment of seed residues 34
Automatic comparison of 3-D structures 35
Multiple structural comparisons 41
5 The comparison of unrelated structures 42
Background 42
6 Large-scale comparisons of protein structures 46
References 48

3 Multiple alignments for structural, functional, or phylogenetic

analyses of homologous sequences 51
L Duret and S. Abdeddaim
1 Introduction 51
2 Basic concepts for multiple sequence alignment 53
Homology: definition and demonstration 53
Global or local alignments 54
Substitution matrices, weighting of gaps 54
3 Searching for homologous sequences 56
4 Multiple alignment methods 57
Optimal methods for global multiple alignments 59
Progressive global alignment 61
Block-based global alignment 63
Motif-based local multiple alignments 65
Comparison of different methods 65
Particular case: aligning protein-coding DNA sequences 68
5 Visualizing and editing multiple alignments 69
Manual expertise to cheek or refine alignments 71
Annotating alignments, extracting sub-alignments 71
Comparison of alignment editors 72
Alignment shading software, pretty printing, logos, etc. 72
6 Databases of multiple alignments 72
7 Summary 73
References 74

4 Hidden Markov models for database similarity searches 77

Ewan Birney
1 Introduction 77
2 Overview 78
3 Using profile and profile-HMM databases 79
Pfam 80
Prosite profiles 80
SMART 81
Other resources and future directions 81
Limitations of profile-HMM databases 81
4 Using PSI-BLAST 81

xii
 CONTENTS

5 Using HMMER2 82
Overview of using HMMBR 83
Making the first alignment 83
Making a profile-HMM from an alignment 84
Finding homologues and extending the alignment 84
6 False positives 85
7 Validating a profile-HMM match 85
8 Practical issues of the theories behind profile-HMMs 86
Overview of profile-HMMs 86
Statistics for profile-HMM 87
Profile-HMM construction 89
Priors and evolutionary information 89
Technical issues 90
References 91

5 Protein family-based methods for homology detection and

analysis 93
Steven Henikoff and jorja Henikoff
1 Introduction 93
Expanding protein families 93
Terms used to describe relationships among proteins 93
Alternative approaches to inferring function from sequence alignment 94
2 Displaying protein relationships 95
From pairwise to multiple-sequence alignments 95
Patterns 96
Logos 97
Trees 97
3 Block-based methods for multiple-sequence alignment 98
Pairwise alignment-initiated methods 98
Pattern-initiated methods 99
Iterative methods 99
Implementations 300
4 Position-specific scoring matrices (PSSMs) 301
Sequence weights 102
PSSM column scores 102
5 Searching family databases with sequence queries 103
Curated family databases: Prosite, Prints, and Pfam 105
Clustering databases: ProDom, DOMO, Protomap, and Prof_pat 105
Derived family databases: Blocks and Proclass 106
Other tools for searching family databases 107
6 Searching with family-based queries 108
Searching with embedded queries 108
Searching with PSSMs 108
Iterated PSSM searching 109
Multiple alignment-based searching of protein family databases 110
References 110

xiii
CONTENTS

6 Predicting secondary structure from protein sequences 113

Jaap Heringa
1 Introduction 113
What is secondary structure? 113
Where could knowledge about secondary structure help? 114
What signals are there to be recognized? 114
2 Assessing prediction accuracy 118
3 Prediction methods for globular proteins 120
The early methods 120
Accuracy of early methods 122
Other computational approaches 122
Prediction from multiply-aligned sequences 123
A consensus approach: JPRED 129
Multiple-alignment quality and secondary-structure prediction 131
Iterated multiple-alignment and secondary structure prediction 132
4 Prediction of transmembrane segments 133
Prediction of a-helical TM segments 134
Orientation of transmembrane helices 136
Prediction of p-strand transmembrane regions 136
5 Coiled-coil structures 137
6 Threading 138
7 Recommendations and conclusions 138
References 139

7 Methods for discovering conserved patterns in protein sequences

and structures 143
Inge jonossen
1 Introduction 143
2 Pattern descriptions 144
Exact or approximate matching 144
PROSITE patterns 145
Alignments, profiles, and hidden Markov models 146
Pattern significance 148
Pattern databases 150
Using existing pattern collections 153
3 Finding new patterns 154
A general approach 154
Discovery algorithms 155
4 The Pratt programs 156
Using Pratt 157
Pratt: Internal search methods 159
Scoring patterns 161
5 Structure motifs 162
The SPratt program 162
6 Examples 164
7 Conclusions 164
References 165

xiv
 CONTENTS

8 Comparison of protein sequences and practical database

searching 167
Golan Yona and Steven E. Brenner
1 Introduction 167
2 Alignment of sequences 168
Rigorous alignment algorithms 169
Heuristic algorithms for sequence comparison 171
3 Probability and statistics of sequence alignments 173
Statistics of global alignment 174
Statistics of local alignment without gaps 175
Statistics of local alignment with gaps 177
4 Practical database searching 178
Types of comparison 178
Databases 179
Algorithms 181
Filtering 181
Scoring matrices and gap penalties 182
Command line parameters 185
5 Interpretation of results 187
6 Conclusion 188
References 188

9 Networking for the biologist 191

R. A. Harper
1 Introduction 191
2 The changing face of networking 192
Networking in Europe 194
The way we were . . . e-mail servers for sequence retrieval 195
Similarity searches via e-mail 199
Speed solutions for similarity searches 201
3 Sequence retrieval via the WWW 203
Entrez from the NCBI 205
SRS from the EBI 205
4 Submitting sequences 208
Bankit at NCBI 209
Sequin from NCBI 209
Webin from EBI 210
Sakura from DDJB 212
5 Conclusions 212
References 213

10 SRS—Access to molecular biological databanks and Integrated

data analysis tools 215
D. P. Kreil and T. Etzold
1 Introduction 215
SRS fills a critical need 215
History, philosophy, and future of SRS 216

XV
2 A user's primer 217
A simple query 239
Exploiting links between databases 220
Using Views to explore query results 221
Launching analysis tools 223
Overview 225
3 Advanced tools and concepts 225
Refining queries 225
Creating custom Views 230
SRS world wide: using DATABANKS 232
Interfacing with SRS over the network 233
4 SRS server side 236
User's point of view 236
Administrator's point of view 238
5 Where to turn to for help 240
Acknowledgements 241
References 241

List of suppliers 243

Index 247
 Protocol list

The comparison of protein structures

Features used for the comparison of protein 3-D structures 19
Rigid-body structural comparisons: translations and rotations 22
The alignment: determination of equivalent pairs 27
Root mean squared deviations (RMSD) 30
The comparison of Identical proteins
Similarities among different structures of identical proteins 32
The comparison of homologous structures: example methods
Finding initial seed residues 34
Semi-automatic methods 34
Structural comparisons seeded from sequence alignments 35
GA_FTT (ref. 26, 27) 36
Local similarity search by VERTAA 37
Structure comparison by DALI (11) 39
Structure comparison by SSAP (10) 40
Structure comparison by DEJAVU (30) 40
Multiple structural alignments from pairwise comparisons 41
The comparison of unrelated structures
Structure comparison by SARF2 (41) 45
GENFIT (22) 45
Block-based methods for multiple-sequence alignment
Finding motifs from unaligned sequences and searching sequence databanks 100
Assessing prediction accuracy
Jackknife testing 119
Recommendations and conclusions
Predicting secondary structure 139
The changing face of networking
Using [email protected] for the retrieval of sequences and software 198
Using [email protected] for sequence retrieval 198
Blast similarity search e-mail server at NCBI 199
FASTA similarity search e-mail server at EBI 201

xvii
PROTOCOL LIST

A user's primer
Performing a simple SRS query 219
Applying a link query to selected entries 220
Displaying selected entries with one of the pre-defined views 223
Launching an external application program for selected entries 223
Advanced tools and concepts
Browsing the index for a database field 227
Search SRS world wide 232

xviii
 Abbreviations

AACC amino acid class covering

API application programming interface
CASP critical assessment in structure prediction
3-D three-dimensional
EBI European Bioinformatics Institute
EM expectation-maximization
e-value expectation value
EVD extreme value distribution
PDF fast data finder
FN false negative
HMM hidden Markov model
ILP inductive logic programming
LAMA local alignment of multiple alignments
MAST multiple alignment searching tool
MDL minimum description length
MP membrane protein
NCBI National Centre for Biotechnology Information
NCGR National Centre for Genomic Research
NNSSP nearest neighbour secondary structure prediction
PD pattern driven
PDB protein data bank
PHD profile secondary structure predictions from Heidelberg
PPV positive predictive value
PSSM position-specific scoring matrices
QOS quality of service
RMSD root mean square deviation
RTT round trip time
SAP structure alignment program
SD sequence driven
SP sum of pairs
SRS sequence retrieval system

xix
ABBREVIATIONS

SSE secondary structure element

TM transmembrane
TP true positive
WWW World Wide Web
 Chapter 1
Threading methods for protein
structure prediction
David Jones and Caroline Hadley
Department of Biological Sciences, University of Warwick, Coventry CV4 7AL, UK.

1 Introduction
As the attempts to sequence entire genomes increases the number of protein
sequences by a factor of two each year, the gap between sequence and structural
information stored in public databases is growing rapidly. In stark contrast to
sequencing techniques, experimental methods for structure determination are
time-consuming, and limited in their application, and therefore will not be able
to keep pace with the flood of newly characterized gene products. The develop-
ment of practical methods for predicting protein structure from sequence is
therefore of considerable importance in the field of biology.
Several different approaches have been used to predict protein structure from
sequence, with varying degrees of success. Ab initio methods encompass any
means of calculating co-ordinates for a protein sequence from first principles—
that is, without reference to existing protein structures. Little success has been
seen in this area, with more theory produced than actual useful methodology.
Comparative (or homology) modelling, attempts to predict protein structure on
the strength of a protein's sequence similarity to another protein of known
structure (following the theory that similar sequence implies similar structure).
Some success has been achieved, but several limitations to this method, not
least of which are its dependence on alignment quality and the existence of a
good sequence homologue, indicate it is not applicable to a large fraction of pro-
tein sequences. The third main category of protein structure prediction, falling
somewhere between comparative modelling and ab initio prediction, is fold
recognition, or threading.

2 Threading methods
The term 'threading' was first coined in 1992 by Jones et al. (1), but the field has
grown considerably since then with many different methods being proposed:
for example, Godzik and Skolnick (2); Ouzounis et al. (3); Abagyan-et al. (4);
Overington et al. (5); Matsuo et al. (6); Madej et al. (7); Lathrop and Smith (8);

1
DAVID JONES AND CAROLINE HADLEY

Figure 1 An example of a pair of protein structures in the same family, (a) Human myoglooin
[2mml], |b) pig haemoglobin, alpha chain [2pghA]. At the family level, proteins have higher
sequences identity (in this case, 32%) and have highly similar structures. Figures created
using Molscript (19).

Figure 2 A pair of structures within the same superfamily. (a) A. denitrificans azurin [lazcA],
(b) poplar plastocyanin [1plc]. Members of the same superfamily may have insignificant
sequence identity (16% in this case), but still share most features of the protein fold,
reflecting a common evolutionary origin.

Taylor (9) amongst others. The idea behind threading came about from the
observation that a large percentage of proteins adopt one of a limited number of
folds (Figures 1-3), In fact, just 10 different folds (the 'superfolds') account for 50%
of the known structural similarities between protein superfamilies (18). Thus,
rather than trying to find the correct structure for a protein from the huge
number of all possible conformations available to a polypeptide chain, the cor-
rect (or close to correct) structure is likely to have already been observed and
already stored in a structural database. Of course, in cases where the target
protein shares significant sequence similarity to a protein of known 3-D struc-
ture, the 'fold recognition' problem is trivial—simple sequence comparison will
identify the correct fold. The hope was, however, that threading might be able
to detect structural similarities that are not accompanied by any detectable
sequence similarity, and this has subsequently been proven to be the case.

2
 THREADING METHODS FOR PROTEIN STRUCTURE PREDICTION

Figure 3 A pair of analogous folds, (a) Chicken triosephosphate isomerase [1timA],

(b) E. colt fructose bisphosphate aldolase [1dosA]. Members of the same fold family have
the same major secondary structure elements with the same arrangement and connectivity.
Very low sequence identity and large variations in the details of the structures reflects the
lack of common ancestry between analogous folds. Although the aldolase structure contains
additional helices, the TIM barrel fold is obviously present in both proteins. The TIM barrel is
one of 10 'superfolds' identified by Orengo et al. (18).

Figure 4 shows an outline of a generic fold recognition method. Firstly, a library

of unique or representative protein structures needs to be derived from the data-
base of all known protein structures. Different groups use different selection
criteria for their fold libraries: in some cases, complete protein chains are used
in the library, but in other cases, structural domains or even conserved proteins
cores are used. Each fold from this library is then considered in turn and the
target sequence optimally fitted (or aligned) to each library fold (allowing for
relative insertions and deletions in loop regions). Many different algorithms
have been proposed for finding this optimal sequence-structure alignment, with
most groups using some form of dynamic programming algorithm (including
the examples described below), but other algorithms such as Gibbs sampling (7)
or branch-and-bound searching (8) have also been used with some success.
Finally, some kind of objective function is needed to determine the goodness of
fit between the sequence and the template structure. It is this objective function
which is optimized during the sequence-structure alignment. Again opinions
differ as to the form of this objective function. Most groups use some kind of
'pseudo energy' function based on a statistical analysis of observed protein struc-
tures, but other more abstract scoring functions have also been proposed (see
ref. 20 for a recent review). The final result of a fold recognition method is a
ranking of the fold library in descending order of goodness of fit', with the best
fitting fold (typically the lowest energy fold) being taken as the most probable
match.

3
DAVID JONES AND CAROLINE HADLEY

Figure 4 This is an outline of the fold recognition approach to protein structure prediction,
and identifies three clear aspects of the problem that need consideration: a fold library, a
method for modelling the object sequence on each fold, and a means for assessing the
goodness-of-fit between the sequence and the structure.

No matter what algorithm or scoring function is used, fold recognition is not

without its limitations, and some progress must be made before it can be con-
sidered a routine protein structure prediction tool. Several different aspects of
this method are particularly open to improvement, namely the question of
potential functions (i.e. the calculations used to determine the energy of a
particular sequence once fitted onto a template fold), improvements in align-
ments (i.e. correctly aligning the sequence onto the template fold, to produce
the best fit), and the need for progress in post-processing the results (i.e. from

4
 THREADING METHODS FOR PROTEIN STRUCTURE PREDICTION

the energy calculations, etc., choosing the best 'fit'). Significant progress may
also arise from improvements in the threading library used (i.e. the templates
upon which the sequences will be threaded).
To get some idea of the variety of methods which have been developed, four
distinct approaches to the fold-recognition problem will be described. Virtually
all fold-recognition methods are similar to at least one of these methods, and
some newer methods incorporate concepts from more than one.

2.1 1-D-3-D profiles: Bowie et al. (1991)

The first true fold recognition method was by Bowie, Luthy, and Eisenberg (10),
where they attempted to match sequences to folds by describing the fold in
terms of the environment of each residue in the structure. The environment was
described in terms of local secondary structure (3 states: a, p, and coil), solvent
accessibility (3 states: buried, partially buried, and exposed), and the degree of
burial by polar rather than apolar atoms. The basic idea of the method is the
assumption that the environment of a particular residue thus defined is expected
to be more conserved than the actual residue itself, and so the method is able to
detect more distant sequence-structure relationships than purely sequence-
based methods. The authors describe this method as a 1-D-3-D profile method,
in that a 3-D structure is translated into a 1-D string, which can then be aligned
using traditional dynamic programming algorithms. Bowie et al. have applied
the 1-D-3-D profile method to the inverse folding problem and have shown that
the method can indeed detect remote matches, but in the cases shown the hits still
retained some weak sequence similarity with the search protein. Environment-
based methods appear to be incapable of detecting structural similarities between
extremely divergent proteins, and between proteins sharing a common fold
through convergent evolution—environment only appears to be conserved up
to a point. Consider a buried polar residue in one structure that is found to be
located in a polar environment. Buried polar residues tend to be functionally
important residues, and so it is not surprising then that a protein with a similar
structure but with an entirely different function would choose to place a hydro-
phobic residue at this position in an apolar environment. A further problem
with environment-based methods is that they are sensitive to the multimeric
state of a protein. Residues buried in a subunit interface of a multimeric protein
will not be buried at an equivalent position in a monomeric protein of similar
fold.

2.2 Threading: Jones et al. (1992)

The method which introduced the term 'threading' (1) went further than the
method of Bowie, Luthy, and Eisenberg in that instead of using averaged residue
environments, a given protein fold was modelled in terms of a 'network' of
pairwise interatomic energy terms, with the structural role of any given residue
described in terms of its interactions. Classifying such a set of interactions into
one environmental class such as 'buried alpha helical' will inevitably result in

5
DAVID JONES AND CAROLINE HADLEY

the loss of useful information, reducing the specificity of sequence-structure

matches evaluated in this way. Thus, in true threading methods, a sequence is
matched to a structure by considering detailed pairwise interactions, rather than
averaging them into a crude environmental class. However, incorporation of
such non-local interactions means that simple dynamic programming string-
matching methods cannot be used. There is therefore a trade-off to be made
between the complexity of the sequence-structure scoring scheme and the
algorithmic complexity of the problem.
Jones et al. (1) proposed a novel dynamic programming algorithm (now com-
monly known as 'double' dynamic programming) to the problem of aligning a
given sequence with the backbone co-ordinates of a template protein structure,
taking into account the detailed pairwise interactions. The problem of matching
pairwise interactions is somewhat similar to the problem of structural com-
parison methods. The potential environment of a residue i can be defined as being
the sum of all pairwise potential terms involving i and all other residues j=i.
This is an analogous definition to that of a residue's structural environment, as
described by Taylor and Orengo (11). In the simplest case, structural environ-
ment of a residue i may be defined as the set of all inter-Ca distances between
residue i and all other residues j=i. Taylor and Orengo propose a novel dynamic
programming algorithm for the comparison of such residue structural environ-
ments, and this method proved to be effective for the comparison of residue
potential environments. A detailed description of the algorithm has recently
been published (12).
For a sequence-structure compatibility function, Jones et al. chose to use a set
of statistically derived pairwise potentials similar to those described by Sippl
(13). Using the formulation of Sippl, short (sequence separation, k < 10),
medium (11 < k < 30), and long (k > 30) range potentials were constructed
between the following atom pairs: Cp -> Cp, Cp -> N, Cp -> 0, N -> Cp, N->O,
0 —> Cp, and O —> N. For a given pair of atoms, a given residue sequence
separation and a given interaction distance, these potentials provide a measure
of energy, which relates to the probability of observing the proposed interaction
in native protein structures. In addition to these pairwise terms, a 'solvation
potential' was also incorporated. This potential simply measures the frequency
with which each amino acid species is found with a certain degree of solvation,
approximated by the residue solvent accessible surface area.
By dividing the empirical pair potentials into sequence separation ranges,
specific structural significance may be tentatively conferred on each range. For
instance, the short range terms predominate in the matching of secondary
structural elements. By threading a sequence segment onto the template of an
alpha helical conformation and evaluating the short range potential terms, the
possibility of the sequence folding into an alpha helix may be evaluated. In a
similar way, medium range terms mediate the matching of super-secondary
structural motifs, and the long range terms, the tertiary packing.
Recent features added to the method allow sequence information and pre-
dicted secondary structure information to be considered in the fold-recognition

6
 THREADING METHODS FOR PROTEIN STRUCTURE PREDICTION

process. Sequence information is weighted into the fold recognition potentials

using a transformation of a mutation data matrix (12). By carefully selecting the
weighting of the sequence components in the scoring function it is possible to
balance the influence of sequence matching with the influence of the pairwise
and solvation energy terms. In contrast to this, secondary structure information
is not incorporated into the sequence-structure scoring function. In this case,
secondary structure information is used to mask regions of the alignment path
matrix so that the threading alignments do not align (for example) predicted B
strands with observed a helices. A confidence threshold is applied to the second-
ary structure prediction data so that only the most confidently predicted regions
of the prediction are used to mask the alignment matrix.

2.3 Protein fold recognition using secondary structure

predictions: Rost (1997)
Although most fold recognition methods employ potentials of one kind or
another, it is quite easy to design a useful fold recognition approach that at first
sight does not employ potentials of any kind. Although not the first example of
this approach, as a good recent example the PHD secondary structure prediction
service (14) has recently been extended to offer a fold recognition option. In this
case the system predicts the secondary structure and accessibility of each residue
in the protein of interest, encodes this information in the form of a string
(similar to the scheme employed by Bowie et al.) (10) and then matches this
string against a library of strings computed from known structures. A number of
other similar methods are also in development in other labs, though all based on
the initial prediction of secondary structure by PHD. Clearly no explicit potentials
are being employed in these methods, but potentials are implicitly coded into
the neural network weights used to predict secondary structure in the first
place.

2.4 Combining sequence similarity and threading: Jones

(1999)
Jones (15) has recently proposed a hybrid fold recognition method which is
designed to be both fast and reliable, and is particularly aimed at automated
genome annotation. The method uses a sequence profile-based alignment
algorithm to generate alignments which are then evaluated by threading tech-
niques. As a last step, each threaded model is evaluated by a neural network in
order to produce a single measure of confidence in the proposed prediction. The
speed of the method, along with its sensitivity and very low false-positive rate
makes it ideal for automatically predicting the structure of all the proteins in a
translated bacterial genome. The method has been applied to the genome of
Mycoplasma genitalium, and analysis of the results shows that as many as 46%
(now 51%) of the proteins derived from the predicted protein coding regions
have a significant relationship to a protein of known structure. The fact that
alignments are generated by a sequence alignment step means that the method

7
DAVID JONES AND CAROLINE HADLEY

Figure 5 Hypothetical applicability of different categories of fold-recognition methods to the

Open Reading Frames of small bacterial genomes. At present sequence-based fold
recognition (e.g. GenTHREADER) is successful for around 50% of the ORFs. Structures for a
further 15% of ORFs can probably be assigned by full threading methods such as
THREADER, and the remaining 35% cannot currently tie recognized either because the fold
has not yet been observed, or because the ORF encodes a non-globular protein (e.g. a
transmembrane protein).

is only expected to work for family or superfamily level similarities between the
target and template proteins. This is both a positive and negative feature of the
method. The negative aspect is that, of course, many purely structural similari-
ties will not be detected by the method. The positive aspect is that superfamily
relationships produce the most reliable results, and also allow some aspects of
the function of the target protein to be inferred from the matched template
structure. This latter point is particularly useful when annotating unknown
genome sequences. Figure 5 shows the current applicability of different types of
fold recognition method to a genome such as that of M, genttalmm,
Unlike full threading methods, which require a great deal of computer power
to run, this type of method can be made readily available to the public via a
simple Web server. The GenTHREADER method is available from the following
URL:
http://globin.bio.warwick.ac.uk/psipred

3 Assessing the reliability of threading methods

Although the published results for the fold recognition methods can look im-
pressive, showing that threading is indeed capable of recognizing folds in the
absence of significant sequence similarity, it tan be argued that in all cases the
correct answers were already known and so it is not clear how well they would

8
 THREADING METHODS FOR PROTEIN STRUCTURE PREDICTION

perform in real situations where the answers are not known at the time the
predictions are made. It was not until these methods were tested in a set of blind
trials—the Critical Assessment in Structure Prediction experiments (CASP)—
that it became clear how powerful these methods could be when used without
prior knowledge of the correct answer. The CASP experiment has now been run
three times (CASP1 in 1994, CASP2 in 1996, CASP3 in 1998): and in the last meet-
ing results from over 30 methods were evaluated by the independent assessors.
Up to date information on all of the CASP experiments can be obtained from the
following Web address:
http://predictioncenter.llnl.gov

3.1 Alignment accuracy

Most published methods are evaluated solely on the basis of fold assignment, i.e.
the method is evaluated on its ability to correctly pick the correct fold. However,
in practice, fold assignment is not sufficient in its own right. Given a correct fold
assignment the next step is of course to generate an accurate sequence structure
alignment and to use this alignment to generate an accurate 3-D model for the
target protein. In cases where a fold has been assigned, the alignment can be
passed to an automatic comparative modelling program (e.g. MODELLER3) so
that loops and side chains can be built in.
The accuracy of alignments that can be produced by fold recognition methods
can be measured in terms of the Root Mean Square Deviation (RMSD) between
the implied prediction model and the observed experimental structure. Analysis
of the results of the CASP experiments has shown that alignment accuracy
correlates strongly with the degree of evolutionary and structural divergence
between the available template structures and the target protein. The degree of
model accuracy that can be expected can be broken down into three categories
of structure relationship:
(a) Family (e.g. Figure 1). Evident sequence similarity. Threading models will be
almost entirely accurate, with an RMSD of between 1.0 and 3.0 Angstroms,
depending on the degree of sequence similarity.
(b) Superfamily (e.g. Figure 2). No significant sequence similarity, but evident
common ancestry between the template and target structure. Models for this
class of similarity will be partially correct (mostly in active site regions) and
will have an RMSD of between 3.0 and 6.0 Angstroms typically (though some-
times more depending on the accuracy of the alignment produced).
(c) Analogy (e.g. Figure 3). No apparent common ancestry between the template
and target structure. Low quality models are expected for this category of
similarity. RMSD is not a good way to evaluate models of this quality as very
large shifts in the alignment produce virtually random RMSD values. At best,
alignments in this class are 'topologically correct', in that the correct elements
of secondary structure are equivalenced, but frequently shifts in the align-
ment are so large as to render the models entirely incorrect

9
DAVID JONES AND CAROLINE HADLEY

3.2 Post-processing threading results

Perhaps one of the most significant observations that came from the CASP2
prediction experiment was that a great deal of success in fold recognition can be
achieved purely from a deep background knowledge of protein structure and
function relationships. Alexey Murzin (one of the authors of the SCOP protein
structure classification scheme) identified a number of key evolutionary clues
which led him to correctly assign membership of some of the target proteins to
known superfamilies (16). Also in two cases he was able to confidently assign a
'null prediction' to targets with unique folds purely by considering their pre-
dicted secondary structure. These feats are quite remarkable, but not easily
reproduced by non-experts in protein structure and function. Despite this, it is
very clear that an important future development of practical fold-recognition is
to take both structure and function into account when ranking the sequence-
structure matches.
Even without new developments in fold recognition algorithms, information
on function and other sources of information can be applied to the results of a
threading method as a 'post-processing' step. Rather than simply taking the top
scoring fold to be the assumed correct answer, a fold from, say, the top 10
matches can be selected by human intervention. Such intervention might
involve visual inspection of the proposed alignment, inspection of the proposed
3-D structure on a graphics workstation, comparison of proposed secondary
structure with that obtained from secondary structure prediction or even con-
sideration of common function between the target and template proteins.

3.3 Why does threading work?

Although many different formulations of energy function have been used for
fold recognition, it has been shown that the principal factor in the most success-
ful of these empirical potentials essentially encodes the general 'hydrophobic
effect', rather than specific interactions between specific side chains, (e.g. the
interaction potential between like charges is the same as that derived for unlike
charges, reflecting not the specific interaction between side chains, but their
overall preference to lie on the surface). Despite this observation that specific
pair interactions are not vital to successful fold recognition, threading methods
based on pairwise interactions do seem to work better than profile methods (as
evidenced for example in the predictions made during the CASP experiments).
This might at first sight seem contradictory, particularly as it is apparent that
specific pairwise interactions are not conserved between analogous fold families
(17). Nevertheless, threading methods do seem to be picking up signals which
are not detected by simple 1-D profile methods. Why might this be the case?
One reasonable explanation may be that profile-based fold-recognition methods
make the assumption that the pattern of accessibility between two divergent
protein structures is perfectly conserved, and it is this assumption that results in
their relatively poor performance. Threading methods, on the other hand, are
able to model the environment of a residue by summing the hydrophobic pair

10
 THREADING METHODS FOR PROTEIN STRUCTURE PREDICTION

interactions surrounding a particular residue. These pair interaction environments

of course change as the threading alignment changes, and it is this sensitivity of
residue environments to changes in the sequence-structure alignment that
results in the increased predictive power of threading methods. Although this
explains why threading works even when specific contacts are not being
conserved, it also explains why sequence-structure alignments are generally of
poor quality when compared with known structure-structure alignments.

4 Limitations: strong and weak fold-recognition

What are the limitations of current fold-recognition methods? Let's consider
two forms of the fold-recognition problem. In the first form of the problem we
seek a set of potentials (and a method for performing the sequence-structure
alignment) which will reliably recognize the closest matching fold for a given
sequence from the thousands of alternatives—as many as 7000 naturally occur-
ring folds have been estimated (18). This form of the problem is referred to as
the 'strong' fold-recognition problem. It is possible that the strong fold-recog-
nition problem is actually insoluble because, quite simply, the real protein free
energy function is itself almost certainly incapable of satisfying this require-
ment. In other words, given the physical 'unreality' of threaded models, there
may exist no energy function which is capable of uniquely recognizing the
correct fold in all cases. One possible avenue for moving towards this goal may
be to consider simulated folding pathways for each fold in the fold library, but
for the time being, perfect fold recognition is but a distant dream.
The 'weak' fold-recognition problem is a far more practical formulation of the
problem. Here the goal is to recognize and exclude folds which are not com-
patible with the given sequence with the eventual aim of arriving at a shortlist of
possible conformations for the protein being modelled. At first sight this may not
seem different from the goal of strong fold recognition, but the distinction is
quite important. Even without a sophisticated fold-recognition method, weak
recognition can be achieved by the application of simple common-sense rules.
For example, if it is known that a protein is comprised entirely of alpha-helices
(which might be known from circular dichroism spectroscopy, for example) then
a large number of possible folds can be eliminated immediately (the correct fold
could not be the all-beta immunoglobulin fold, for example). By applying a set of
such rules, the 7000 or so possible folds could quickly be whittled down to a
shortlist of say 10.
In reality, most, if not all of the published fold recognition methods really im-
plement weak fold recognition. In the hands of an experienced user, however,
who can make use of functional or structural clues in the prediction experi-
ment, even weak fold recognition can be very powerful.

4.1 The domain problem in threading

Perhaps the main practical limitation of most 'weak' threading methods is that
they are aimed at recognizing single globular protein domains, and perform very

11
DAVID JONES AND CAROLINE HADLEY

poorly when tried on proteins which comprise multiple domains. Unfortunately,

threading cannot be used for identifying domain boundaries with any degree of
confidence, and indeed the general problem of detecting domain boundaries
from amino acid sequence remains an unsolved problem in structural biology. If
the domain boundaries in the target sequence are already known, then of
course the target sequence can be divided into domains before threading it, with
each domain being threaded separately. Predictions can be attempted on very
long multi-domain sequences, but in these cases the results will not be reliable
unless it is clear that the matched protein has an identical domain structure to
the target. For example, the periplasmic small-molecule binding proteins (e.g.
leudne-isoleucine-valine binding protein) are two domain structures (two doubly
wound parallel alpha/beta domains), but they all have identical domain organ-
ization. Proteins within this superfamily can thus be recognized by threading
methods despite their multi-domain structure.

5 The future
One major difference between the academic challenge of protein structure
prediction and the practical applications of such methods is that in the latter
case there is an eventual end in sight. As more structures are solved, more target
sequences will find matches in the available fold libraries—matched either by
sequence comparison or threading methods. In terms of practical application,
the protein-folding problem will thus begin to vanish. There will of course still
be a need to better understand protein-folding for applications such as de novo
protein design, and the problem of modelling membrane protein structure will
probably remain unsolved for some time to come, but nonetheless, from a
practical viewpoint, the problem will be effectively solved. How long until this
point is reached? Given the variety of estimates for the number of naturally
occurring protein folds, it is difficult to come to a definite conclusion, but taking
an average of the published estimates for the number of naturally occurring
protein folds and applying some intelligent guesswork, it seems likely that when
threading fold libraries contain around 1500 different domain folds it will be
possible to build useful models for almost every globular protein sequence in a
given proteome. At the present rate at which protein structures are being
solved, this point is possibly 15-20 years away. However, pilot projects are now
underway to explore the possibility of crystallizing every globular protein in a
typical bacterial proteome. If such projects get fully under way, which seems
likely, then a complete domain fold library may be only five years away.

References
1. Jones, D. T., Taylor, W. R., and Thornton, J. M. (1992). A new approach to protein fold
recognition. Nature, 358, 86.
2. Godzik, A. and Skolnick, J. (1992). Sequence-structure matching in globular proteins:
Application to supersecondaiy and tertiary structure determination. Proc. Nail. Acad.
Set USA, 89,12098.

12
 THREADING METHODS FOR PROTEIN STRUCTURE PREDICTION

3. Ouzounis, C, Sander, C., Scharf, M., and Schneider, R. (1993). Prediction of protein
structure by evaluation of sequence-structure fitness. Aligning sequences to contact
profiles derived from three-dimensional structures.J. Mol. Biol., 232, 805.
4. Abagyan, R., Frishman, D., and Argos, P. (1994). Recognition of distantly related
proteins through energy calculations. Proteins: Struct, Funct. Genet., 19, 132.
5. Overington, J., Donnelly, D., Johnson, M. S., Sali, A., and Blundell, T. L (1992).
Environment-specific amino-acid substitution tables—tertiary templates and
prediction of protein folds. Protein Sci., 1, 216.
6. Matuso, Y., Nakamura, H., and Nishikawa, K. (1995). Detection of protein 3-D-l-D
compatability characterised by the evaluation of side-chain packing and electrostatic
interactions. J. Biochem. (japan), 118, 137.
7. Madej, T., Gilbrat, J.-F., and Bryant, S. H. (1995). Threading a database of protein cores.
Proteins, 23, 356.
8. Lathrop, R. H. and Smith, T. F. (1996). Global optimum protein threading with gapped
alignment and empirical pair score functions.J. Mol. Biol., 255, 641.
9. Taylor, W. R. (1997). Multiple sequence threading: An analysis of alignment quality
and stability.J. AM Biol., 269, 902.
10. Bowie, J. U, Luthy, R., and Eisenberg, D. (1991). A method to identify protein
sequences that fold into a known three-dimensional structure. Science, 253,164.
11. Taylor, W. R. and Orengo, C. A. (1989). Protein structure alignment.J. Mol. Biol.,
208, 1.
12. Jones, D. T. (1998). THREADER : Protein Sequence Threading by Double Dynamic
Programming. In Computational methods in molecular biology (ed. S. Salzberg, D. Searls,
and S. Kasif). Elsevier, Amsterdam.
13. Sippl, M. J. (1990). Calculation of conformational ensembles from potentials of mean
force. An approach to the knowledge-based prediction of local structures in globular
proteinsJ. Mol. Biol, 213, 859.
14. Rost, B. (1997). Protein fold recognition by prediction-based threading. J.. Mol. Biol,
270, 1.
15. Jones, D. T. (1999). GenTHREADER: An efficient and reliable protein fold recognition
method for genomic sequences. J. Mol. Biol., 287, 797.
16. Murzin, A, G. and Bateman, A. (1997). Distant homology recognition using structural
classification of proteins. Proteins Suppl., 1, 105.
17. Russell, R. B. and Barton, G. J. (1994). Structural features can be unconserved in
proteins with similar folds—an analysis of side-chain to side-chain contacts,
secondary structure and accessibility. J. Mol. Biol., 244, 332.
18. Orengo, C. A., Jones, D. T., and Thornton, J. M. (1994). Protein superfamilies and
domain superfolds. Nature, 372, 631.
19. Kraulis, P. J. (1991). Molscript—a program to produce both detailed and schematic
plots of protein structures. J . Appl. Crystallogr., 24, 946.
20 Jones, D. T., and Thornton, J. M. (1996). Potential energy functions for threading. Curr.
Opin. Struct. Biol., 6, 210.

13
This page intentionally left blank
 Chapter 2
Comparison of protein
three-dimensional structures
Mark S. Johnson and Jukka V. Lehtonen
Department of Biochemistry and Pharmacy, Abo Akademi University, Tykistokatu
6 A, 20520 Turku, Finland.

1 Introduction
In this chapter we define the different types of questions that may be asked
through the comparison of the three-dimensional (3-D) structures of proteins,
how to make the comparisons necessary to answer each question, and how to
interpret them. We shall focus on the different strategies used, and the assump-
tions made within typical computer programs that are available.
Protein structure comparisons are often used to highlight the similarities and
differences among related—homologous—3-D structures. Homologous proteins are
descended from a common ancestral protein, but have subsequently duplicated,
evolved along separate paths, and thus changed over time. The independent
evolution of related proteins with the same function, orihologous proteins, which
are found in different species, and the paralogaus proteins, which have evolved
different functions, all retain information on the original relationship. The amino
acid sequences change over time reflecting the mutations, insertions and de-
letions that occur in their genes during evolution, and for many proteins the
sequences themselves are so similar that common ancestry is apparent. For
others, the sequences can be so dissimilar that the case for homology may be diffi-
cult to make on the basis of the primary structure. Nonetheless, comparing the
3-D structures when they are available can identify homologous proteins. This is
possible since the evolution of proteins occurs such that their folds are highly
conserved even though the sequences that encode them may not be recognizably
similar.
Homologous proteins are often compared in order to highlight features
(typically the amino acids and their relative orientations to one another), which
have come under strong evolutionary pressure not to change because of struc-
tural and functional restraints placed on them. Conversely, differences in an
otherwise conserved active site or binding site are used to explain differences in
observed function.
Dayhoff and coworkers (1) long ago predicted that about 1000 different protein

15
MARK S. JOHNSON AND JUKKA V. LEHTONEN

families should exist in nature, and it has become clear over recent years that
most newly-solved 3-D structures do fall into an existing family of structures
(2, 3). The approximately 100 000 proteins encoded in the human genome, whose
sequences will be known early in this century, will fall within this limited
number of families. Thus, one key bioinfbrmational goal has been to compare
and classify all proteins and their component domains into family groups, and
one immediate goal is to solve at least one representative structure for each
sequence family that is not obviously connected to any existing structural family.
This single representative structure can then be used in knowledge-based
modelling (4) to estimate the 3-D structures for other members of the family.
Comparisons are also made among non-homologous proteins to try and high-
light structural features that are locally similar, but whose present-day sequences
have not arisen as a consequence of evolutionary divergence from a common
ancestor. Classic examples include the active site similarities among serine
proteinases, subtilisins and serine carboxypeptidase II (5), each of which invoke
the participation of histidine, serine and an aspartic acid in their proteolytic
mechanism of action. The folds are different and the relative positions of these
key amino acids along the sequence are different too. In the 3-D structures, how-
ever, the residues are similarly positioned to reproduce a common catalytic
mechanism that has been exploited by nature on at least three separate occasions.
Comparisons among non-homologous proteins can highlight structural units
that are common features of the protein fold and comparisons have been made
to classify amino acid conformations, regular elements of secondary structure
(helices, strands, turns), supersecondary structure, and cofactor and ligand bind-
ing sites.
The comparison of protein structures can be achieved in many different ways.
In this chapter, we present several of the basic procedures used in the wide
variety of programs that have been developed over the years. These methods
range from rigid-body comparisons, to methods more typical of sequence com-
parisons—dynamic programming, and to those methods that employ Monte
Carlo simulations, simulated annealing and genetic algorithms to find solutions
for combinatorially-complex structural comparison problems. We will describe
methods that demand partial solutions as input to the procedure, as well as
strategies for automatic hands-off solutions; and approaches to both homologous
and non-homologous structural comparisons.

2 The comparison of protein structures

2.1 General considerations
The optimal superposition of two identical 3-D objects can be determined ex-
actly. This only requires the calculation of (a) a translation vector to place one
copy of the object over the other at the origin of the co-ordinate system and (b) a
rotation matrix that describes the rotations needed to exactly match the two
copies of the object. The translation vector describes movements along the x, y,

16
 COMPARISON OF PROTEIN THREE-DIMENSIONAL STRUCTURES

and z directions in the co-ordinate system. The rotation matrix describes the a, B,
and -y rotations in the three orthogonal planes. One of the main tasks of many
super-positioning procedures is to define these values and then to apply them to
the co-ordinates of the objects and they will then be superposed on each other.
Two identical objects will have all points superposed exactly.
The major difficulty with non-identical objects, such as a pair of protein
structures, is that they typically have different numbers of amino acids, different
amino acids with different numbers, types and connectivities of atoms. Further-
more, amino acids present in one structure can be missing in the other: inser-
tions and deletions—the gaps seen in a sequence alignment. Thus, except in the
case of one protein co-ordinate set being compared with itself, no two proteins
will have atoms in exactly identical positions. A protein whose structure has
been solved several times will also vary with overall differences in the main
chain co-ordinates of no more than about 0.3 A, but they will be different.
The superposition of most protein structures as rigid-bodies, therefore, is not
straightforward, and several different considerations need to be resolved in
advance of the comparison. These include:

(a) Which atoms will be compared between the molecules?

(b) How will the dissimilarity or similarity between relative positions of matched
atoms be taken into account?
(c) Should the structures be compared as rigid-bodies (in most cases, resulting in a
partial alignment of the most similar regions, which can be displayed graphic-
ally)? Or have significant structural shifts occurred that require a procedure
that can accommodate these changes (typically providing the complete align-
ment of the sequences, including gap regions, on the basis of the structural
features compared)?
(d) How will one define what constitutes an equivalent matched set of co-
ordinates between non-identical objects where exact matching of atoms will
only rarely be seen?
(e) How will the program be initially seeded? Many methods need to be supplied
with co-ordinates of a set of equivalent atoms at the onset of the comparison,
a minimum of three matched pairs, thus requiring some information on the
likely superposition of the two structures in advance of comparison.
(f) How will the quality of the structural comparison that results be assessed?

2.2 What atoms/features of protein structure to compare?

Depending on what question you wish to answer by comparing a pair of struc-
tures, the choice of which atoms' co-ordinates will be superposed can be crucial.
For example, to look at similarities/differences surrounding a bound cofactor
common to two proteins, you may choose to superpose all or some of the atoms
of the cofactor, apply the translation vector and rotation matrix to the entire
co-ordinate file—protein and cofactor included. Alternatively, the backbone

17
MARK S. JOHNSON AND JUKKA V. LEHTONEN

co-ordinates of the proteins could be superposed and the relative positions of

the cofactors examined after the superposition.
It is usually not very useful to compare atoms of amino acid side chains when
making global structural comparisons. Different amino acid types have different
number of atoms and different connectivities that can preclude their direct
comparison. Residues, even identical ones, will have different conformations,
especially when they are located at the solvent-exposed surface of the proteins.
However, there are situations where the local comparison of side chains can be
very useful, for example, in the comparison of residues lining an active or bind-
ing site especially when different ligands are bound to the same or similar
structures.
For most general methods, which aim to superimpose two proteins over the
maximum number of residues, the Ca-atom co-ordinates are typically employed
(all atoms of the protein backbone and even the side chain Cp-atom, but exclud-
ing the more positionally-variable carbonyl 0, can also be used). (Except where
noted, we will consider Ca-atom co-ordinates in the protocols described herein.)
Whereas the side chain conformations can vary wildly between matched
positions in two structures, the Ca-atom or backbone trace of the fold is typically
well conserved, with regular elements of secondary structure, the a-helices and

Table 1 Examples of features' of proteins that can be used in comparisons

Properties
(a) Residues (b) Segments
Identity Secondary structure type
Physical properties Amphipathicity
Local conformation Improper dihedral angle
Distance from gravity centre Distance from gravity centre
Number of neighbours in vicinity Average Ca density
Position in space Position in space
Global direction in space Global direction
Main chain accessibility Main chain accessibility
Side chain accessibility Side chain accessibility
Main chain orientation Orientation relative to gravity centre
Side chain orientation
Main chain dihedral angles

Relations
(a) (b)
Disulfide bond Relative orientation of two or more segments
Vectors" to one or more nearest neighbours Vectors" to one or more nearest neighbours
Distances to one or more nearest neighbours Distances to one or more nearest neighbours
(e.g. atom pairs or contact maps)
Change in number of neighbours in vicinity
Ionic bond
Hydrogen bond
Hydrophobic cluster
8
See refs 7, 8, 10, 11.
" Vector defines both distance and direction in the local reference frame.
 COMPARISON OF PROTEIN THREE-DIMENSIONAL STRUCTURES

fS-strands, matching closely and sequentially along the fold of the two struc-
tures. Differences in Ca-atom traces arc more often seen at loop regions that
connect the strands and helices in proteins: Frequently these loop regions are
exposed to the solvent at the surface of the protein and thus have fewer
constraints placed on their conformations.
For more dissimilar protein structures, rigid body movements and other
structural changes can occur in one structure relative to the other. When this
happens, rigid-body comparisons of the 3-D structures can often lead to poorly
matched structures, although the folds are the same. If these changes are not
large, then dynamic programming procedures (6) that consider only Cut-Co: atom
distances or other structural properties of the amino acids (Table 1) after an
initial rigid-body comparison can be quite effective in matching all residues
from the protein structures (7-9), Others have described automated procedures
that involve the comparison of structural relationships that require special
techniques to solve these problems of combinatorial complexity (7, 8, 10, 11).

Features used for the comparison of protein 3-D

structures
Distances between atomic co-ordinates are often used (a) for more similar proteins where
rigid-body shifts of one structure relative to another are not a significant factor, (b) to
illustrate the degree of structural change, or (c) where a local comparison of a site of
interest—active site or binding site—is desired for visualization purposes. Where signifi-
cant changes to the structures have occurred, other structural features, which are not as
sensitive to these relative structural shifts, can be compared in addition to atomic co-
ordinates.
Rigid-body structural comparisons
1 Choose the atoms for comparison that are appropriate for the question to be asked.
Most often, but not necessarily, the Ca-atom co-ordinates are used by default.
2 Comparisons will then be based on the distances between atoms that are con-
sidered to be equivalent. For rigid-body methods, a distance cut-off is used to define
equivalent matched positions. Typically, the cut-off value is on the order of 3 A,
although values between 2.5 A and 4.5 A have been used. Lower values are more
restrictive and will lead to fewer aligned positions in more dissimilar structures.
Structural feature comparisons
1 Features of individual atoms, residues or segments of residues, both properties of
and relationships between individual atoms, residues or segments, are considered
either separately or in combination with each other as a basis for structural com-
parisons (see Table 1 and refs 7, 8).
2 Comparisons will be based on differences/similarities between potential matched
regions in the two structures in terms of the features compared. An alignment

19
MARK S. JOHNSON AND JUKKA V. LEHTONEN

algorithm is used to give the best 'sequence alignment' based on the structural
features that have been supplied.
(a) Property comparisons may require an initial alignment (e.g. rigid-body).
(b) Relationships can be aligned by a variety of methods, e.g. Monte Carlo simula-
tions (11), simulated annealing (7), double dynamic programming (10), genetic
algorithms.
3 the structures can subsequently be superposed according to the matches in the
alignment, but a single global superposition may be meaningless when large move-
ments, such as domain movements, have taken place. In that case, each domain
should be superposed separately.

Dynamic programming methods can align structures on the basis of differ-

ences/similarities between any number and combination of properties—which are
features of individual residues or segments of residues contiguous in sequence.
In order to compute the difference or similarity between positions in a struc-
ture, for example on the basis of Ca-Ca distances, an alignment is required to
give an estimate of the distances between atoms in the structures. Other struc-
tural properties, such as residue solvent accessibility, can be used with dynamic
programming directly, but may provide less useful information for the com-
parison. Relationships—features of multiple non-sequential residues (Table I): e.g.
patterns of hydrogen bonding, hydrophobic clusters, Ca-atom contact maps-
can also be compared. Monte Carlo simulations (11), simulated annealing (7), and
double dynamic programming (10) have all been used to equivalence relation-
ships among residue sets from structures. Each of these methods gives an align-
ment of the structures in the form of a sequence alignment, but to visualize the
results of the comparison, a rigid-body comparison would still be required. This
could be made over all matched positions or over those positions that matched
'best' according to the comparison criteria used. The global rigid-body super-
position based on the alignment may also be unsatisfactory if large structural
changes have taken place. To accommodate very large changes, such as domain
movements, the domains can be superposed separately.

2.3 Standard methods for finding the translation vector and

rotation matrix
For methods that compare the relative atomic positions in two structures, A and
B, and produce the superposed co-ordinates as output, it is necessary to deter-
mine a translation vector and the rotation matrix that, when applied to the
original co-ordinates, will generate the new co-ordinates for the superposed
proteins. Firstly, the centre-of-mass of the each protein is translated to the origin
of the co-ordinate system. Secondly, one of the structures is rotated about the
three orthogonal axes in order to achieve the optimal superposition upon the
other structure. Because the atoms chosen for comparison will not match
 COMPARISON OF PROTEIN THREE-DIMENSIONAL STRUCTURES

exactly in terms of their relative atomic positions after superposition, a least-

squares method is typically used to achieve the optimal superposition.
The following function minimizes the residual 8, which is expressed math-
ematically as:

where 5R is the rotation matrix being sought that minimizes the differences
between a total of N equivalent co-ordinate sets A"1 from the first protein and Beq
from the second protein; w is a weighting that can be applied to each ith pair of
equivalent positions.
Numerous methods have been developed to solve this pairwise least-squares
problem in a variety of different ways (12-16). Others have described more gen-
eral methods suitable for the least-squares comparison of more than two three-
dimensional structures (17, 18) In our experience, the method of Kearsley (19)
is a straightforward and simple means to obtain the optimal rotation matrix
for a set of equivalent co-ordinates. We will only consider this procedure here
(Protocol 2).
The major obstacle to solving the least-squares problem is that matched atom
pairs from the two structures to be compared need to be specified to the
algorithm at the beginning of any calculations. Thus, the computer program
requires some idea of the final alignment before it can proceed. There are
common situations where the comparisons would be made over a pre-defined
set of residues: for example, (a) comparisons over residues that line an active site
or binding site—to highlight similarities and differences over those positions; (b)
comparisons of independent structure solutions for the same protein. In these
cases, the atomic positions to be compared are usually known a priori, and a
single round of rigid-body comparison is sufficient to obtain the optimal match.
Frequently, however, global comparisons are made between proteins where the
best-matched positions are not obvious in advance. In the case of similar protein
structures, the requirement of an initial set of matches to seed the comparison
is inconvenient at best, requires the preanalysis of the proteins involved, and in
the case of more dissimilar proteins, may be difficult to define. Additionally, we
have often observed that when part of the answer is specified at the beginning
of the comparison, then the final solution can be prejudiced to give a final result
that is not necessarily the optimal one: The comparison was locked into a set of
possible solutions by the information supplied to seed the procedure. Despite
these criticisms, there are many good methods that employ this strategy.
For example, Sutcliffe et al (16) specify a set of at least 3 Ca-atoms common to
the two structures (3 positions define a unique plane in each structure). Good
candidates for these common residues, supplied a priori, can be conserved resi-
dues at an active site or ligand binding site, be positions conserved in terms of
the sequence similarity, or can be equivalent positions observed to form part of
the common fold when the proteins are examined on a graphics device. This
and other similar methods use an iterative procedure to progress towards better

21
MARK S. JOHNSON AND JUKKA V. LEHTONEN

and better solutions that incorporate more and more equivalent atom pairs. (Later
in this chapter we will detail several automatic strategies that have heen used to
get around this need for predetermining a set of equivalent atoms at the onset
of the structural comparison.)
In the equation describing the residual (above), Ae<l and B1'1* contain the x, y,
and z axes co-ordinates for exactly the same number of atoms from each of the
two structures. These atoms are termed equivalent positions, and are those aligned
positions that the superposition will now be calculated for. All other atoms in
the molecules are ignored in determining the superposition, but the translation
vector and the rotation matrix determined on the basis of these equivalent
positions is subsequently applied to all atoms in the co-ordinate file, including
any bound ligand, metal ions, and water molecules. Here, we will detail how to
calculate the translation vector for each protein and describe one simple yet
elegant method for determining the rotation matrix, developed by Kearsley (19,
20), which we use as the method of choice for our own procedures (Protacd 2).

Rigid-body structural comparisons: translations and

rotations
This protocol details the steps required to optimally superpose the equivalent atom co-
ordinates from two proteins.
Data required
The co-ordinates of all atoms in the proteins' co-ordinate file (minimum of the Ca-atom
co-ordinates) and the matched equivalent atoms in the two proteins.
The translation vector
1 Calculate the centre of mass from the x, y, and z co-ordinates for each set of equiva-
lent atoms Ae<i and $"* from the two structures. For N atoms in the equivalent set of
the first protein:

In other words, sum all of the x co-ordinates together and divide by N to give the
average x co-ordinate for the equivalent set of atoms; repeat for the y and z co-
ordinates. Repeat for the corresponding N equivalent atoms in the second structure:

Thus, the centre of mass is a single x, y, and z co-ordinate set for each of the proteins.
2 Translate both structures, all atoms in the file, so that their centres of mass (accord-
ing to the set of equivalent atoms used) are located at the origin of the co-ordinate
system. For every atom i in the first structure:

22
 COMPARISON OF PROTEIN THREE-DIMENSIONAL STRUCTURES

In other words, subtract the x, y, and z co-ordinate values for the centre of mass
from the x, y, and z co-ordinate values for every atom in the co-ordinate file. Repeat
for the second structure:

The rotation matrix: the Kearsley method (ref, 19) minimizes the average
difference between sets of atoms using quaternion algebra
1 Generate a symmetric 4 X 4 matrix by adding selected combinations of differences
and sums of co-ordinates calculated for each matched pair of equivalent atoms to
the elements of the matrix (19). These are the co-centred co-ordinates, but only the
coordinates of equivalent matched atom pairs, A^ and B^ are used at this stage.
2 Diagonalize the 4 X 4 matrix in order to obtain its eigenvalues and eigenvectors
(see ref. 21 for general procedures).
3 Select the lowest eigenvalue and use elements of the corresponding eigenvector to
construct the 3 x 3 rotation matrix 9? (see ref. 19 for details).
4 Multiplication of each co-ordinate in the second structure B by R will produce the
superposition of the entire structure onto protein A, where the average distance
between matched atoms of the equivalent set is a minimum: B (trans.,rot.) = R x
Bf (trans.).
5 The selected eigenvalue divided by the number of atom pairs in the equivalent set is
equal to the square of the RMSD after rotation. R, calculated above, leads to the
superposition whose RMSD is a minimum for these sets of equivalent atoms.

2.3.1 Structural alignment of sequences

In Figure 1, is shown the loss of superposed Ca-atoms in globin comparisons as
the percentage sequence identity decreases. As an alternative to rigid-body struc-
tural comparisons, especially when the rigid-body structural similarity is reduced
due to modest structural alterations, other methods have been developed that
provide the alignment of the sequences of the structures. Nonetheless, rigid-
body comparisons are often used in combination with these other procedures or
for visualisation of the results.
For example, the dynamic programming algorithm described below can make
comparisons on the basis of Ca-Ca atom distances, as well as other features (see
Table 1).
(a) As we have stated above, a rigid-body comparison is often needed in order to
make comparisons of structural properties suitable for dynamic program-
ming alignment.
(b) The dynamic programming method is often used in conjunction with rigid-
body super-positioning methods in order to efficiently assign equivalent
matches.

23
MARK S. JOHNSON AND JUKKA V. LEHTONEN

Figure 1 Reduction in the extent of the common equivalent matches in pairwise structural
superpositions as a function of decreasing percentage sequence identity. Traces of the
backbones are shown for Ca-positions within 2.5 A after rigid-body superposition with the
computer program MNYFIT (16). The haemoglobin a-chain of Pagothenia bemacchii (Protein
Data Bank (PDB, ref. 51) code: 1PBX) is aligned in (a) with the a-chain of equine
haemoglobin (2MHB) and in (b) with the B-chain of human haemoglobin (2HHB). (c) The
human haemoglobin B-chain (2HHB) aligned with the sea lamprey globin (2LHB). (d) The
erythrocruorin of Chironomous thummi thummi (1ECD) aligned with the leghaemoglobin of
Lupinus luteum (1LH1). (From ref. 4, with permission.)

(c) The dynamic programming algorithm produces a full alignment of all posi-
tions in the structures (residues are aligned with each other or with gaps),
while the rigid-body methods align fewer and fewer potions in the structures
as the sequence similarity decreases (Figure 1).
(d) Dynamic programming algorithms do not give a superposition of the
structures suitable for visualization. This can be obtained from the alignment
by applying the rigid body method to the defined matched pairs.
(e) Dynamic programming can often lead to alignments of the structures where
rigid-body movements have occurred in the structures themselves. For
example, the large movements of the entire domains seen in the liganded
and unliganded structures of the periplasmic bacterial lysine-arginine-
ornithine binding protein (Figure 2). Rigid-body comparisons can be applied,

24
 COMPARISON OF PROTEIN THREE-DIMENSIONAL STRUCTURES

Figure 2 Two different conformations of the 3-D structure for the same protein, the
lysine-arginine-omithine binding protein from Salmonella typbimurium. Left: the structure of
the protein in complex with lysine (1LST), lysine not shown. Right, the uncomplexed structure
(2LAO). The smaller domain on the upper part of the figures is in same orientation and the
arrow pointing to the Cot-atom of Glu 216 illustrates the magnitude of the movement of the
larger domain at the bottom of the figure,Figureprepared with MOLSCRIPT (52).

however, to the domains separately to pinpoint any changes within each

domain that have occurred upon ligand binding; while superpositioning on
one domain can be used to highlight the relative movements that have
occurred between the domains upon binding.

2.4 Standard methods to determine equivalent matched

atoms between structures
There is no exact definition of topological equivalence, and the criteria used can
vary from method to method. In rigid-body superposition methods, a distance
cut-off between equivalent atoms is frequently used. In methods were other
structural features are considered, alt aligned positions might be considered to
be topologically equivalent between two structures, or they may be assigned
according to the degree of positional similarity of features used to make the
comparison.

2.4.1 Definitions of structural equivalence: the alignment

In determining a set of equivalent atom sets, distance criteria are often used.
After one structure has been superposed on another, topological equivalent
atoms can be limited to those atom types under consideration that are within a
distance cut-off value. The Euclidean distance, D, between two points is:

In rigid body comparisons, where the Ca-atoms of the protein backbone have
been used as a basis for comparison, a distance cut-off typically in the range 2.5 A

25
MARK S. JOHNSON AND JUKKA V. LEHTONEN

to 4.5 A has been used. Values above 3 A lead to more multiple matches to a
single atom: the distance between two consecutive Ca-atoms along the protein
backbone is around 3.5 A. Lower values will reduce the number of equivalent
matches when more dissimilar proteins are compared. Distances or dissimilarity
measures will also be required for the comparison of other structural features,
both properties and relationships, see ref. 7 for example. Common to both rigid-
body methods, which rely on simple distance data, and other methods, which
incorporate other types of information into the alignment process, is the need
to determine the matching of locations between the structures to be superposed
(Protocol 3). This can be part of an iterative procedure to provide a new set of
equivalent atoms that are then used to determine a new translation vector and
rotation matrix in order to improve a match. This is also one of the final steps in
any comparison procedure, where the resultant alignment is determined. Three
basic approaches have been used: (a) dynamic programming, (b) graph theoret-
ical match list handling and clique detection methods, and (c) methods more
suitable for solving combinatorially-complex matching problems.
The Needleman and Wunsch (6) method is a convenient fast method for
aligning proteins. By scoring all possible pairs of matches between two struc-
tures, the method insures that the optimal scoring solution is found for the
scoring scheme employed. The method accommodates a loss of elements in one
structure relative to another—the gaps corresponding to insertions and de-
letions. Thus, the method provides a full alignment where every residue posi-
tion in each protein is matched to either a residue position in the other protein
or a gap. Thus, this method can efficiently resolve the multiple matching and
many combinatorial problems seen with the list sorting procedure. Once struc-
tural relationships have been equivalenced between a pair of structures, this
information can also be used within the dynamic programming method.
With the match list sorting procedure, for example ref. 22, possible equiva-
lent matches between the proteins are tabulated: matches of protein B to each
position in protein A in one list, and matches of protein A to protein B in a
second list. These lists contain both authentic matches of conserved structure,
chance matches that need to be eliminated from the lists and multiple matches
between one element in one protein to several different elements in the other
protein. The challenge, then, is to cull these lists by keeping the best matches
(i.e. matches that can extend a series of previous matches, have a good matching
score or give a good fit), removing structurally unlikely matches (matches that
are not co-linear—are out of sequence with other matches—and isolated matches
that do not extend further other matches), and by reducing multiple matches to
single matches.
A more elaborate approach was introduced by Mitchell et al. (23). Their
method does not filter out extraneous matches, but instead tests each com-
bination of matches to find the optimal equivalent set. As a result, a 'clique', the
maximum sub-graph common to two graphs representing the structures is
found. The clique detection algorithm is based on graph theory and offers a way
to find similar parts of structures that have not been superimposed. The basic

26
 COMPARISON OF PROTEIN THREE-DIMENSIONAL STRUCTURES

idea is to represent each structure as a graph of nodes and vertices. Each node
corresponds to either an atom, piece of main chain, secondary structure ele-
ment, or similar definite piece of structure. Each vertex is a relation between two
nodes in a structure: the distance between the atoms, vector from one atom to
another (both distance and direction in some co-ordinate frame), distance and
angle between two secondary structure elements, or more a more complicated
distance measure involving other properties of the nodes. If two structures
contain a similar substructure, then the nodes belonging to that substructure are
connected in both structures by very similar vertices. The task is to find the
maximal common sub-graph from the set of possible common sub-graphs. While
this is a NP-complete task, it is feasible due (a) efficient search algorithms evolved
within graph theory, and (b) the use of (few) secondary structure elements (SSEs)
as the compared pieces of the structures instead of (many) atoms. Several other
programs have been described that use a very similar approach (see ref. 24 and
citations therein); the main differences are in the ways structures are repre-
sented and in the method used to reduce the search space for efficiency.
The comparison of relationships among features in one structure relative to
another is a powerful addition to any structural comparison procedure (see ref. 7
for an excellent discussion). Relationships, such as patterns of hydrogen bond-
ing, involve the comparison of a minimum of two residue positions for eveiy
hydrogen bond in both structures. In certain cases, e.g. in the method of Taylor
and Orengo (10), relationships—in this case inter-atomic vectors, are compared
using their novel double dynamic programming method. More often, the
matching of relationships is treated as a combinatorially-intensive task. There
are lots of candidate pairs of hydrogen bonds in each structure and matching
them relies on methods such as simulated annealing, Monte Carlo simulations
and genetic algorithms.

The alignment: determination of equivalent pairs

Methods used to find the optimal match between entire structures or between parts of
structures consisting of the best matching regions of the structures. Equivalent or matched
positions are defined by the user (i.e. property distances within a cut-off value) or by the
strategy of the method employed (e.g. all matched positions produced by dynamic pro-
gramming methods).
Dynamic programming methods (6)
1 Construct a matrix with dimensions equivalent to the lengths of the structures to
be compared.
2 Each cell in the matrix corresponds to a residue in the first protein matching a
residue in the second protein. The matrix accommodates all possible alignments.
3 Cells are filled in with a score relating each two matched positions. These scores
may be distances between Ca-atoms, for example, distance scores based on other

27
MARK S. JOHNSON AND JUKKA V. LEHTONEN

features of the protein (see Table t), or similarity scores derived from distances. In
this description, we will refer to a matrix filled with similarity scores derived from
distances.
4 Beginning at one corner (ammo-terminal end or carbon-terminal end of the
sequences) of the matrix and heading towards the opposite corner, sum diagonal
values to the current position if they are the best score (a residue-residue match), or
sum with an off-diagonal score minus a penalty (indicates a possible gap in one
protein or the other).
5 The largest value found at one edge of the matrix specifies the first two aligned
positions and gives the optimal alignment score for the comparison.
6 The full alignment that produced the optimal score can be traced beginning at the
highest vahie and progressing towards the opposite side of the matrix by following
the next best score in the matrix. Wben the next highest value is on the diagonal,
residues are matched in sequence; when an off diagonal score (less a penalty) is the
next best choice, then a gap is indicated.
7 This method produces the full alignment including gap regions, but elements within
a cut-off value can be used to determine the rigid-body superposition of the structures.
Clique detection methods (23, 25)
1 Represent each structure as a graph of nodes (Ca-atoms or secondary structure
elements) and vertices connecting the nodes. Each vertex is a distance between the
connected two nodes (atoms).
2 List for each vertex in structure A all such vertices in structure B, which are similar
within an error threshold (Le. vertices connecting the same kind of nodes with
similar distances).
3 Find the maximal common sub-graph (largest set of nodes and vertices, which exists
in both structure graphs) using a tree search algorithm, Monte Carlo simulation, or
a genetic algorithm. Each vertex in the common sub-graph corresponds uniquely to
one vertex in both structures A and B.
4 The nodes included in the sub-graph are equivalent for the two structures. If the
nodes are atoms, the superposition can be made directly (see Protocol 2). Also, the
secondary structure elements can be superimposed as if they were atoms of a rigid
molecule, or the Co-atoms within the SSEs can be superimposed.
Match list approaches (22)
This method is a variation of the clique detection method, which assumes that the
structures are initially superimposed, but equivalent matches are not known.
1 In the case of Ca-Ca distance comparisons, create two lists, one for each protein A
and B.
(a) In one list, tabulate all Cot-atoms in protein B with matches within a cut-off
distance, say 3.5 A, to a position in protein A.

28
 COMPARISON OF PROTEIN THREE-DIMENSIONAL STRUCTURES

(b) In a second list, tabulate all Ca-atoms in protein A with matches within a cut-off
distance to a position in protein B.
2 Filter from the list the poorest matches to reduce the number of matches to a
unique set of equivalent matches:
(a) Remove matches that are not part of a contiguous run of at least 4 Ca-atoms.
(b)f Reduce multiple matches from one protein to a single Ca-atomin the other pro-
tein, e.g. does one of the matches extend a contiguous run of existing matches?
(e) If there are still multiple matches remaining, then the match with shortest
distance is kept and the others are removed.
Comparisons of relationships (7, 10, 11, 21)
1 The matching of relationships among features of one structure with relationships
among features of another structure is accomplished using one of several different
techniques.
(a) Monte Carlo simulations (11. 21).
(b) Simulated annealing (7, 21).
(c) Double dynamic programming (10)
(d) Genetic algorithms (22. 26, 27) can also be used.
2 The matched relationships may be insufficient in themselves to accurately align the
34) structures, and thus would be combined with the feature comparisons within a
dynamic programming procedure, for example, to give the final alignment (7).

2.5 Quality and extent of structural matches

Once a structural alignment has been made, a score or scores can be assigned to
the alignment that give an indication of the quality and the extent of matching
between the two structures. With methods that iteratively improve a. structural
comparison, an evaluation score is necessary to monitor the improvement at
each cycle of comparison, and to indicate when the program should stop be-
cause no further improvement in the alignment could be obtained. The final
alignment scores can be used to compare different protein comparisons within
a family and provide useful indications of the phyletic ancestry of the proteins
(e.g. 8, 28), Among the most frequently used key indicators of the'goodness' of a
structural comparison include the root mean squared deviation (RMSD), the
number of topologically-equivalent atoms matched in the comparison, and the
alignment score that is obtained,

2.5.1 Root mean squared deviations

The RMSD is commonly used to indicate the goodness of fit between two sets of
co-ordinates. Often, but not always, the RMSD value is quoted for only those
matched Ca-atoms that are within a specified distance cut-off, say atoms within
3.0 A of each other after the proteins have been superposed. In this case and

29
MARK S. JOHNSON AND JUKKA V. LEHTONEN

given the cut-off value of 3 A, the RMSD obtained and each of the Cu-Ca atom
distances contributing to the RMSD will be less than the 3 A. Alternatively, the
RMSD can be calculated over all matched Cot-atom pairs, regardless of the dis-
tance between the superposed atoms. Of course, the RMSD can also be calcu-
lated between sets of any type of superposed atoms, not just Ca-atom pairs as
illustrated in Protocol 4.

Root mean squared deviations (RMSD)

The RMSD gives a measure of the average level of deviations over the matched atoms that
are included in the calculation. Given the same number of equivalent atom pairs, a
smaller value indicates a better superposition than does a larger value.
Data required
• Co-ordinates of the equivalent sets fyq (trans.), Bf (trans., rot.).
Method
1 Calculate the Euclidean distance between each pair of equivalent atoms Af* (trans.)
and B^ (trans. ,rot.).
2 Take the sum of all squared distances D, and divide by the number of pairs, N, to
give the mean.
3 Calculate the square root of the mean squared distance to obtain the RMSD,
4 Thus, the

2.5.2 Topological equivalent atoms pairs

Another criterion that is used to gauge the extent or quality of a superposition
of two structures is the number of atom pairs that superpose within a distance
cut-off. Structure comparison methods usually try to maximize the number of
superposed equivalent atoms while minimizing the RMSD over those equivalent
atoms.
Note that two different sets of superposed structures, given the same cut-off
value, can have the same number of equivalent matches, but with different
RMSD values over those matches. The match with the lower RMSD would be
considered the more similar pair. Conversely, one structural comparison, for
example, may produce 121 matches with an RMSD of 2.1 A, while a second
comparison matches 50 atom pairs with an RMSD of 1.2 A: the comparison with
the 121 matches would be considered the better match.

2.5.3 Structural alignment scores

For structural alignment methods that employ dynamic programming in order
to produce a complete alignment of the structures, including gaps, a key measure

30
 COMPARISON OF PROTEIN THREE-DIMENSIONAL STRUCTURES

of the alignment quality is the alignment score corresponding to the overall

optimal structural superposition. This value includes scores for matching all
positions and penalties for every gap that appears in the alignment. The align-
ment score is composed of the values placed into the matching matrix during
the dynamic programming procedure. In the case of Ca-atom based compari-
sons, the residue-residue matching scores would be the distances between the
atoms. In the case of procedures that consider other criteria, e.g. those features
listed in Table 1, the alignment score would include the scores attributed to
matches of residue positions according to those features.
The raw alignment score is useful during iterative procedures to provide an
indication of the progress of the superposition. Within a family of homologous
3-D structures, the alignment score, normalized for the length of the smaller
protein or for the number of matched residues along the sequences, can be com-
pared to give an idea of the mutual structural relationships among the family
members.

3 The comparison of identical proteins

3.1 Why compare identical proteins?
The simplest type of comparison of 3-D structures involves the comparison of
two (or more) sets of co-ordinates for the same protein. Self-comparisons are
often used to reveal:
(a) Similarities/differences between independent solutions of crystal structures.
(b) Similarities/differences among sets of structures, generated using distance
geometry, and consistent with distance information obtained in NMR
spectroscopy.
(c) Similarities/differences between structures obtained using X-ray diffraction
and NMR spectroscopy.
(d) Similarities/differences that occur between apo- and holo-protein structures:
alterations in structure that occur upon binding ligands, cofactors, metal
ions, etc.
(e) Similarities/differences of two structures after superposing on an identical
ligand or subset of residues or co-ordinate positions.

3.2 Comparisons
In the comparison of identical proteins that have 3-D structures that differ to
varying degrees, it is needed to compare the structures using a rigid-body
approach one time only (Protocol 5). No iteration is necessarily required to
achieve the best result, since one would typically supply all atom positions in
the structure for comparison. Likewise, no pre-comparison is necessary to supply
a seed set of residues for the comparison. In practice, iterative procedures are
used. Again, if big differences in the structures are anticipated, e.g. the relative
domain movements in Figure 2, then this approach may not be appropriate

31
 except to provide an RMSD value that is an indication of the relative changes to
the structures.

Similarities among different structures of Identical

proteins
Finds regions of high structural similarity between different solutions of the structure of
the same protein.
Data required
• Coordinates, minimum ofCa-atoms, for all structures
Method
1 No alignment between the proteins is necessary* since the proteins are identical
and each position maps 1:1 in sequence along the protein.
2 A single application of a comparison algorithm (see Protocol 2) is sufficient to
obtain the optimal result over all of the compared atoms.
3 Calculate the RMSD over all atoms or those within the cut-off distance, as desired
(see Protocol-*).
4 Iterative methods (see below), seeded by some key positions, can be used also.
5 By adjusting the cut-off value used to define equivalent matched atoms to lower
values, the most similar structural regions may be identified and hence, the differ-
ences pinpointed too.
"Note that different data sets from different sources do not necessarily contain the same
atnino acids or atoms for the same protein.

4 The comparison of homologous structures:

example methods
4.1 Background
Most comparison programs are designed to compare non-identical homologous
structures, but they can be also used to superpose structures for the same pro-
teins as described in Section 3. There are a large number of different programs
and strategies that have been published and we have necessarily had to select
just a few as illustrations—our apologies to any author who feels that we have
neglected their own work. In general, the methods fall into two different groups:
(a) Those that require the advance definition of pairs of suspected 'equivalent'
atoms in order to seed the alignment. An iterative procedure is then used to
maximize the number of equivalent matched atom pairs while minimizing
the RMSD.

32
 COMPARISON OF PROTEIN THREE-DIMENSIONAL STRUCTURES

(b) Those methods that sample the realm of possible solutions and, as a result,
automatically find optimal alignments without specifying an initial starting
alignment.

Some of these procedures involve rigid-body comparisons and others generate a

full alignment of the sequences on the basis of the structures. InFigure3, we
show the extreme differences in results for the same proteins obtained with a
multiple sequence alignment, a rigid-body structure comparison, and a pro-
cedure (7) that combines the comparison of properties and relationships to
derive the structural matching.

Figure 3 The differences in alignments of the aspartic proteinase amino- and carboxyl-
terminal domains (labelled with an 'N' or 'C', respectively) from (a) multifeature (7) and from
(b) multi-sequence comparisons. Asterisks in (a) indicate those positions among the
structures that were found to be equivalent under rigid-body superposition with the computer
program MNYFIT (16). PDB codes: 4APE, endothiapepsin; 2APP, penicillopepsin; 2APR,
rhizopuspepsin. (From ref. 8, with permission.)

33
MARK S. JOHNSON AND JUKKA V. LEHTONEN

4.2 Methods that require the assignment of seed residues

As we have already discussed above, a set of seed matches between a pair of
structures is often needed by methods in order to initiate the comparison of
structures, because some residue properties, such as Cn-Ca distances, require a
partially correct alignment in order to calculate these distances. Once seeded,
the alignment improves over several rounds of comparison. Obvious candidates
for seed residues are listed in Protocol 6.

Finding initial seed residues

Required data
• Sequences and/or co-ordinates of the proteins to be compared

Method
1 Supply a minimum of three conserved residues from a sequence-based alignment,
or
2 Supply key residues implicated in a conserved binding or catalytic motif, or
3 Supply segments corresponding to secondary structure elements observed on a
graphics device to be conserved between the structures.

In Protocol 7, we present a general procedure for the alignment of two structures

using rigid-body comparisons, which requires a seed set of matches between the
two 3-D structures.

Semi-automatic methods
Required data
• Co-ordinates of the proteins to be Initial set of equivalent atom pairs to seed
compared the alignment procedure

Method
1 Calculate translation vector based on seed residues, translate all co-ordinates to the
origin and calculate the rotation matrix for the seed residues (see Protocol 2).
2 Apply the rotation matrix to all atoms of the second protein to achieve the first
superposition (see Protocol 2).
3 Obtain the alignment using dynamic programming or clique analysis (see Protocol
3).

34
 For all matched residue pairs in the alignment, calculate the Euclidean distance.
Those matched pairs within the distance cut-off value will form the new updated
set of equivalent atom pairs for the next round of super-positioning.
Repeat steps 1-4 until convergence is obtained: the calculated RMSD (Protocol 4)
does not decrease and the number of equivalent atom pairs matched in the two
proteins does not increase.
A rigid-body comparison has been used as a starting point for more detailed structural
comparisons involving multiple structural features (e.g. the program COMPARER
described in ref. 7).

4.3 Automatic comparison of 3-D structures

To get around the requirement for an initial set of equivalent seed matches,
alternative methods have been developed. Here, several of the many published
methods are described to illustrate the different strategies that have been
employed:

(a) Methods that supply seed matches automatically to a rigid-body approach

after making a sequence-based alignment (Protocol 8).
(b) Methods that use a genetic algorithm (Protocol 9) or 'spectra'-comparison
method (Protocol 10) to find the optimal rigid-body comparison.

Structural comparisons seeded from sequence

alignments
Automatic alignment of two homologous protein structures without the need to
specify initial equivalent atoms pairs. Method can fail for proteins of low sequence simi-
larity.

Required data
• Ca-atom co-ordinates of the compared structures and their sequences

Method
1 Align the amino acid sequences with a dynamic programming algorithm (Protocol
3, but using sequence-matching scores to produce the alignment).
2 Superimpose the structures according to Protocol 7 using the most conserved portions
of the sequence alignment as the initial set of seed residues.

35
MARK S. JOHNSON AND JUKKA V. LEHTONEN

(c) Methods that do not make rigid-body comparisons directly, but instead make
comparisons on the basis of similarities in structural properties and/or
relationships (Protocols 11-14).

4.3.1 Structural comparisons seeded from sequence alignments

Russell and Barton (9) developed a method that first makes an alignment of the
sequences and then uses equivalent matches defined in the alignment to seed a
structural comparison. We have given a general protocol for such an approach
above {Protocol 8). The procedure should work well for proteins where portions of
the sequence alignment can be trusted; when the sequence similarity is low and
the alignment is not correct, then the method may not be useful.

4.3.2 Rigid-body comparisons using a genetic algorithm

Genetic algorithms (29) describe the solution to a problem within a numerical
string. A large number of strings are originally assigned random values as their
solutions, and the genetic algorithm seeks"to evolve this initial set towards better
and better solutions by exchanging partially good solutions among strings and by
mutating the strings. Protocol 9 describes the general procedures used by May and
Johnson (26, 27) to automatically compare one or more structures. The approach
is time-consuming but has been successfully adapted to parallel processors
(Lehtonen and Johnson, unpublished).

GA_FIT (ref. 26, 27)

Automatic rigid-body alignment of two protein structures without the need to specify
initial equivalent atoms pairs.

Required data
• Cat-atom coordinates of compared structures

Method
1 Create a large random set of superpositions for the pair of structures,
2 Assign equivalent matches (Ca-atoms within a specified distance cut-off) using
dynamic programming and score each alignment (see Protocol 3).
3 Create a new set of superpositions by crossing-over and mutating the existing
solutions.
4 Repeat steps 2-3 until a close to final solution is achieved.
5 Optimize the best found superposition/alignment by least squares minimization
(see Protocol 2).
6 Calculate the final alignment with the dynamic programming algorithm (see
Protocols).

36
 COMPARISON OF PROTEIN THREE-DIMENSIONAL STRUCTURES

4.3.3 Rigid-body comparisons using the density of Ca-atom packing

and spectral alignment
VERTAA, described in Protocol 10 (Lehtonen and Johnson, unpublished method),
compares and aligns spectra equal to the Ca-atom density in each structure as a
function of the position along the sequence of each protein (Figure 4). This
method is a rapid and automatic means for comparing structures.

Local similarity search by VERTAA

Fast, automatic alignment of two protein structures without the need to specify initial
equivalent atoms pairs.
Required data
• Ca-atom co-ordinates of the two structures.

Method
1 VERTAA, for each of two structures, plots the number of Cot-atoms within a
given radius (14.0 A) from each Cot-atom in the structure. Other properties can be
used too.
2 These 'spectra' are scaled and overlapping segments are aligned. More than one
alignment method is available:
(a) The dynamic programming algorithm (Protocol 3). Fast and robust if the input
values are properly scaled.
(b) The Fourier correlation (21). The values can be considered as a function over a
limited range and a correlation function obtained with the fast Fourier trans-
form to bring the spectra into register. Dynamic programming is then used to
define equivalent matches (Protocol 3).
3 Superimpose the structures (see Protocol 2) based on equivalent matches denned in
step 2.
4 Define a new alignment with dynamic programming and the Ca-Ca distances of the
superimposed structures within 3.5 A (see Protocol 3).
5 While the alignment and superimposition improve, repeat steps 3 and 4,

4.3.4 Structural comparisons based on matching Ca-atom contact

maps
Holm and Sander (11) make comparisons by comparing Ca atom-Ca atom con-
tact maps (by contacts, we mean nearby in space) constructed from each protein
structure (Protocol 11).

37
MARK S. JOHNSON AND JUKKA V. LEHTONEN

38
 COMPARISON OF PROTEIN THREE-DIMENSIONAL STRUCTURES

Figure 4 Plots of Co-atom densities, alignment of plots, and the corresponding superposition
of the structures, (a) Cu-atom densities of residues in -y-chymotrypsin A (PDB code 2GCH).
(b) C«-atom densities of residues in Streptomyces griseus proteinase B (PDB code 3SGB,
chain E). For both spectra, the average density is set equal to 0. (c) The parts of the two
plots from (a) (dark) and (b) (light), which correspond to each other according to the alignment
of their spectra, (d) The superpositioned 3-D structures (2GCH dark, 3SGB light) based on the
alignment specified in (c). The side chains of the catalytic triad are shown and the ctosely
matching parts are drawn as ribbon diagrams. This superposition was made with the
computer program VERTAA (Lehtonen and Johnson, unpublished results) and contains 118
residues within 3.5 A with an RMSD of 1.8 A, Figure (d) was prepared with MOLSCRIPT (52).

Structure comparison by DALI (11)

Automatic alignment by finding the optimal clique for contact maps obtained from the
structures (Protocol 3).
Required data
• Coc-atom co-ordinates of compared structures

Method
1 Calculate a distance matrix for each protein A. Element (i, j) of the matrix contains
the intramolecular distance between the ith and fth Ca-atom in A. Likewise, calculate
a distance matrix for protein B.
2 List from each distance matrix all possible 6 by 6 sub-matrices.
3 Reduce the number of sub-matrices by clustering similar ones and using the mean of
each cluster as the contact pattern. Sort contact patterns by intra-pattern distance.
4 Compare each pair of two contact patterns from A with all pairs of sub-matrices
from B. Compare each pair of two contact patterns from B with all pairs of sub-
matrices from A. List all pair-pair matches.
5 Remove redundancy from the list of matches and son it by match quality, which is
a function of the differences between the sub-matrices from A and from B,
6 Find the most extensive, non-exclusive collection of matches from the list. DALI
uses a Monte Carlo simulation to search the best 40000 matches. The simulation
tries to extend the matches by combining matches that contain a common contact
pattern in both distance matrices. The random element of the simulation is used to
find the best scoring combination from mutually exclusive possibilities.

4.3.5 Comparisons using double dynamic programming

Taylor and Orengo (10) have developed a novel use of dynamic programming in
order to facilitate the comparison of relationships. Dynamic programming is
used once to compare structural relationships in the two proteins thus providing
scores for a second round of dynamic programming where the two structures
are aligned (Protocol 12),

39
MARK S. JOHNSON AND JUKKA V, LEHTONEN

Structure comparison by SSAP (10)

Fast automatic alignment of two protein structures using double dynamic programming.
Required data
• Cot-atom co-ordinates of compared structures

Method
1 Calculate a distance matrix for protein A Element (J, j) of the matrix contains the
intramolecular vector from the 4th to the j"1 Ca-atom in A. The vector is in the co-
ordinate frame defined by the covalent bonds of A's 1th Ca-atom. Likewise, calculate
a distance matrix for protein B.
2 Calculate intramolecular difference matrices for each pair of rows from the two
distance matrices. Thus, element (i,j) of the matrix constructed from row h of A's,
and row k of B's distance matrix will contain the difference of the magnitude of the
tt CO

vectors Aw and B^ converted to a similarity value.

3 Low level alignments of the local structure are made first using a dynamic program-
ming algorithm (see Protocol 3) to find the best scoring path through the intra-
molecular difference matrices (see ref. 10 for details). Scores along the path will
contribute to a separate 'summed scoring matrix1 from which the final alignment
will be determined.
4 Use a dynamic programming algorithm (see Protocol 3) to trace an alignment path
through the summed scoring matrix. This higher level alignment defines the equiva-
lent matches between the structures.

4.3.6 Structural alignments based on secondary structure element

{SSE} matching
Kleywegt and Jones (30) describe a method for structural comparisons based on
the alignment of elements of regular secondary structure (Protocol 13}.

Structure comparison by DEJAVU (30)

Automatic alignment of protein structures by finding the optimal clique on the basis of
secondary structure comparisons (Protocol 3),
Required data
• Ca-atom co-ordinates of the two structures or SSE templates of the structures
Generation of SSEs wfth YASSPA in O (see ref. 30 for details)
Search the structures and tabulate main chain fragments that are similar to templates of
typical a-helices and (S-strands.

40
 COMPARISON OF PROTEIN THREE-DIMENSIONAL STRUCTURES

Comparison of structures with DEJAVU (see ref. 30 for details)

1 Check that both structures have the required number of SSEs.
2 Check that there exists at least one SSE of the same type (same length—number of
Ca-atoms) in the second structure for each SSE in first structure.
3 Find the most extensive, non-exclusive collections of matched SSEs. DEJAVU does a
depth-first tree search to find all sets of matching SSE pairs, where all pairs in a set
are matching also in 3-D space. The tree contains all possible combinations of pairs.
If the path from the root to a node already has too many mismatches, the sub-tree
below the node is not searched, saving time.
4 Report the matched SSEs and the Ca-atoms for the best scoring alignment
5 The output can be directed to external programs for refinement of the super-
position and visualization.

4.4 Multiple structural comparisons

Multiple structural comparisons can be made using several different strategies.
Sutcliffeetal. (16} constructed multiple rigid-body structural alignments by com-
paring each structure to an average representation of the structures (in practice,
one of the structures was chosen for this purpose at the beginning of the com-
parisons). More frequently, multiple alignments are assembled from pairwise
structural alignments according to the topology of a tree estimated on the basis
of sequence alignments (9, 27, inter alia). This (Protocol 14) follows the strategy
first introduced by Barton and Sternberg (31) and Feng and Doolittle (32) for the
efficient multiple alignment of protein sequences.

Multiple structural alignments from pairwise comparisons

Multiple alignments assembled from pairwise comparisons,
Required data
• Co-atom co-ordinates of the structures in PDB format

A general approach
1 Use a sequence alignment procedure to align the proteins and to cluster them as a
bifurcating tree (see refs 31-34 and several chapters in ref. 35).
2 Use a pairwise structural alignment method to align clusters according to the tree
topology. This will involve comparing pairs of structures, one structure with a set of
previously aligned structures, and aligned structures with aligned structures, until
all clusters have been coalesced into a final alignment involving all of the proteins.

41
MARK S. JOHNSON AND JUKKA V. LEHTONEN

5 The comparison of unrelated structures

5.1 Background
Non-homologous protein structures have frequently been compared to high-
light features of protein structure that are common across many families. It has
been less often recognized, however, that proteins with different folds can also
share similarities that can extend to a fairly large organization of their struc-
tures, for example about common ligand and cofactor binding sites. The ele-
ments contributing to these similarities are likely to involve fragments of each
structure that do not map along the protein chains in any predetermined way
(Figure 5 and ref. 22). Thus, despite similar local structure, the segments con-
tributing to the similarity can be both rearranged and discontinuous with
respect to each other (e.g. 36-38). Such similarities are particularly difficult to
recognize even if a hint of a common functional requirement is present, like a
common cofactor. Nonetheless, the recognition of local similarities can provide
evidence about the rules governing the structure-function relationship suitable
for protein modelling, the prediction of structure from sequence and computer-
based drug design. For example, Kobayashi and Go (39) have reported a local
motif about the ATP binding site common to cyclic-AMP dependent protein
kinase and D-Ala:D-Ala ligase involving 4 equivalent residues. Comparisons
using the computer program GENFIT (22) automatically and repeatedly found
up to 60 matches (36) that includes an extensive supersecondary structure
organization used to position polar and nonpolar residues that interact with the
similarly oriented cofactor, bound metal and bound water molecules (Figure 6).
Given two unrelated protein structures, A and B, the goal of a computer pro-
gram is to find the largest equivalent subset of the two structures. Because the
proteins are not derived from a common ancestor, the matches providing
equivalent structural interactions:
(a) Are not necessarily sequential along the two sequences (Figure 5).
(b) Can involve matched elements of secondary structure whose chain directions
are opposite to each other (Figure 5) but can still provide equivalent inter-
actions, for example, with bound ligand.
Here we describe two different approaches that have been successfully used
to find similarities among unrelated protein structures, SARF2 (40, 41) and
GENFIT (22). SARF2 (41) considers SSEs, finds the maximal common sub-graph
for two structures; and systematically creates different alignments, tries to im-
prove them, evaluates them and reports the best found alignments (Protocol 15).
GENFIT (22) considers matched segments of Ca-atoms and employs a genetic
algorithm to randomly sample large numbers of possible alignments and uses a
match-list approach to assign equivalent segments of structure, which are sub-
sequently used to make a local rigid-body superposition for each alignment
(Protocol 16). GENFIT, by virtue of the genetic algorithm, will find and report
different equally likely superpositions in different runs (Figure 7).
Both of these approaches establish equivalent matches between objects,

42
 COMPARISON OF PROTEIN THREE-DIMENSIONAL STRUCTURES

Figure 5 Illustration of the matching of local structural similarities from non-homologous

proteins versus homologous proteins, (a) Topological diagrams show four protein structures
(a, b, c, and rf] with similar local structural elements. Topologies a and b represent two
homologous structures with the same fold, while c- has a different topology than a and b, yet
has the same core structure. Topology d illustrates a different fold that still has the
structurally equivalent segments of polypeptide chain in same place, but some segments
may have opposite chain directions. In (b), the correspondence found in the structural
alignment is shown at the sequence level. Note that only the sequences of a and b have a
straightforward linear correspondence. (From ref. 22, with permission.)

segments of Ca-atoms (GENFIT) or SSEs (SARF2). GENHT starts with 'too many'
equivalent matches and reduces them until a maximal, but non-conflicting set,
is obtained. This is done for each of the many parallel comparisons being made,
but a single optimal result will be obtained in any one run: the parallel com-

43
MARK S. JOHNSON AND JUKKA V. LEHTONEN

Figure 6 Local similarity of ATP cofactor binding site seen in the pairwise superposition of
riPonucleotide reductase (PDB code: 3R1R, chain A; light grey) with cAMP-dependent protein
kinase (1CDK; dark grey) (a and b), and with D-Ala:D-Ala ligase (HOW; dark grey) (c). The
bound ATP molecules of the structures are shown as stick models. In (a), the four common
segments are drawn as ribbon diagrams, (b and c) The environment around the cofactor is
illustrated by showing the equivalent hydrogen bonds (dashed lines) and equivalent Cn-atoms
(spheres) forming hydrophobic contacts to the cofactor. (From ref. 37, with permission.)

parisons converge towards that result. SARF2 searches among a large set of
matches between the structures and finds the largest non-conflicting subset of
matches. Both methods are free from restraints on the order and chain direction
of objects along the sequence, but optional restraints can be applied.

44
 COMPARISON OF PROTEIN THREE-DIMENSIONAL STRUCTURES

Structure comparison by SARF2 (41)

Local similarity alignment of non-homologous structures.

Required data
• Ca-atom co-ordinates of the two structures

Method
1 Search for and tabulate main-chain fragments from the structures that are similar
to five-residue long templates of typical a-helices and p-strands,
2 Create a list of SSE pairs from the first structure that match SSE pairs from the
second structure. Distance and angular criteria between the SSE's in both structures
is important to the determination of a match.
3 Combine matches to find the largest collection of SSE's that can be aligned. SARF2
uses an exhaustive, recursive search algorithm to find possible solutions (see ref. 41
for details).
4 For the best solutions found, superimpose the matched SSE's and then add nearby
Ca-atoms to matched regions using the dynamic programming method. Iteratively
repeat the superpositions of Ca-atoms until the maximum number of matched
atoms have been found.
5 A list of superpositions, ranked according to an alignment score, result.

GENFIT (22)
Automatic alignment of two locally similar protein structures using a genetic algorithm.
This implementation has been designed for parallel processing environments.

Required data
• Ca-atom co-ordinates of the two structures

Method
1 Create a large random set of superpositions for the pair of structures.
2 Assign equivalent matches using the match list algorithm (see Protocol 3). Criteria
for a match include:
(a) Ca-atorn matches must be within a user specified distance cut-off.
(b) Matches must include a minimum of four consecutive Ca-atoms.
(c) The direction of the main chain for matched segments is unimportant by default.
(d) Matches do not need to be co-linear (i.e. the location of a match along the
sequence relative to other matches is unimportant).

45
MARK S. JOHNSON AND JUKKA V. LEHTONEN

3 Calculate an alignment score for each superposition and create a new set of
superpositions by crossing-over and mutating existing ones (see ref. 22 for details).
4 Repeat steps 2 and 3 until convergence has been achieved.
5 Optimize the best superposition/alignment by least-squares rigid-body minimization
(Protocol 2).
6 Recalculate the alignment with the match list algorithm (Protocol 4).
7 If the number of equivalent matches has increased or the fit has improved, repeat
steps 5 and 6 with the current alignment.
8 Repetitive runs can produce different results showing that equally likely alternative
results exist.

6 Large-scale comparisons of protein structures

One of the straightforward goals in bioinfbrmatics today is to compare, cluster
and classify both sequences and the known 3-D structures. Initially, this means
categorizing each existing sequence or structure in a data bank. Then, when
new entries are made to sequence and structure data banks, each new entry will
need to be compared against the existing classifications.
The methods described in this chapter can and have been applied to such
analyses. For example, both MNYFIT (16) and COMPARER (7) have been used to
accurately align all families of 3-D structures containing two or more structures
(43-45), that can be accessed in a public database: http://vww-CTyst.bioc.com.ac.uk/
cgi-binljoy.cgi. Other available databases include FSSP (46) created using DALI (11):
7ittp://wvw.eW.uc.ufc/ddi//ssp/; and CATH (47) created in part using SSAP (10):
http:Jlwww.biochem.ud.Qc.uk/tem/rath/. Several other data banks worth mentioning
include MMDB (48): http://www.ticbi.nlm.nih.gov/Stnjrture/ and SCOP (49): hop:
Jlscop, mrc- Imb. cam. ac. uk/scopl
In Figure 8, we present a classification of structures from several different
families that belong to the all-p structural classification. This classification was
made by comparing the structures on the basis of their secondary structures and
then clustering them according to the pairwise structural similarity (44, 50),
Figure 7 Two examples of differing alignments of locally similar structures, (a and b)
Superposition of UDP-ga lactose 4-e pirn erase chain B (2UDP) and DMA methyltransferase
(1HMY) showing similarity between the larger domains. In (b), 1HMY has been rotated by
180 degrees around the axis of the p-sheet in comparison to (a). The symmetry of the nearly
planar p-sheet allows for several different, but similarly-scoring alignments, (c and d)
Superposition of cyclic-AMP-dependent protein hinase (1CDK) and glutamine synthetase
(1LGR) showing local similarities about the ATP-binding sites and the differences seen from
matching fewer longer segments (a-helices) or many shorter segments (B-sheets). In (c), the
antiparallel B-sheets are aligned and the cofactors overlap, while in (d) the ex-helices are
matched, but the p-sheets and the cofactor do not superpose well. The superpositions have
been made with program GEN FIT (22). (From ref. 22, with permission.!

46
MARK S. JOHNSON AND JUKKA V. LEHTONEN

Figure 8 Dendogram of clusters of protein structures composed primarily from B-strands.

Each cluster, a family of proteins, is distinguished from the others by its unique fold. (From
ref. 44, with permission.)

References
1. Dayhoff, O. M, Barker, W. C., and Hunt, L. T. (1983). In Methods in enzymology (ed.
C. H. W. Hirs and S. W. Timashefi). Vol. 91, p. 524. Academic Press, London.
2. Chothia, C. (1992). Nature, 357, 543.
3. Blundell, T. L and Johnson, M. S. (1993). Protein Sri., 2, 877.
4. Johnson, M. S., Srinivasan, N., Sowdhamini, R., and Blundell, T. L. (1994). Crit. Rev.
Biochem. Mol. Biol., 29, 1.
5. Robertus, J. D., Alden, R. A., Birktoft, J. J., Kraut, J., Powers, J. C., and Wilcox, P. E.
(1972). Biochemistry, 11, 2449.
6. Needleman, S. B. and Wunsch, C. D. (1970). J. Mol. Biol., 48, 443.
7. Sali, A. and Blundell, T. L. (1990). J. Moi. Biol., 212, 403.
8. Johnson, M. S., Sali, A., and Blundell, T. L (1990). In Methods in enzymology (ed. R. F.
Doolittle), Vol. 183, p. 670. Academic Press, San Diego.
9. Russell, R. B. and Barton, G. J. (1992). Proteins, 14, 309.
10. Taylor, W. R. and Orengo, C. A. (1989). J. Mol BioL, 208, 1.
11. Holm, L. and Sander, C. (1993). J. MoJ. Biol., 233, 123.
12. McLachlan, A. D. (1972). Acta Crystallogr., A28, 656.
13. McLachlan, A. D. (1979). J. Mol. Biol., 128, 49.
14. McLachlan, A. D. (1982). Acta Crystallogr., A38, 871.
15. Diamond, R. (1988). Acta Crystallogr., A44, 211.
16. Sutcliffe, M. J., Haneef, L, Carney, D., and Blundell, T. L. (1987). Protein Eng., 1, 377.
17. Diamond, R. (1992). Protein Sri., 1, 1279.

48
 COMPARISON OF PROTEIN THREE-DIMENSIONAL STRUCTURES

18. Shapiro, A., Botha, J. D., Pastor, A., and Lesk, A. M. (1992). Acto Crystallogr., A48, 11.
19. Kearsley, S. K. (1989). Acta Crystallogr., A45, 208.
20. Kearsley. S. K. (1990). J. Comput. Chem., 11, 1187.
21. Press, W. H., Teukolsky, S. A., Vetterling, W. T., and Flannery, B. P. (ed.) (1992).
Numerical recipes in C. The an of scientific computing. (2nd edn). Cambridge University
Press, Cambridge.
22. Lehtonen, J. V., Denessiouk, K., May, A. C. W., and Johnson, M. S. (1999). Proteins, 34,
341.
23. Mitchell, E. M., Artymiuk, P. J., Rice, D. W., and Willett, P. (1990). J. MoL Biol., 212,
151.
24. Gibrat, J.-F., Madej, T., and Bryant, S. H. (1996). Curr. Opin. Struct. Btol., 6, 377.
25. Grindley, H. M., Artymiuk, P. J., Rice, D. W., and Willet, P. (1993). J. Mol. Biol., 229 (3),
707.
26. May, A. C. W. and Johnson, M. S. (1994). Protein Eng., 7,475.
27. May, A. C. W. and Johnson, M. S. (1995). Protein Eng., 8, 873.
28. Johnson, M. S., Sutcliffe, M. J., and Blundell, T. L (1990) J. Mol. Bvol., 30,43.
29. Goldberg, D. E. (ed.) (1989). Genetic algorithms in search, optimization, and machine learning.
Addison-Wesley, Reading, MA.
30. Kleywegt, G. J. and Jones, T. A. (1997). In Methods in enzymology (ed. C. W. Carter and
R. M. Sweet), Vol. 277, p. 525. Academic Press.
31. Barton, G. J. and Steinberg, M. J. (1987)./. Mol. Biol., 198, 327.
32. Feng, D. F. and Doolittle, R. F. (1987). J. Mol. Bvol., 25, 351.
33. Johnson, M. S. and Overington, J. P. (1993). J. Mol. Biol., 233, 716.
34. Johnson, M. S., May, A. C. W., Rodionov, M. A., and Overington, J. P. (1996). In Methods
in enzymology (ed. R. F. Doolittle, Vol. 266, p. 575. Academic Press.
35. Doolittle, R. F. (ed.) (1996). In Methods in enzymology. Vol. 266, p. 711. Academic Press,
San Diego.
36. Denessiouk, K., Lehtonen, J. V., Korpela, T., and Johnson, M. S. (1998). Protein Set, 7,
1136.
37. Denessiouk, K., Lehtonen, J. V., and Johnson, M. S. (1998). Protein Set, 7,1768.
38. Denessiouk, K., Denesyuk, A. I., Lehtonen, J. V., Korpela, T., and Johnson, M. S. (1999).
Proteins, 35, 250.
39. Kobayashi, N. and Go, N. (1997). Nature Struct. Biol, 4, 6.
40. Alexandrov, N. N., Takahashi, T., and Go, T. (1992)./. Mol. Biol., 225, 5.
41. Alexandrov, N. N. and Fischer, D. (1996). Proteins, 25, 354.
42. Sali, A., Overington, J. P., Johnson, M. S., and Blundell, T. L. (1990). Trends Biochem. Set.,
15, 235.
43. Overington, J. P., Johnson, M. S., Sali, A., and Blundell, T. L (1990). Proc. R Soc. Lond. B,
241,132.
44. May, A. C. W., Johnson, M. S., Rufino, S. D., Wako, H., Zhu, Z.-Y., Sowdhamini, R., et ol.
(1994). Phil. Trans. R. Soc. Lond. B, 344, 373.
45. Mizuguchi, K., Deane, C. M., Blundell, T. L, Johnson, M. S., and Overington, J. P.
(1998). Bioin/ormotics, 14, 617.
46. Holm, L and Sander, C. (1998). Nucleic Acids Res., 26, 316.
47. Orengo, C. A., Michie, A. D., Jones, S., Jones, D. T., Swindells, M. B., and Thornton, J. M.
(1997). Structure, 5,1093.
48. Marchler-Bauer, A., Addess, K. J., Chappey, C., Geer, L, Madej, T., Matsuo, Y., etal.
(1999). Nucleic Acids Res., 27, 240.
49. Murzin, A. G., Brenner, S. E., Hubbard, T., and Chothia, C. (1995). J. MoL Biol., 247,
536.
50. Rufino, S. D. and Blundell, T. L. (1994). J. Comput. Aided Mol. Des., 8, 5.

49
MARK S. JOHNSON AND JUKKA V. LEHTONEN

51. Bernstein, F. G, Koetzle, T. F., Williams, G. J. B., Meyer, E. J. Jr, Brice, M. D., Rodgers,
J. K., etal. (1977).J. Mol. Biol, 112, 535.
52. Kraulis, P. J.(1991).J. Appl. CrystaUogr., 24, 946.

50
 Chapter 3
Multiple alignments for
structural, functional, or
phylogenetic analyses of
homologous sequences
L. Duret
Laboratorie de Biome'trie, Genetique et Biologie des Populations, Universite
Claude Bernard, France.

S. Abdeddaim
Departement d'lnformatique de Rouen, Universite de Rouen, France.

1 Introduction
Understanding the structure, function and evolution of genes is one of the main
goals of genome sequencing projects. Classically, gene function has been investi-
gated experimentally through the analysis of mutant phenotypes. More recently,
comparative analysis of homologous sequences has proved to be a very efficient
approach to study gene function (this approach has been coined 'comparative
genomics' or 'phylogenomics'). Indeed, the evolution of living organisms may be
considered as an ongoing large-scale mutagenesis experiment. For more than
three billion years, genomes have continuously undergone mutations (substitu-
tions, insertion, deletions, recombination, and so on). Deleterious mutations are
generally rapidly eliminated by natural selection, while mutations that have no
phenotypic effect (neutral mutations) may, by random genetic drift, eventually
become fixed in the population. Globally, advantageous mutations are very rare,
and hence residues that are poorly conserved during evolution generally corres-
pond to regions that are weakly constrained by selection (1). Thus, studying
mutation patterns through the analysis of homologous sequences is useful not
only to study evolutionary relationships between sequences, but also to identify
structural or functional constraints on sequences (DNA, RNA, or protein).
The alignment of homologous sequences consists of trying to place residues
(nucleotides or amino acids) in columns that derive from a common ancestral
residue. This is achieved by introducing gaps (which represent insertions or de-
letions) into sequences. Thus, an alignment is a hypothetical model of mutations

51
L. DURET AND S. ABDEDDAIM

(substitutions, insertions, and deletions) that occurred during sequence evolution.

The best alignment will be the one that represents the most likely evolutionary
scenario. Generally, this best alignment cannot be unambiguously established.
Firstly, because of the computational complexity of this problem, alignment
algorithms that are usable in practice cannot guarantee to find the best solution
(see Section 4.1). Secondly, even with an ideal algorithm, finding the best align-
ment would not be guaranteed because current knowledge of the probability of
occurrence of the different types of mutation is still limited (see Section 2.3).
However, as long as homologous sequences are not too divergent, fast approxi-
mate algorithms may be used to provide reliable alignments. In practice, such
alignments are commonly used in molecular or evolutionary biology. Typical
examples of usage of multiple alignments are indicated in Table 1.

Table 1 Examples of usage of multiple alignments

• Identification of functionally Important sites

Multiple alignments allow the identification of highly conserved residues are likely to
correspond to essential sites for the structure or function of the sequence and may thus be
useful to design mutagenesis experiments.

• Demonstration of homology between sequences (see Section 2.1)

• Molecular phylogeny
Molecular phylogenetic trees rely on multiple alignments (protein or DNA) to infer mutation
events from which it is possible to retrace evolutionary relationships between sequences.
Such trees are useful to reconstruct the history of species or multigenic families, and notably
to identify gene duplication events to distinguish orthologues from paralogues. It is important
to note that unreliable parts of alignments should not be used to build phylogenetic trees
since they do not reflect the real pattern of mutations that occurred during evolution and may
lead to artifactual results.

• Search for weak but significant similarities In sequence databases

The sensitivity of sequence similarity search may be improved by weighting sites according to
their degree of conservation. Thus, multiple alignments of homologous sequences are used
by methods such as profile searches (see the chapter by Henikoff in this volume) or PSI-
BLAST (24) to identify distantly related members of a family.

• Structure prediction
The use of multiple alignments increases significantly the efficiency of protein secondary
structure prediction. Moreover, the identification of covariant sites (or compensatory
mutations) in alignments (protein or RNA) is a strong argument to suggest that these sites
interact in the molecule In vivo. Finally, alignments are commonly used for homology
modeling, i.e. for the structure prediction of sequences by comparison with homologues of
known structure.

• Function prediction
The three-dimensional (3D) structure of homologous proteins or RNA is often much more
conserved than their primary sequence. Similar shape usually implies similar function. Thus, if
a new gene is found to be homologous to an already characterized gene it is possible to infer
the likely function of the new gene from the known one. Such inferences should however be
used with great caution.
• Design of primers for PCR (polymerase chain reaction) Identification of related genes

52
MULTIPLE ALIGNMENTS FOR STRUCTURAL, FUNCTIONAL, OR PHYLOGENETIC ANALYSES

The general procedure to compute a multiple alignment of homologous

sequences consists of three steps:
(a) Search for homologues in sequence databases.
(b) Compute alignments.
(c) Check and edit alignments.
In this chapter, we will focus, essentially, on steps (b) and (c). Firstly, we will
define some general concepts underlying multiple alignment methodology. We
will then describe and compare different methods that have been developed to
align sequences. As far as possible, we will indicate WWW sites where these tools
are available, so that they may be used from any computer with an appropriate
WWW browser software and internet connection. The list of WWW sites that
we provide here is also available at the following address:
http://pbil.univ-lyonl.fr/alignment.html
Some problems such as contig assembly, related to multiple alignment will
not be treated in this chapter. We will only describe methods intended for the
alignment of homologous sequences and, in particular, we will not deal with the
problem of finding common motifs in a set of unrelated sequences. Note that
there is not an absolute difference between motif search and multiple align-
ments: when homologous sequences have diverged too much there may remain
only a few short conserved fragments, separated by regions of variable length.
Motif-based methods have been developed to identify and align such conserved
fragments within highly divergent sequences. In Section 4.4 we will mention
some of these methods. However, for a more exhaustive review on this topic, see
Chapter 7 by Jonassen in this volume.

2 Basic concepts for multiple sequence alignment

2.1 Homology: definition and demonstration
Two sequences are said to be homologous if they derive from a common ancestor.
Generally, homology is inferred by sequence similarity. It should be stressed,
however, that similarity does not necessarily reflect homology: similarity be-
tween short sequence fragments may result from evolutionary convergence (2),
or may simply occur by chance. Moreover many sequences contain relatively long
fragments of very biased nucleotide or amino acid composition (e.g. CA-repeats in
DNA, proline-rich domains in proteins) (3). Generally, similarities between such
'low complexity regions' do not reflect evolutionary relationship. However, in the
absence of such compositional bias, similarity over an extended region usually
implies homology. Statistical tests can be used to evaluate the chances that an
observed similarity occurred purely by chance and thus accept or reject the
hypothesis of homology (4). Such tests are now generally provided by similarity
search programs.
Multiple alignments may be useful to help demonstrate homology: a weak

53
L. DURET AND S. ABDEDDAIM

similarity which would be considered as non-significant in a pairwise sequence

comparison may prove to be highly significant if the same residues are con-
served in other distantly related sequences. It should be emphasized that if
sequences have diverged too much, homology may not be recognizable on the
basis of sequence similarity alone.

2.2 Global or local alignments

In the above paragraph, we implicitly considered sequences that are homologous
over their entire length. However, in many cases, homology is restricted to a
limited region of the sequences. Indeed, many proteins consist of a combination
of discrete 'modules' that have been shuffled during evolution. It is clear that
many protein-coding genes result from recombination between different frag-
ments of other genes. This modular evolution has played a major role in protein
evolution and has been particularly facilitated in eukaryotes thanks to the
presence of introns within genes (5).
Multiple copies of a given module may be repeated within a sequence, and a
set of modules may occur at different relative positions in different genes. In
such cases, it is not possible to align sequences over their whole length (global
alignment) and it is thus necessary to perform alignments only on homologous
modules (local alignment) See Figure 1 for an illustration.

2.3 Substitution matrices, weighting of gaps

As indicated earlier, searching for the best alignment consists of searching for
the one that represents the most likely evolutionary scenario. Thus, the prob-

Flgure 1 Global versus local alignment, (a) Conserved regions occur in the same order in all
sequences. They can be represented in a single global alignment, (b) Some conserved
regions are duplicated or occur in a different order along sequences. It is necessary to
perform local alignments to display similarities between all conserved regions.

54
MULTIPLE ALIGNMENTS FOR STRUCTURAL, FUNCTIONAL, OR PHYLOGENETIC ANALYSES

ability of occurrence of the different mutational events during evolution must

be taken into consideration when computing a multiple alignment. In align-
ments, three types of mutations are considered: substitutions, insertions or
deletions (the two latter events are often indistinguishable, and are commonly
referred as 'indels').

2.3.1 Substitutions
The probability of substitution of one amino acid by another depends on the
structure of the genetic code (i.e. on the number of mutations necessary to pass
from one codon to another) and also on the phenotypic effect of that mutation.
Substitutions of one amino acid by another with similar biochemical properties
generally do not greatly affect the structure and hence the function of the
protein. Thus, during evolution, such conservative substitutions are relatively
frequent compared to other substitutions. It is important to note that the prob-
ability of substitution of one amino acid by another depends on the evolutionary
distance between sequences. At short evolutionary distances, probabilities of
substitution mainly reflect the structure of the genetic code, whereas at larger
distances, probabilities of substitution depend essentially on biochemical simi-
larities between amino acids. Various methods have been proposed to build
series of matrices that give estimates of probabilities of all possible substitutions
for different evolutionary distances (6-8). The most commonly used are the PAM
and BLOSUM substitution matrices. PAM matrices suitable for increasing
evolutionary distances are indicated by increasing indices (e.g. PAM80, PAM120,
and PAM250). The opposite convention has been used for the BLOSUM series (e.g.
BLOSUM80 for short evolutionary distances, BLOSUM45 for large evolutionary
distances). Generally, alignment programs allow users to choose which substitu-
tion matrix to use. In the CLUSTAL W program (9) (see Section 4.2) substitution
matrices are automatically selected and varied at different alignment stages
according to the divergence of the sequences to be aligned.
Probabilities of substitutions also vary along sequences according to the local
environment of amino acids in the folded protein. Thus, several environment-
specific substitution matrices have been developed (e.g. for ot-helix, or p-sheet)
(10). However, to our knowledge, these matrices are rarely used for multiple
alignments.
At the DNA level, probabilities of substitution vary according to the bases.
Notably, transitions (substitutions between two purines—A, G—or two pyri-
midines—C, T) are generally more frequent than transversions (substitutions
between a purine and a pyrimidine). Thus, multiple alignment programs
generally propose a parameter to weight more heavily transversions than
transitions. Probabilities of nucleotide substitution also depend on neighbouring
bases (e.g. in vertebrates, C in CG dinucleotides is hypermutable) (11, 12).
However, currently available alignment programs do not make use of such
information.

55
L. DURET AND S. ABDEDDAIM

2.3.2 Insertions, deletions

The probability of occurrence of an indel depends on its length. Thus, when
computing an alignment, penalties (p) associated with gaps are often estimated
using a linear or 'affine' model such as:

where L is the length of the gap, a the gap opening penalty, and b the gap
extension penalty. However, analyses of alignment of homologous sequences
have shown, both for protein and nucleic sequences, that this model under-
estimates the probability of long indels (7, 13, 14). Indeed, more realistic indel
penalties can be estimated with models of the following form:

However, because of computational complexity, such models have not been im-
plemented in commonly used alignment programs. Fortunately, other approaches
have been proposed to align sequences with large indels (see Section 4.3).
The probability of occurrence of indels in proteins also depend on the degree
of divergence between sequences (7, 13). Thus, as for amino-acids substitution
matrices, indel penalty parameters should ideally be varied according to the
divergence of the sequences to be aligned. The probability also depends on the
nature of the sequences: protein, structural RNA, non-coding DNA (in which
transposable elements may be inserted), etc. Moreover, probabilities of indel
may vary along sequences. In proteins notably, indels are more frequent within
external loops than in the core of the structure. Thus, knowledge on the
structure of proteins can be used to weight indels. For example, the CLUSTAL W
program uses residue specific indel penalties and locally reduced indel penalties
to encourage new gaps in potential loop regions rather than in regular second-
ary structure. In cases where secondary structure information is available, indel-
penalty masks can also be used to guide the alignment.
It is important to note that, in most programs, default parameters for gap
penalties have been set for typical globular proteins. These may not be optimal
for other sequences.

3 Searching for homologous sequences

The first step in the analysis of a family of homologous sequences consists of
searching for all available members of that family. Published sequences are
stored in databases: GenBank (15) or EMBL (16) for nucleic acid sequences and
SWISSPROT-TREMBL (17) or PIR (18) for protein sequences. Retrieval systems
such as Entrez (19), SRS (20), or ACNUC (21) have been developed to query those
databases and extract sequences according to the associated annotation (e.g.
keywords, taxon, authors). Some WWW addresses for commonly used database
retrieval systems are shown in Table 2. Unfortunately, it is not possible to rely on
the annotation to identify in a database all homologous sequences belonging to
a given family. Presently, the most efficient way to identify those homologues

56
Table 2 Websites for text-based searches in sequence databases

Entrez at NCBI http://www.ncbi.nlm.nih.gov/Entrez/

SRS at EBI http://srs.ebi.ac.uk/
WWW-QUERY at PBIL http://pbil .univ-lyonl.fr/
ExPASy http://www.expasy.ch/sprot/
DBGET at GenomeNet http://www.genome.ad.jp/

Table 3 Websites for sequence similarity searches in databases

BLAST at NCBIa http://www.ncbi.nlm.nih.gov/BLAST/
WU-BLAST at EMBLb http://dove.embl-heidelberg.de/Blast2e/
FASTA at EBI http://www2.ebi.ac.uk/fasta3/
Smith-Waterman search at EBI http://www2.ebi.ac.uk/bic_sw/
BCM search launcher http://gc.bcm.tmc.edu:8088/search-launcher.html
BLAST at PBILC http://pbil.univ-lyonl.fr/BLAST/blast.html

' Possibility to select BLAST output results by taxa.

* Performs multiple alignment on homologous sequences detected by BLAST.
c
Possibility to select BLAST output results by taxa or keyword.

consists in taking one member of the family and comparing it to the entire
database with a similarity search program such as FASTA (22) or BLAST (23, 24).
To guarantee a more exhaustive search, one may repeat this procedure with
several distantly related homologues identified in the first step. See the review
by Altschul, et al. (25) for a comprehensive discussion of sequence similarity
searches.
The sensitivity of a sequence similarity search may be improved by weighting
sites according to their degree of conservation. Thus, once several homologous
sequences have been identified, it is possible to use methods such as profile
searches (see Chapter 5 in this volume) or PSI-BLAST (24) that rely on a multiple
alignment to identify more distantly related members of the family. A list of
some similarity search WWW servers is presented Table 3.

4 Multiple alignment methods

Once homologous sequences have been identified, which program should be
preferentially used to align them? Several multiple alignment methods (algo-
rithms) have been developed, but none of them is ideal. Thus, it is important to
have an idea of what these algorithms try to solve, in order to make an informed
choice of the most appropriate method(s) for a particular problem. The multiple
alignment problem is algorithmically hard: methods that guarantee to find the
best alignment (for a given measure of alignment score and for a given set of
substitution matrix and gap penalty parameters) require so much time and
space (memory) that they cannot be used in practice with, say, more than 10 to
15 sequences of length 100. Thus, alternative algorithms have been developed

57
L. DURET AND S. ABDEDDAIM

using heuristics to gain speed and limit space requirements. Although these
heuristics do not guarantee to find the optimal alignment, they are very useful
in practice and often give results very close to the exact solution. In the follow-
ing we will focus on four families of multiple alignment algorithms:
(a) Algorithms that guarantee to find the optimal alignment for a given scoring
scheme; these algorithms can be used only for a limited number of short
sequences.
(b) Heuristic algorithms that are based on a progressive pairwise alignment
approach.
(c) Heuristic algorithms that build a global alignment based on local alignments.
(d) Heuristic algorithms that build local multiple alignments.
It should be noted that this list is not exhaustive. Other multiple alignment
methods such as those based on hidden Markov models (26) or genetic algorithms

Table 4 Websites for multiple alignments

• Optimal global multiple alignment
MSA at IBC http://www.ibc.wustl.edu/ibc/msa.html
• Progressive global multiple alignment
ClustalW at EBIa http://www2.ebi.ac.uk/clustalw/
ClustalW, Multalin at PBILb http://pbil.univ-lyonl.fr/
MAP, ClustalW at BCM http://kiwi.imgen.bcm.tmc.edu:8088/search-
launcher/launcher.html
Multalin at INRAb http://www.toulouse.inra.fr/multalin.html
ClustalW at Pasteur0 http://bioweb.pasteur.fr/seqanal/alignment/intro-uk.html
ClustalW at DDBJ http://www.ddbj.nig.ac.jp/searches-e.html
MAP http://genome.cs.mtu.edu/map.html
• Block-based global multiple alignment
DCA at BiBiServ http://bibiserv.techfak.uni-bielefeld.de/dca/
DIALIGN2 at BiBiServ http://bibiserv.TechFak.Uni-Bielefeld.DE/dialign/
DCA at Pasteurc http://bioweb.pasteur.fr/seqanal/alignment/intro-uk.html
DIALIGN2 at Pasteurc http://bioweb.pasteur.fr/seqanal/alignment/intro-uk.html
ITERALIGN at Stanford http://giotto.stanford.edu/~luciano/iteralign.html
• Motif-based local multiple alignment
MEME at SDSC http://www.sdsc.edu/MEME/
MEME at Pasteur http://bioweb.pasteur.fr/seqanal/motif/meme/
MATCH-BOX http://www.fundp.ac.be/sciences/biologie/bms/matchbox_
submit.html
BLOCK Maker at FHCRC http://www.blocks.fhcrc.org/blockmkr/make_blocks.html
PIMA at BCM http://kiwi.imgen.bcm.tmc. edu:8088/search-
launcher/launcher.html
PIMA II at BMERC http://bmerc-www.bu.edu/protein-seq/pimall-new.html
a
Possibility to display and edit alignment with the JALVIEW JAVA applet.
b
Coloured alignments.
c
In combination with many WWW tools for molecular phylogeny.

58
MULTIPLE ALIGNMENTS FOR STRUCTURAL, FUNCTIONAL, OR PHYLOGENETIC ANALYSES

Table 5 Software for multiple alignments

ClustalW (UMPV)a ftp://ftp-igbmc.u-strasbg.fr/pub/ClustalW/
ClustalX (UMPV)a ftp://ftp-igbmc.u-strasbg.fr/pub/ClustalX/
(ClustalW + graphical interface)
Multalin http://www.toulouse.inra.fr/multalin.html
a
MSA (U) http://www.ibc.wustl.edu/ibc/msa.html
blALIGN(U)" http://www.gsf.de/biodv/dialign.html
DCA (U)a http://bibiserv.techfak.uni-bielefeld.de/dca/
RIW/DNR (U)a ftp://ftp.genome.ad.jp/pub/genome/saitama-cc/
MACAW (MP)a ftp://ftp.bio.indiana.edu/molbio/align/macaw/
a
Availability: U = UNIX , M = Macintosh, P = PC, V = VMS.

(27) can also be used. For a review of multiple alignment algorithms see refer-
ence (28).
Many of the programs reviewed here can be used directly through the WWW
(see Table 4) or downloaded over the Internet to be installed on a local computer
(see Table 5).

4.1 Optimal methods for global multiple alignments

In this section, we will mention several methods that are said to be optimal,
because they guarantee to find the 'best' multiple alignment among all possible
solutions for a given scoring scheme. It should be stressed that the term
'optimal' is taken here in its mathematical meaning. Whether a mathematically
optimal alignment corresponds or not to the biologically correct alignment (i.e.
the alignment that represent the most likely evolutionary scenario) will depend
on the choice of parameters (weighting of substitutions and of indels, see Section
2.3) and on the way the multiple alignment is scored.

4.1.1 Scoring schemes for multiple alignments

In principle, the score of a multiple alignment should reflect its likelihood
(according to a given evolutionary model). There are different ways to measure
the score (or cost) of a multiple alignment. In the following we consider that a
sequence is an ordered set of letters taken from an alphabet X. An alignment of n
sequences S1, . . . ,Sn can be defined as a matrix a(S1, . . . ,Sn) = A, where each entry
Aij is either a letter from 2 or a null symbol (the gap symbol, usually denoted
by -). The row i from A is the sequence Si, after gaps are removed.
In the simplest model, the cost of an alignment of n sequences is defined as
the sum of the cost of its columns. However, this model is crude because each
column of the alignment is considered independently of its context (i.e. a gap of
length I is considered as corresponding to L independent indels).
In more realistic models, a gap is interpreted as one single mutational event
(a deletion or an insertion of I residues) and associated with a cost that depends
on its length (see Section 2.3.2). With such models, pairwise alignment costs are
defined as the sum of substitution and gap costs. However, the definition of the

59
L. DURET AND S. ABDEDDAIM

multiple alignment cost is more complex. One possible solution, known as the
Sum of Pairs (SP) alignment cost (29), consists of calculating the multiple
alignment cost from pairwise alignment costs. A multiple alignment a(S1,... ,Sn)
contains n(n-l)/2 pairwise alignments a(Si.,Sj) where 1 < i < j < n. Each projection
a(Si.,Sj) is the pairwise alignment built from a(S1,... ,Snj by removing all the rows
except the rows i and j, and then by removing all the columns that contains two
null letters. The SP multiple alignment cost is defined as the sum of all its
projections costs (29).
Simple SP alignment cost may, however, be inappropriate when some groups
of sequences are heavily over or under-represented in a family. This drawback
may be corrected by introducing a proper weighting system (30, 31) which
assigns a weight to each sequence. This can be used to give less weight to
sequences from overrepresented groups. Another solution consists of using a
cost function based on an evolutionary tree. The tree leaves are the sequences we
want to align, and the internal nodes are their hypothetical ancestral sequences.
For a given tree, the cost of an alignment a(Sl, ... ,Sn) is the sum of all its
projections afSi.,SjJ on adjacent sequences Si and Sjin the tree (32).

4.1.2 Algorithmic complexity of optimal multiple alignment

methods
The optimal alignment is the one with the maximal score (or the minimal cost).
Needleman and Wunsh (33) proposed an efficient algorithm, based on dynamic
programming, to compute this minimal cost for pairwise alignments. This
dynamic programming approach can be easily generalized to more than two
sequences. However, computing the minimal alignment cost of n sequences, each
of length 1 requires o(2n I") time and o(Fj space (i.e. time proportional to 2" 1" and
computer memory proportional to 1") and the complexity is even higher if gap
cost are not linear (see Section 2.3.2). Such an algorithm cannot be used, in practice,
for much more than three sequences. For example, to align ten sequences of
length 100, on a very fast computer that would need 10~9 sec to compute the score
for one column of a multiple alignment, it would take approximately three
million years (210 1001010~9 = 1014 sec) to compute the alignment. This assumes
we have approximately ten billion giga-bytes of memory.
Carrillo and Lipman (29) proposed a branch and bound algorithm to compute
a minimal SP cost alignment. This algorithm uses an upper bound of the
alignment SP cost to limit the space and time used by dynamic programming.
This approach is implemented in the program MSA (34). A new version of MSA
with substantial improvements in time and space usage is available (35). Despite
these improvements, MSA cannot easily be used for more than about 10 short
sequences.
As stated previously, cost functions based on an evolutionary tree are, in
principle, better than SP alignment costs to measure the likelihood of an align-
ment. However, the alignment problem under an evolutionary tree is even harder
than the SP alignment problem, as the algorithm has to find the alignment, the

60
MULTIPLE ALIGNMENTS FOR STRUCTURAL, FUNCTIONAL, OR PHYLOGENETIC ANALYSES

tree, and the ancestral sequences such that the alignment cost is minimal. The
problem remains hard even if the tree is given (36).

4.2 Progressive global alignment

Progressive alignment is the most commonly used method to align biological
sequences. This heuristic approach is very rapid, requires low memory space and
offers good performance on relatively well-conserved, homologous sequences
(37, 38).

4.2.1 Description of progressive alignment methods

Progressive alignment consists of building a multiple alignment using pairwise
alignments in three steps:
(a) Compute the alignment scores (or distances) between all pairs of sequences.
(b) Build a guide tree that reflects the similarities between sequences, using the
pairwise alignment distances.
(c) Align the sequences following the guide tree. Corresponding to each node in
the tree, the algorithm aligns the two sequences or alignments that are
associated with its two daughter nodes. The process is repeated beginning
from the tree leaves (the sequences) and ending with the tree root.
Depending on the algorithms, steps (b) and (c) are done separately, or merged in
one step where the tree topology is deduced from the progressive alignment
process.
figure 2 illustrates the progressive alignment process. SI is first aligned
with S2 following the given tree, S3 is then aligned with S4, then the two
alignments a(Sl,S2) and a(S3,S4) are aligned together, and finally S5 is aligned with
a(Sl,S2,S3,S4). Notice that even if a(Sl,S2) and ot(S3,S4j are optimal alignments
computed by dynamic programming, the progressive alignment approach does
not guarantee that a(Sl,S2,S3,S4) is optimal for a multiple alignment cost func-
tion (SP cost or the tree cost for example).
A great number of tools that use a progressive alignment approach have been
proposed, they differ by the methods used in at least one of the three steps.
In the first step (a) the pairwise alignment cost can be computed by dynamic
programming, or by heuristic algorithms. The multiple alignment program
CLUSTAL W (9) for example allows one to choose either dynamic programming
or a heuristic method. Dynamic programming gives more accurate scores but is
slower than heuristic methods.
Different algorithms can be used to build a tree (step b) given a distance
matrix between sequences. Following Feng and Doolittle (37), early versions of
CLUSTAL (39) used the UPGMA algorithm (40). However, UPGMA is notorious for
giving incorrect branching orders when rates of substitution vary in different
lineages. Therefore, CLUSTAL W (9) now uses the Neighbor-Joining (41) algorithm
to build the guide tree.
The main problem in the third step (c) consists of aligning two alignments.

61
L. DURET AND S. ABDEDDAIM

Figure 2 Progressive alignment process, (a) All sequences are compared to each other S2.
(b) A guide tree is calculated from the pairwise distance matrix, (c) Sequences are
progressively aligned following the guide tree.

The simplest method for this problem reduces each alignment to a consensus
sequence, and uses a pairwise alignment algorithm to do the work. In the con-
sensus sequence, each column of the alignment is represented by its most
frequent letter. Consensus alignment was used in the first version of CLUSTAL.
In most programs, each alignment is considered as a profile (see Chapter 5). hi a
profile, a column is reduced to a distribution giving the frequency of each letter.
Two profiles are aligned as two sequences by dynamic programming without
major modification of the algorithm. The alignment of two profiles of length I
takes ofa2!2), where a is the alphabet size. CLUSTAL W uses profile alignment
with position-specific gap penalties (see Section 2.3.2).

62
MULTIPLE ALIGNMENTS FOR STRUCTURAL, FUNCTIONAL, OR PHYLOGENETIC ANALYSES

4.2.2 Problems with progressive alignment methods

An important problem with this progressive alignment approach stems from
the 'greedy' nature of the algorithm: any mistakes that appear during early
alignments cannot be corrected later as new sequence information is added. For
example, suppose that we have to align three sequences (x, y, z). Consider a short
fragment of these sequences for which the optimal alignment is:
X ACTTA
y A-GTA
z ACGTA

Suppose that the guide tree based on pairwise comparison of entire sequences
indicates that we should first align sequence x with sequence y, followed by the
alignment of sequence z with the first two (already aligned together). At the first
step, there are three possible alignments of x and y giving exactly the same
score:
x ACTTA x ACTTA X ACTTA
y A-GTA y AGT-A y AG-TA

At the later step, the gap that was introduced cannot be changed. Thus adding
sequence z could give the following three alignments:
X ACTTA x ACTTA x ACTTA
y A-GTA y AGT-A y AG-TA
z ACGTA z ACGTA z ACGTA

Only the first of these alignments is optimal. At the first step, only one of the
three possibilities will be used. If it is the wrong one, we cannot correct this
later.
To avoid that problem, iterative optimization strategies such as RIW or DNR
(42) have been proposed. These methods are reported to perform better than
CLUSTAL W (42). However, although these methods are much faster than
optimal algorithms, they are still to slow for large dataset.
Another limitation of the progressive approach described above is that it
requires computing pairwise distances between all sequences to calculate the
guide tree. One may sometimes have to align set of homologous sequences that
include some non-overlapping fragments (e.g. partial protein sequences). When
sequences are non-overlapping they are obviously completely unrelated and
thus the guide tree generated may be totally false. The alignment produced in
this case can be unpredictable.

4.3 Block-based global alignment

The sequences to be compared may share conserved blocks, separated by non-
conserved regions containing large indels. In such cases, the result of optimal or
progressive global alignment methods will depend greatly on the choice of gap
penalty parameters. An alternative to these approaches consists of searching for

63
 Figure 3 (a) Consistent set of blocks, (b) Non-consistent set of blocks.

conserved blocks that will be used as anchors in order to align the sequences.
Blocks are alignments of fragments (segments) of sequences (local alignments).
Most methods consider gap-free blocks. Depending on the programs used, the
blocks allowed can be exact (composed of identical segments) or not exact and
they may be uniform (found in every sequence) or not. The selected set of blocks
must be consistent, i.e. the blocks can occur together in a multiple global
alignment (figure 3). Once blocks have been computed, it is possible to use a
classical approach to align regions between blocks (e.g. ref. 43).
The first multiple block alignment program (44) used a sorting algorithm in
order to compute uniform exact blocks. Faster algorithms based on suffix trees
(45), or equivalent data structures, can also be used to compute exact blocks.
However, homologous regions are rarely exactly conserved. ASSEMBLE (46) per-
forms a dot matrix analysis on all pairs of sequences and then compares these
dot matrices to find uniform blocks that are not necessarily exact. In practice, it
often happens that some blocks are not present in all sequences. Thus, a further
improvement has consisted of developing methods that allow blocks that are
not necessarily uniform. DIALIGN (47,48) is based on computing gap-free blocks
between pairs of segments (diagonals).
A set of uniform blocks is consistent when each pair of blocks is ordered (they
do not cross each other). Using this observation, selecting an optimal consistent
set of blocks can be reduced to a classic optimal-path algorithm in a graph (44).
The optimal-path algorithm requires o(M2) time for M blocks. Faster algorithms
(sub-quadratic) have been proposed in order to compute an optimal consistent
uniform set of blocks (49, 50). However, finding an optimal consistent set of non-
uniform blocks is an intractable problem (51). Indeed, the consistency of non-
uniform blocks cannot be reduced to a binary relation between them. A set of
three non-uniform blocks, such that all its three pairs of blocks are consistent, is
not necessarily consistent. To compute a 'good' consistent set of diagonals,
DIALIGN uses a heuristic algorithm in which diagonals are incorporated by de-

64
MULTIPLE ALIGNMENTS FOR STRUCTURAL, FUNCTIONAL, OR PHYLOGENETIC ANALYSES

creasing score order into a consistent set of diagonals. Diagonals not consistent
with the set of selected diagonals are rejected In order to check if a new diagonal
is consistent or not with the set of selected diagonals, DIALIGN maintains a data
structure in o(kL2) time, for k sequences of total length I. This makes it slower
than progressive alignment programs. This computation time can however be
reduced to o(k2L + I2) (52) and even, thanks to recent developments, to o(L2).
Thus, faster versions of block-based alignment methods should be available in
the near future.

4.4 Motif-based local multiple alignments

The sequences to compare may share similar regions, without necessarily being
globally related. These homologous modules may occur in different relative
positions and may be duplicated in different sequences. In such cases it is not
possible to compute a global alignment, but one may look for 'good' local align-
ments of segments taken in the sequences. Calculating local alignments consists
of finding approximate repeated patterns in a set of sequences. Dynamic pro-
gramming has been adapted in order to find the maximal diagonal score for
pairwise comparison (53). For more than two sequences the problem is hard and
heuristics are needed as for the global multiple alignment problem.
PRAUGN (54) computes consensus words for a given word length. For each
possible word w of length k one may define the neighbourhood of w as the set of k
length words whose score with w is higher than a given cut-off. The score of w is
then the sum of all the scores with his neighbours that occur in the given
sequences. PRALIGN tries to compute the best score words (consensus words) of
fixed length. The main problem with this program is its space requirement: for
a fixed length k the space used is proportional to 20k (for proteins). This space
requirement could be much reduced using automatons as it is done in BLAST.
The MACAW method (55) combines pairwise comparisons in order to com-
pute multiple local alignments. In a first step, MACAW marks, for each pair of
sequences, all the diagonals with significant scores. The diagonals are then merged
into local alignments. MACAW is generally considered too time-consuming for a
general local alignment method, as it needs o(L2) time for the first step (I is the
sum of the sequence lengths).
Most recent local alignment programs are based on statistical methods.
Statistical methods use computationally efficient heuristics in order to solve
optimization problems. GIBBS (56) uses iterative Gibbs sampling in order to find
blocks. The computation time of this approach grows linearly with the number
of input sequences. GIBBS is available in the programs MACAW and Block Maker
(57). The tool MEME (58) uses an expectation-maximization (EM) algorithm (59,
60) to locate repeated patterns.

4.5 Comparison of different methods

When sequences are similar (say more that 50% pairwise identity for proteins,
70% for DNA) and are homologous over their entire length, all global alignment

65
L. DURET AND S. ABDEDDAIM

methods give more or less correct results. Moreover, in such cases, any reason-
able set of parameters (substitution matrix, and usually, gap opening and gap
extension penalties) will give similar alignments. However, when at least two
sequences in a given family share less identity, or if homologous regions are
interrupted by large gaps of different sizes, the result of alignment may vary
considerably according to programs and parameters used.
Several comparative analyses of multiple alignment programs have been
published (42, 48, 61, 62). These comparisons are based on the ability to detect
motif patterns on several protein families or based on reference alignments
derived from three-dimensional protein structures. Comparative analysis can
also be based on the effect of the multiple alignment programs on phylogeny.
Such a study was done on 18S rDNA from 43 protozoan taxa (63). These com-
parisons must be taken only as indications. Indeed, the parameter values (sub-
stitution matrices, gap penalty, etc.) used in these comparisons may not be
optimal for other sequence families (61). In addition these parameters are not
really comparable, even if the programs use the same strategies. For example a
gap opening score of 5 does not have the same meaning in CLUSTAL W (9) as it
does in MULTAL (38), as the value 5 will be modified in the programs (multiplied
by constants for example). For these reasons and because no known method
guarantees to find the correct alignment, it is still necessary to combine different
methods from different families of algorithms and human expertise to obtain
satisfactory alignments.
Figure 4 summarizes indications to guide users in their choice according to
the sequences they have to align. For the alignment of two sequences, one
should use an optimal pairwise alignment method (for example LALIGN or SIM
(64), see Table 6). For more than two sequences, one generally has to use heuristic
approaches. As a first step, the user should try to compute the multiple
alignment with a progressive alignment program. These programs are rapid, do
not demand large memory capacity and may thus be run on large dataset even
on micro-computers. Among programs using this approach, we recommend
CLUSTAL W (or its graphical user interface version: CLUSTAL X) (65, 66). This
includes useful features such as automatic selection of amino-acid substitution
matrix during alignment and lower weighting of gaps in potential protein loops.
If this first alignment shows that all sequences are related to each other over
their entire lengths, it is unlikely that any other method will give a better result
(Figure 4a).
However, if there are some highly divergent sequences, large gaps, or poorly
conserved regions it is -ecommended to compare the results of different methods
and/or sets of parameters. Figure 4b shows homologous sequences sharing con-
served blocks separated by non-conserved regions of varying size. This situation,
which is frequently observed in practice (e.g. in genomic DNA sequences and in
many protein families), is particularly error prone for progressive alignment
methods, notably because the linear weighting of gaps tends to over-penalize
long indels. Block-based global methods (e.g. DIALIGN, ITERALIGN) (47, 48, 67)
are not sensitive to these long gaps and are particularly appropriate for such

66
 Comment Appropriate approach

Sequences are related over their Progressive global alignment

entire length. method (e.g. CLUSTAL W).

Sequences share conserved blocks, Block-based global alignment

separated by non-conserved method (e.g. D1ALIGN,
regions containing large indels. 1TERAL1GN). Compare
Blocks are consistent (i.e. in the alignments produced by different
same order) but not necessarily programs (including progressive
uniform (i.e. some blocks may be methods).
missing in some sequences).

Sequences contain a non-consistent Motif-based local alignment

set of conserved blocks (i.e. some method (e.g. MEME). Compare
blocks are duplicated or occur in a alignments produced by different
different order along sequences). programs.

Figure 4 Choice of multiple alignment methods according of the nature of the sequence set.
L. DURET AND S. ABDEDDAIM

Table 6 Websites for pairwise alignments

LFASTA at PBILa http://pbil.univ-lyonl.fr/lfasta.html
SIM at ExPASy3 http://wvm.expasy.ch/sprot/sim-prot.html
BLAST two sequences at NCBI http://www.ncbi.nlm.nih.gov/gorf/bl2.html
LALIGN at CRBM http://www2.igh.cnrs.fr/bin/lalign-guess.cgi
SIM, GAP, NAP, LAP http://genome.cs.mtu.edu/align/align.html
a
Possibility to visualize pairwise alignments with LAIN VIEW (83).

cases. Moreover, one drawback of progressive methods (but also of optimal

global alignment methods) is that an alignment is produced even if sequences
are not related, possibly of random origin. DIALIGN and ITERALIGN on the con-
trary do not attempt to generate a global alignment if sequences are only locally
related. Another interesting feature of these programs is that they indicate the
significance of the alignment: in DIALIGN for example, regions that are not con-
sidered to be aligned (e.g. a non-conserved region between two aligned blocks)
are printed as lower-case letters whereas aligned residues are in upper-case.
Global methods (optimal, progressive, or block-based) are appropriate only if
all conserved blocks are consistent (see Figure 3). If, as presented in Figure Ac, some
.domains are duplicated, or ordered differently along sequences it is necessary to
use a local multiple alignment method to align all related domains. The WWW
version of the MEME tool (see Table 4) provides a graphical representation of the
motifs found in sequences which proves to be very helpful to analyse the
domain organization of proteins.

4.6 Particular case: aligning protein-coding DNA sequences

It is sometimes necessary to align protein-coding DNA sequences rather than
proteins. Two examples are the design of primers to identify related genes by
PCR or for molecular phytogenies relying on the measure of substitution rates at
synonymous (Ks) or non-synonymous (Ka) sites of codons. Due to the degeneracy
of the genetic code, it is generally more difficult to align coding DNA sequences
than their protein translation. Moreover, some ambiguities in DNA alignments
may be solved when considering the protein translation. For example, the two
DNA alignments below have exactly the same similarity score:

However, the second alignment can be rejected unambiguously taking into

account the protein translation. Thus, the procedure commonly used to align
protein-coding DNA sequences is the following:
(a) Extract coding DNA sequences and the corresponding protein translation.

68
MULTIPLE ALIGNMENTS FOR STRUCTURAL, FUNCTIONAL, OR PHYLOGENETIC ANALYSES

(b) Align protein sequences.

(c) Back-translate the protein alignment into a nucleic acid alignment.
The program PROTAL2DNA (C. Letondal, unpublished) has been written for this
purpose, and is available at the Pasteur WWW server:
http://bioweb.pasteur.fr/seqanal/interfaces/protal2dna.html
Note that the WWW-QUERY server (see Table 2) may be used to extract both
coding DNA sequences and their corresponding protein translation (taking into
account species- or organdie-specific genetic codes).

5 Visualizing and editing multiple alignments

Results of multiple alignment programs are generally saved simply as text files.
There is presently no standard format for multiple alignments. However, the
MSF output format (Figure 5) is provided by most of popular alignment programs
and is recognized by many programs that require alignments as an input (e.g.
molecular phylogeny, profile searches). The MASE format presents the advantage
of allowing the inclusion of annotation regarding the whole alignment or specific
to each sequence (Figure 6). Textual representation of multiple alignments is,

Figure 5 Example of alignment In MSF format.

69
L. DURET AND S. ABDEDDAIM

Figure 6 MASE format. This format is used to store nucleotide or protein multiple
alignments along with annotations relative to the whole alignment (indicated in the header),
or specific to each sequence. The beginning of the file must contain a header containing at
least one line (but the content of this header may be empty). The header lines begin by ';;'.
The body of the file has the following structure: First, each entry begins with one (or more)
annotation lines. Annotation lines begin by the character';'. Again, this annotation line may
be empty. After the annotations, the name of the sequence is written on a separate line. At
last, the sequence itself is written on the following lines.

however, poorly informative. Therefore, graphical interfaces have been developed

to manipulate and edit multiple alignments. Generally, these interfaces allow
users to colour or shade residues (amino acids or nucleotides) according to
various criteria such as physico-chemical properties, degree of conservation
within the alignment, hydrophobicity or secondary structure. The use of colours
is very helpful to interpret a multiple alignment. It gives a much more com-
prehensive view of the information embedded in a multiple alignment than a
simple textual representation. Besides, these interfaces propose several interest-
ing facilities detailed below. A list of such graphical interfaces is given in Table 7.

70
MULTIPLE ALIGNMENTS FOR STRUCTURAL, FUNCTIONAL, OR PHYLOGENETIC ANALYSES

Table 7 Multiple alignment viewers and editors

Jalview (J)ab http://www2.ebi.ac.uk/~michele/jalview/contents.html

CINEMA 2.1 (J)M http://www.biochem.ucl.ac.Uk/bsm/dbbrowser/CINEMA2.l/
SEAVIEW (U)a http://pbil.univ-lyonl.fr/software/seaview.html
MPSA (UM)a http://www.ibcp.fr/mpsa/
Se-AI (M)a http://evolve.zps.ox.ac.uk/Se-AI/Se-AI.html
a
ClustalX (UMPV) ftp://ftp-igbmc.u-strasbg.fr/pub/ClustalX/
(ClustalW + graphical interface)
DCSE (U)a http://indigo2.uia.ac.be:80/~peter/dcse/
a
Availability: U = UNIX , M = Macintosh, P = PC, V = VMS, J = JAVA applet.
" Links to sequence databases.
c
Possibility to download alignments from the PRINTS database.

5.1 Manual expertise to check or refine alignments

Whatever the quality of the software, it is necessary to examine the alignment
to check that there are no obvious errors. Alignments of sequences with large
length differences, or with duplicated domains are particularly error prone,
even if the sequences are not very divergent. A good control consists in verifying
that local similarities detected by pairwise sequence comparisons are preserved
in the multiple alignment. For such purposes, one may use the results of simi-
larity searches (BLAST, FASTA, etc.), or run pairwise local alignment software
(see Table 6) or use a dot-plot representation. Pairwise alignments can be
computed directly from JALVIEW. SEAVIEW (68) includes a dot-plot utility that
can be used to drive the alignment, semi-automatically. CINEMA (69) allows user
to run a BLAST search or a dot-plot on selected sequences.
In some cases, it may be necessary to refine part of the alignment. Experi-
enced users are often able to recognize residues that have been misaligned. In
some cases, external information (e.g. known interactions or 3-D structures)
may also reveal alignment errors. SEAVIEW and CLUSTALX allow users to run
CLUSTALW on a specified region and/or a specified set of sequences, without
changing the rest of the alignment.
Alignment editors (except CLUSTALX) also allow users to manually add or
remove gaps in the alignment. In some interfaces (e.g. JALVIEW or SEAVIEW), it
is possible to define groups in order to edit a subset of sequences. In the absence
of objective criteria, manual alignment editing should, however, be used with
caution.

5.2 Annotating alignments, extracting sub-alignments

The SEAVIEW software allows users to annotate alignments (e.g. to indicate the
location of relevant features such as enzyme active sites or RNA splicing signals).
The locations of annotations are correctly preserved after indels are inserted or
moved. This software also allows one to define groups of sequences and blocks
in the alignment and thus to extract sub-alignments. This feature is particularly

71
L. DURET AND S. ABDEDDAIM

useful when building phylogenetic trees where one needs to exclude unreliable
parts of alignments (i.e. regions for which the alignment is ambiguous). It is also
useful to select particular domains for profile searches. Definitions of groups
and blocks can be saved along with the alignment in MASE format (Figure 6).

5.3 Comparison of alignment editors

Each of the editors presented in Table 7 has some specific useful features, some
of which have been mentioned above. Programs written in JAVA (JALVIEW,
CINEMA) present two advantages. First, they can be used from any computer
and run directly from a WWW browser (although, depending on the network
load, the time necessary to download the JAVA applet through the internet
sometimes limits considerably their usefulness). Secondly, thanks to the network
communication facilities provided by JAVA, these programs allow users to directly
access information stored in sequence databases available on the internet.
CLUSTALX is a graphical interface to the CLUSTALW program and not simply an
alignment viewer. However, it does not allow manual editing of alignments.
MPSA is dedicated to protein secondary structure prediction. SEAVIEW is par-
ticularly suited for phylogenetic analyses and can notably be used in combina-
tion with the PHYLOWIN graphical interface dedicated to molecular phylogeny
(68).

5.4 Alignment shading software, pretty printing, logos, etc.

To publish the results of such analyses, it is generally useful to prepare a high
quality colour figure of the multiple alignment. Some of the above editors (e.g.
JALVIEW, CINEMA) can be used to save or print coloured alignments in a format
suitable for publication. Other programs, some of which are available on the
WWW, have been developed specifically for that purpose (see Table 8). The
program LOGO (70) is intended to give a visual representation of a consensus
sequence, along with possible variants.

6 Databases of multiple alignments

Databases of precompiled multiple alignments have been developed, essentially
for protein sequences (71-79) but also for rRNA (80-82) and some other nucleic
acid sequences (see Table 9). The approach used to cluster together homologous
protein sequences varies according to databases. Some intend to classify to-
gether proteins homologous over their entire length (protein families), whereas
others focus on the classification of protein domains (see Table 9). For example,

Table 8 Pretty printing, shading, logos, etc.

BOXSHADE http://ulrec3.unil.ch/software/BOX_form.html
WebLogo http://www.bio.cam.ac.uk/cgi-bin/seqlogo/logo.cgi
Mview http://mathbio.nimr.mrc.ac.uk/nbrown/mview/
AMAS http://barton.ebi. ac.uk/servers/amas_se rver.html

72
MULTIPLE ALIGNMENTS FOR STRUCTURAL, FUNCTIONAL, OR PHYLOGENETIC ANALYSES

Table 9 Databases of multiple alignments

• Protein families
PIRALN http://www-nbrf.georgetown.edu/nbrf/getaln.html
HOVERGEN http://pbil.univ-lyonl.fr/databases/hovergen.html
PROTOMAP http://www.protomap.cs.huji.ac.il/
Megaclass http://www.ibc.wustl.edu/megaclass/

• Protein domains
ProDom http://protein.toulouse.inra.fr/prodom.html
PRINTS http://www.biochem.ucl.ac.uk/bsm/dbbrowser/PRINTS/
PRINTS.html
DOMO http://www.infobiogen.fr/~gracy/domo/
PFAM http://genome.wust! .edu/Pfam/
BLOCKS http://blocks.fhcrc.org/

•RNA/DNA
Ribosomal Database Project http://www.cme.msu.edu/RDP/
The rRNA WWW server http://rrna.uia.ac.be/
3
ACUTS http://pbil.univ-lyonl.fr/acuts/ACUTS.html

' Ancient Conserved UnTranslated Sequences.

the HOVERGEN database compiles multiple alignments and phylogenetic trees

for all families of vertebrate protein-coding genes along with the corresponding
GenBank annotations (79). This database provides all the data necessary to
decipher the orthology/paralogy relationships among vertebrate multigenic
families and is thus particularly useful for phylogenetic studies or for
comparative analysis of vertebrate genes. However, this approach is limited to
relatively well-conserved sequences alignable over their entire length. Conversely,
databases of protein domains may achieve to cluster very distantly related
sequences and are useful to analyse the structure, function, and evolution of
modular proteins. For some complex families, it may be useful to consult
specialized databases such as those available for immunoglobulins or HOX
proteins (for a complete list, see the WWW page maintained by Amos Bairoch:
http://www.expasy.ch/alinks.html).

7 Summary
In this chapter, we describe methods commonly used to align homologous
sequences. Searching for the best alignment consists of finding the one that
represents the most likely evolutionary scenario (substitutions, insertion, and
deletion). Different alignment algorithms have been developed, but none of
them is ideal. Because of time and memory requirements, algorithms that
guarantee to find the best alignment for a given evolutionary model can be used
in practice only with a very limited number of short sequences. Therefore, non-
optimal algorithms based on heuristics have been proposed to gain speed and

73
L. DURET AND S. ABDEDDAIM

limit memory requirements. We discuss the choice between these different

methods (progressive global alignment, block-based global alignment, motif-
based local multiple alignment) according to the nature of the sequences to
align. We also describe graphical tools that have been developed to visualize and
edit multiple alignments. Finally, we mention several databases that compile
multiple alignments of homologous protein or nucleotide sequences. All inter-
net addresses where the tools and resources described here are available are
listed in the following WWW page:
http://pbil.univ-lyonl.fr/alignment.html

References
1. Kimura, M. (1983). The neutral theory of molecular evolution, Cambridge University Press,
Cambridge.
2. Doolittle, R. F. (1994). Trends Biochem. Sci., 19,15.
3. Wootton, J. C. and Federhen, S. (1993). Computers Chem., 17,149.
4. Altschul, S. F. and Gish, W. (1996). In Methods in enzymology (ed. R. F. Doolittle), Vol. 266,
p. 460. Academic Press, London.
5. Patthy, L. (1996). MatrixBioJ., 15, 301.
6. Schwartz, R. M. and Dayhoff, M. O. (1978). In Atlas of protein sequence and structure
(ed. M. O. Dayhoff), p. 353. Nat. Biomed. Res. Found., Washington DC.
7. Gonnet, G. H., Cohen, M. A., and Benner, S. A. (1992). Science, 256, 1443.
8. Henikoff, S. and Henikoff, J. G. (1992). Proc. Natl. Acad. Sci. USA, 89,10915.
9. Thompson, J. D., Higgins, D. G., and Gibson, T. J. (1994). Nucleic Acids Res., 22, 4673.
10. Overington, J., Donnelly, D., Johnson, M. S., Sali, A., and Blundell, T. L. (1992). Protein
Sci., 1, 216.
11. Bains, W. (1992). Mutat. Res., 267, 43.
12. Hess, S. T., Blake, J. D., and Blake, R. D. (1994). J. Mol. Biol., 236, 1022.
13. Benner, S. A., Cohen, M. A., and Gonnet, G. H. (1993). J. Mol. Biol, 229, 1065.
14. Gu, X. and Li, W. H. (1995). J. Mol. Evol, 40, 464.
15. Benson, D. A., Boguski, M. S., Lipman, D. J., Ostell, J., Ouellette, B. F. F., Rapp, B. A.,
et al. (1999). Nucleic Acids Res., 27, 12.
16. Stoesser, G., Tuli, M. A., Lopez, R., and Sterk, P. (1999). Nucleic Acids Res., 27, 18.
17. Bairoch, A. and Apweiler, R. (1999). Nucleic Acids Res., 27, 49.
18. Barker, W. C., Garavelli, J. S., McGarvey, P. B., Marzec, C. R., Orcutt, B. C., Srinivasarao,
G. Y., et al. (1999). Nucleic Acids Res., 27, 39.
19. Schuler, G. D., Epstein, J. A., Ohkawa, H., and Kans, J. A. (1996). In Methods in
enzymology (ed. R. F. Doolittle), Vol. 266, p. 141. Academic Press, London.
20. Etzold, T. and Argos, P. (1993). CABIOS, 9, 49.
21. Gouy, M., Gautier, C., Attimonelli, M., Lanave, C., and Di-Paola, G. (1985). Camp. Appl.
Biosri., 1,167.
22. Pearson, W. R. and Lipman, D. J. (1988). Proc. Natl. Acad. Sci. USA, 85, 2444.
23. Altschul, S. F., Gish, W., Miller, W., Myers, E. W., and Lipman, D. J. (1990)./. MoZ. Biol.,
215, 403.
24. Altschul, S. F., Madden, T. L, Schaffer, A. A., Zhang, J. H., Zhang, Z., Miller, W., et al.
(1997). Nucleic Acids Res., 25, 3389.
25. Altschul, S. F., Boguski, M. S., Gish, W., and Wootton, J. C. (1994). Nature Genet, 6,119.
26. Hughey, R. and Krogh, A. (1996). Comput. Appl. Biosci., 12, 95.
27. Notredame, C. and Higgins, D. G. (1996). Nucleic Acids Res., 25, 4570.

74
MULTIPLE ALIGNMENTS FOR STRUCTURAL, FUNCTIONAL, OR PHYLOGENETIC ANALYSES

28. Chan, S. C., Wong, A. K. C, and Chiu, D. K. Y. (1992). Bull. Math Bid., 54, 563.
29. Carillo, H. and Lipman D. (1988). SIAMJ. Appl. Math., 48,1073.
30. Altschul, S. F., Carroll, R. J., and Lipman, D. J. (1989). J. Mo!. Bid., 207, 647.
31. Gotoh, O. (1995). Comput. Appl. Biosci,, 11, 543.
32. Sankoff, D. (1975). SIAMJ. Appl. Math, 78, 35.
33. Needleman, S. B. and Wunsh C. D. (1970). J. Mo!. Bio!., 48, 443.
34. Lipman, D. J., Altschul, S. F., and Kececioglu, J. D. (1989). Proc. Natl. Acad.. Sri. USA, 86,
4412.
35. Gupta, S., Kececioglu, J. D., and Schaffer, A. (1995). J. Comput. Biol., 2, 459.
36. Jiang, T., Lawler, E. L, and Wang, L. (1994). ACM Sympos. Theory Comput., 26, 760.
37. Feng, D. F. and Doolittle, R. F. (1987). J. Mol. Evol., 25, 351.
38. Taylor, W. R. (1988). J. Mol. EvoL, 28, 161.
39. Higgins, D. G. and Sharp, P. M. (1988). Gene, 73, 237.
40. Sneath, H. A. and Sokal, R. R. (1973). Numerical taxonomy. W. H. Freeman. San Francisco.
41. Saitou, N. and Nei, M. (1987). Mol. Biol. EvoL, 4, 406.
42. Gotoh, O. (1996). J. Mol. Biol., 264, 823.
43. Vingron, M. and Argos, P. (1989). Comput. Appl. Biosci., 5, 115.
44. Sobel, E. and Martinez, H. (1986). Nucleic Acids Res., 14, 363.
45. McCreight, E. M. (1976). J. ACM, 232, 262.
46. Vingron, M. and Argos, P. (1991). J. Mol. Biol., 218, 33.
47. Morgenstern, B., Dress, A., and Werner T. (1996). Proc. Natl. Acad. Set. USA, 93, 12098.
48. Morgenstern, B., Atchley, W. R., Hahn, K., and Dress, A. (1998). Ismb, 6, 115.
49. Zhang, Z., Raghavachari, B., Hardison, R., and Miller, W. (1996). J. Comput. Biol., 1, 217.
50. Myers, E. and Miller, W. (1995). In Proc. 6th ACM-SIAM Symposium On Discrete Algorithms,
p. 38.
51. Zhang, Z., He, B., and Miller, W. (1996). J. Disc. AppZ. Math., 71, 337.
52. Abdeddairn, S. (1997). In lecture notes in computer science, Vol. 1264, p. 167.
Springer-Verlag.
53. Smith, R. F. and Waterman, M. S. (1981). J. Mol. Bio!., 147,195.
54. Waterman, M. S. and Jones, R. (1990). In Methods in enzymology (ed. R. F. Doolittle),
Vol. 183, p. 221. Academic Press, London.
55. Schuler, G. D., Altschul, S. F., and Lipman, D. J. (1991). Proteins, 9,180.
56. Lawrence, C. E., Altschul, S. R, Boguski, M. S., Liu, J. S., Neuwald, A. F., and Wootton,
J. C. (1993). Science, 262, 208.
57. Henikoff, S., Henikoff, J. G., Alford, W. J., and Pietrovoski, S. (1995). Gene-COMBIS, Gene,
163,17.
58. Bailey, T. L and Elkan, C. (1995). Ismb, 3, 21.
59. Lawrence, C. E. and Reilly, A. A. (1990). Proteins, 7, 41.
60. Cardon, L. R. and Stormo, G. D. (1992). J. Mol. Bio!., 223,159.
61. McClure, M. A., Vasi, T. K., and Fitch, W. M. (1994). Mo!. Biol. EvoL, 11, 571.
62. Briffeuil, P., Baudoux, G., Lambert, C., De Bolle, X., Vinals, C., Feytmans, E., et al.
(1998). Bioinformottcs, 14, 357.
63. Morrison, D. A. and Ellis, J. T. (1997). Mol. Bio!. EvoL, 14,428.
64. Huang, X. and Miller, W. (1991). Adv. Appl. Math., 12, 337.
65. Thompson, J. D., Gibson, T. J., Plewniak, F., Jeanmougin, F., and Higgins, D. G. (1997).
Nucleic Acids Res., 25, 4876.
66. Jeanmougin, F., Thompson, J. D., Gouy, M., Higgins, D. G., and Gibson, T. J. (1998).
Trends Biochem. Sci., 23, 403.
67. Brocchieri, L. and Karlin, S. (1998). J. Mol. Bio!., 276, 249.
68. Galtier, N., Gouy, M., and Gautier, C. (1996). Comput. Appl. Biosci., 12, 543.

75
L. DURET AND S. ABDEDDAIM

69. Parry-Smith, D. J., Payne, A. W., Michie, A. D., and Attwood, T. K. (1998). Gene, 221,
GC57.
70. Schneider, T. D. and Stephens, R. M. (1990). Nucleic Acids Res., 18, 6097.
71. Srinivasarao, G. Y., Yeh, L. S. L, Marzec, C. R., Orcutt, B. C, Barker, W. C., and Pfeiffer,
F. (1999). Nucleic Acids Res., 27, 284.
72. Yona, G., Linial, N., Tishby, N., and Linial, M. (1998). Ismb, 6, 212.
73. Corpet, R, Gouzy, J., and Kahn, D. (1999). Nucleic Adds Res., 27, 263.
74. Attwood, T. K., Flower, D. R., Lewis, A. P., Mabey, J. E., Morgan, S. R., Scordis, P., et al.
(1999). Nucleic Acids Res., 27, 220.
75. Gracy, J. and Argos, P. (1998). Bioinfarmatics, 14,164.
76. Gracy, J. and Argos, P. (1998). Bioinformatics, 14,174.
77. Bateman, A., Birney, E., Durbin, R., Eddy, S. R., Finn, R. D., and Sonnhammer, E. L. L.
(1999). NucIeicSVidsKes., 27, 260.
78. Henikoff, J. G., Henikoff, S., and Pietrokovski, S. (1999). Nucleic Acids Res., 27, 226.
79. Duret, L, Mouchiroud, D., and Gouy, M. (1994). Nucleic Acids Res., 22, 2360.
80. Maidak, B. L, Cole, J. R., Parker Jr, C. T., Garrity, G. M., Larsen, N., Li, B., et al. (1999).
Nucleic Acids Res., 27,171.
81. Van de Peer, Y., Robbrecht, E., de Hoog, S., Caers, A., de Rijk, P., and de Wachter, R.
(1999). Nucleic Acids Res., 27, 179.
82. De Rijk, P., Robbrecht, E., de Hoog, S., Caers, A., Van de Peer, Y., and de Wachter, R.
(1999). Nucleic Acids Res., 27, 174.
83. Duret, L, Gasteiger, E., and Perriere, G. (1996). Comput. Appl. Biosci., 12, 507.

76
 Chapter 4
Hidden Markov models for
database similarity searches
Ewan Birney
The Sanger Centre, Wellcome Trust Genome Campus, Cambridge, UK.

1 Introduction
Despite the huge number of genes in an organism, the protein coding genes are
thought to be made from a limited number of basic protein structures. Evolution
has reused these protein structures, combining them to form different proteins,
and altering them in different genes to achieve different functions. The diversity
of species, each with its own copies of genes made from the limited number of
building blocks, means that, for a protein of interest, a number of different re-
lated proteins may be found. In this chapter, I will discuss one set of techniques
which can be used to take advantage of this diversity of protein sequence. These
techniques are all related to the use of profiles, which are also discussed in
Chapter 5. hi this chapter, the emphasis will be on the use of hidden Markov
models (HMMs) for profile analysis. Some practitioners consider profiles to be a
type of HMM.
It is important to realize that a protein might be related to another protein in
a variety of different ways. It could be that the entire protein is homologous
(that is, derived from a common ancestor) to another, such as human and mouse
src protein (see Figure 1) or the human src2 protein which is a paralog to the src
protein. Alternatively only a portion of the protein might be derived from a
common ancestor, such as the jyn protein, which shares a common C terminal
region with a divergent N terminus to the src protein. Finally only a small region
might be conserved, such as the SH3 domain which is also found hi the Grb2
protein (along with many other proteins) with no other organization conserved
between the two proteins. This last type of conservation, conservation of a
domain generally corresponds to a structural domain of the protein which can
fold independently and, in most cases, function independently of other regions.
Figuring out when you have really defined a domain rather than a more exten-
sive piece of conservation is one of the challenges for a researcher. Profile
analysis is useful for all these different types of conservation. It is especially
useful for domain analysis as this is the hardest feature to define using other
methods.

77
EWAN BIRNEY

Figure 1 Three different types of relationship are shown. The grey ovals indicate regions
which are conserved, whereas the lines and other boxes show regions which are not related.
(1) Two very closely related genes, where the entire protein sequence in each gene is
conserved. (2) Two genes where the C termini are related but the N termini are unique.
(3) Two genes which share one domain but the other regions are entirely different.

2 Overview
For people coming from outside the field, the use of profiles and profile-HMMs
can require confronting much confusing jargon and cryptic computer programs.
This chapter is meant to demystify this type of analysis. The first point to
emphasize is that the programs are, basically, just employing some concept of a
'consensus'. This follows intuitively from the observation that if some sequences
have an Aspartate before a critical catalytic residue and others a Glutamate then
a new enzyme can be expected to have either an Aspartate or a Glutamate at
this position. This sort of simplistic rule is recast into a mathematically con-
venient form: resulting in some idea of a probability for each possible amino
acid at a different position, called a profile. The difficulty lies, as in many areas in
sequence analysis, that there may be different numbers of amino acids between
conserved residues. A consequence of the differing lengths is that there are
usually a number of different ways of providing a match to a 'consensus', and
some way of choosing the 'best' one must be decided. The variable lengths be-
tween conserved residues also makes the statistical behaviour of the technique
very hard to handle using conventional statistical analysis.
This chapter will concentrate first on using databases of profile-HMMs
through the World Wide Web (WWW), which is by far the easiest way of using
them. Then we will concentrate on PSI BLAST (1) which is the easiest do-it-
yourself profile method, also available through the Web. The final example will

78
 hIDDEN MARKOV MODELS FOR DATABASE SIMILARITY SEARCHES

Table 1 Some useful Web addresses

Protein family sites

Pfam (Europe) http://www.sanger.ac.uk/Software/Pfam
Pfam (USA) http://pfam.wustl.edu/
Prosite Profiles http://www.isrec.isb-sib.ch/profile/profile.html
Prints http://www.biochem.ucl.ac.uk/dbbrowser/PRINTS
BLOCKS http://www.blocks.fhcrc.org/
SMART http://coot.embl-heidelberg.de/SMART/
Software tools
PSI-BLAST http://www.ncbi. nlm.nih.gov/cgi-bin/BLAST/nph-psi_blast
HMMER2 http://hmmer.wustl.edu/

cove? the use of the HMMER2 (2) package which I find to be the most effective
profile-HMM package available. It is UNIX based and relatively easy to use,
though it is not currently available through the Web. Finally I will outline some
of the theories behind profile HMMs from the point of view of how it impacts on
their practical use.
The reader should be aware that there are many other profile-HMM packages.
I would draw your attention in particular to the Meta-MEME package (3) and the
PROBE package (4) as well thought out solutions. There are also a number of
other profile packages (5, 6) which are more focused on the use of the package
by their own groups. Finally, a number of commercial solutions exist (7-9), and
you may well have access to them. If you know someone on site who is already
skilled in using one of these packages, it is best to use that local knowledge and
treat this chapter as more of an introduction to the concepts involved. In
addition, it is likely that when you are reading this chapter, that new methods
or new presentations of old methods will become available. To keep up to date,
use the web, and try the URLs in Table I to find the most up-to-date resources.
Finally, I would like to warn users that I have a strong bias towards using a
probabilistic framework to explain and justify the methods: this fits easiest with
the HMM formalism and the use of Bayesian statistics (a branch of probability
analysis). Other researchers are less zealous about using this sort of framework
to explain the results. In either case the most important question is whether
these methods are biologically useful, whatever the theories say.

3 Using profile and profile-HMM databases

The starting-point of this sort of analysis is usually a protein sequence which
you might have derived from your own sequencing project of a gene of interest.
The aim is to use a pre-made profile-HMM, previously constructed by another
group to highlight regions of your protein which have homology to already well
characterized domains. All these resources focus on domains, being the con-
served building blocks of proteins, so the end result of the analysis will be to
return which regions of your protein look as if they have particular domains.

79
EWAN BIRNEY

Once you have a protein sequence it is probably best to put it into Fasta
format (see Section 8.5), though many resources will allow you to use other for-
mats. Then connect to one of the resources shown in Table 1 and find the page
for searching with your sequence.
Choose the 'search' page. Then use the 'file-upload' button on the forms to
submit your own sequence, and click 'submit' or 'run analysis'. The search against
the database will probably take a little over a minute, and should not take more
than 10 minutes. Each resource returns its own particular format of results, but
what is generally reported is the type of the domain, the position in your pro-
tein as a start- and end-point, and some indication of how confident your can be
of the hit. They all provide a nice graphical representation of the domain on
your sequence as a cartoon of the sequence with different coloured or shaped
regions indicating the different domains. Clicking on the graphic will take you
usually to an in-depth description of the domain, which in many cases will
contain links to other resources and literature references. How to interpret the
precise results varies from resource to resource.

3.1 Pfam
Pfam (10) is a database of protein families and corresponding profile-HMMs. Pfam
uses the HMMER2 package to provide tools for making the HMMs in the 'first
place and then for searching them. A search against Pfam will provide you with
three ways of deciding confidence in the matches. The first is a classical e-value
(expectation value) which generally is considered significant if it is below 1.0 for
individual searches. The second is the Bits score which is derived from the
underlying scoring scheme used to score the match between the sequence and
the profile-HMMs. It is related to the Bayesian inference of the probability of the
match (for a deeper explanation of the statistics read Section 8). A final check is
provided by a manually derived cut-off which an 'expert' has chosen to separate
the true examples from false examples. These cut-offs are chosen conservatively
so that, to the researcher's knowledge, they do not misclassify any protein. This
can mean that, in some cases, known trues are missed using this cut-off.
At the time of writing the Pfam database (Version 3.3) had 1344 protein
families, which covered 57% of the protein primary sequence database. In new
genome projects over one third of proteins had at least one hit to a protein
family.

3.2 Prosite profiles

Prosite profiles (11) are an addition to the Prosite resource to define protein
domains using profile-HMM technology. Prosite profile reports a classical e-value
type statistic which is presented on a log scale. Scores above 5 can be considered
significant and scores above 7 very significant. The raw score is not a meaningful
number except for different examples of the same domain the better the score
the better the match. There are no manually set cut-offs.
At the time of writing, there were 205 prosite profiles. There is no reliable

80
 HIDDEN MARKOV MODELS FOR DATABASE SIMILARITY SEARCHES

way of estimating the coverage of prosite profiles. However, a researcher can

combine a prosite profile search with a Pfam search, allowing the two resources
to be combined in the same submission, with a common output. You are advised
to make sure the prosite form is using the most up-to-date Pfam release.

3.3 SMART
SMART (12) is currently based on conventional, non HMM profile technology.
The raw score is meaningless, rather you must trust the manually set cut-offs
provided internally. At the time of writing there were 302 profiles in SMART.
SMART is not focused on coverage but rather on providing very accurate align-
ments and resources of the domains of interest. It is likely by the time of reading
this that SMART has switched to using the HMMER2 package rather than old
style profiles.

3.4 Other resources and future directions

There are a number of other resources which provide access to ready made
homology databases; in particular PRINTS (13) and BLOCKS (14) (see Table 1 and
Chapter 5). Both these resources are less focused on finding individual domains
and instead focus on finding smaller 'motifs'. They may give less clear-cut
answers, and for difficult domains may be less sensitive, but come with a number
of useful utilities and options.
By the time of reading this chapter, it is likely that a number of resources will
be using a common documentation resource (Interpro). This will provide more
consistent documentation between the different resources, and is likely to be a
forerunner to further integration between the resources.

3.5 Limitations of profile-HMM databases

The obvious limitation of a profile HMM database is that if the domains in your
protein are not represented in the database then the databases will (hopefully)
return nothing. The only option here is to start your own profile analysis using
one of the techniques listed below. In addition, it might be that there is an error
in the database, giving the wrong start/end points or misclassifying a region. In
all the above databases, the most likely error will be that you miss a true domain
in your protein. You can lower the thresholds for determining whether a domain
exists or not, but be careful that you do not simply accept a false match due to
'noise'. Use the e-value statistic to decide whether this domain is justified on
statistical grounds and read the information in section 7.0 on validating matches.

4 Using PSI-BLAST
PSI-BLAST (1) is a profile building and searching package which is fast, accessible
through the Web, and aimed at a less expert audience than the other profile
packages. This makes it ideal for occasional use or quick investigations of a
particular protein sequence.

81
EWAN BIRNEY

In many ways PSI-BLAST follows the same methodology as using the HMMER
package below, just that this is done behind the scenes. PSI-BLAST starts from a
single sequence, which is then searched against a database using the fast BLAST
method. The resulting matches are aligned back to the query sequence, and this
derived multiple alignment is used to estimate a profile. The profile is then used
to search the database, collect homologues, and align back to the profile, and so
the process iterates onwards until it stabilizes or some cut-off is exceeded.
A URL to start the process off is given in Table 1. You load in your protein
sequence and launch the first search. At the end of each search you have the
option of including or rejecting each sequence for the next iteration. This gives
you the chance to eliminate potential false positives and include weak but true
matches from your knowledge of the biology. An e-value statistic is provided to
give an automatic selection of the next round of sequences, which should guide
you in your selection.
Many of the problems inherent in using PSI-BLAST are also present when
using HMMER, and so I would encourage you to read sections 6.0 and 7.0 care-
fully. Crucially you must be aware that the statistic to quote for the significance
of a match is the first one in which it appears in the profile: once a particular
sequence has been included in the set which makes the profile, it will, un-
surprisingly, score very well against the resulting profile.
The other problems of PSI-BLAST are less to do with the method and more to
how it is used. Because it starts with a single sequence, it is tempting to put in
an entire sequence of interest and simply start iterating. If the sequence con-
tains one common domain, although PSI-BLAST will find all the homologues of
the sequence, both including the domain and excluding it your results will be
dominated by this domain and become unmanageable. As you focus your effort
on a particular region, it is better to excise that region and use that as a starting
point for further analysis.

5 Using HMMER2
HMMER2 (2) is a package of UNIX command line programs which make and use
profile HMMs. If you have no experience of the UNIX command line, then using
HMMER2 is going to be a struggle. I suggest taking a short course in UNIX first.
In addition to the HMMER2 software, you will need a number of other reason-
ably standard bioinformatics resources. In particular:
(a) A copy of an up-to-date protein database in fasta format as a single file.
(b) A method of retrieving sequences from this database, preferably with the
ability to retrieve only a portion.
(c) A multiple alignment program such as Clustal W (17).
(d) A specialized multiple alignment editor.
It may also be useful to have some experience of a text reformatting language
such as Perl or Python, or access to someone who can write small glue programs

82
 HIDDEN MARKOV MODELS FOR DATABASE SIMILARITY SEARCHES

for you. Installing the resources is best done with the co-operation of the systems
support group for the UNIX machine you are using.

5.1 Overview of using HMMER

Figure 2 gives the basic flow of profile analysis. The main steps are to create a
multiple alignment of the region of interest and from this multiple alignment
make a profile-HMM. The profile-HMM is searched against a protein database: a
number of new protein matches may be identified and these can then be in-
corporated into the multiple alignment. There are two places where human
knowledge can make a large difference in the analysis; firstly manual editing of
the multiple alignment can produce dramatically better results, secondly which
sequences are included or not into new alignments can be vetted using
biological knowledge of the process.

5.2 Making the first alignment

To start the whole procedure off one needs to both identify potential homo-
logues and produce the first multiple alignment. The first potential homologues
are usually found using single sequence searches. Then these homologues are
aligned using a multiple alignment program such as clustal w. This is discussed
in depth in Chapter 3. An important issue to realize is that in many cases you
will be attempting to make a multiple alignment of a domain common to a
number of different proteins. To successfully make such a multiple alignment
you will need to determine the rough boundaries of the domain in each protein
from the single sequence searches. You will then need to excise the regions of
the protein with a small (10 residue or so) leeway on each side: hopefully your
database retrieval program will have this ability built in. Once the multiple
alignment has been made you will probably want to edit it.

Figure 2 A flow diagram of how profile-HMMs are commonly used. The programs in the
HMMer package which are used to provide the different transformations are given beside the
arrows. PSI-Blast uses the same principles although much of the mechanics are then hidden
from the user

83
EWAN BIRNEY

5.3 Making a profile-HMM from an alignment

Thankfully the latest version of HMMER has sensible defaults for making
profileHMMs. The program hmmbuild is used to make the HMM, which is run
with defaults by typing at the UNIX prompt: 'hmmbuild HMM Alnfile' where
alnfile is the name of a file containing a multiple alignment. The defaults which
are run are as follows:
(a) Tree weighting. A tree is calculated from the multiple alignment and
sequences weighted by how many close neighbours they have. This means
that overrepresentation by one sub family (for example, many haemoglobin
alpha chains for a globin profile-HMM) does not violently bias the profileHMM
towards that subfamily.
(b) Dirchlet mixtures. A concept of accepted amino acid conservation patterns,
such as Valine, Isoleucine, and Leucine being common replacements for each
other is provided as Dirchlet mixtures. For more information about Dirchlet
mixtures, see Section 8.4.
(c) The effective number of sequences is estimated, meaning that alignments
which are of very close homologues will be assumed to represent fewer
sequences than alignments of more distantly related sequences (see Section 8.4).
(d) The placement of which columns in the multiple alignment are match
columns as opposed to insert columns.
These defaults are typical for the needs of most HMMs. As you get used to the
package, you may want to change them (options are provided on the command
line; to get a full list of options type 'hmmbuild -help'), though even experts
tend not to deviate too much from these settings.
The only parameter which is worth altering routinely is the local/global mode
switch. By default the profile-HMMs are built in global mode, where the entire
HMM must be matched against the sequence. An alternative is local mode, by
using the -f flag on hmmbuild, which allows only a portion of the HMM to
match. If the global mode is used, the HMM becomes far more sensitive for
finding distant members, but is unsuitable for finding fragments. If you have
misdefined the region of interest, perhaps making it too long at the N or C
termini, then the global mode will penalize sequences that do not fit the entire
model, preventing them being found. Generally global mode is a better choice,
but at the start of the analysis, local mode can be a more sensible choice as one
is not always very confident about the locations of the ends of a domain.
Having made the profile-HMM, you need to-calibrate it by typing 'hmmcali-
brate HMM'. The calibration step calculates statistical parameters for the HMM
by generating a random database. This statistical parameterization is crucial for
its effective use. Calibration takes around 10 minutes or so.

5.4 Finding homologues and extending the alignment

The HMM can be searched against a database of sequences using the program
hmmsearch. This takes the HMM file as the first argument and the database file

84
 HIDDEN MARKOV MODELS FOR DATABASE SIMILARITY SEARCHES

as the second argument. The database file is in Fasta format. The results are
printed on standard output, so you usually need to redirect the output to save
that information in a file. The results give you the following information:
(a) A list of sequences which the HMM hit, ranked from most significant to least.
(b) A list of domains contained in the sequences, ranked from most significant to
least.
Notice that a particular sequence can contain more than one domain. In par-
ticular, although each of the domain scores might, on their own, not be signifi-
cant, the combined score of multi-domain match might easily be so.
Both the per-sequence and per-domain matches are provided with two
statistics: a bits score and an e-value (see Section 8.2). The more reliable score is
the e-value; e-values down to 1.0 can be considered significant (an e-value of 1.0
means that, by chance, 1 random sequence is expected to get this score in the
database of the size which you used). The bits score is helpful as it is indepen-
dent of database size.
Having chosen a significance level one would then like to make a new multi-
ple alignment of all the protein sequences found by the HMM. At the moment,
this is the most labour intensive step, as the HMMER package does not provide
all the functionality for this task. Somehow, one needs to extract all the sequences
which are hit and truncate them to the correct start/end points. This is best done
by a perl script or similar device. Once you have all the sequences which were
hit, as a Fasta file, the program hmmalign will provide a multiple alignment of
the proteins on the basis of the HMM. This multiple alignment can then be used
to make a new HMM for the next round.

6 False positives
One of the problems inherent to the iterative procedures, both PSI-blast and the
use of HMMER outlined above is that if a false positive is added to the alignment,
itself and any close relatives will score highly against the profile-HMM. For ex-
ample, if you inadvertently add a globin sequences to a protein kinase align-
ment the resulting HMM will match globin sequences surprisingly well.
This ability to start collecting false positives at will means that a researcher
should ideally by very vigilant as the iterations progress. Indications that the
profile might be picking up noise are:
(a) Low complexity regions occurring in alignment.
(b) A region overlapping a known domain, where it is clear that the multiple
alignment is not a divergent subfamily for this domain.
(c) Biological information that indicates that this match is false.

7 Validating a profile-HMM match

Once a researcher has found a suggested domain, how can they validate this?
The score of the sequences to the HMM from the final alignment is not the

85
EWAN BIRNEY

correct measure of the significance of the match, as it includes all the sequences
you wish to score, and they will all score well. In fact the problem of justifying a
grouping of sequences is not well handled by the current statistics, in particular
when an iterative strategy is used. The following lines of evidence may be used
to give a researcher confidence that the similarity they observe is not by chance.

(a) See whether all the sequences can be connected together by significant single
sequence scores (e.g. from programs such as BLAST2). Ideally one should be
able to show this with the full length proteins (just taking the domain
improves the statistics considerably).
(b) Quote the significance of the 'new' sequences for the first time they provided
a significant score against the profile-HMM.
(c) To show that A is related B, show that by starting from either A or B one can
produce a profile which finds the other sequence using criterion (b).
(d) Provide biological justification that the relationship makes sense (e.g. com-
mon mode of enzymatic action). Conversely, biological information which
indicates that they should not be related should lessen the researcher's belief
in the result.

8 Practical issues of the theories behind

profile-HMMs
8.1 Overview of profile-HMMs
A profile is a sequence of conserved positions, each conserved position having a
score for each amino acid. For practical purposes, one should neither assume
that a particular sequence has all the conserved positions in a profile, nor that a
particular sequence will not introduce additional amino acids between two con-
served positions in the profile. These two possibilities are the two types of gaps,
that is a deletion of part of the profile, or an insertion of residues relative to the
profile. A number of ad hoc methods have been produced to solve the gap
problem in ways that seem biologically sensible.
The ad hoc nature of profiles was replaced by a stronger theory based around
HMMs which still essentially produced the same sort of profile as before. An
HMM is a mathematical model which produces a stream of some observable
information in a probablistic manner. In the case of protein sequences, the
observable information will be the amino acids. The probablistic nature of the
process of producing the amino acids means that a particular HMM can produce
more than one set of protein sequences, and also that different sequences are
produced with different probabilities. The hidden part of a HMM is that one
does not know which model made the particular amino acids one is looking at,
and, if one did know the model, which amino acids were made by which part of
the model. It is these two problems which one wants to solve, and they corre-
spond to the two questions: 'does this sequence have an example of this HMM in

86
 HIDDEN MARKOV MODELS FOR DATABASE SIMILARITY SEARCHES

it' and 'if this sequence does have an example of the HMM, what is the align-
ment of the HMM to the sequence'.
The HMMs which have been used in this field deliberately mimic the profile
model described above. For each conserved column in the multiple alignment,
three possible states are permitted: a match state which indicates that a single
residue is being aligned to the position, a delete state, which indicates that no
amino acid in this protein is present for this position, and an insert state, which
allows any number of amino acids to be inserted after the match position. A full-
length HMM will have some 500 or so different states, broken down into triplets
representing conserved column positions. The behaviour of these states is
governed by probabilities for the production of different amino acids from
match and insert states and probabilities for the transitions to the neighbouring
states. These probabilities are analagous to the scores of the profile and the gap
penalities in the profile respectively. Indeed, for practical purposes, the prob-
ability representation of a profile-HMM is rarely used. Instead, the probabilities
are transformed into sensibly sized integers via a log transformation. In this
logged representation, adding the numbers is equivalent to multiplying the
underlying probabilities, making the correspondence between profiles and
profile-HMMs all the more clear.
Given a particular profile-HMM, the questions 'does this sequence have an
example of my HMM' and, given that the last question is true, 'what is the align-
ment of the sequence to the HMM' can be easily answered using some well
known algorithms. These two questions are essentially what hmmsearch and
hmmalign provide answers for in the HMMER package.

8.2 Statistics for profile-HMM

Every sequence will match the profile HMM in some manner: some sequences
will match the profile HMM better (in the sense that the probability that the
HMM would produce such a sequence is higher) than others. How does this
statistic (called a likelihood) allow you to say whether this match really is due to
the presence of this domain or is it just by chance?
As the underlying basis for HMMs is a probabilistic model, this question is
easily answered by Bayesian statistical methods. Non-mathematicians usually
have not been exposed to Bayesian statistics previously. Bayesian statisitics try
to answer such questions by assigning a probability to its being true or not. In
contrast classical statistics answers the question by assuming a particular hypo-
thesis, which is rejected or not by the data. Both approaches are guaranteed to
give the same answer when enough data is taken into account, but will often
provide different answers for practical problems.
To provide a Bayesian interpretation of the profile-HMM, all possible different
models of how the sequence was produced, have to be defined: for profile-
HMMs, two models will be considered, the model of the profile-HMM domain
and a null model of random amino acids drawn from the frequencies found in a
large protein database. The probability of seeing the observed sequence under
the assumption of the two different models are given, each being a likelihood.

87
EWAN BIRNEY

What is quoted is the log-likelihood ratio of the two models: when the base of
the log is 2, this statistic (the log-likelihood ratio) is called a Bits score. A bits score
of 0 is when the likelihood ratio is 1, and hence each model is equally likely to
have produced the sequence. Depending on how the ratio is quoted, either more
negative or more positive scores indicate that the desired model is more likely.
In the HMMER package, the more positive the bits score the better the match to
the profile-HMM.
The likelihood ratio does not provide quite enough information to allow an
estimate of the probability of the profile-HMM occurring, given the sequence
seen. An additional piece of information, being the probability of the profile-
HMM occurring without seeing the sequence data needs to be defined. As this
information has to be defined without seeing any sequence, it is called prior
information. Mathematically this is the same idea as the prior information
which will be introduced in the next section, but in practice it is used in a very
different aspect. Sensible priors include 1/d where d is the size of the data base
or one could use the probability of a random sequence having this domain in a
genome, say 1/10 000. Finally, to be confident of the match, the probability of
the profile-HMM occurring should be over 0.95. These two extra manipulations
—the prior information and the need for a significant probability translate into
a bits cut-off above which one considers matches to be significant. 25 bits trans-
lates to sensible choices of prior and significance, and so matches over 25 bits
can be considered to be significant.
There is another statistic that can be used to estimate whether the match is
significant or not. This is a classical (or frequentist) statistic and is one that most
users will be more familiar with. To provide a frequentist statistic, one needs to
assume that the match is random, derive the probability that a random match
would produce the score, and reject the assumption if this seems very unlikely
to have occurred. The problem with this sort of analysis was that it was clear
that the distribution of scores of random sequences against a profile-HMM was
not normally distributed, and so estimation of probability was very difficult. In
recent years the field has produced theoretical and empirical evidence that the
distribution is closely related to an Extreme Value Distribution (EVD) (15). PSI-
Blast assumes that for a particular way of making a profile, all profiles have the
same EVD parameters, regardless of content. PSI-Blast therefore tabulates this
information for all possible profile construction mechanisms, and uses the tabu-
lated parameters. HMMER uses a separate calibration step, where the profile-
HMM is compared to a large random database, and an EVD is fitted to the
resulting distribution. The parameters from this fitting are stored in the HMM so
they can be reused for individual sequence searches.
The natural way of reporting the classical statistic is as an expectation value
(e-value). This is the number of sequences expected to get this score by chance,
and is simply dr where d is data base size and r is the probability that a random
sequence will get this score. An e-value of 1.0 is therefore where you expect to
start seeing random sequences: e-values less than this are significant.
Which statistic to use: bits score or e-value? It is clear that the e-value statistic

88
 HIDDEN MARKOV MODELS FOR DATABASE SIMILARITY SEARCHES

is more robust and more sensitive in the HMMER2 package, and it is what I
would recommend. However, the e-value has some less desirable properties, in
particular, it changes as the database size changes, unlike the bits score—of
course, if you decide the prior on the bits score should be 1/d (d is database size),
the cut-off for significance of the bits score will change with database size.
Quoting both in publications is very sensible.

8.3 Profile-HMM construction

The use of a particular profile-HMM is well understood. A harder problem to
solve is 'what profile-HMM best represents my collection of known family mem-
bers'. By analogy with other fields, such as speech recognition, this problem can
be answered by expectation maximization, or similar methods, where an HMM
is constructed that maximizes the probability of producing all the sequences
known to belong to a certain family. In theory this can work from just the
sequences, and no multiple alignment, effectively both aligning the sequences
and making the profile-HMM at the same time. In practice, training an HMM
from unaligned sequences does not work well, principally because of local
minima problems. A better solution is to train the HMM from already aligned
sequences, as inFigure2. When already aligned sequences are presented to the
training program, the only aspect which the program must estimate is which
columns to consider as conserved positions, and which columns should be
collapsed into an insert state of the preceding conserved position. Indeed, if you
so wish, you can provide this information directly. Once this is known, it is
relatively easy to estimate the probabilities for the HMM from the observed
sequences using standard theories.

8.4 Priors and evolutionary information

The final problem in HMM construction is that, in general, one is not interested
in finding proteins which are only slightly different from the examples one
already knows. Rather one wants to find a new subfamily related to the sub-
family one has already gathered. This usually means that a sequence from the
new subfamily will have some features which are not present in any of the
sequences one has already gathered, and yet because of the pattern of conserva-
tion and knowledge of behaviour of proteins (for example, a conserved valine
position is more likely to have a leucine in a distant subfamily member at this
position, than an arginine) one can recognize it as being a related member.
The introduction of extra knowledge into the process of estimating an HMM
is called prior information, indicating that it is known before any sequence data
is seen. Ideally one would like to represent all the knowledge about protein
evolution and protein structure in some manner which would allow the profile-
HMM construction machinery to use it. In practice a number of assumptions
have to be made to allow the mathematics of profile construction to work.
1. Profile-HMM building does not currently work with a concept of the observed
sequences being related on an evolutionary tree. Therefore, if you present a

89
EWAN BIRNEY

profile-HMM construction method with 10 near-identical alpha globins and one

beta globin, the resulting HMM would be predominantly alpha globin. This is
solved by weighting the sequences by a tree before applying the profile- building
machinery. In the above example, each alpha globin sequence might get a
weight of 1/10, and the beta globin sequence a weight of 1.

2. The estimation of amino acid probabilities, taking into account protein evolu-
tion, has a stronger theoretical backing. The problem is phrased as an under
sampling problem, where although one has a column of, say, 10 amino acids at
this position the frequencies of amino acids represented by observation is not an
ideal way to estimate the underlying probabilities; clearly not all the amino
acids can be represented even once! This problem occurs in many other situa-
tions and has been well studied. A good solution is to provide the estimation
machinery with prior knowledge of what sort of ammo-acid frequencies one ex-
pects in columns, for example, one with high leucine, valine, and isoleucine
probabilities, and another with a high probability of arginine and lysine, but low
probability for hydrophobic amino acids. These distributions are represented in
a complicated mathematical form called Dirchlet mixtures (16) and, by using
them, the estimation of probabilities for amino acid positions can take into
account evolutionary information. A Dirchlet mixture is a just a convenient
mathematical form for this information; there is nothing special about them
except that they make the downstream mathematics far easier to handle. The
Dirchlet mixture can be thought of rather like a protein comparison matrix
used in ad hoc profile methods.

3. The final twist to profile-HMM construction is how to balance the informa-

tion from the Dirchlet mixtures (the prior) with the information from the
observed multiple alignment. The problem here is that the aim is to find new
sub families so, the fact that one has seen over 1000 different alpha globins,
does not mean that the observed amino acid frequencies on their own will make
a good HMM for finding beta globin sequences. This is solved in the HMMER
package by estimating how many effective sequences one has observed (a
collection of divergent sequences will count more than a collection of close
relatives).

8.5 Technical issues

In Fasta format the first line starts with the greater-than sign (>) and is followed
by the name of the first protein, which should be composed of only characters
from the alphabet, or the underscore symbol (_) or numbers. Most programs
(though not all) will allow any non space character in the name. Following this
line the next lines are the protein sequence in one letter code. The sequence
stops at either the end of the file or the next greater-than (>) sign (which marks
the start of the next sequence name if there is more than one sequence in the
file).

90
 HIDDEN MARKOV MODELS FOR DATABASE SIMILARITY SEARCHES

References
1. Altschul, S. F., Madden, T. L, Schaffer, A. A., Zhang, J., Zhang, Z., Miller, W., et al.
(1997). Nucleic Acids Res., 25, 3389.
2. Eddy, S. R. (1998). Bioinformatics, 14, 755.
3. Grundy, W. N., Bailey, T., Elkan, C. P., and Baker, M. E. (1997). Comput. Appl. Biosci., 13,
397.
4. Neuwald, A. F., Liu, J. S., lipman, D. J., and Lawerence, C. E. (1997). Nucleic Acids Res.,
25,1665.
5. Krogh, A., Brown, M., Mian, I. S., Sjolander, K., and Haussler, D. (1994). J. Mol. Biol,
235,1501.
6. Bucher, P., Karplus, K., Moeri, N., and Hoffmann, K. (1996). Camp. Chem., 20, 3.
7. http://www.gcg.com/
8. http://www.netid.com/
9. http://www.compugen.com/
10. Bateman, A., Birney, E., Durbin, R., Eddy, S. R., Finn, R. D., and Sonnhammer, E. L.
(1999). Nucleic Acids Res., 27, 260.
11. Hofmann, K., Bucher, P., Falquet, L, and Bairoch, A. (1999). Nucleic Acids Res., 27, 215.
12. Ponting, C. P., Schultz, J., Milpetz, P., and Bork, P. (1999). Nucleic Acids Res., 27, 229.
13. Attwood, T. K., Flower, D. R., Lewis, A. P., Mabey, J. E., Morgan, S. R., Scordis, P., et al.
(1999). Nucleic Adds Res., 27, 220.
14. Henikoff, J. G., Henikoff, S., and Pietrokovski, S. (1999). Nucleic Acids Res., 27, 226.
15. Altschul, S. F. and Gish, W. (1996). In Methods in enzymology (ed. R. F. Doolirtle), Vol.
266, p. 460. Academic Press.
16. Brown, M., Hughey, R., Krogh, A, Mian, I. S., Sjolander, K., and Haussler, D. (1993) In
Proceedings of the First International Conference on Inteligent Systems for Molecular Biology (ed.
L. Hunter, D. Searls, and J. Shavlik), p. 47. AAAI Press, Menlo Park, CA, USA.
17. Thompson,J.D., Higgins, D. G., and Gibson, T. J. (1994). Nucleic Acids Res., 22, 4673.

91
This page intentionally left blank
 Chapter 5
Protein family-based methods
for homology detection and
analysis
Steven Henikoff and Jorja Henikoff
Howard Hughes Medical Institute, Fred Hutchinson Cancer Research Center,
Seattle, USA.

1 Introduction
1.1 Expanding protein families
Most methods for homology detection have traditionally relied upon pairwise
comparisons of protein sequences, and in recent years, several improvements in
pairwise methods have been introduced (see Chapter 8). But, with sequence data
becoming available at an accelerating rate, there is an increasing opportunity to
use multiple related sequences for improved homology detection. Even when
functional information is lacking for known members of a protein family, these
members can be aligned and the alignments used in searches. Protein multiple
alignments have been shown to improve performance of secondary structure
prediction methods by identifying constraints on positions (1,2), and so it seems
reasonable to expect that improvements will be likewise obtained by using
multiple alignments for homology detection and analysis. In this chapter we re-
view some of the numerous methods that are aimed at achievement of this goal.

1.2 Terms used to describe relationships among proteins

Any region of shared similarity between sequences may be referred to as a 'motif.
To call something a motif does not necessarily imply that shared similarity re-
flects shared ancestry. For example, the well-studied helix-turn-helix DNA bind-
ing motif is found in proteins belonging to apparently unrelated families with
different origins, and this suggests convergence towards a common structure.
Confusion often arises from the use of the structural term 'domain' to de-
scribe regions of sequence similarity. A separately-folded domain may be obvious
from looking at the structure of a protein, but, without seeing a structure, it
may not be possible to decide from an alignment what is the limit of a domain.

93
STEVEN HENIKOFF AND JORJA HENIKOFF

Furthermore, domains need not be contiguous along a sequence, and it is com-

mon for proteins to fold starting with one domain, continue on to fold into
another domain, then return to the original domain further along the sequence.
For sequence analysis applications, a useful concept is that of 'module'. A
module can be thought of as a sequence segment that may be found in different
contexts in different proteins, the result of mobility during protein evolution.
Modules may correspond to separately folded domains, such as the C2H2 zinc
finger motif, and they may be repeated within a sequence. Unlike domains,
modules are necessarily contiguous along a sequence. Nevertheless, readers
should be aware that modules identified by sequence similarity are typically
referred to as 'domains' without confirming structural evidence, and the term
'multi-domain' is commonly applied to any chimeric protein.
'Family' is a generic term used to describe proteins (or genes) with sufficiently
high sequence similarity that common ancestry may be inferred A multi-domain
protein might have modules that belong to several different families. Confusion
can arise from the use of terms such as 'superfamily' and 'subfamily', which are
not precisely defined. For these terms to be useful, some sense of what is meant
by a family is required. Thus, if we refer to the opsins, the beta-adrenergic re-
ceptors and the olfactory receptors as separate families, even though they are
related to one another, then we would refer to the G-protein coupled receptor
superfamily to describe them all. Sometimes, proteins fold similarly, even
though no sequence similarity between them is detected. For instance, the TIM
barrel fold has been found for dozens of separate superfamilies, and it is not
certain as to whether they share common ancestry. Conversely, sequence simi-
larity may be evident, even though common ancestry is doubtful, as in the case
of coiled-coil regions of proteins. As a practical matter, the methods described in
this chapter are most useful for families and modules, where alignment-based
methods can provide profound functional insights.

1,3 Alternative approaches to inferring function from

sequence alignment
Opposing views of sequence alignment problems have resulted in two different
classes of comparison tools for sequence analysis of protein families. Motif-based
tools consider aligned protein sequences to consist of nuggets of alignment
information (blocks) separated by regions that have no certain alignment. To
proponents of this view ('blockers'), the task is to first find these conserved
nuggets. 'Gappers' agree that there are nuggets worth finding, but that these
will be best found by determining where to place the gaps in each sequence
such that the blocks correctly align. Both blockers and gappers agree that align-
ing conserved nuggets is worthwhile, but they use different methods for accom-
plishing this. Blockers favour motif-based methods that first find regions of con-
servation. Such block-based methods as the BLAST family of searching programs
and the BLOSUM amino acid substitution matrices continue to be favoured for
many comparative sequence analysis applications (Chapter 8). Gappers favour

94
 PROTEIN FAMILY-BASED METHODS FOR HOMOLOGY DETECTION AND ANALYSIS

methods that decide upon gap placement (described in Chapter 3) and use
gap-based tools, especially dynamic programming and hidden Markov models
(described in Chapter 4), for database probing. As is so often the case, the truth
lies somewhere between the extremes. So although we blockers prefer to
reduce the protein alignment problem to finding a set of ungapped blocks to
represent a protein family or module, we recognize that insertions and deletions
occur occasionally within conserved regions, and this is challenging for block-
based methods.
Alignment usefulness is the major driving force in developing methodology.
Obtaining a correct alignment is more important for some applications than for
others. The ability to find corresponding residues and local regions that have
similar functions is of unquestionable value, and the better the conservation of
a residue or local region in a sequence, the more likely it is that common func-
tion can be inferred. Regions of uncertain alignment, such as those that differ-
ent alignment programs using various score parameters disagree on, have little
if any value for drawing functional inferences. However, so much alignment
information is present in conserved regions that it might make sense to align
beyond what can be done with confidence in order that more nuggets are
captured. We suspect that this accounts for the success of many gap-based
approaches: gapped alignments may have a high degree of uncertainty, but the
proportion that is aligned successfully is sufficient to identify extensive shared
regions of sequence similarity in database searches, even to the point of dis-
covering correct folds more successfully than structure-based threading (3).
Practical utility requires ready availability to the general public. Nowadays,
this means access via the World Wide Web using a browser, and so nearly all
methods highlighted here (Table 1) can be performed without any special soft-
ware, hardware, or computational expertise. Some potentially powerful tools
are too computationally intensive to be made available in this way. Additionally,
some tools require a specialist's knowledge and are not sufficiently automated
for the average biologist to use them wisely. We believe that such tools should
be avoided if possible: sequence alignment is fraught with hazards, and erron-
eous conclusions drawn from naive use of powerful sequence analysis tools
abound (4).

2 Displaying protein relationships

2.1 From pairwise to multiple-sequence alignments
Depictions of pairwise sequence alignments are not easily extended to multiple
alignments. For displaying pairwise alignments, identities and conservative
replacements are typically emphasized using symbols between the aligned
sequences. However, adding just a third sequence below the first two leaves
open the question of how to represent similarities between the third sequence
and the first, and addition of more sequences becomes increasingly complex.
Dot matrix representations of pairwise alignments present the same problem.

95
STEVEN HENIKOFF AND JORJA HENIKOFF

Table 1 URLs
1. Displaying alignments
Boxshade http://www.ch.embnet.org/software/BOX_form.html
Logos http://blocks.fhcrc.org/about_logos.html
Trees http://blocks.fhcrc.0rg/about_trees. html

2. Finding alignments
BCM launcher http://dot.imgen.bcm.tmc.edu:9331/multi-align/multi-align.html
MACAW ftp:/ncbi. nlm.nih.gov/repository

3. Searching family databases

Prosite http://www.expasy.ch/prosite/
Blocks http://blocks.fhcrc.org/blocks_search.html
Prints http://www.bioinf.man.ac.uk/fingerPRINTScan/
bin/attwood/SearchPri ntsForm2.pl
http://blocks.fhcrc.org/blocks_search.html
ProDom http://www.toulouse.inra.fr/prodom/doc/blast_form.html
Pfam http://pfam.wustl.edu/hmmsearch.shtml
http://www.sanger.ac.uk/Pfam/search.shtml
Proclass http://www-nbrf.georgetown.edu/gfserver/geneflnd.html
ProfileScan http://www.isrec.isb-sib.ch/software/PFSCANJtorm.html
Identify http://dna.Stanford.EDU/identify/
Recognize http://dna.stanford.edu/ematrix/
Prof_Pat http://wwwmgs.bionet.nsc. ru/mgs/programs/prof_pat/

4. Searching with multiple alignments

MAST http://meme.sdsc.edu/meme/website/mast.html
COBBLER http://blocks.fhcrc.org/
PSI-BLAST http://www.ncbi.nlm.nih.gov/cgi-bin/blast/psiblast.cgi
LAMA http://blocks.fhcrc.org/LAMA_search.html

Such displays not only become complex, but also they fail to represent shared
similarities. Because of these limitations, multiple alignment representations
that emphasize regions of high similarity have been introduced.
Traditional displays of multiple sequence alignments show aligned sequences
one above the next, highlighting identical or similar residues in a column using
boxes, shading, or colour. These displays can be complex, especially when repre-
senting protein families that consist of large numbers of sequences that group
into distinct subfamilies. Position-based representations greatly simplify the
display of multiple alignments, because relationships between successive amino
acids in a sequence are not shown. Indeed, computer programs that utilize multi-
ple sequence alignment information in searches likewise consider all positions
in aligned sequences to be independent of one another, and so position-based
representations depict approximately what a searching program examines.

2.2 Patterns
The simplest position-based representations of multiple alignments are patterns,
which display only key conserved residues. The Prosite database (5) is a com-

96
 PROTEIN FAMILY-BASED METHODS FOR HOMOLOGY DETECTION AND ANALYSIS

pilation of sequence families that provides one or more patterns representing

each family. An example of a Prosite pattern is C-x(2,4)-C-x(3)-[LIVMFYWC]-x(8)-
H-x(3,5)-H, which is read as: cysteine, followed by 2 to 4 amino acids of any type,
followed by cysteine, followed by any 3 amino acids, followed by one of the
following—leucine, isoleucine, valine . . . , and so on. Although Prosite patterns
are manually derived from multiple sequence alignments, the process of deter-
mining patterns from alignments has been automated (6, 7). Pattern-based
methods are described in detail in Chapter 7.

2.3 Logos
Sequence logos (8) are vivid graphical displays of multiple sequence alignments
consisting of ordered stacks of letters representing amino acids at successive
positions (Figure 1). The height of a letter in a stack increases with increasing
frequency (or probability) of the amino acid, and the height of a stack of letters
increases with increasing conservation of the aligned position. Stack heights are
displayed in bit units. One bit is the answer to a yes-or-no question, where yes is
as likely as no. About 4 bits are required to fully specify a residue at a given
position, because the first question narrows the field from 20 residues to 10, the
second to 5, etc. The most probable amino acid is at the top of the stack, making
it more visible, and below it is the next most probable residue, and so on. Logo
colours or shades are chosen to emphasize similar amino acid properties. Logos
can be scaled such that the stack height is proportional to the observed
frequency of a residue divided by the frequency with which the residue is
expected to occur by chance (odds ratio).

2.4 Trees
The most serious drawback of position-based displays is that they show only
alignment information in common among the sequences in a family, not the

Figure 1 Sequence logo depicting the chromodomain block (BL00598 in Blocks v. 11.0).
Each alignment position is represented as a stack of letters, where the height of the stack
and the height of each letter is measured in bit units.

97
STEVEN HENIKOFF AND JORJA HENIKOFF

differences that distinguish between sequences or groups of sequences. In con-

trast, trees are designed to discriminate between individual sequences by pro-
viding an intuitive diagram of relationships drawn from an alignment. Although
trees were introduced as phylogenetic tools, they have become increasingly
popular for displaying protein families. Trees can generally be used to distin-
guish orthologs from paralogs, because orthologs will branch in a manner that
is consistent with species phylogeny, whereas paralogs may deviate. Paralogous
proteins often have distinct biological activities, and so trees can guide experi-
mental investigations.
For using trees to draw inferences about function and not to infer phylogeny,
some aspects of tree construction and interpretation that matter to phylo-
genetics may be relatively unimportant. Debate continues about how a tree
should be constructed from an alignment, whether to use parsimony, distance,
or maximum-likelihood methods. However, we are unaware of evidence that
the choice of a tree-making program matters much for distinguishing paralogs
from orthologs or for deciding whether one branch has a function that is
comparable to the function of another branch. That is, we are not using a tree to
distinguish whether bakers yeast is closer to fruit flies than to maize, so we can
ignore the details at the leaves of a tree and focus on the separation of one group
of yeast, fly, and maize proteins from paralogous groups of proteins. In our
experience, the quality of the alignment might be important in making such
distinctions, but different tree-making programs draw trees that are sufficiently
similar for our purposes. Distance methods such as neighbour-joining (9) do
have the advantage of being fast and can then be applied to large numbers of
proteins and thus are suitable for making trees to analyse protein families.

3 Block-based methods for multiple-sequence

alignment
Searching methods that utilize multiple sequence alignment information can be
block-based or gap-based. Gap-based methods for finding multiple alignments
are described in detail in Chapter 3, and gap-based methods for searching with
them, such as hidden Markov models (HMMs), are described in Chapter 4. In this
section, we describe block-based strategies for finding multiple-sequence align-
ments that are then used for database searching by many of the methods
described in subsequent sections.

3.1 Pairwise alignment-initiated methods

One general approach to finding motifs involves performing pairwise compari-
sons between sequences and then asking which high-scoring local regions are in
common for most or all of the sequences in the group. Where aligned segment
pairs overlap, they are multiply aligned. However, determining which segments
truly overlap can be challenging, and different methods have been introduced
(10-12). In the MACAW program, overlapping segment pairs that exceed a

9a
 PROTEIN FAMILY-BASED METHODS FOR HOMOLOGY DETECTION AND ANALYSIS

threshold score are combined into an ungapped block (12). The extent of the
block is limited by the requirement that each column have some minimum de-
gree of homogeneity. Blocks that are separated by the same number of residues
in all sequences may be fused, and so blocks can contain both conserved and
diverged positions. MACAW is an interactive program that allows users to choose
a set of blocks from among candidates. The threshold score for block searching
can be relaxed by the user in order to find new blocks in regions between blocks
that were found in the first pass.
Starting with pairwise alignments presents the same potential drawback as
for gap-based hierarchical multiple sequence alignment programs (Chapter 5),
which is that information in common for all of the sequences might not be
represented in the pairwise alignments. In addition, the number of pairwise
comparisons needed is n2 for n sequences, and this can become somewhat im-
practical for large protein families and long sequences. Simultaneous methods
for finding motifs, described below, can potentially avoid these problems.

3.2 Pattern-initiated methods

The rapidity with which amino acid 'words' can be scanned exhaustively through
a set of related sequences has motivated pattern-based motif finders (13-17). An
example of this approach, Motif (15), examines all sequences for the presence of
spaced triplets of the form aa1 d1 aa2 d2 aa3 where d1 and d2 are fixed distances
between the amino acids. So Ala-Ala-Ala is one triplet, Ala-x-Ala-Ala is another,
and Leu-x[16]-Ala-x[7]-Val is another. An exhaustive search is carried out for all
such triplets in the full set of related sequences using all combinations of d1 and
d2 out to a reasonable maximum distance (about 20). The rationale is that true
motifs will typically include one or more sets of spaced triplets in all of the
sequences in the group. Because some true motifs do not contain aa1, aa2, and
aa3 in the full set of sequences, the number of sequences required to contain a
triplet (the 'significance level') can be reduced. In such cases, the block con-
taining the triplet is scanned along each of the sequences that lack the triplet to
find the best segment based on maximizing an overall score for the block. Each
sequence is then rescanned to maximize the score. ASSET generalizes the search
for patterns by scanning sequences for shared flexible patterns that occur in
multiple sequences at a statistically significant level (17).

3.3 Iterative methods

Other approaches avoid limiting the motif search to a predetermined list with-
out becoming computationally explosive by detecting motif'seeds' that occur in
as few as two sequences, then asking whether any of these seeds can mature to
include other sequences in the group (18-21). Both Expectation-maximization
(EM) and 'Gibbs sampling' (20) start with a block of specified width, then align
random positions within all but one sequence. In EM, this sequence is scanned
along the block, and the segment that maximizes a block score is chosen. In
Gibbs sampling, the segment is chosen by a random sampling procedure, where

99
STEVEN HENIKOFF AND JORJA HENIKOFF

the probability of being chosen is proportional to the block score. Other

sequences are then sampled in the same way to further improve the significance
of the alignment. Successive rounds of EM or Gibbs sampling continue until no
further improvement is seen.

3.4 Implementations
Some of these methods are conveniently available over the internet, and
sequences in FASTA format may be submitted by either pasting into a window
or by file browsing. Because many real motifs can be subtle and as short as a few
residues, sensitive methods may return alignments for sequences that are not
based upon true relationships. A simple experiment (Protocol !) demonstrates
that even sequences chosen at random from a database can be aligned to yield
motifs that appear convincing and will easily detect the parent sequences and
their homologues from sequence databanks. Furthermore, even gross misalign-
ments can be masked by the existence of significant similarity among just a
fraction of sequences, and visual examination is notoriously unreliable (Figure 2).
One solution is to report a reliability measure for each position (22), and one
measure is implemented in Match-Box (23). BlockMaker's solution (24) is to
apply two very different motif finders with different scoring systems, Motif and
Gibbs sampling. In each case, a block assembly algorithm (25) is used to
determine a best set of blocks representing a protein family, and the two sets
are compared by the user: blocks with similar alignments obtained by the two
methods may be trusted, but those that differ require scrutiny. Both Match-Box
and BlockMaker require that the blocks be in order along the sequences, and so
repeats might be missed. However, the EM-based program, MEME (21), does not
impose an ordering criterion, and MEME finds repeats and displays them within
blocks. BlockMaker, MEME and Match-Box are available from the BCM multiple
alignment search launcher, which allows successive searches of a single query
with several tools, both traditional and motif-based. Performance evaluation of
the methods available over the Web show that there are trade-offs between
sensitivity and reliability (23), and so it is worthwhile to try several methods on
any particular set of sequences and compare the results.

Finding motifs from unallgned sequences and searching

sequence databanks
1 Go to the Swiss-Prpt Random-entry retriever (hdp://www,expasy,ch/sprot/get-random-
entry.html) and successively extract 10 Swiss-Prot sequences of length > 300 aa
residues in FASTA format (look for the 'FASTA format' link at the bottom of the page).
2 Copy and paste these sequences into the large box of the BCM alignment launcher
(http://dot.imgen.bcmtmc.edu:9331/multi^lign^milti-align.htinl), Choose BlockMaker
and submit.

100
 PROTEIN FAMILY-BASED METHODS FOR HOMOLOGY DETECTION AND ANALYSIS

3 From the BlockMaker results page, examine the alignments in both sets of blocks
(from Motif above and Gibbs below), and choose the set of blocks that has the most
total residues. Click on the MAST direct link (to http://meme.sdsc.edu/meme/
website/mast.html) above the chosen set. MAST search results will be returned by
e-mail.
4 Compare the names of your submitted sequences to the significant hits from your
MAST search. Other significant hits may be homologues of your submitted sequences.
5 Now that you have done the necessary control, you are ready to use your own
sequences. Return to the BCM alignment launcher, copy and paste your own
sequences into the box and successively click on the various choices of multiple
aligners.

Figure 2 A typical example of a set of MOTIF-generated blocks obtained using Protocol 1.

Boxshade was used to highlight 'conserved' positions among the randomly chosen
sequences. Using these blocks as input to MAST for searching Swiss-Prot each of the
sequences represented in the block was detected with an E-value between 1.6 x 10-16 and
7.4 x 10 6. The first unrelated sequence was detected E = 0,212.

The interactive MACAW program provides an excellent alternative to auto-

mated web-based multiple aligners. The program allows a user to choose either
MACAW or Gibbs sampling for making blocks, and to make parameter choices
at different stages in the alignment process. The program is available to run
under popular computer operating systems.

4 Position-specific scoring matrices (PSSMs)

Alignments, patterns, logos, and trees provide useful visual displays, but for
searching databases, score-based representations are most widely used. These
were Introduced by McLachlan (26) and popularized by Gribskov et at. (27) who
coined the term 'position-specific scoring matrix' (or PSSM, pronounced
'possum'). A PSSM consists of columns of weights for each amino acid derived
from corresponding columns of a multiple sequence alignment. Other terms
have been used to describe this basic idea, including weight matrix, profile and
HMM. Profiles are PSSMs constructed using the average score method (27),
although the term has also been used to describe matrices representing a string
of local environments for successive residues in a structure (28). Profile HMMs,
described in Chapter 4, are PSSMs that are constructed using an iterative prob-
abilistic algorithm for determination of position-specific gap penalties (29-31). A

101
STEVEN HENIKOFF AND JORJA HENIKOFF

simple PSSM has as many columns as there are positions in the alignment, and
20 rows, one for each amino acid. In some applications, a PSSM consists of rows
that correspond to successive positions in the alignment (27), rather than
columns, and in some, there are position-specific gap scores.
Because they consist of numbers, PSSMs are useful for computer-based align-
ment and database searching methods but not for visual display. However, logos
are computed from PSSMs, and rules can be applied to convert PSSMs to pat-
terns (6) or consensus sequences (32). The construction of PSSMs from multiple
alignments has improved over the years, and as a result, we are better able to
detect weak similarities in searches (33). To construct effective PSSMs, two
major issues, described below, must be addressed.

4.1 Sequence weights

PSSM performance can be improved by differentially weighting sequences to
reduce redundancy resulting from non-representative sampling of sequences
(34-37). Very similar sequences get low weights and more diverged sequences
get higher weights in order to make a PSSM more representative of the family as a
whole. Several different strategies have been used to arrive at sequence weights.
Some methods start with a tree and find a root, where the weight of a sequence
is proportional to its distance from the root (e.g. 35). Pairwise distance methods
calculate a weight from the average distance of a sequence to all other sequences,
either to the observed sequences or to imaginary sequences derived by sampling
residues from the observed sequences (e.g. 36). Position-based sequence weights
are calculated by determining the weight of a residue within its column in an
alignment and adding residue weights for all positions (37). In the maximum
discrimination method, weights are chosen to best discriminate between true
positives and background (31). Comprehensive empirical evaluation of sequence
weighting methods has revealed that weighting sequences is much better than
not weighting at all (37). However, no single method stands out, and at least one
variant of each of these strategies provided excellent results.

4.2 PSSM column scores

Pairwise alignment methods utilize amino acid substitution matrices to provide
a set of scores for each aligned residue. Current applications utilize log-odds
scores computed from alignment data, such as PAM (38), JTT (39), or BLOSUM
(40) substitution matrices, and theory advocates the suitability of log-odds scores
for pairwise alignments (41). However, scoring a multiple alignment against a
sequence is more complex, requiring a scoring scheme that is able to utilize the
observed occurrences of residues in a column corresponding to an alignment
position. That is, the column of an alignment should be modelled in a way that,
when aligned with each of the 20 amino acids, a meaningful score can be
obtained. The original average score method (27) simply extends the use of
pairwise scores by averaging them. For instance, when aligned with a serine, a
position represented by an alanine and 3 cysteines would get a score equal to

102
 PROTEIN FAMILY-BASED METHODS FOR HOMOLOGY DETECTION AND ANALYSIS

the serine-alanine score plus 3 times the serine-cysteine score divided by 4

(ignoring sequence weights which would alter the relative contribution of each
occurrence to the sum).
Unfortunately the average score is insensitive to the number of sequences in
the multiple alignment. The average score for a serine aligned with an alanine
and 3 cysteines is identical to that for a serine aligned with 10 alanines and 30
cysteines. The problem with this situation is that a serine might be expected to
occur frequently given only 4 observations of such similar residues, but after 40
observations without seeing a serine, we would expect to see one only rarely.
Therefore, the average score method becomes less and less realistic as the
number of different sequences in an alignment increases. An effective way of
dealing with this problem is to add 'pseudocounts' to the observed counts of
residue occurrences (42-44). Intuitively, this is equivalent to adding hypothetical
sequences to those that have been observed, and for each sequence, the choice
of residues at each aligned position is governed by what might be expected for
real related sequences not yet seen. So if we have already observed an alanine
and 3 cysteines, we might expect to see more cysteines and alanines, but also
occasional serines but maybe not arginines. Hypothetical occurrences can be
added to real occurrences as fractional pseudocounts. Notice that if we add the
same pseudocounts when 10 alanines and 30 cysteines have been seen, then the
relative proportion of real observations to pseudocounts increases 10-fold; this
conforms with our intuition that we are much more certain that the observed
occurrences adequately model future occurrences when we have a large num-
ber of independent observations. Comprehensive evaluations demonstrate that
using pseudocounts modeled on alignment data, much better overall perform-
ance is obtained than using the average score method (33,44).

5 Searching family databases with sequence

queries
For any protein sequence of interest, a search of the latest databanks is the first
and often the most important step toward understanding function, and the
identification of homologues in this way has been a major driving force in both
academic biology and in the growing genomics industry. A second step should
involve searching protein family databases. There are several reasons for this:
Making sense of dozens or hundreds of hits in the sequence databanks can be
challenging, whereas hits in protein family databases provide immediate classi-
fication and entries to the literature. The different regions of multi-domain pro-
teins are readily classified using family databases, whereas in searches of
sequence databanks, modules can be missed if hits to family members contain-
ing them are low on the list. Searches of family databases can be more sensitive
than searches of sequence databanks because multiple alignment information is
utilized. The much smaller size of family databases, typically only ~1% the size
of sequence databanks, reduces noise.

103
STEVEN HENIKOFF AND JORJA HENIKOFF

Currently, there are several choices of family databases and searching options
available over the internet (Table 1). An illustrative example is depicted in Figure 3,
which shows how well the different methods detected key features of a protein
that we recently described, a cytosine-5 DNA methyltransferase homologue with
an embedded chromodomain module, called a 'chromomethylase', which is
encoded by the Arabidopsis thaliana CMT1 locus (45). In addition to being the
subject of current experimental work in our group, we chose this sequence
because chromomethylases are not yet present in any of the family databases,
although both the cytosine-5 DNA methyltransferases and the chromodomains
are represented in all of them, and because this example reveals strengths and
weaknesses of the different methods especially well. Both the DNA methyltrans-
ferase and the chromodomain represent novel subfamilies of their respective
families, and so detection in their entirety can be challenging for a protein
family classification method that does not generalize well from known examples.
This is an anecdotal example, and overall performance can only be judged using

Figure 3 Classification of the 791 aa A. thaliana chromomethylase by family databases. The

horizontal line indicates the length of the protein from the amino (N) to the carboxyl (C) end,
the closed boxes show the extent of cytosine-5 DNA methyltransferase regions detected and
the open boxes show chromodomain regions. For methods that report E- or p=values, a 0.05
level of significance was considered to be the threshold for detection, and this exceeded the
level of the highest-scoring false positives. For methods that do not report E- or p-value
statistics (Prosite, Printscan, Prof_pat, Proclass) or for those that report multiple levels of
stringency (Identify and Recognize), the threshold level of detection was considered to be
just above the first false positive hit.

104
 PROTEIN FAMILY-BASED METHODS FOR HOMOLOGY DETECTION AND ANALYSIS

comprehensive empirical evaluations. However, because coverage of different

databases varies widely, such rigorous direct comparisons have not been carried
out.

5.1 Curated family databases: Prosite, Prints, and Pfam

Prosite is the original family database, introduced in 1989. Prosite provides ex-
cellent documentation and carefully crafted patterns for searching (Section 2.2).
In cases where patterns are difficult to find, Prosite provides a profile PSSM
(Section 4). Prosite 15 (July, 1998) has 1020 documentation entries, mostly repre-
senting families, and 1358 patterns based on sequences in Swiss-Prot. Searching
a query sequence against Prosite patterns is strictly a hit-or-miss affair, and no
statistics are provided. The chromomethylase example illustrates this vividly.
Prosite reported the chromodomain, which is a highly diverged module that is
relatively difficult to detect. However, Prosite failed to detect the DNA methyl-
transferase, even though some of the conserved regions are very easily detected
by standard searching programs (Figure 3) and this family is represented by two
patterns in the database. Indeed, a comprehensive empirical evaluation showed
that even standard BLAST searching outperforms searching of Prosite patterns
(46).
Prints, introduced in 1993, is similar to Prosite in providing excellent docu-
mentation. Rather than patterns, Prints provides carefully crafted 'fingerprint'
multiple alignments (ordered sets of blocks), that can be searched using pattern
or PSSM methods. Prints 20 (October, 1998) has 990 fingerprint entries for 5701
blocks based on sequences in the OWL protein database. Printscan detected all 3
blocks in the fingerprint representing the upstream and central conserved
regions of the DNA methyltransferase, but did not detect the chromodomain.
Maintaining curated databases and crafting patterns or fingerprints is made
especially difficult because of the rapid expansion of protein families in recent
years. Pfam (47), introduced in 1996, addresses this problem by using seed align-
ments that are manually constructed, and HMM (hidden Markov model) PSSMs
from the seeds are then used to automatically extract and align new sequences
from databanks. Unlike Prosite and Prints, Pfam does not provide documentation
beyond a family name and links to source databases and does not delineate con-
served regions within entries. Pfam 3.2 (October, 1998) has 1344 entries repre-
senting families and modules based on sequences in Swiss-Prot/TrEMBL. HMM
PSSMs are used to search Pfam. For the chromomethylase, all of the conserved
regions of the DNA methyltransferase and the chromodomain were detected.

5.2 Clustering databases: ProDom, DOMO, Protomap, and

Prof_pat
An alternative to curation is to search a database against itself, then duster similar
sequences into families automatically. Although the procedure sounds simple,
in practice it is fraught with difficulties owing to the complexity of proteins and
protein families and to the need to avoid chance similarities when comparisons

105
STEVEN HENIKOFF AND JORJA HENIKOFF

are carried out on such a large scale. The first public database of this type was
introduced in 1990 (48), and several have been introduced over the years, only
some of which are extant. ProDom, which was introduced in 1994 (49), has been
continually maintained and enhanced (50); version 36 (August, 1998) contains
17 777 entries from Swiss-Prot with more than 2 sequences. ProDom entries vary
from short single motifs to longer stretches of similarity that might encompass
nearly entire sequences. ProDom is searched with multiple alignments or con-
sensus sequences. Using either option, ProDom detected the central and down-
stream conserved regions of the DNA methyltransferase, missing the upstream
region and the chromodomain.
Recently, three new clustering databases have been introduced. DOMO (51),
which is based on Swiss-Prot and PIR, is similar to ProDom, although it uses
different methodology to generate the database. DOMO clusters tend to be
longer and fewer in number than ProDom clusters. At present, DOMO does not
allow user-supplied sequences to be searched for classification. Protomap (52),
which is based on Swiss-Prot, does not yield multiple alignments as do ProDom
and DOMO, but rather provides a graphical tree-like view of the clustering. To
classify a protein sequence with Protomap, a Smith-Waterman search of Swiss-
Prot is performed, and each individual cluster that contains a sequence hit is
reported. For the chromomethylase, Protomap detected the chromodomain and
the central and downstream conserved regions of the DNA methyltransferase,
missing the upstream region of conservation. Prof_pat (53) extracts patterns
from clustering Swiss-Prot/TrEMBL, and these can be searched. Prof_pat did not
detect either the DNA methyltransferase or the chromodomain above false
positives.

5.3 Derived family databases: Blocks and Proclass

Intermediate between the curated and automated databases are those that utilize
protein family groupings provided by other resources. The Blocks Database,
which was introduced in 1991, uses the automated Protomat system for finding
blocks (ungapped regions of local conservation) representing a protein family.
Starting with Swiss-Prot sequences listed in Prosite family entries, alignment
blocks are found (patterns or profiles provided with Prosite are not used) and
concatenated into a database. Blocks 11.0 (August, 1998) contains 994 families
and 4034 blocks based on Swiss-Prot and is searched using the PSSM-based
BlockSearch method that reports single and multiple block hits along a
sequence. Whereas other protein family searchers on the internet require a
protein sequence, Blocks can be searched with a DNA sequence query, in which
case hits from all three frames on each strand are assembled. For searching, the
current default database is Blocks+, a superset of families from Blocks, Prints,
Pfam, ProDom and DOMO. Blocks+ (Nov. 1998) includes 8388 blocks repre-
senting 1922 families. Except for Prints, where fingerprint blocks are utilized
directly for searching, Protomat is used to make blocks for entries from each
database, and families that have block regions in common are removed to avoid

106
 PROTEIN FAMILY-BASED METHODS FOR HOMOLOGY DETECTION AND ANALYSIS

redundancy. BlockSearch detected the chromodomain and all of the conserved

regions of the DNA methyltransferase. When the Prints database was searched
with BlockSearch, all 3 upstream and central DNA methyltransferase motifs and
all 3 chromodomain motifs in Prints were now detected at highly significant
levels.
Proclass (54), introduced in 1997, also combines families from different sources:
Prosite, PIR superfamilies and families automatically discovered using the GenFind
program (55). Proclass v. 3 (March, 1998) contains 1275 Prosite groups and 3979
PIR superfamilies and is searched using a neural network-based system. To our
knowledge, Proclass searching is the only system that detects sequence similarity
using methodology that is not alignment-based. When the chromomethylase
sequence was searched, Proclass reported the Prosite chromodomain pattern
and both of the Prosite DNA methyltransferase patterns, which were missed by
the Prosite scanner.

5.4 Other tools for searching family databases

Identify (7) searches sequences versus pattern-based representations of individual
blocks and fingerprints derived from the Blocks and Prints databases. Because
patterns can be searched much more rapidly than scored-based representations
of multiple alignments (see Chapter 7), Identify search results are returned
within a second or so. Identify detected the chromodomain and only one down-
stream DNA methyltransferase blocks above all false positives. Using Recognize,
which is a score-based version of Identify, the central and other downstream
regions were detected as well.
A collection of profile PSSMs from Prosite, Pfam and other sources is available
for searching using generalized HMM-like profile PSSMs at the ProfileScan site
(56). ProfileScan reported the chromodomain and the central and downstream
DNA methyltransferase conserved regions but missed the conserved upstream
region.
In summary, there are numerous protein family searching tools available for
sequence classification. None is perfect, and as illustrated by the chromomethy-
lase example, it is worthwhile to try several of them for analysing a sequence of
interest. Pairwise sequence tools also varied in their ability to confidently detect
features of the chromomethylase. GAP-BLAST detected the chromodomain and
all DNA methyltransferase conserved regions, although subsequent iterations of
PSI-BLAST caused the chromodomain to be lost at the expense of the DNA
methyltransferase that surrounds it. FASTA failed to detect the upstream con-
served region of the DNA methyltransferase, and the chromodomain was re-
ported, but at a non-significant level. As a practical matter, the chromodomain
would have gone unnoticed or assumed to be a chance hit because it is preceded
by ~100 higher-scoring DNA methyltransferase sequences in databanks, and
indeed its presence was not noted in the original sequence entry (GenBank/
EMBL U53501). By using family databases for classification, this potential problem
can be minimized.

107
STEVEN HENIKOFF AND JORJA HENIKOFF

6 Searching with family-based queries

Finding homologues in sequence databanks underlies much of the recent pro-
gress in functional genomics, both in academia and industry, and pairwise
methods, such as BLAST searching, currently dominate. However, as more and
more sequences fall into families, opportunities increase for using family in-
formation for identifying modules and new family members. Progress in making
better PSSMs described in Section 3.3 has resulted in improvements in searching
performance, and practical tools have become available for taking advantage of
protein family information in searching sequence databanks.

6.1 Searching with embedded queries

A potential drawback to block-based approaches is that regions of uncertain
alignment are not scored, and the loss of this alignment information can poten-
tially reduce searching sensitivity. This problem arises because even with
effective motif-finding systems, the 'edges' of blocks are often uncertain, and
they might be chosen differently for different subsets of proteins in a family
(51, 57, 58). This problem has been addressed by implementation of a simple
'embedding' strategy: a consensus is determined for a set of related sequences,
the sequence that is closest to the consensus is chosen, and blocks are embedded
into that sequence (46). Because interblock regions of uncertain alignment are
represented as a single sequence, they cannot be misaligned (this would reduce
the specificity of a PSSM), while multiple alignment information in block regions
is retained. Embedding of PSSMs using this system has not been implemented on
the internet for general database searching, although the basic idea has been
incorporated into PSI-BLAST (described below). As an approximation, using the
COBBLER (Consensus Biasing By Locally Embedding Residues) system, a con-
sensus residue is determined for each position of all the blocks. A single
sequence is chosen as the one closest to the consensus over all block positions,
and these consensus residues are then substituted for the real residues in the
chosen sequence. This consensus-biased sequence can then be used to search
sequence databanks using available single sequence querying tools, such as
BLAST and PSI-BLAST. The improved overall performance that results is especi-
ally useful for identifying known modules in unexpected places: For instance,
the chromodomain in the A. thaliana chromomethylase (Figure 3) was initially
identified using a COBBLER-embedded sequence to search the nr protein
databank with BLAST (45).

6.2 Searching with PSSMs

Using PSSMs to search sequence databanks is computationally demanding, and
the availability of services is relatively limited. The Multiple Alignment Search-
ing Tool (MAST) program (57) searches block-based multiple alignments against
the standard sequence database sets, which are updated daily. MAST output
provides excellent statistics for both individual and multiple block hits with

108
 PROTEIN FAMILY-BASED METHODS FOR HOMOLOGY DETECTION AND ANALYSIS

block maps for intelligent interpretation of search results. MAST accepts PSSMs
directly from MEME and BlockMaker. Additionally, the Blocks server provides a
processor that can be used to convert other multiple alignments into efficient
PSSMs for sending directly to the MAST server.

6.3 Iterated PSSM searching

Several of the concepts highlighted above have been incorporated into the
Gap/PSI-BLAST searcher, an elegant extension of the popular BLAST database
searching program (59). The first round of searching employs Gap-BLAST, a new
pairwise method for detecting family members in the traditional way. From the
significant hits detected in the first round, a PSSM is constructed and this is used
to search the databank again, a process that can be repeated multiple times
until no further hits are reported above a chosen level of significance.
Gap-BLAST is especially notable because it represents a successful block-based
approach to the pairwise searching problem. When databanks are searched,
computational speed is an important factor, because finding an 'optimal' align-
ment using traditional methods for placing gaps is too slow to be practical on a
large scale with standard hardware. Speedy methods, such as FASTA and BLAST,
begin by searching exhaustively for matches or short motifs shared by two
sequences, extending these and stringing them together to find high scoring
alignments. However, the gain in speed is accomplished at the expense of re-
duced searching performance (60). Gap-BLAST combines speed with near-
optimal searching performance by starting with short motifs, but accepting only
those that define opposite ends of a high scoring ungapped alignment. This
alignment is extended, and only if it exceeds a threshold score is a gapped
alignment sought, that is, gapping is employed to optimize alignment of highly
similar regions. Searching performance of Gap-BLAST is nearly indistinguishable
from that of an optimal gap placement method (Smith-Waterman dynamic
programming) when the same scoring parameters are used. This is one inroad of
blocker concepts into the gapper realm for pairwise alignment; another is the
realization that pairwise alignment can be generally improved by allowing
highly dissimilar regions to be skipped over (61, 62).
PSSM construction in PSI-BLAST is similar to that described in Section 4, employ-
ing position-based sequence weights (37) and pseudocounts that are modelled
upon amino acid substitution probabilities (33, 44). The embedding concept
described above is generalized in PSI-BLAST to deal with the complication that
for any position in the query sequences, there may be a variable number of
database sequences that align. Thus, the final PSSM provides position-specific
scores that represent as few as one (the query sequence alone) and as many as all
of the sequences detected in the previous round of the search. The high sensi-
tivity to distant relationships provided by PSI-BLAST, and the enjoyment that a
user may get by iteratively searching for homologues in real time, can lead to its
overenthusiastic use, and serious errors may result. This is because any chance
hit that is included in the developing PSI-BLAST PSSM will almost inevitably pull

109
STEVEN HENIKOFF AND JORJA HENIKOFF

out its neighbours in subsequent rounds, and this can lead to erroneous infer-
ences of homology. A defence against this type of error is to use conservative
levels of statistical significance for addition of sequences to the PSSM. However,
because proteins are not comprised of random sequences of residues, the
random statistical model that underlies the BLAST programs can be unreliable
(63), and so novel conclusions drawn from iterative searches should be viewed
with appropriate caution.

6.4 Multiple alignment-based searching of protein family

databases
The effectiveness of utilizing family-based information to search databases
encourages the use of multiple alignments for searching multiple alignment
databases. LAMA (for local alignment of multiple alignments) is a program that
searches ungapped blocks versus family databases (64). In LAMA, PSSM columns
are scored against one another by calculating a correlation coefficient, and a
high scoring alignment is one in which the sequence-weighted distribution of
residues is highly similar overall between aligned columns. The high sensitivity
of LAMA for locally aligned regions has led to its use in discovering subtle
similarities, such as those shared by helix-turn-helix DNA binding motifs found
in unrelated modules.
Tools such as LAMA, which thrive on abundant alignment data, are likely to
become more widely used as protein families expand in size. Because of its high
sensitivity, LAMA or its descendants should become increasing valuable for
modelling 3-D structures of sequences by facilitating local alignment to family
members of known structure. As the percentage of unclassified proteins
dwindles, a major alignment-based problem facing biologists will be to deter-
mine which subfamily a protein belongs to, and from this, more precise struc-
tural and functional inferences may be made. We anticipate the development of
a next generation of computational tools to deal with this problem.

References
1. Rost, B. (1996). In Methods in enzymdogy (ed. R. F. Doolittle), Vol. 266, p. 525. Academic
Press.
2. Gamier, J., Gibrat, J.-F., and Robson, B. (1996). In Methods in enzymology, Vol. 266, p. 540.
3. Moult, J., Hubbard, T., Bryant, S. H., Fidelis, K., and Pedersen, J. T. (1997). Proteins:
Struct. Funrt. Genet., Suppl. 1, 2.
4. Henikoff, S. (1991). NewBiol, 3,1148.
5. Bairoch, A., Bucher, P., and Hofmann, K. (1997). Nucleic Acids Res., 25, 217.
6. Jonassen, I., Collins, J. F., and Higgins, D. G. (1995). Protein Sci, 4,1587.
7. Nevill-Manning, C. G., Wu, T. D., and Brutlag, D. L (1998). Proc. Natl Acad. Sci USA, 95,
5865.
8. Schneider, T. D. and Stephens, R. M. (1990). Nucleic Acids Res., 18, 6097.
9. Saitou, N. and Nei, M. (1987). Mol. Biol. Evol, 4,406.
10. Vingron, M. and Argos, P. (1991). J. Mol. Biol., 218, 33.

110
 PROTEIN FAMILY-BASED METHODS FOR HOMOLOGY DETECTION AND ANALYSIS

11. Boguski, M. S., Hardison, R. C., Schwartz, S., and Miller, W. (1991). New Biol, 4, 247.
12. Schuler, G. D., Altschul, S. F., and Lipman, D. J. (1991). Proteins: Struct. Funct. Genet, 9,
180.
13. Sobel, E. and Martinez, H. M. (1986). Nucleic Acids Res., 14, 363.
14. Posfai, J., Bhagwat, A. S., Posfai, G., and Roberts, R. J. (1989). Nucleic Acids Res., 17, 2421.
15. Smith, H. O., Annau, T. M., and Chandrasegaran, S. (1990). Proc. Natl. Acad. Sri. USA, 87,
826.
16. Depiereux, E. and Feytmans, E. (1992). CABIOS, 8, 501.
17. Neuwald, A. F. and Green, P. (1994). J. Mol. Biol, 239, 698.
18. Bacon, D. J. and Anderson, W. F. (1986) J. Mol. Biol., 191,153.
19. Stormo, G. D. and Hartzell, G. W. 3rd (1989). Proc. Natl. Acad. Sri. USA, 86,1183.
20. Lawrence, C. E., Altschul, S. F., Boguski, M. S., Liu, J. S., Neuwald, A. F., and Wootton,
J. C. (1993). Science, 262, 208.
21. Bailey, T. and Elkan, C. (1994). In Proceedings of the Second International Conference on
Intelligent Systems for Molecular Biology, pp. 28-36. AAAI Press, Menlo Park, CA.
22. Notredame, C., Holm, L, and Higgins, D. G. (1998). Bioinformotics, 14, 407.
23. Briffeuil, P., Baudoux, G., Lambert, C., De Bolle, X., Vinals, C., Feytmans, E., et al.
(1998). Bioinformatics, 14, 357.
24. Henikoff, S., Henikoff, J. G., Alford, W. J., and Pietrokovski, S. (1995). Gene, 163, GC17.
25. Henikoff, S. and Henikoff, J. G. (1991). Nucleic Adds Res., 19, 6565.
26. McLachlan, A. D. (1983). J. Mol. Biol., 169, 15.
27. Gribskov, M., McLachlan, A. D., and Eisenberg, D. (1987). Proc. Natl. Acad. Sri. USA, 84,
4355.
28. Bowie, J. U., Luthy, R., and Eisenberg, D. (1991). Science, 253, 164.
29. Krogh, A., Brown, M., Mian, I. S., Sjolander, K., and Haussler, D. (1994). J. Mol. Biol.,
235, 1501.
30. Baldi, P., Chauvin, Y., Hunkapiller, T., and McClure, M. A. (1994). Proc. Natl. Acad. Sri.
USA, 91,1059.
31. Eddy, S. R., Mitchison, G., and Durbin, R. (1995).J. Comput. Biol, 2, 9.
32. Patthy, L. (1987). J. Mol. Biol., 198, 567.
33. Henikoff, J. G. and Henikoff, S. (1996). CABIOS, 12, 135.
34. Luthy, R., Xenarios, I., and Bucher, P. (1994). Protein Sri., 3,139.
35. Thompson, J. D., Higgins, D. G., and Gibson, T. J. (1994). CABIOS, 10, 19.
36. Sibbald, P. R. and Argos, P. (1990) J. Mol. Biol., 216, 813.
37. Henikoff, S. and Henikoff, J. G. (1994). J. Mol. Biol., 243, 574.
38. Dayhoff, M. (1978). Atlas of protein sequence and structure, Vol. 5, suppl. 3, pp. 345-58.
National Biomedical Research Foundation, Washington, DC.
39. Jones, D. T., Taylor, W. R., and Thornton, J. M. (1992). CABIOS, 8, 275.
40. Henikoff, S. and Henikoff, J. G. (1992). Proc. Natl. Acad. Sri. USA, 89,10915.
41. Altschul, S. F. (1991). J. Mol. Biol., 219, 555.
42. Dodd, I. B. and Egan, J. B. (1987). J. Mol. Biol., 194, 557.
43. Brown, M. P., Hughey, R., Krogh, A., Mian, I. S., Sjolander, K., and Haussler, D. (1993).
In Proc. First Int. Conf. on Intelligent Systems for Molecular Biology (ed. L. Hunter, D. Searls,
and J. Shavlik), pp. 47-55. AAAI Press, Washington DC.
44. Tatusov, R. L, Altschul, S. F., and Koonin, E. V. (1994). Proc. Natl. Acad. Sri. USA, 91,
12091.
45. Henikoff, S. and Comai, L. (1998). Genetics, 149, 307.
46. Henikoff, S. and Henikoff, J. G. (1997). Protein Sci., 6, 698.
47. Sonnhammer, E. L., Eddy, S. R., and Durbin, R. (1997). Proteins: Struct. Funct. Genet., 28,
405.

Ill
STEVEN HENIKOFF AND JORJA HENIKOFF

48. Smith, R. F. and Smith, T. F. (1990). Proc. Natl. Acad. Sri. USA, 87,118.
49. Sonnhammer, E. L. L. and Kahn, D. (1994). Protein Sci., 3, 482.
50. Corpet, F., Gouzy, J., and Kahn, D. (1998). Nucleic Acids Res., 26, 323.
51. Gracy, J. and Argos, P. (1998). Bioinformatics, 14, 174.
52. Gona, G., LInial, N., Tishby, N., and Linial, M. (1998). ISMB, 6, 212.
53. Bachinsky, A. G., Yarigin, A. A., Guseva, E. H., Kulichkov, V. A., and Nizolenko, L. P.
(1997). CABIOS, 13, 115.
54. Wu, C. H., Zhao, S., and Chen, H. L. (1996). J. Comput. Biol, 3, 547.
55. Wu, C. H., Shivakumar, S., Shivakumar, C. V., and Chen, S. C. (1998). Bioinformatics, 14,
223.
56. Bucher, P., Karplus, K., Moeri, N., and Hofmann, K. (1996). Comput. Chem., 20, 3.
57. Bailey, T. L. and Gribskov, M. (1997). J. Comput. Biol, 4,45.
58. Nicodeme, P. (1998). Bioinformatics, 14, 508.
59. Altschul, S. F., Madden, T. L., Schaffer, A. A., Zhang, J., Aneng, Z., Miller, W., et al
(1997). Nucleic Acids Res., 25, 3389.
60. Pearson, W. R. (1995). Protein Sci., 4, 1145.
61. Altschul, S. F. (1998). Proteins: Struct. Funct. Genet., 32, 88.
62. Alexandrov, M. M. and Luethy, R. (1998). Protein Sci., 7, 254.
63. Brenner, S. E., Chothia, C., and Hubbard, T. J. (1998). Proc. Natl. Acad. Sri. USA, 95, 6073.
64. Pietrokovski, S. (1996). Nucleic Acids Res., 24, 3836.

112
 Chapter 6
Predicting secondary structure
from protein sequences
Jaap Heringa
National Institute for Medical Research, The Ridgeway, Mill Hill,
London NW7 1AA, UK

1 Introduction
Protein structure is intrinsically hierarchic in its internal organization. The
highest level in this hierarchy is constituted by complete proteins or assemblies of
such proteins, which become subdivided through domains via super-secondary
structure to secondary structure at the lowest hierarchical level.
At higher levels within in this hierarchy, especially from the domain level
upwards, the connectivity of the polypeptide backbone between substructures
becomes less important. A protein thus can retain a stable structure irrespective
of the sequential arrangement of domains and presence of fragments linking
them together. Such linker regions often constitute exposed surface loops that
do not disrupt the folds of the domains they connect (1).
At the level of protein secondary structure, however, the elements are not
only crucially dependent on their amino acid compositions, but, unlike domain
and higher-order structures, are also very much context dependent; i.e. they rely
critically on the substructures in their environment. It is because of this context
dependency, that predicting protein secondary structure is a very difficult task,
which after three decades of research has not attained the accuracy on which
further prediction of tertiary structure can be based. It must be stressed,
however, that some successful predictions of higher-order structure, based on a
knowledge of the secondary structure, have been achieved (e.g. ref. 2).
This chapter covers some background aspects of secondary structure pre-
diction and describes recent and successful prediction methods, most of which
are available through the World Wide Web and so can be used by virtually every
biologist who likes to find out about the secondary structure associated with a
particular protein query sequence.

1.1 What is secondary structure?

Perhaps a suitable definition in the context of this chapter for a secondary
structure is that it is a consecutive fragment in a protein sequence, which
corresponds to a local region in the associated protein structure showing

113
JAAP HERINGA

Figure 1 Ribbon diagram of the flavodoxin

fold [PDB code 5null belonging to the
Q/B doubly-wound fold family.

distinct geometrical features. The two basic secondary structures are the a-helix
and the B-strand. Both show distinct structural features and are easily recogniz-
able in a protein structure (Figure 1). Other secondary structure types occurring
in protein structures are more difficult to classify as they are less regular than
a-helices or B-strands. Such structures are defined in the context of" most
prediction methods as coil; i.e. leftover secondary structures that cannot be
considered in the a-helical or pi-stranded conformation,
In general, about 50% of the amino acids fold into a-helices or B-stands, so
that roughly half the protein structures are irregularly shaped. The primary
reason for the regularity observed for helices and strands is the inherent polar
nature of the protein backbone, which contributes a polar nitrogen and oxygen
atom for each amino acid. To satisfy energetical constraints, the parts of the
main-chain buried in the internal protein core need to form hydrogen-bonds
between those polar atoms. The a-helix and B-strand conformations are optimal
as each main-chain nitrogen atom can associate with an oxygen partner (and via1
versa) whenever they adopt one of these two secondary structure types. It must
be stressed that, in order to satisfy their hydrogen-bonding constraints, p-strands
need to interact with other B-strands, which they can do in a parallel and anti-
parallel fashion, thus forming a B-pleated sheet. p-strands thus depend on crucial
long-range interactions between residues remote in sequence. They therefore
are more context dependent than a-helices, which would be more able to fold
'on their own'. The fact that the vast majority of prediction methods have
greatest difficulty in delineating B-strands correctly is believed to be due to their
pronounced context dependency,

1.2 Where could knowledge about secondary structure help?

Experimental evidence on early protein folding intermediates has shown that
secondary structural elements form at early stages during the folding process
(for a review, sec rcf. 3). These results support the significance of the so-called
'framework' model of protein folding (4, 5), where two or more secondary stnic-

11 4
 PREDICTING SECONDARY STRUCTURE FROM PROTEIN SEQUENCES

tural elements would associate early during folding to provide a structural frame
to which subsequently other substructures could attach. Therefore, knowledge
of protein secondary structural regions along the sequence is a prerequisite to
model the folding process or kinetics associated with it. Also for tertiary model
building, the ability to predict the secondary structure from the sequence alone
is crucial, as it allows for docking experiments to be carried out on the predicted
a-helices and B-strands.
On the architectural side of protein structure, it is possible to recognize the
three-dimensional topology by comparing the successfully predicted secondary
structural elements of a query protein with a database of known topologies.
Successful prediction here means parts of those helices and strands essential for
the topology would have been predicted, without necessarily accurate predict-
ion of the edges of those structures or the detection of non-essential secondary
structures. An example of topologically essential secondary structures for the
flavodoxin fold is given in Figure 2. The figure shows a schematic representation

Figure 2 TOPS diagrams for four flavodoxin structures and their basic topology. The
essential secondary structures are given in the basic topology diagram.

115
JAAP HERINGA

of the secondary structures as provided by the TOPS server (http://tops.ebi.ac.uk).

In a TOPS diagram (6), a a-helix is represented by a circle and a B-strand by a
triangle. The flavodoxin fold belongs to the class of a/B-folds with the essential
secondary structures distributed over the sequence as [Ba]5. The five strands fold
into a single p-pleated sheet ordered topologically as B2-B1-B3-B4-B5, where the
numbers indicate their relative position in the sequence and hyphens the
hydrogen bonded and spatial interactions between the strands. The five a-
helices, each following a p-strand, shield the B-sheet from the solvent and
therefore are of an amphipathic nature (see below). From the topologies of a few
different flavodoxin structures (Figure 2) can be seen that varying substructures
can be added on to the basic structure, albeit they do not disrupt the fold of the
topologically essential secondary structures. Therefore, proper prediction of the
sequential order of the topologically essential helices and sheets often allows
the recognition of the fold type associated with the protein sequence con-
sidered, thereby conferring the information pertaining to that fold. Further-
more, active sites of enzymes typically comprise amino acids positioned in
loops, so that, for example, identically conserved residues at multiple alignment
sites predicted to be in loop regions (i.e. not predicted as a-helix or B-strand),
could be functional and together elucidate the function of the protein (or
protein family) under scrutiny.

1.3 What signals are there to be recognized?

A number of observations on secondary structures as found in the large collec-
tion of protein structures deposited in the Protein Data Bank (PDB) (7), could be
summed up for each of the secondary structures a-helix, B-strand, and loop as
follows.

a-helix:
(a) As the number of residues per turn is 3.6 in the ideal case and helices are often
positioned against a buried core, they have one phase contacting hydro-
phobic amino acids, while the other phase interacts with the solvent. Such
amphipathic helices (8) thus show a periodicity of three to four residues in
hydrophobicity of the associated sequence stretch (Figure 3).
(b) Proline residues do not occur in middle segments as they disrupt the a-helical
turn. However, they are seen in the first two positions of a-helices.

B-strand:
(a) B-Strands mostly fold into so-called B-pleated sheets which have two strands
forming either edge. Therefore the hydrophobic nature of edge strands is dif-
ferent from that of strands internal to a B-sheet. As side-chains of constituent
residues along a p-strand alternate the direction in which they protrude,
edge strands of p-sheets can show an alternating pattern of hydrophobic-
hydrophilic residues, while buried strands tend to contain merely hydro-
phobic residues (Figure 3).

116
 PREDICTING SECONDARY STRUCTURE FROM PROTEIN SEQUENCES

(b) As B-strand is the most extended conformation (i.e. consecutive Ca atoms are
farthest apart), it takes relatively few residues to cross the protein core with a
strand. Therefore, the number of residues in a B-strand is usually limited and
can be anything from two or three amino acids, whereas helices shielding
such strands from solvent comprise more residues.
(c) The B-strands can be disrupted by single residues that induce a kink in the
extended structure of the main-chain. Such so-called 3-bulges are often
comprised of relatively hydrophobic residues.

Coil:
(a) Multiple alignments of protein sequences often display gapped and/or highly
variable regions, which would be expected to be associated with loop regions
rather than the two basic secondary structures.
(b) Loop regions contain a high proportion of small polar residues like Ala, Gly,
Ser, and Thr. Glycine residues are seen in loop regions due also to their
inherent flexibility.
(c) Proline residues are often seen in loops as well. They are not observed in
helices and strands as they kink the main-chain, although they can occur in
the N-terminal two positions of a-helices as mentioned above.
In addition to the positional requirements in hydrophobicity, there are also
general compositional differences between helix, strand and coil conformations
and this is the signal used in many of the early prediction methods (see below)
for single sequences. Methods that utilize multiple alignments can also exploit
the fact that the amino acid exchange patterns are different for the three
secondary structure states.
A few additional rules can help in clarifying the structure or function of a
protein sequence, once the secondary structure is predicted:
(a) Hydrophobic and particularly conserved hydrophobic residues are normally
buried in the protein core.
(b) More than 95% of all so-called B-a-B motifs; i.e. a B-strand followed in sequence
by a a-helix and another B-strand, show a right-handed chirality. The afore-
mentioned flavodoxin family (Figure 2) indeed shows only right-handed B-a-B
motifs. This fact can be used to build a topology for the secondary structures
of the sequence(s) considered.

Figure 3 Hydrophobic patterns along secondary structures.

117
JAAP HERINGA

(c) Helices often cover up a core of B-strands. Therefore, if both a-helices and (3-
strands are predicted, an attempt should be made to distribute the helices
evenly at either phase of a tentative B-sheet in topology modelling.
(d) As mentioned, strictly conserved residues in different regions of a multiple
alignment can be predicted with great confidence to be responsible for the
catalytic functions, particularly if they are polar and predicted to be in loop
structures hence unlikely to be buried. As active site residues are positioned
together in a protein 3-D structure, the coil structures they constitute should
be brought together in a topology model.

2 Assessing prediction accuracy

The most widely used way to assess the quality of an alignment is by calculating
the overall per residue three-state accuracy, called the (Q3:

where N is the total number of residues predicted and PS is the number of

correctly predicted residues in state S (S = H, B, or C). Some researchers use the
so-called Matthews' correlation coefficient as it more stringently estimates the
prediction accuracy for each structural state:

where Ps and Ns are respectively the number of positive and negative cases cor-
rectly predicted for the structural state considered, and ~PS and ~NS the num-
bers of false positives and negatives, respectively. Three-state predictions would
thus yield three Matthews' correlation coefficients. If overprediction or under-
prediction occurs for any of the structural states, this is more dramatically re-
flected in the Matthews' correlations than in the Q3 percentage. A third way to
assess prediction accuracy is by weights of evidence, defined for each secondary
structural type S as:

Although this measure is relatively robust to different sampling frequencies of

the structural states, the interpretation of the resulting values is not as straight-
forward as for the other two measures. Because understanding the Q3 measure
is the easiest and its use leads to just one percentage, it is the measure most
frequently used in the literature to report prediction accuracy.
A very important issue in assessing performance is the notion of sustained
accuracy. Knowledge about the average accuracy of a given method over a set of
predicted proteins is not meaningful if unaccompanied by the variance of those
predictions. It is important to know what worse case predictions can be ex-
pected from a method, even if its mean accuracy is quite high.
A standard scenario to assess prediction accuracy is the jackknife test carried
out over a large set of test proteins (see Protocol 1). This ensures that no infor-

118
 PREDICTING SECONDARY STRUCTURE FROM PROTEIN SEQUENCES

mation about a query sequence or multiple alignment is used in training the

method. Nonetheless, unnoticed but systematic tuning of the method to the
database might still occur, so that the most rigorous test of any method is the
prediction of test cases that have no homologues in the database and have not
been seen during the development of the method.
Notwithstanding the importance of the measures for accuracy as listed above,
the real success of a secondary structure prediction depends on how the know-
ledge is being used. An example is the aforementioned fold recognition, where
correct prediction of the edges of secondary structural elements is not essential,
but missing structures that are crucial for the basic topology is costly. However,
all above measures, equally penalize, for example, missing two residues at either
side of a seven-ammo acid strand or missing a complete topologically essential
strand of four residues.

Jackknife testing
1 Take out one protein of the complete set of N proteins,
2 Train the method on the remaining N-l proteins (the training set).
3 Predict the secondary structure for the protein taken out.
4 Repeat step 1-3 for all N proteins and assess the average accuracy.
It is possible to test the method by averaging the predictions over all combinations of x
proteins (1<x<N), each time using the method trained on the remaining N-x proteins.
This provides an impression of the influence of different training sets on the sustained
accuracy of a single protein being predicted. As the number of combinations grows
rapidly with x. the training phase of most methods is too slow for extensive testing using
this mode. It can, however, also be used to save computation time if the database is split
evenly in test groups of sequences (e.g. 9), as each sequence within a test group is
associated with a single training set, thus saving training overhead.

An additional problem in secondary structure prediction is the standard of

truth. Most prediction methods are assessed in accuracy by using known tertiary
structures from the protein data bank (7) with their secondary structural ele-
ments assigned using the DSSP method of Kabsch and Sander (10). Colloc'h et al
(11) compared three such secondary structure determination algorithms, among
which was the DSSP method, and found significant differences in their second-
ary structural assignments. This ambiguity in secondary structural assignments
can be dramatic for particular proteins where agreement of the methods can be as
low as 65% (12, 13). Moreover, in structurally equivalenced sets of homologous
proteins with known tertiary structure, the corresponding secondary structural
elements can vary in length or show shifts of one to a tew residues, and hence a
realistic maximum prediction accuracy per residue would be in the range 80-
100% (14). Many researchers have suggested that prediction evaluation should be

119
JAAP HERINGA

based on the overlap of predicted and observed segments rather than on indi-
vidual positions (15-20). A recent secondary structure assignment program that
combines many of the features of earlier methods, such as checking hydrogen
bonding patterns and stereochemical characteristics, is the knowledge-based
method STRIDE (21), claimed to yield assignments in close agreement to those
made by crystallographic experts.

3 Prediction methods for globular proteins

3.1 The early methods
Attempts to predict protein secondary structure began more than four decades ago
(e.g. 22, 23), while the first computer algorithms appeared a quarter of a century
ago (24-26). The algorithms of Nagano (22) and Chou and Fasman (25) were based
on statistical information, whereas Lim's method (26) was stereochemically
oriented and relied on conserved hydrophobic patterns in secondary structures
such as amphipathicity in helices (8). Secondary structure prediction has gen-
erally been formulated for three states, helix, strand, and coil. This holds also for
recent versions of the early and popular GOR method (27, 28), which considers
the influence and statistics of flanking residues on the conformational state of a
selected amino acid to be predicted. The popular early methods by Lim (26) and
Chou-Fasman (25) as well as the GOR method (27, 28) will be described in more
detail.

3.1.1 Lim
Lim (26) developed a set of complicated stereochemical prediction rules for a-
helices and B-sheets based on their packing as observed in globular proteins. Apart
from being the most successful early method (see below), Lim's stereochemical
rules are quite important for understanding protein folding. An example is the
set of hydrophobicity rules for a-helices with terminal hydrophobic pairs at
sequence positions i and i + 1, hydrophobic pairs in middle helical segments
positioned at (i, i + 4) and middle hydrophobic triplets positioned at (i, i + 1, i + 4)
or (i, i + 3, i + 4) (see also Figure 3). The Lim method never gained widespread
popularity because a computer implementation has not been available until
recently.

3.1.2 Chou-Fasman
The most widely used pioneering method is the one by Chou and Fasman (25), in
which predictions are based on differences in residue composition for three
states of secondary structure: a-helix, B-strand, and turn (i.e. neither a-helix nor
B-strand). Chou and Fasman performed a statistical analysis over a number of
crystallographically determined protein tertiary structures and determined the
frequency of each amino acid type in the three states. The position of turn
residues was included in the frequency calculations given significant positional
differences in residue type occurrences at turn sites. The frequencies were

120
 PREDICTING SECONDARY STRUCTURE FROM PROTEIN SEQUENCES

normalized to amino acid type preferences for each of the structural states by
dividing each by that found in all positions of the known structures. For helix
and strand, effects of neighbouring residues in the protein sequence were taken
into account by averaging the preferences over three residues for a-helix pre-
dictions and over two for B-strands. Secondary structures were initiated accord-
ing to the higher preference values and minimum nucleation lengths required
for each structural state. Extensions were effected as long as preferences re-
mained high and certain residues were not encountered (e.g. proline in a -
helix). The Chou-Fasman method has owed its early popularity to the straight-
forward underlying statistics that are easy to understand.

3.1.3 GOR
The GOR method quickly became the standard for a decade after its first
appearance. Although the initial versions GOR I and GOR II predicted four states
by discriminating between coil and turn secondary structures, GOR III (28) and
the most recent version, GOR IV (29) perform the common three-state pre-
diction. The GOR method relies on the frequencies observed for residues in a 17-
residue window (i.e. eight residues N-terminal and eight C-terminal of the
central window position) for each of the three structural states. The amino acid
frequencies are exploited using an information function based on conditional
probabilities defined as:

where S one of the structural states H, E, or C, and R is one of the 20 residue

types. The factor P(S|R) denotes the conditional probability of a secondary
structural state for a sequence position given that it is occupied by residue type
R. Rewriting the formula for frequencies gives:

where fS,R is the frequency of residue type R in state S,fR the general frequency of
residue type R, and fs/n that of structural state S. Significant in this formula is
that the information of a particular residue type in one of the structural states is
not only based on the normalized frequency, but shows an extra weighting
based on the inverse fraction of all residues in that state. In the GOR method,
this formula is used to calculate the information difference between the various
states defined as I(AS; R) = I(S; R)—I(!S; R) with !S denoting all other states (not S).
The information difference formula then becomes:

The above formula is defined for a single sequence position, but can be easily
extended to the GOR 17-residue window by, for example, writing R17 instead of
R. Unfortunately, it is not feasible to sample all possible 17-residue fragments
directly from the PDB (as there are 1720 possibilities). The subsequent versions of

121
JAAP HERINGA

the GOR method over the years have explored increasingly detailed approxi-
mations of this sampling problem, along with the increase of data in the PDB:
(a) GOR I just treated the 17 positions in the window independently, and so single-
position information could be summed over the 17-residue window.
(b) GOR II did the same but sampled over a larger database.
(c) GOR III (28) refined by including pair frequencies derived from 16 pairs be-
tween each non-central and the central residue in the 17-long window. As
the PDB at the time was not large enough to provide sufficient data, dummy
frequencies were calculated (28).
(d) The current version, GOR IV (29) uses pairwise information over all possible
paired positions in a window (there are 17 X 16/2 possibilities), albeit with a
relatively small weight as compared with the GOR I-type single-position
information (a) which is included as well.
The theoretical principles used in the GOR method are statistically sound and no
ad-hoc rules or artificial variables are invoked, which makes it one of the most
elegant methods with a high accuracy given its single sequences prediction.
However, as in many other methods (vide infra), a post-processing step was intro-
duced for the GOR IV method to refine the predictions. Helices are required to
be at least four residues in length and strands should consist of two or more
residues. If a shorter helix or strand fragment is initially predicted, the method
assesses the probabilities of extending the fragment to the minimum associated
length or deleting it (i.e. changing it to coil).

3.2 Accuracy of early methods

The Chou-Fasman, GOR III, and Lim methods were assessed to show accuracies
of 50%, 53%, and 56% respectively (30). Version IV of the GOR method, however,
raises the single sequence prediction accuracy to 64.4% (29), as assessed through
jackknife testing (see Protocol 1) over a database of 267 proteins with known
structure. Random prediction would yield about 40% correctness given the
observed distribution of the three states in globular proteins (with roughly 30%
helix, 20% strand, and 50% coil). Although they are significantly beyond the
random level, these single-sequence prediction accuracies are not sufficient to
allow the successful prediction of the protein topology.

3.3 Other computational approaches

The Chou-Fasman and GOR methods both exploit compositional biases exhibited
by the three types of secondary structures. Information derived from single
sequences have been explored as well in the form of sequence pattern matching
(16, 31-34).
On the algorithmic side, researchers have integrated novel computational
concepts to optimize the implementation of observed patterns in mapping the

122
 PREDICTING SECONDARY STRUCTURE FROM PROTEIN SEQUENCES

primary on to the secondary structure and to thus enhance the success rate of
prediction. These include:
(a) Neural network applications (9, 35).
(b) Nearest-neighbour methods (36-39).
(c) Linear discriminant analysis (40).
(d) Inductive logic programming (ILP) (41).
Examples of the first three formalisms will be described in the following section.
The latter computational concept (ILP) is designed for learning structural relation-
ships between objects. Muggleton et al. (41) used the ILP computer program
Golem to automatically describe qualitative rules for residues in the a-helix con-
formation and central in a 9-residue window. The rules made use of the physico-
chemical amino acid characterizations of Taylor (42) and were established during
iterative training steps over a small set of 12 known a/a protein structures. With
the thus obtained set of rules, a-helices in four independent a/a proteins were pre-
dicted with an accuracy of 81% on a per residue basis (Q3). The Golem algorithm
is of limited use because it is only able to predict helices in all-helical proteins.

3.4 Prediction from multiply-aligned sequences

In 1987, Zvelebil et al. (43) for the first time exploited multiple alignments to
predict secondary structure automatically by extending the GOR method and
reported that predictions were improved by 9% compared to single sequence
prediction. Also Levin et al. (44) quantified the effect and observed 8% increased
accuracies when multiple alignments of homologous sequences with sequence
identities of a 25% were used. As a result, the current state-of-the-art methods
all use input information from multiple sequence alignments.

3.4.1 Neural network methods

Neural networks are organized as interconnected layers of input and output
units, and can also contain intermediate (or 'hidden') unit layers (for a review,
see ref. 45). Each unit in a layer receives information from one or more other
connected units and determines its output signal based on the weights of the
input signals. A neural network can be regarded as a black box, which is trained
to optimize the grouping of a set of input patterns into a set of output patterns
by adjusting the weights of the internal connections. Therefore, neural nets are
learning systems based upon complex non-linear statistics.

PHD
The PHD method (Profile network from HeiDelberg) (9) combines the added
information from multiple sequence information with the optimization strength
of the neural network formalism. The method makes use of three consecutive
complete neural networks:
(a) The first network produces the first raw 3-state prediction for each alignment
position. It takes as input the fractions of the 20 amino acids at each multiple

123
JAAP HERINGA

alignment position together with the two 6-residue flanking regions; i.e. a 13-
residue window (w = 13) is used to predict each alignment position with the
central residue in the middle position. The output of the first network for
each alignment position is three probabilities for three the states (helix,
strand, and coil).
(b) A second network refines the raw predictions of the first level by filtering the
3-state probabilities for each alignment position based on the probabilities of
the flanking positions. It takes as input the output of the first network and
processes the information using a 17-residue window. The output of the
second network comprises for each alignment position the three adjusted
state probabilities. This post-processing step for the raw predictions of the
first network is aimed at correcting unfeasible predictions and would, for
example, change (HHHEEHH) into (HHHHHHH).
(c) The first two networks perform the basic prediction of the secondary struc-
ture associated with a query multiple alignment. However, as the networks
can be trained in various ways, PHD employs a number of separately trained
consecutive network pairs ((a) and (b)) and feeds their predictions (3-state
probabilities) into a third network for a so-called jury decision.
The predictions obtained by the jury network undergo a final filtering to delete
predicted helices of one or two residues and changing those into coil. The method
was trained on a non-redundant set of 130 alignments from the HSSP database
(46), each containing one sequence with a known structure. The method showed
an overall prediction accuracy of 70.8% in a jackknife test over 126 alignments
(4 of 130 alignments were transmembrane protein families), which for computa-
tional reasons were divided in 7 groups (see Protocol 1). Although this count is not
the highest accuracy reported, the PHD method to date shows the most sustained
performance as compared with all other methods available on the Web.
If the PHD webserver is given a single sequence for prediction, it performs a
BLAST-search to find a set of homologous sequences and aligns those using the
MAXHOM alignment program (46). The resulting alignment is then fed into the
actual PHD neural net algorithm.

Pred2ary
Another accurate profile and neural net-based prediction method is Pred2ary
(35) which was assessed with an accuracy of 74.8% and balanced prediction over
the three structural states. The method employs a second neural net to filter the
raw predictions of the first net, as does the PHD method (9). A recent extended
version, which combines in a jury decision the outputs of a massive number of
120 networks individually trained, is claimed to predict 75.9% ± 7.9% accurately.
This is achieved by constructing a priori probabilities of correctly predicting the
structural state at each query sequence position for all combinations of network
output weighs for helix and strand These probabilities are then used for a final
state prediction corresponding to the highest of the a priori probabilities for each
of the three states.

124
 PREDICTING SECONDARY STRUCTURE FROM PROTEIN SEQUENCES

3.4.2 k-nearest neighbour methods

As with neural network methods, the application of a k-nearest neighbour
method requires an initial training phase in which a large pool of so-called
exemplars is established. In the context of secondary structure prediction, this
pool typically consists of sequence fragments of a certain length derived from a
database of known structures, so that the central residue of such fragments
(exemplars) can be assigned the true secondary structural state as a label. Then,
a window of the same length is slid over the query sequence and for each
window the k most similar fragments from the pool of exemplars are deter-
mined using a similarity criterion. The distribution of the k secondary structure
labels is then used to derive propensities for the three states. In the methods
covered below, k is in the range 25-100.

Yi and Lander
Yi and Lander (36) were the first to use nearest-neighbour classifiers for pre-
diction of secondary structure. A database of 110 proteins with known tertiary
structure was used to derive a large collection of 19-residue exemplars for which
only the environmental states were noted; i.e. the residue type information was
discarded. As a label for each exemplar the secondary structural state of the
central residue was taken. For each 19-residue window of the query protein, 50
nearest neighbour exemplars were identified using the amino acid environ-
mental scoring system of Bowie et al. (47), which includes as parameters the
secondary structure state, accessible surface area and polarity; and scores the
likelihood of a residue type to be in a particular state (or range) over these three
parameters. As a score, the average was taken of 19 residues within a query
window matched with the 19-position exemplar considered. During training, for
each exemplar a cut-off score was determined, which should be met by the
query fragment compared to it in order to count the exemplar as a neighbour:
The cut-off score can be viewed as a reliability check for the predictive value of
the exemplars. The 50 thus obtained nearest neighbours showed a distribution
of the associated secondary structure labels, from which probability estimates
for the three structural states were derived for the query fragment considered.
Yi and Lander explored various scoring systems and found that the best performer
included 15 environmental classes (3 secondary structures times 5 different
accessibility/polarity classes) combined with an amino acid exchange score from
the Gonnet et al. matrix (48). Note that for this final scoring system, the amino
acid types of the exemplars were taken into account. This scenario resulted in a
prediction accuracy of 67.1%. Using a neural network for a jury decision over six
different scoring systems led to the final accuracy of 68%, as assessed through
jackknife testing (Protocol 1).

NNSSP
The NNSSP (Nearest Neighbour Secondary Structure Prediction) (37) method
adopts the nearest neighbour approach of Yi and Lander (36) for single sequence
prediction. Differences with the Yi and Lander method are:

125
JAAP HERINGA

(a) Predictions are made for multiple alignments.

(b) N- and C-terminal positions of helices and strands; and B-turns are explicitly
taken as additional secondary structure types.
(c) When predicting, the database of exemplars (see above) is restricted to
sequences similar to the query sequence. This reduces computation and leads
to more biologically related nearest neighbours.
(d) Alignment regions with insertions/deletions are explicitly taken into account.
Salamov and Solovyev (37) explored various window lengths and finally choose
predictors combining window sizes of 11,17, or 23; nearest neighbour numbers
of 50 or 100, and balanced or non-balanced training (i.e. 3 X 2 x 2 = 12 pre-
dictors). A simple majority rule over the 12 predictors increased the accuracy by
0.9%. A few simple filters were effected to refine the thus obtained predictions as
follows:
(a) Helices predicted to consist of 1 or 2 residues are deleted (changed to coil),
but (EHE) becomes (EEE).
(b) Strands of length 1 or 2 are deleted, but (HEEH) becomes (HHHH).
(c) Helices of length 4 or less are deleted. This rule is applied after a full cycle of
rule (a) and (b).
The overall accuracy of the method is 72.2%, which results from a jackknife test
over the database of 126 proteins by Rost and Sander (9).

PREDATOR
The PREDATOR method of Frishman and Argos (38, 39) owes its accuracy mostly
to the incorporation of long-range interactions for B-strand prediction and attains
68% prediction accuracy for single sequence prediction which was assessed
using a one-at-a time jackknife test (see Protocol 1) over the protein set of Rost
and Sander (RS) (9). Using a k-nearest neighbour approach (with k = 25 and 13-
residue windows), propensities for the general three states (PH, PE, and Pc) were
determined for each residue. Using pairwise potentials involving long-range
interactions, two more propensities for B-strand were determined. This was
done by assessing the likelihood for all pairwise 5-residue fragments (separated
by more than six amino acids) to form parallel or anti-parallel B-bridges, based
on summing residue hydrogen bonding propensities obtained from known
structures (two sets of propensities for anti-parallel and one for parallel bridges).
As the final parallel and anti-parallel B-strand propensity for each residue (Ppar
and pAntipar), the maximum scoring window pair was taken with the residue con-
sidered at the N-terminal position in one of the windows. Pairwise hydrogen
bonding potentials were also determined for a-helical residues at a sequence
separation of four residues. Their sum was calculated over a 7-residue window to
arrive at an extra helix propensity for the residue N-terminal in the window
(PHelix). The last additional propensity concerned p-turns (pTurn) and was obtained
by summing single-residue propensities in classic p-turn positions 1-4 (49) using

126
 PREDICTING SECONDARY STRUCTURE FROM PROTEIN SEQUENCES

a four-residue window. For each of the thus obtained seven independent

propensities, threshold values (T) were calculated and used in the following
five rules applied consecutively to arrive at a three-state prediction for each
residue:
1. If (PPar > Tpar or pAntipar > TAntipar) and PHelix < Thelix, then predict B-strand; other-
wise, if PHelix > Thelix, then predict a-helix, otherwise predict coil.
2. If Pc > Tc, then predict coil.
3. If PE > TE, then predict B-strand.
4. If PH > TH,then predict a-helix.
5. If PTurn > TTurn, then predict coil.
Apart from the novel scheme to employ long-range interaction to aid strand
prediction, the PREDATOR method can also use information from multiple
sequences to enhance predictions. However, PREDATOR does not use or con-
struct a multiple alignment, but rather compares the sequences using pairwise
local alignments (50). The current method is not able to extract local alignments
from a multiple alignment provided by the user, while leaving the multiple
alignment intact, but it is planned to realize this option in a future release
(Frishman, personal communication). As predictions by PREDATOR are carried
out for a single base sequence, a set of highest scoring local alignments is com-
piled through matching the base sequence with each of the other sequences. A
weight is then compiled for each matched local fragment based on the
alignment score and length of the local alignment. For each residue in the base
sequence, the weighted sum over all stacked fragments (see Figure 4) is compiled
independently for the seven propensities and subjected to the above five rules to
arrive at a three-state prediction. The extra information conferred by the multiple
sequences resulted in a Q3 of 74.8% (39), as assessed using one-at-a-time jackknife
testing over the RS protein set. As for the Pred2ary method showing identical
accuracy, this Q3 is the highest reported in the literature.

3.4.3 Linear discriminant analysis: the DSC method

The DSC method combines the compositional propensities from multiple
alignments with a set of concepts important for secondary structure prediction

Figure 4 Usage of local alignments in the PREDATOR algorithm. For details, see text.

127
JAAP HERINGA

(see Section 1.3). This information is processed using linear statistics. Apart from
the conformational propensities, the following concepts are used:
• N-terminal and C-terminal sequence fragments are normally coil.
• Moments of hydrophobicity (see Figure 3).
• Alignment positions comprising gaps are indicative for coil regions.
• Moments of conservation.
• Autocorrelation.
• Residue ratios in the alignment.
• Feedback of predicted secondary structure information.
• Simple filtering.
The relative importance of these concepts was determined in five runs, which
successively relied on increased information as follows:
(a) Run 1: The GOR method was used on each of the aligned sequences and the
average GOR score for each of the three states was compiled for each align-
ment position.
(b) Run 2: For each position in the query multiple alignment, a so-called attribute
vector was compiled, consisting of 10 attributes: three averaged GOR scores
for H, E, and C; distance to alignment edge; hydrophobic moment assuming
helix; hydrophobic moment assuming strand; number of insertions; number
of deletions; conservation moment assuming helix and that assuming strand.
(c) Run 3: Positional 20-attribute vectors were determined consisting of the
above 10 attributes and those in a smoothed fashion.
(d) Run 4: Positional 27-attribute vectors were compiled comprising the 20
attributes of the preceding round, combined with fractions of predicted ot-
helix and B-strand, and fractions of the five most discriminating residue
types; His, Glu, Gln, Asp, and Arg.
(e) Run 5: A set of 11 filter rules were employed for a final prediction, such as,
for example, ([E/C]CE[H/E/C][H/C]) -> C. These filter rules were found auto-
matically using machine learning.
For run (b) to (d), a linear discrimination function was determined for each of
the three secondary structural states. A linear discrimination function is effect-
ively a set of weights for the attributes in the positional vector, so that the
secondary structure associated with the highest scoring discrimination function
is assigned to the alignment position considered.
The DSC predictions are based on the information arising from the five above
runs. The Q3 was assessed for successively increasing numbers of runs (run 1,
runs 1 and 2, runs 1-3, 1-4, 1-5) for the five runs based on the Rost-Sander
protein set and comprised 63.5%, 67.8%, 68.3%, 69.4%, and 70.1% (DSC), respect-
ively. The DSC method performs especially well for moderately sized proteins in
the range 90-170 residues. A special feature of the DSC technique is that it

128
 PREDICTING SECONDARY STRUCTURE FROM PROTEIN SEQUENCES

accepts predictions by the PHD algorithm as input and attempts to refine those
using the above concepts. The Q3 of this PHD-DSC combinatorial procedure was
evaluated at 72.4% (40).

3.4.4 SSPRED: a secondary structure specific exchange method

The SSPRED method (51) exploits an alternative aspect of the positional infor-
mation provided by multiple alignment, in that it uses the amino acid pairwise
exchanges observed for each multiple-alignment positions. Using the 3D-ALI
database (52) of combined structural and sequence alignments of distantly
homologous proteins, three amino acid exchange matrices were compiled for
helix, strand, and coil, respectively. Each matrix contains preference values for
amino acid exchanges associated with its structural state as observed in the 3D-
ALI database. They are used to predict the secondary structure of a query align-
ment through listing the unique observed residue exchanges for each alignment
position and adding the corresponding preference values over each of the three
exchange matrices. The fact that each exchange type is counted only once for
each alignment position provides implicit weighting of the sequences, thus
avoiding predominance of related sequences. The secondary structure associated
with the matrix showing the highest sum is then assigned to the alignment
position. Following these raw predictions, three simple cleaning rules are applied
and completed in three successive cycles:
(a) Single position interruptions: if a sequence site is predicted in one struc-
tural state and the two flanking positions in another, the position is changes
into that of the consistent flanking sites, for example (H[E/C]H) becomes
(HHH) where [E/C] indicates E or C.
(b) Double position interruptions: if in five consecutive positions two middle
sites are of another type than the three flanking sites, the middle positions are
changed to the flanking types. For instance, (HH[E/C][E/C]H) or (H[E/C][E/C]HH)
becomes (HHHHH).
(c) Short fragments: helices predicted less than or equal to 4 and strands less
than or equal to 2 in length are changed into coil predictions.
The accuracy of the method was assessed over one-at-a-time jackknife testing and
amounted to 72%, albeit over a relatively small test set of 38 protein families.

3.5 A consensus approach: JPRED

The JPRED server at the EMBL-European Bioinformatics Institute (Hinxton, UK)
conveniently runs state-of-the-art prediction methods such as PHD (9), PREDATOR
(38, 39), DSC (40), and NNSSP (37), while also ZPRED (43) and MULPRED (Barton,
unpublished) are included. The NNSSP method has to be activated explicitly, as
it is the slowest of the ensemble. The server accepts a multiple alignment and
predicts the secondary structure of the sequence on top of the alignment: Align-
ment positions showing a gap for the top sequence are deleted. A single sequence
can also be given to the server. In the latter case, a BLAST-search is performed to

129
 PREDICTING SECONDARY STRUCTURE FROM PROTEIN SEQUENCES

Figure 5 Secondary structure prediction for chemotaxis protein cheY (3chy). The top
alignment block represents the multiple alignment of the 3chy sequence with 13 distant
flavodoxin sequences by the method PRALINE. The middle block is the same sequence set
aligned by CLUSTALX. Under each of the alignments are given the alignments by five
secondary structure prediction methods. The bottom block depicts consensus secondary
structures determined by Jpred over the five methods used, respectively for the PRALINE and
CLUSTALX alignments, as well as for a set of 32 homologous sequences aligned by
CLUSTALX (cons HOMOLOGS). Vertical bars (|') under each of the consensus predictions
indicate correct predictions. The bottom line identifies the standard of truth as obtained from
the 3chy tertiary structure by the DSSP program. (10) The secondary structure states
assigned by DSSP other than 'H' and 'E' were set to ' '(coil) for clarity.

find homologous sequences, which are subsequently multiply aligned using

CLUSTALX and then processed with the user-provided single sequence on top in
the alignment. If sufficient methods predict an identical secondary structure for
a given alignment position, that structure is taken as the consensus prediction
for the position. In case no sufficient agreement would be reached, the PHD
prediction is taken. This consensus prediction is somewhat less accurate when
the NNSSP method is not invoked or completed in the computer time slot
allocated to the user. An example of output by the JPRED server for the signal
transduction protein cheY (PDB code 3chy) is given in Figure 5 (vide infra).

3.6 Multiple-alignment quality and secondary-structure

prediction
Multiple-alignment protocols use heuristics to overcome the combinatorial ex-
plosion that arises when all possible alignments would be tested exhaustively.
Most global alignment methods therefore establish an order in which the
sequences are aligned progressively based on the alignment scores of all possible
pairwise alignments (the number is N x (N - l)/2 with N the number of
sequences). Although most methods show a comparable overall quality in align-
ment construction for sequences showing residue identities of 30% or higher,
significant differences can arise in individual cases, particularly when evolution-
ary distant sequences are included. As the currently most successful secondary
structure prediction methods all employ positional information from multiple
alignments, it is clear that alignment quality is crucial for accurate prediction.
As an example, the popular multiple alignment program CLUSTALX and the
recently developed method PRALINE (53) (see below) were used to automatically
construct a multiple alignment for the signal transduction protein cheY (PDB
code 3chy) and 13 distant flavodoxin sequences. The 3chy structure adopts a
flavodoxin fold (see Figure 2) despite very low sequence similarities with genuine
flavodoxins. Figure 5 shows both alignments with secondary structure predictions
by the JPred server as well as JPred consensus predictions for the two alignments.
The difference in accuracy of the two consensus predictions amounts to more
than 30%, an order of magnitude more than the increase in prediction accuracy
obtained over the last five years. It must be stressed that the flavodoxin sequences
are evolutionary distant from the cheY sequence, such that the alignments were

131
JAAP HERINGA

only included to illustrate their crucial importance for secondary structure

prediction rather than to argue in favour of any of the two used alignment
programs based on this single example. The Jpred server was also given the single
3chy sequence, after which the Jpred server constructed an evolutionary related
set of homologs through a BLAST search and aligned the 32 resulting sequences
using CLUSTALX. The accuracy of the consensus secondary structure prediction
by Jpred was 3% higher than that obtained for the PRALINE alignment of the
cheY-flavodoxin set (Figure 5). Moreover, it successfully delineated the second B-
strand of the 3chy structure, which was missed by the predictions based on both
the CLUSTALX and PRALINE alignments.

3.7 Iterated multiple-alignment and secondary structure

prediction
As mentioned, most reliable secondary structure prediction methods utilize
sequence information in multiple alignments and their prediction accuracy is
crucially dependent on the quality of a multiple alignment used. If in turn a
multiple alignment would be guided by the predicted secondary structure, an
iterative scheme would be possible that optimizes both the multiple-alignment
quality and secondary structure prediction. This procedure is implemented in
the PRALINE multiple alignment method (53). A multiple alignment is con-
structed initially without information about the secondary structure (Figure 6a).
Then, the secondary structure is predicted (for which any of the aforementioned
methods could be used) and iteratively a new alignment is constructed, now
using the predicted secondary structure. PRALINE employs dynamic program-
ming to progressively construct a multiple alignment for a query set of
sequences and therefore relies on an ammo acid exchange weights matrix and a
pair of gap penalties (for a review, see ref. 54). The initial alignment is con-
structed using a default residue exchange matrix (e.g. the BLOSUM62 matrix)
and gap penalties. After secondary structure prediction, resulting in a tentative
secondary structure for each sequence if a single sequence-based method is used
or in a single secondary structure if a method reliant on a multiple alignment is
effected (Figure 6a), PRALINE utilizes the thus obtained secondary structure infor-
mation as illustrated in Figure 6b. At each alignment step during the progressive
alignment, pairs of sequences (and/or profiles representing already aligned
sequence blocks) are matched using three secondary structure-specific residue
exchange matrices (55) and associated gap penalties. As shown in Figure 6b, the
residue exchange weights for matched sequence positions with identical second-
ary structure states is taken from the corresponding residue exchange matrix.
Sequence positions with inconsistent secondary structure states are treated with
the default exchange matrix. The secondary structure information is thus used
in a conservative manner based upon the assumption that consistent secondary
structure predictions are indicative for their reliability when performed for each
individual sequence (Figure 6a). In this way, the multiple alignments guide the
secondary structure predictions, which in turn guide the alignment.

132
 PREDICTING SECONDARY STRUCTURE FROM PROTEIN SEQUENCES

Figure 6 Iterative multiple-alignment and secondary-structure prediction by the PRALINE method.

4 Prediction of transmembrane segments

Membrane proteins (MP) form a distinct topological class due to the presence of
one or more transmembrane (TM) sequence segments. In contrast to globular
proteins where all possible mutual orientations of individual structural elements
are in principle possible, MP transmembrane segments are subjected to severe
restrictions imposed by the lipid bilayer of the cell membrane.
There is a considerable lag in structures available for MPs relative to the large
and vastly growing numbers of soluble proteins, as little X-ray or NMR data
regarding the tertiary structure of MPs have been available until recently (56).
The most frequently observed secondary structure in transmembrane segments
is the a-helix, but also transmembrane structures based on B-strands that
constitute a p-barrel have been encountered. The initial idea that TM segments
are either completely a-helical or consist of B-strands exclusively, was disrupted
by electron microscopy data for the nicotinic acetylcholine receptor (57), which
was interpreted as a central five-helix bundle surrounded by B-strands, albeit
based on preliminary data with low resolution.

133
JAAP HERINGA

Fortunately, the location of the transmembrane segments in primary

structure of the MP is relatively easy to predict due to the rather strong tendency
of certain hydrophobic amino acid types with their special physico-chemical
properties to occur in membrane spanning regions. Thus, efforts concerning the
theoretical analysis of MPs over the past two decades have been focused on the
determination of the membrane sequence segment boundaries and their
tentative orientation with respect to the membrane, although mostly assuming
a-helical structures.

4.1 Prediction of a-helical TM segments

The following considerations form the basis of transmembrane subsequence
prediction.
(a) Amino acids immersed in the lipid phase are likely to be hydrophobic.
Therefore, any physical measure of amino acid hydrophobicity derived from
physical calculations and/or experimental data can serve as a measure of
likeliness for a residue type to occur in a membrane-spanning segment.
(b) The propensities of amino acids to reside in the lipid bilayer can be inferred
from abundant but not very precise experimental data on the boundaries of
the transmembrane segments acquired from site-directed mutagenesis, en-
zymatic cleavage, immunological methods and the like. This contrasts with
the standard secondary structure prediction methods for soluble proteins
where statistical propensities of different amino acids to form one of the
major secondary structure elements are derived from more accurate protein
tertiary structural data from X-ray crystallography and NMR spectroscopy.
(c) Transmembrane segments are believed to adopt in most cases a-helical con-
formation. An a-helix is the most suitable local arrangement because, in the
absence of water molecules inside the membrane, all main-chain polypeptide
donors and acceptors must mutually satisfy each other through formation of
hydrogen bonds as occur in an a-helix. This energetic argument is supported
by experimental evidence where polypeptide chain tends to adopt helical
conformation in a non-polar medium (58). Therefore, a-helical propensities
of amino acids derived from the analysis of globular proteins can be con-
sidered in MP structure prediction.
Although the globular protein interior is less apolar than the lipid bilayer, ex-
tensive usage of these data has been made for MP structure prediction, particu-
larly with the classical hydrophobicity scale of Kyte and Doolittle (59). Other
techniques are more specifically aimed at searching MP transmembrane regions
(60-63). Hydrophobic scales can be used to build a smoothed curve, often called
a hydropathic profile, by averaging over a sliding window of given length along
the protein sequence to predict transmembrane regions. Stretches of hydro-
phobic amino acids likely to reside in the lipid bilayer appear as peaks with
lengths corresponding to that expected for a transmembrane segment, typically
16-25 residues. The choice of window length should correspond to the expected

134
 PREDICTING SECONDARY STRUCTURE FROM PROTEIN SEQUENCES

length of a transmembrane segment. Given that the average membrane thick-

ness is about 30 A, approximately 20 residues form a helix reaching from one
lipid bilayer surface to another. A further threshold is also required to deter-
mine the exact boundaries of a membrane spanning segment. Kyte and Doolitle
(59) early on set their limit by examining the hydropathic character of just a few
membrane proteins. Later, a much larger learning set was used by Klein et al.
(64) through discriminant analysis. Rao and Argos (65) suggested a minimum
value for the peak hydrophobicity and a cut-off value at either end of the peak to
terminate the helix.
Although many of the above techniques constitute an essential part of all
major sequence analysis packages, the relatively simple physical considerations
forming the basis of these methods do not exhaust the whole variety of possible
situations.
If a membrane protein has more than one transmembrane helix, the relative
orientation of the helices and the interaction of the corresponding sidechains
are also important for structure prediction. The structure of the membrane
proteins determined to date and also some theoretical evidence (66) support the
view that a-helices in membranes form compact clusters. While the residues
facing the lipid environment conform to the preferences described above, the
interface residues between different helices do not necessarily have contact
with the membrane, and, can therefore, behave differently. It is possible that
charged residues occur in the helices in a coordinated fashion such that
positively charged sidegroups on one helix will have their negatively charged
counterparts on another helix. These charges could, for instance, constitute a
membrane channel. Intra-membrane a-helices can thus have an amphipathic
character (see above). In such cases hydropathic profiles can work poorly in
detecting transmembrane segments. In certain cases where the number of
transmembrane segments is large (more than 20 as in some channel proteins),
the inner helices of the transmembrane helical bundle can completely avoid
contact with the lipid bilayer and, therefore, any restrictions on their amino
acid content—or even length—could be artificial.
Eisenberg et al. (67) introduced a quantitative measure of helix amphipathicity
called the 'hydrophobic moment', and defined as a vector sum of the individual
amino acid hydrophobicities radially directed from the helix axis. The hydro-
phobic moment provides sufficient sensitivity to distinguish amphipathic ex-
helices of globular, surface, and membrane proteins. Many methods for
amphipathic analysis were developed based on Fourier analysis of the residue
hydrophobicities (68, 69) and the average hydrophobicity on one helix face (70).
Several prediction methods have emerged which utilize multiple factors,
complex decision rules, and large learning sets. Von Heijne (63) proposed a syn-
thetic technique in which a standard hydrophobicity analysis is supplemented
by charge bias analysis (see Section 4.2). Other methods include the joint usage
of several selected hydrophobicity scales (71) or the application of optimization
techniques with membrane segments as defined by X-ray analysis serving as
reference examples (72).

135
JAAP HERINGA

Persson and Argos (73) incorporated sequence information from multiple

alignments to aid TM prediction. The propensities of amino acids to reside in
either the central or the flanking regions of a transmembrane segment were
calculated using more then 7500 individual helical assignments contained in the
SWISS-PROT sequence databank. These values were then used to build a pre-
diction algorithm wherein, for each segment of a multiple sequence alignment,
and for each sequence, average values of the central and flanking propensities
are calculated over windows of respectively 15 and 4 residues long. If the peak
values for central transmembrane regions exceed a certain threshold, this region
is considered as a possible candidate to be membrane spanning. The algorithm
then attempts to expand this region in both sequence directions until a flanking
peak is reached or the central propensity averages fall below a certain value.
Additional restraints are imposed on the possible length of the tentative trans-
membrane segments. The optimal window length was found to be about 15
residues. Due to the increased amount of information utilized by the technique,
more accurate prediction results were achieved as compared with earlier
methods. The gain in sensitivity is due to the usage of multiple alignments as
well as the introduction of a second propensity for flanking regions.
Neural networks (see Section 3.4.1) have also been applied to the TM prediction
problem. Early attempts involved training on secondary structural elements of
globular proteins (74). Rost et al. (75) trained the PHD method on multiple
alignments for 69 protein families with known transmembrane helices and
achieved 95% prediction accuracy using the jackknife test.

4.2 Orientation of transmembrane helices

Another aspect of transmembrane segment prediction is prediction of mem-
brane sidedness, or orientation. For bacterial membrane proteins it was found
that intracellular loops in between transmembrane helices contain arginine and
lysine residues much more frequently than the extracellular exposed loops (76,
77). This pattern has been shown to apply also to eukaryotic membrane pro-
teins, but to a lesser extent (78). An additional observation, made for eukaryotic
proteins, is that the difference in the total charge of the approximately 15
residues flanking the transmembrane region on both sides of the membrane
also coincides with the orientation of the protein (79). If the C-terminal portion
of the protein adjacent to this segment is more positive in charge then the N-
terminal portion, the C-terminus will reside in the cytosol, and vice versa. Non-
random charge distribution may also play an important role in membrane
insertion of the protein. These findings, collectively known as the 'positive
inside rule', aid prediction schemes for MP topology. However, the positive
inside rule is only applicable to a-helical TM regions.

4.3 Prediction of B-strand transmembrane regions

As the methods described above all predict TM segments assuming the a-helical
conformation, transmembrane segments constituted by B-strands are not likely

136
 PREDICTING SECONDARY STRUCTURE FROM PROTEIN SEQUENCES

to be predicted successfully. Four different families of B-barrel membrane

proteins are known to date (porins, OmpA, FhuA, and FepA). For example,
porins form voltage-dependent membrane channels and have a B-barrel fold
constituted by 16 B-strands (80). Hydrogen bonds are formed only between
adjacent p-strands. Most of the outer surface of the barrel faces the lipid en-
vironment whereas the internal part serves as an aqueous pore. Each individual
B-strand could therefore be expected to be amphipathic with a period of two
residues. However, while every second residue facing the lipid bilayer is hydro-
phobic, those side chains protruding towards the interior of the barrel display
no definitive tendency, thus lowering the amphipathic signal. Another compli-
cation is the fact that the number of amino acid residues in extended confor-
mation needed to span the membrane is much smaller than that for the helical
conformation, typically about 10. Consequently, smoothed hydropathic profiles
are likely to miss such short stretches.

5 Coiled-coil structures
If a protein is predicted to contain a-helices, higher-order information as well as
increased confidence in predictions made could be gained from testing the
possibility that a pair of helices adopts a superhelical twist resulting in a coiled-
coil conformation. The left-handed coiled-coil interaction involves a repeated
motif of seven helical residues (abcdefg). The a and d positions are normally
occupied by non-polar residues constituting the hydrophobic core of the
helix-helix interface, whereas the other positions display a high likelihood to
comprise hydrophilic residues. The e and g positions in addition are often charged
and can form salt-bridges to each other. The program COILS2 (81, 82) exploits
this information and compares a query sequence with a database of known
parallel two-stranded coiled-coils. A similarity score is derived and compared to
two score distributions, one for globular proteins (without coiled-coils) and one
for known coiled-coil structures, and a probability is then calculated for the
query sequence to adopt a coiled-coil conformation. As the program assumes the
presence of heptad repeats, the probabilities are derived using windows of 14,
21, and 28 amino acids. However, the program offers the option to include user-
defined window lengths two allow the handling of cases with extreme coiled-
coil lengths. A recently updated scoring matrix which includes new structures
with known coiled-coils and contains amino acid type propensities at the various
positions in the heptad repeats, led to increased recognition of coiled-coils
elements. The COILS2 method accurately recognises left-handed two-stranded
coiled coils but loses sensitivity for coiled-coil structures composed of more than
two strands. It is not able to recognize right-handed or buried coiled-coil helices
and therefore is not applicable to transmembrane coiled-coil structures known
to basically show the similar coiled-coil conformations as soluble proteins,
albeit with dramatically different and more hydrophobic constituent amino
acids (56).

137
JAAP HERINGA

6 Threading
If a homologous protein with known structure is available for a query sequence,
this structure can then be aligned to the query sequence using the threading
technique (83). Treading methods test the feasibility for a given sequences to
adopt a particular fold, based on assessing the likelihood for the amino acids in
the query sequence to occur in the local residue environments within the known
tertiary structure. The optimal fit of the query sequence through the tertiary
structure effectively leads to an alignment, which can be used to copy the
secondary structure of the known fold to the query sequence. Although the
incorporation of tertiary structure information should lead to better alignment
and recognition of related sequences, the increased sensitivity of available
threading methods as compared with conventional sequence alignments is not
always clear. Jones et al. (84) discuss various threading methods available and
also how their results should be interpreted.

7 Recommendations and conclusions

Table 1 lists WWW addresses of some of the available prediction methods dis-
cussed in this chapter and in Protocol 2 some recommendations are given to
maximize the chances of an accurate prediction of the secondary structure
associated with a protein query sequence. In cases where a multiple alignment
is used, it is generally important to test the consistency and quality of the align-
ment constructed, as this can have dramatic consequences for the prediction
accuracy of multiple-alignment-based methods. In testing the consistency of the
currently most accurate prediction methods and determining a consensus pre-
diction, the positional reliability indices offered by some of prediction methods
should be included. Furthermore, the general accuracies for predicting each of
the three secondary structural states that are published for a number of the
methods can be used to weight the contribution of their positional predictions

Table 1 Websites of various secondary structure prediction methods and related services

Service Reference URL

GOR4 27, 28, 29 http://absalpha.dcrt.nih.gov:8008/gor.html
PHDa 9 http://dodo.cpmc.columbia.edu/predictprotein/b
Pred2ary "35 http://yuri.harvard.edu/~jmc/2ary.html
a
NNSSP 37 http://dot.imgen.bcm.tmc. edu:9331/pssprediction/pssp.html
PREDATOR3 38, 39 http://www.embl-heidelberg.de/cgi/predator_serv. pl
DSCa 40 http://bonsai.lif.icnet.uk/bmm/dsc/dsc_read_align.html
Zpreda 43 http://kestrel.ludwig.ucl.ac.uk/zpred.html
Jpred http://barton.ebi. ac.uk/servers/jpred.html
COILS2 81,82 http://www.isrec.isb-sib.ch/coils/COILS_doc.html
a
Method can also be run using the Jpred server.
b
Mirror websites for PHD can be found here as well.

138
 PREDICTING SECONDARY STRUCTURE FROM PROTEIN SEQUENCES

in a consensus. Specifically, the PREDATOR method should be included in the

trials as it is the only method relying on multiple local rather than global
alignment of the query sequences. It is important to realize that there is no
single best prediction method so that the degree of consistency over a variety of
methods is crucial for getting an idea about the prediction accuracy. Attempts to
recognize higher-order structure, such as the fold the protein might adopt or
the likelihood for coiled-coil structures, could enhance the confidence in pre-
dictions made or help correcting possible mispredictions. Easily recognizable
errors might be disruptions in alternating a-helix/B-strand predictions in a likely
a/p protein fold or the occurrence of a single B-strand within a tentative a-
helical protein. In general, it is likely that the accuracy of computerized pre-
diction methods can be enhanced further if such reasoning with higher order
structure is formalized and incorporated in the prediction mechanisms. Some
easy benefits will come from the steadily increasing structural protein data that
can be used to better train and tune the statistical methods. The current
availability of the prediction methods optimizes the chance for development of
sensitive consensus methods. It is clear from the ongoing interest and activity in
both the application and development of secondary structure prediction
methods that the end of the three decades of research efforts is not in sight.

Predicting secondary structure

1 Get a balanced and non-redundant set of homologous sequences for a given protein
query sequence.
2 Try a number of multiple alignment routines to obtain a consistent multiple
alignment.
3 Check the alignment carefully by eye using any additional information (e.g. active
site residues, disulfide bridges, etc,).
4 Use as many good secondary structure prediction methods as possible and con-
struct a consensus prediction (a convenient aid is the Jpred server).
5 Try to recognize super-secondary or higher-order structural features from the
predicted secondary structure elements and try to interpret and correct prediction
results (e.g. the missed second B-strand in the flavodoxin example) (see Section 3.6),

References
1. Heringa .J. and Taylor, W. R. (1997). Curr.Opin.'.Struct. Biol., 7,416.
2. Springer, T. A. (1997). Prof. Natl. Acad. Sri. USA, 94, 65.
3. Baldwin R, L, and Roder H. (1991). Curr. Biol., 1,218.
4. Goldenberg, D. P., Frieden, R. W,, Haack, J. A.., and Morrison. T. B. (1989). Nature, 338,
127.
5. Baldwin, R. L. (1990), Nature, 346. 409.

139
JAAP HERINGA

6. Flores, T. P., Moss, D. S., and Thornton, J. M. (1994). Protein Eng., 7, 31.
7. Bernstein, F. C., Koetzle, T. F., Williams, G. J., Meyer, E. F., Brice, M. D., Rodgers, J. R.,
et al (1977).J. Mol. Biol, 112, 535.
8. Schiffer, M. and Edmundson, A. B. (1967). Biophys.J., 7,121.
9. Rost, B. and Sander, C. (1993) .J. Mol. Biol., 232, 584.
10. Kabsch, W. and Sander, C. (1983). Biopolymers, 22, 2577.
11. Colloc'h, N., Etchebest, C., Thoreau, E., Henrissat, B., and Mornon, J.-P. (1993). Protein
Eng., 6, 377.
12. Sklenar, H., Etchebest, C., and Lavery, R. (1989). Proteins: Struct. Funct. Genet., 6, 46.
13. Woodcock, S., Mornon, J.-P., and Henrissat, B. (1992). Protein Eng., 5, 629.
14. Russell, R. B. and Barton, G. J. (1993). J. Mol. Biol., 234, 951.
15. Taylor, W. R. (1984). J. Mol. Biol., 173, 512.
16. Cohen, F. E., Abarbanel, R. M., Kuntz, I. D., and Fletterick, R. J. (1986). Biochemistry, 25,
266.
17. Cohen, F. E. and Kuntz, I. D. (1989). In Prediction of protein structure and the principles of
protein conformation (ed. G. D. Fasman), pp. 647-706. Plenum, New York, London.
18. Sternberg, M. J. E. (1992). Curr. Opin. Struct. Biol., 2, 237.
19. Benner, S. A., Cohen, M. A., and Gerloff, D. (1993). J. Mol. Biol., 229, 295.
20. Rost, B., Sander, C., and Schneider, R. (1994) J. Mol.Biol., 235, 13.
21. Frishman, D. and Argos, P. (1995). Proteins: Struct. Funct. Genet., 25, 633.
22. Szent-Gyorgyi, A. G. and Cohen, C. (1957). Science, 126, 697.
23. Periti, P. F., Quagliarotti, G., and Liquori, A. M. (1967).J. Mol. Biol., 24, 313.
24. Nagano, K. (1973). J. Mol.Biol., 75, 401.
25. Chou, P. Y. and Fasman, G. D. (1974). Biochemistry, 13, 211.
26. Lim, V. I. (1974). J. MoL Biol., 88, 857.
27. Garnier.J., Osguthorpe, D. J., and Robson, B. (1978).J. MoL Biol., 120, 97.
28. Gibrat, J.-F., Gamier, J., and Robson, B. (1987). J. Mol. Biol., 198, 425.
29. Gamier, J. G., Gibrat, J.-F., and Robson, B. (1996). In Methods in enzymology (ed. R. F.
Doolittle), Vol. 266, pp. 540-53. Academic Press.
30. Schultz, G. A. (1988). Annu. Rev. Biophys. Chem., 17,1.
31. Cohen, F. E., Abarbanel, R. M., Kuntz, I. D., and Fletterick, R. J. (1983). Biochemistry, 25,
4894.
32. Taylor, W. R. and Thornton, J. M. (1983). Nature, 354, 105.
33. Rooman, M. J., Wodak, S., and Thornton, J. M. (1989). Protein Eng., 3, 23.
34. Presnell, S. R., Cohen, B. I., and Cohen, F. E. (1992). Biochemistry, 31, 983.
35. Chandonia, J.-M. and Karplus, M. (1998). Proteins: Struct. Funct. Genet., 35, 293.
36. Yi, T.-M. and Lander, E. S. (1993). J. Mol. Biol., 232,1117.
37. Salamov, A. A. and Solovyev, V. V. (1995). J. Mol. Biol., 247,11.
38. Frishman, D. and Argos, P. (1996). Protein Eng., 9,133.
39. Frishman, D. and Argos, P. (1997). Proteins: Struct. Funct. Genet., 27, 329.
40. King, R. D. and Sternberg, M. J. E. (1996). Protein Sci., 5, 2298.
41. Muggleton, S., King, R., and Sternberg, M. J. E. (1992). Protein Eng., 5, 647.
42. Taylor, W. R. (1986). J. Theor. Biol., 119, 205.
43. Zvelebil, M. J., Barton, G. J., Taylor, W. R., and Sternberg, M. J. E.(1987).J. Mol. Biol.,
195, 957.
44. Levin, J. M., Pascarella, S., Argos, P., and Gamier, J. (1993). Protein Eng., 6, 849.
45. Minsky, M. and Papert, S. (1988). Perceptrons. MIT Press, Cambridge, MA.
46. Sander, C. and Schneider, R. (1991). Proteins: Struct. Funct. Genet., 9, 56.
47. Bowie, J. U., Luthy, R., and Eisenberg, D. (1991). Science, 253, 164.
48. Gonnet, G. H., Cohen, M. A., and Benner, S. A. (1992). Science, 256, 1443.

140
 PREDICTING SECONDARY STRUCTURE FROM PROTEIN SEQUENCES

49. Hutchinson, E. G. and Thornton, J. M. (1994). Protein Sci., 3, 2207.

50. Smith, T. F. and Waterman, M. S. (1981). J. Mol. Bicil, 147, 195.
51. Mehta, P. K., Heringa, J., and Argos, P. (1995). Protein Sci., 4, 2517.
52. Pascarella, S. and Argos, P. (1992). Protein Eng., 5, 121.
53. Heringa, J. (1999). Camp. Chem., 23, 341.
54. Heringa, J., Frishman, D., and Argos, P. (1997). In Proteins: a comprehensive treatise, Vol. 1,
principles of protein structure (ed. G. Allen), pp. 171-277. JAI Press, New York.
55. Liithy, R., McLachlan, A. D., and Eisenberg, D. (1991). Proteins: Struct. Func. Genet., 10,
229.
56. Langosch, D. and Heringa, J. (1998). Proteins: Struct. Func. Genet., 31, 150.
57. Unwin, N. (1993).J. Mol. Biol., 229, 1101.
58. Singer, S. J. (1962). Adv. Protein Chem., 17, 1.
59. Kyte, J. and Doolittle, R. F. (1982).J. Mol. Biol., 157, 105.
60. Argos, P., Rao, M. J. K.,and Hargrave, P. A. (1982). Bur. J. Biochem., 128, 565.
61. Sweet, R. M. and Eisenberg, D. (1983). J. Mol. Biol., 171, 479.
62. Cornette, J. L, Cease, K. B., Margalit, H., Spouge, J. L, Berzofsky, J. A., and DeLisi, C.
(1987). J. Mol. Biol., 195, 659.
63. Von Heijne, G. (1992). J. Mol. Biol., 225, 487.
64. Klein, P., Kanehisa, M. I., and DeLisi, C. (1985). Biochim. Biophys. Acta, 815, 468.
65. Rao, M. J. K. and Argos, P. (1986). Biochim. Biophys. Acta, 869, 197.
66. Wang, J. and Pullman, A. (1991). Biochim. Biophys. Acta, 1070, 493.
67. Eisenberg, D., Weiss, R. M., and Terwilliger, T. C. (1982). Nature, 299, 371.
68. Eisenberg, D., Weiss, R. M., and Terwilliger, T. C. (1984). Proc. Natl Acad. Sci. USA, 81,
140.
69. Finer-Moore, J. and Stroud, R. M. (1984). Proc. Natl. Acad. Sci. USA, 81, 155.
70. Vogel, H., Wright, J. K., and Jahnig, F. (1985). EMBO J., 4, 3625.
71. Esposti, M. D., Crimi, M., and Venturoli, G. (1990). Eur.J. Biochem., 190, 207.
72. Edelman, J. (1993).J. Mol. Biol., 232, 165.
73. Persson, B. and Argos, P. (1994). J. Mol. Biol., 237,181.
74. Fariselli, P., Compiani, M., and Casadio, R. (1993). Eur. Biophys. J., 22, 41.
75. Rost, B., Casadio, R., Fariselli, P., and Sander, C. (1995). Protein Sci., 4, 521.
76. von Heijne, G. (1986). EMBOJ., 5, 3021.
77. Boyd, D. and Beckwith, J. (1989). Proc. Natl. Acad. Sci. USA, 86, 9446.
78. Sipos, L and von Heijne, G. (1993). EurJ. Biochem., 213,1333.
79. Hartmann, E., Rapoport, T. A., and Lodish, H. F. (1989). Proc. Natl. Acad. Sci. USA, 86,
5786.
80. Schirmer, T. and Rosenbusch, J. P. (1991). Curr. Opin. Struct. Biol., 1, 539.
81. Lupas, A., van Dyke, M., and Stock, J. (1991). Science, 252,1162.
82. Lupas, A. (1996). In Methods in enzymology (ed. R. F. Doolitle), Vol. 266, pp. 513-25.
Academic Press.
83. Jones, D. T., Taylor, W. R., and Thornton, J. M. (1992). Nature, 358, 86.
84. Jones, D. T., Orengo, C. A., and Thornton, J. M. (1996). In Protein structure prediction: a
practical approach (ed. M. J. E. Sternberg), pp. 173-206. Oxford University Press, Inc.,
New York.

141
This page intentionally left blank
 Chapter 7
Methods for discovering
conserved patterns in protein
sequences and structures
Inge Jonassen
Department of Informatics, University of Bergen HIB, 5020 Bergen, Norway

1 Introduction
The amount of available biomolecular data is exploding: the number of known
protein sequences is increasing rapidly while the number of known structures is
also increasing, though not as rapidly. A very useful observation in this situation
is that common features among proteins can be used to group them into
families, and then study the proteins on a family level. By a family we mean a
set of proteins sharing some definite biological properties in terms of common
function and/or structure, often implying that the proteins have evolved from a
common ancestor, i.e. that they are homologous.
When studying a family, one can compare the sequences and structures (if
known) of the proteins in the family in order to find what sequence or structure
properties are shared by the family members and how these could explain the
biological properties shared by the proteins in the family. A description of
sequence properties is called a sequence pattern, and a description of structure
properties is called a structure pattern. If a pattern is common to a family of
proteins, it is called a motif for the family.
An example protein family is the set of proteins containing the classical zinc-
finger DNA-binding domain. Most of the sequences in this family match the
pattern C-x(2,4)-C-x(3)-[LIVMFYWC]-x(8)-H-x(3,5)-H, thus this is a sequence motif
for this protein family. A sequence matches this pattern if it contains a C
followed by 2-4 arbitrary letters followed by C and 3 arbitrary letters and one of
L, I, V, M, F, Y, W, or C, and so on (this pattern notation is described in detail
below). This particular pattern describes two cysteines and two histidines that
are needed for coordinating the zinc ion in the classical zinc-finger domain,
which means that this particular pattern has a direct biological interpretation.
Additionally, the pattern can be used for classification, since not only do most
sequences in the family match the pattern, but also very few sequences outside

143
INGE JONASSEN

the family match the pattern. So, if one finds a match to this pattern in a new
sequence, the chances are good that this new protein contains a zinc finger
domain and binds to DNA.
A large number of patterns have been compiled and collected in different
protein family databases. An example is the PROSITE database (1) which con-
tains more than 1000 protein families and for most of these it gives a pattern
which occurs in most of the sequences in the family. The patterns are regular
expressions (like the zinc finger pattern given above) or profiles (position specific
scoring matrices). Other databases also use local alignments, profiles, and
Hidden Markov models (HMMs). We describe briefly some of these databases
and how they can be used later in the chapter.
When one knows the structures of some of the proteins in the protein family
under study, the structures can be compared and similarities described as
patterns or motifs. In the same way that protein structures can be described at
different levels (e.g. atom, residue, secondary structure, element level), structure
patterns can describe structure properties at different levels. For example, one
pattern could describe the packing of four alpha-helices and another pattern
could describe the relative position of the cysteines and histidines in the
classical zinc finger.
In this chapter we will use a very broad definition of patterns including both
sequence and structure patterns and all the ways in which these can be defined.
When going into more detail we will focus on sequence patterns which are of
the regular expression type and on one particular type of structure patterns
which describes packing of individual residues. In the following we discuss how
existing databases of patterns can be used for analysing a new protein query
sequence and how to assess the output of such a search. Later we describe
different approaches to finding respectively sequence and structure patterns for
a family.

2 Pattern descriptions
A very simple type of patterns is substring patterns—a sequence matches a
substring pattern if it contains the substring (contiguous word in the sequence).
For example, the substring pattern CDEC is matched by all sequences contain-
ing CDEC as a substring. This very simple type of patterns can sometimes be
useful in the analysis of protein or nucleotide sequences. We will first define the
concept of approximate pattern matching and then describe different general-
izations of substring patterns.

2.1 Exact or approximate matching

When matching a sequence against a substring pattern one may allow for
approximate matching. In this case, a sequence matches the pattern if it con-
tains a substring approximately equal to the pattern. In practice, one defines a
measure of distance between a pattern and a substring and sets an upper limit

144
 METHODS FOR DISCOVERING CONSERVED PATTERNS IN PROTEIN SEQUENCES

on the distance to be allowed. One simple way to measure the distance between
two strings (or a pattern and a string) is to count the number of character
changes needed to transform one into the other. This is called the number of
mismatches or Hamming distance and can measure the distance between two
strings only if they have equal length. For example, the sequence AGCDFCALKW
approximately matches the substring pattern CDEC since the substring CDFC
can be transformed into CDED by substituting the F with an E.
More general distance measures allow for insertion and deletion of characters
in addition to substitutions. The edit distance between two strings (sequences) is
the minimum number of single character insertions/deletions and substitutions
needed to transform one into the other. For example, the sequence AGCDDALKW
approximately matches the pattern CDED when one allows for an edit distance
of more than one, but it does not match if one allows only for one mismatch.
When comparing protein sequences, one may also penalize different substitu-
tions differently since some amino acid replacements are found more often in
equivalent positions in homologous (evolutionarily related) proteins. For example
substitutions can be penalized using a substitution matrix, e.g. PAM-matrices (2)
or BLOSUM matrices (3).
In order to find the edit distance between two sequences, one can use the
dynamic programming algorithm (4) which can also be used when substitution
matrices are used. For substring patterns, the matching problem is very similar
to local pairwise alignment and database searching where speed-ups can be
used, e.g. BLAST (5), Fasta (6).

2.2 PROSITE patterns

Using substring patterns and approximate matching one cannot specify that some
pattern positions are compulsory (for instance, that there must be cysteines or
histidines for binding zinc) while other positions are allowed to vary more
freely. Figure 1 illustrates this by showing four segments (substrings) containing
the zinc finger motif and a consensus sequence (substring pattern) for the four.
If this substring pattern is to match the four segments shown, one needs to
allow up to 17 mismatches. A better way would be to allow no mismatches in
the conserved positions, including the cysteines and histidines, and allow the
remaining positions to be filled by any amino acid.
Some pattern languages allow a description of which amino acids are allowed
at each position. For instance, in the PROSITE database (1) one uses a pattern
language which is a subtype of regular expressions. Each pattern consists of a
sequence of pattern elements that are of the following types:
(a) Single residue: matches one letter identical to the residue in the sequence,
e.g. R matches an R in the sequence.
(b) Set of residues given in square brackets: matches any one sequence
letter contained in the set, e.g. [KER] matches any one of K, E, or R in the
sequence.

145
INGE JONASSEN

Rgure 1 Example of a local alignment of sequence segments containing the classical zinc-
finger motif. The consensus sequence below shows for each position which amino acid is
the most frequent, and to the right it is shown for each segment the number of mismatches
between the segment and the consensus. Grey shading marks the positions conserved in all
four segments. The positions of the conserved cysteines and histidines of the zinc-finger
motif are underlined in the consensus sequence.

(c) Set of residues given in curly brackets: matches any one sequence letter not
in the set, e.g. {KER} matches any letter except K, E, and R in the sequence.
(d) Wildcard x: matches any one letter in a sequence
Additionally:
(a) Single pattern elements can be followed by parentheses (i, j) which means
that the sequence can contain between i and j (inclusive) letters each match-
ing the preceding pattern element. For example, x(3,5) matches between 3
and 5 arbitrary sequence letters.
(b) The pattern can start with '<' meaning that the pattern should match from
the beginning of a sequence.
(c) The pattern can end with '>' meaning that the pattern should match until
the end of a sequence.
(d) The pattern elements are separated by hyphens '-'.
Consecutive pattern elements should be matched by consecutive sequence
symbols, so for example the pattern C-x(2,3)-[DE] matches any sequence con-
taining a C followed by two or three arbitrary letters followed by a D or an E.

2.3 Alignments, profiles, and hidden Markov models

Alternatives to regular expression type patterns are alignments, profiles, and
HMMs. These can also be seen as generalizations of substring patterns. Effect-
ively, for each position in the pattern, one now assigns a score (or a probability)
to each of the 20 amino acids. Additionally, one assigns penalties (or prob-
abilities) to insertions or deletions in each pattern position. Since alignments,
profiles, and HMMs assign a score (probability) to a match to a sequence, they
can be called probabilistic. Regular expression type patterns can be called deter-
ministic patterns—they are deterministic in the sense that a sequence either
matches or does not match the pattern (7).

146
 METHODS FOR DISCOVERING CONSERVED PATTERNS IN PROTEIN SEQUENCES

A probabilistic pattern is normally constructed from a local alignment of a set

of sequences from the family. A local alignment contains one (or more) seg-
ments from each sequence put on top of each other so as to align (put on top of
each other) corresponding sequence positions. One may allow for insertion of
special 'gap-characters' if needed to align corresponding positions. An example
of a local sequence alignment (without gaps) is shown in Figure 1 (see also
Chapter 8). Normally, one would make an alignment of the parts of the sequences
that are the most similar; for example in Figure 1 the zinc-finger domains are
aligned. Local alignments without gaps are used for example in the BLOCKS (8)
and the PRINTS (9) databases. A sequence can be matched with (aligned to) an
existing alignment using dynamic programming to optimize a score. A number
of methods exist for scoring the match between an alignment of a single
sequence letter to an alignment column and for penalizing gaps that may be
inserted. Taking one alignment and a scoring scheme, one can make a scoring
matrix giving a position specific scoring for each amino acid and also position
specific gap penalties. Such descriptions are called profiles and were initially
suggested by Gribskov et al. (10). An alternative approach is to use the frame-
work of Hidden Markov Models (HMMs) (11), which, in the way used in sequence
analysis, provides a scoring scheme similar to that of profiles, but based on
probabilities (see Chapter 4).
This means that the information contained in an alignment is often repre-
sented as a weight matrix, a profile, or a HMM which specifies for each column
in the alignment a score for each of the 20 amino acids when it is aligned with
this position. The scores can be calculated from the distribution of amino acids
in the column as well as using external information from substitution matrices
or Dirichlet mixtures (12). Schemes have been developed to weight the sequences
so as to adjust for biases in the input set (13).
PROSITE is also using profiles as a supplement to regular expression type
patterns, since for some families it is not possible to define one single regular
expression type pattern which matches all family sequences while avoiding
matches in unrelated sequences. In such cases it can be possible to define a
profile matching the sequences in the family with a higher score than all
(known) sequences outside the family. This may be possible since profiles have
more expressive power than the deterministic patterns as they can assign
different scores to each amino acid when matched to each pattern position and
also position-specific gap-penalties. On the other hand, profiles and other prob-
abilistic patterns contain many more parameters to be estimated, and to estimate
their values one needs a large number of examples (family members). Also, if
one is to learn patterns from noisy examples (including unrelated sequences),
the large number of parameters makes it easier to adapt the patterns to match
unrelated sequences. In the context of learning patterns from noisy examples,
therefore, the deterministic patterns can be more appropriate. Also, determin-
istic patterns are very simple, mathematically pure, and the human mind finds
them easy to interpret. Whichever patterns one choose to use, an apparent
problem is how to assess the patterns, to decide between alternative patterns

147
INGE JONASSEN

which is the best one, and to find whether the identified pattern could be the
result of chance.

2.4 Pattern significance

When the quality of a pattern is to be assessed, it depends on what is the
purpose of the pattern. One typical application of patterns is classification; that
is, the patterns are to be used for discriminating between family members and
non-members. Another is to find patterns that describe biologically important
features. Below we discuss some ways of assessing the quality of patterns with
respect to each of the two applications (see also ref. 7).

2.4.1 Characterization
Patterns can be used to describe biologically important features of the proteins,
that is, one wants to describe which features are compulsory in order for the
protein to belong to a particular family and which features are optional. If one
has available a set of sequences (or structures) from the same family, these can
be compared, and it can be determined which residues are conserved through
evolution and therefore likely to be important to the proteins' function and
structure. If the available proteins have undergone little evolution since their
last common ancestor (for example if their sequences are 90% identical), it will
not be easy to find which of the conserved residues are most important for the
biological function of the proteins. Therefore one should try to collect proteins
that are as diverse as possible while avoiding inclusion of unrelated proteins.
Having developed a pattern conserved in a set of sequences, one should find
whether such a pattern is likely to be conserved by chance and therefore not
necessarily biologically important. One common method is to calculate an esti-
mate of the probability that a set of random sequences (equal in number and
lengths to the sequences under analysis) would share a pattern of the same
strength as the identified pattern, as a result of chance. If this probability is very
low, one has a better reason for believing that the identified pattern has some
biological meaning. When evaluating the significance of the discovered pattern,
one should also take into account the number of patterns that have been con-
sidered in the pattern discovery phase (see e.g. ref. 14). For example, a pattern
with probability 10-6 is expected to be found if one million patterns are con-
sidered.
An alternative to calculating pattern probabilities is to measure the patterns'
information content (15). The higher the information content the pattern
possesses, the less likely a random sequence is to match it. The measure was
designed for ranking patterns matching the same number of sequences. Using
the principle of minimum description length (MDL) from machine learning (16),
this has been extended to also take into account the number of sequences
matching each sequence (17).
An alternative approach is to do a series of pattern discovery experiments on
sets of sequences with characteristics similar to the sequences in which the

148
 METHODS FOR DISCOVERING CONSERVED PATTERNS IN PROTEIN SEQUENCES

patterns were found. The sequences should be chosen so that they share no
significant patterns. The result of the pattern discovery experiments will give
information about what type of patterns can be found by chance. For example,
one can repeat x times: shuffle all the sequences, and check which patterns can
be found to match at least the same number of (the shuffled) sequences as the
pattern under analysis. An advantage of this approach is that in assessing the
'background probability' one can use sequences which have the same character-
istics as the original similar sequences (local sequence composition, etc.) for
example by using special shuffling operations.
When evaluating discovered patterns, it is important to take into consider-
ation whether (some of) the sequences under analysis are very closely related.
When calculating the probabilities, the model normally assumes that the
sequences are independently generated by some probabilistic model and the
shuffling would normally also extinguish any close similarities between the
sequences. If some of the sequences are very similar, they will contain many
common patterns, and any pattern matching one of them will probably match
all, and is therefore likely to be deemed as more significant. A scoring scheme
taking this into account, has been proposed (18).

2.4.2 Classification
When a pattern is to be used for classification, it should ideally match all family
members and no other sequences. Most often, however, the pattern fails to
match some member sequences (called false negatives), and it may match some
sequences outside the family (false positives). For an illustration, see Figure 2. The
fewer false negatives, the more sensitive the pattern is said to be, and the fewer
false positives, the more specific it is. Ideally, a pattern should have zero false
positives and negatives.
An estimate of the number of matches in a sequence database can be found
by multiplying the probability that one random sequence matches the pattern
by the number of sequences in the database. In order to calculate the probability
we assume that random sequences are generated using a specific probabilistic
model. Sternberg (19) did this for all the patterns in the PROSITE database and
showed a clear correlation between the expected number of false positives and

Figure 2 Illustration of the concepts of true positives, true negatives, false positives, and
false negatives. The circles are the family members and the squares are non-family
members. A pattern matches the encircled objects, and the status of each object is shown
by its colour (unfilled means 'true' and filled means 'false').

149
INGE JONASSEN

the actual number, i.e. the number of unrelated sequences in the SWISS-PROT
database (20) matching the pattern.
Denoting the number of true positives (sequences in the family matching the
pattern) by TP and the number of false negatives by FN, the sensitivity of a pattern
(21) can be defined as
Sensitivity = TP/(TP+FN)
and measures of how big a proportion of the family sequences are 'picked up by'
(matched by) the pattern. Similarly, the specificity of the pattern can be defined as
Specificity = TN/(TN+FP)
(where TN and FP are respectively the number of true negatives and the number
of false positives) which measures of how big a proportion of the sequences
outside the family are not matched by the pattern. Yet another useful number is
the positive predictive value (PPV) which says how big a proportion of the
sequences matching the pattern are actually in the family,
PPV=TP/(TP+FP)
The value range for all three is from zero to one, one being the best possible.
When evaluating patterns to be used for classification, one needs to use more
than one of the measures. This can be illustrated by two degenerate cases, (1) the
empty pattern matching any protein, and (2) a pattern matching one single
protein being member in the family. Pattern (1) has perfect sensitivity, but very
bad specificity and PPV, while pattern (2) has perfect specificity and PPV, but bad
sensitivity. For a concrete example of the use of these equations, see below. In
practice one often needs to make a trade-off between sensitivity and specificity
when choosing which pattern to use for a family. One way to evaluate a prob-
abilistic pattern's ability to discriminate between family members and other
sequences is to find a cut-off on the score that gives the same number of false
positives and false negatives. Tatusov et al. (22) evaluated alternative ways of
finding weight matrices from local ungapped alignments using this approach.

2.4.3 Discussion
Often the patterns that describe biologically important features will also be good
for classification purposes and vice versa. However, it is possible that a pattern
that gives perfect discrimination can be derived and yet lacks any biological
interpretation. Also, it may be that the features described by a pattern are im-
portant in the family, but not unique to the family, so that it is not specific
enough to be used for classification purposes.

2.5 Pattern databases

A number of different databases for storing information about protein families
and motifs have been established during the last ten years. They differ in a
number of ways. Firstly, they differ in how they represent the patterns for each
family. Secondly, some of them are constructed manually (both sequence group-

150
 METHODS FOR DISCOVERING CONSERVED PATTERNS IN PROTEIN SEQUENCES

Table 1 Summary information about some protein family databases

Database Pattern type URL

PROSITE Reg.exp Http : //expasy . hcuge . ch/sprot/prosite . html
and profiles
BLOCKS Blocks (ungapped Http : //www. blocks . there . org/
local alignments)
Prints Blocks Http : / /www. biochem . ucl . ac . uk/bsm/dbbrowser /
PRINTS/
Identify reg. exp Http: //mot if .stanford.edu/identify/
Pfam HMMs Http : //www. sanger . ac . uk/Pfam/

ing and pattern definition) while others are made (to a varying degree) auto-
matically. For some summary information about a few family databases, see
Table 1.

2.5.1 PROSITE
One of the most widely used databases is PROSITE (1), in which, or each family,
one or several patterns and/or profiles are given in the format described above
(Section 2.2). Profiles are described using the generalised profile syntax (23). Statis-
tics are given which describe the patterns' ability to discriminate between family
members and other sequences given in the SWISS-PROT protein sequence data-
base (20), in the form of the number of false positives and false negatives. Also, a
number of unknowns is given which is the number of sequences in SWISS-PROT
which match the pattern, but for which it is not yet known whether it belongs
to the family or not.
Figure 3 shows a PROSITE entry giving a signature pattern (motif) for the
actinin-type actin-binding domain. We will explain the most important (for our
purpose) parts of the entry. First, the ID and AC lines give, respectively, the
name and the accession number of the PROSITE entry. The pattern is given on
the PA line (the pattern can continue over several PA lines, then end of the
pattern being marked by a period sign). The NR lines give statistics about the
pattern's discriminatory power with respect to the SWISS-PROT database. The
first NR line says that the statistics are with respect to release 35 of SWISS-PROT,
which has 69113 sequence entries. The next NR line gives the number of
matches of different categories (true positives, false positives, false negatives,
etc.) in the SWISS-PROT database. Each is given as x(y) meaning that there are x
matches to the pattern in y different sequences.
The DR lines give references to the corresponding SWISS-PROT entries both
by their names and accession numbers. Each reference is on the. form 'AC, ID,
Status;' where Status is one of T (true positive), N (false negative), F (false positive),
P (partial), and ? (unknown). Finally, the 3-D line gives names of PDB (Protein
Data Bank) entries containing structures of proteins in the family, and the DO
line gives the accession number of the entry in the PROSITE documentation part
corresponding to this entry.

151
INGE JONASSEN

Figure 3 Example of a PROSITE entry taken from release 14.0 (November 1997). See text,
for a detailed explanation.

Referring to the pattern quality measures discussed earlier, this particular

pattern has sensitivity 28/(28 + 0) = 1, specificity (69113 - 28-20)/(69113 - 28)
= 0.9997 (69113-28 is the number of sequences in SWISS-PROT outside the
family), and PPV 28/(28 + 18) = 0.608. The last number means that if a sequence
from SWISS-PROT matches the pattern, the probability that it belongs to the
family is 60.8%. For this particular application, the PPV measure seems more
meaningful than specificity since the number of false positives do not affect the
measure of specificity very much as the number of true negatives most often
will be very much bigger than the number of false positives. Each entry in the
PROSITE documentation part gives a description of the family and explains the
biological significance of the signature patterns. It also gives literature refer-
ences and one or several experts that can be contacted for more information
about the family. The PROSITE database is largely maintained manually; new
families, profiles and patterns are carefully scrutinized.

2.5.2 BLOCKS
Another protein family database is BLOCKS (8) which contains the same families
as PROSITE, but instead of giving patterns or profiles, it gives a set of blocks for
each family. A block is an ungapped local multiple alignment. The database is
constructed fully automatically. For each family the member sequences (as given
in PROSITE) are subjected to pattern discovery and local alignment methods. The
blocks can also be linked in chains. It is recommended that a query sequence is
matched against both PROSITE and BLOCKS, because even though they describe

152
 METHODS FOR DISCOVERING CONSERVED PATTERNS IN PROTEIN SEQUENCES

the same families, the patterns and the blocks in a sense have complementary
strengths when used as classifiers.

2.5.3 PRINTS
The PRINTS database is a collection of fingerprints and is constructed semi-
automatically (9). A fingerprint is defined as a list of motifs, each motif being a
local ungapped alignment. The fingerprints are made by first manually making
an alignment of some family members and then iteratively scanning a database
of protein sequences, adding new members, updating the fingerprint, scanning
again until convergence (no new family members are found). Each entry in
PRINTS, gives the local alignments corresponding to the motif and information
about partial matches etc. On the website of PRINTS, a tool FingerPRINTScan, is
available for scanning a query sequence against the database. The output can be
visualized showing the position of the motif matches in the query sequence.

2.5.4 PFam
Pfam is a database of multiple sequence alignments and HMM-profiles of protein
domains (24). It is partly manually curated. Seed alignments for each family are
made semi-automatically, and these are extended using HMM methods. Version
3.1 (August 1998) contains 1313 families.
On the Web site there are available tools for matching a query sequence
against the database. Also, a special database SWISSPFam is made which shows
the domain organization (according to Pfam) or the sequences in SWISS-PROT
and TrEMBL.

2.5.5 Identify
Identify (25) is a database of patterns of the same form as used in PROSITE but
without flexible length wildcards. It contains patterns for the families in the
BLOCKS and PRINTS databases. The patterns were constructed from the un-
gapped alignments (blocks) in the BLOCKS and PRINTS databases by using a
pattern finding program (EMOTIF). For each block, there can be several patterns
so that each pattern matches a subset of the sequences. The patterns were
generated to have a certain specificity (calculated as the probability that a
random sequence matches the pattern by chance, cf. ref. 19), and patterns were
generated for different specificity levels. The World Wide Web server allows the
user to input a query sequence which then is matched against the pattern
collection and the user is given the list of matching patterns together with links
to the corresponding BLOCKS and PRINTS database entries.

2.6 Using existing pattern collections

Most of the family databases are available on the World Wide Web, and they
also provide on-line tools for (1) matching a sequence against the pattern in the
database, and sometimes (2) matching a new (user-defined) pattern against a
sequence database. For example, on the PROSITE website, the search engine

153
INGE JONASSEN

ScanProsite that can be used for both (1) and (2) with regular expression type
patterns. For scanning against the profiles in PROSITE, the tool ProfileScan can
be used. The ScanProsite program does not return any information about the
probability that the match between the sequence and the pattern could be by
chance, and it does only allow for exact matching between the pattern and the
sequence.
Another very useful tool is PdbMotif (26), which takes as input a protein
structure and finds all matches between the protein's sequence and patterns in
the PROSITE database. It generates a script that can be input to RasMol (27) to
highlight the pattern matches. PdbMotif also outputs a probability that the
sequence should match the PROSITE pattern by chance, which can help to
identify possible false positives.

3 Finding new patterns

Analysing a set of proteins believed to be related one may want to find a motif
describing the set. The goal may be to find which features are common to the
proteins under study helping to better understand the relationships between
sequence, structure, and function of the proteins. Motifs can also help to identify
new possible family members. We may also want to find motifs to be included
in protein family databases where the motif may later be used for classification.
One may develop motifs semi-manually using knowledge about the proteins
under study and for example alignments generated either manually or auto-
matically using a multiple (local or global) sequence alignment program. If
the alignment programs were guaranteed to find the correct alignments, this
approach would be all you needed. However, multiple alignments are often diffi-
cult to obtain and interpret. Therefore, in many cases, direct methods for find-
ing motifs directly from the sequences or structures can lead to better results. In
such cases, the motifs found can also be used to guide the alignment of the
sequences or structures in the family.

3.1 A general approach

There exist a large number of methods for the discovery of patterns from pro-
tein (and DNA) sequences. A survey of these is given by Brazma et al. (7) who
propose a general three-step approach to pattern discovery:
(a) Choose a solution space, i.e. set of patterns that the method potentially can
discover.
(b) Define a fitness function reflecting how well a pattern fits the input
sequences.
(c) Develop an algorithm, which given a set of input sequences, returns the
pattern(s) from the solution space with high (highest) fitness.
The different methods can be classified according to their approach to each of
the three steps. This three-step approach can be used for the discovery of all

154
 METHODS FOR DISCOVERING CONSERVED PATTERNS IN PROTEIN SEQUENCES

types of sequence patterns (including regular expression type patterns, profiles,

and Hidden Markov models) and also for the discovery of motifs in protein
structures. An additional criterion is whether a method is guaranteed, for any
input set, to find the best (as measured by the fitness function) pattern in the
solution space. In the following sections we will illustrate the three steps
approach by describing in some detail some representative methods, especially
focusing on the Pratt and SPratt methods.

3.2 Discovery algorithms

Previous sections have discussed in some detail different solution spaces and
principles of fitness functions. Here we discuss in some more detail the third step,
namely that of finding the patterns in the solution space that have high, and
possibly the highest, fitness. Brazma et. al. (7) identified two main algorithmic
approaches to this problem:
(a) Pattern Driven (PD) methods.
(b) Sequence Driven (SD) methods.

3.2.1 Pattern driven methods

In the simplest form, PD methods enumerate all patterns in the solution space
calculating each pattern's fitness so that the best ones can be output. For ex-
ample, one step of the algorithm by Smith et al. (28) works by enumerating all
patterns of the form Al-x(dl)-A2-x(d2)-A3 having single amino acid symbols and d
values up to some maximum (e.g. 10). For each pattern, the number of matching
input sequences were counted, and the patterns matching the most sequences
were subjected to further analysis.
Some methods have mechanisms for avoiding looking at all patterns in the
solution space, for example by not searching parts of the solution space that
cannot possibly contain patterns scoring higher than already analysed patterns.
Often heuristics are used to discard parts of the solution space that are unlikely
to contain high scoring patterns. Examples of this will be seen when the Pratt
method is described in some more detail below (see Figure 4).

3.2.2 Sequence driven methods

These methods work by comparing (normally pairs of) the input sequences and
expressing local similarities as patterns. By repeating this for different pairs of
sequences, or by subjecting the patterns themselves to pairwise comparison,
patterns are gradually built up that match many or all of the input sequences.
Some of these methods are very closely related to methods for sequence
alignment.
For example, the pattern discovery method proposed by Smith and Smith (29)
first computes the similarity between all pairs of input sequences. The most
similar pair is input to a local pairwise alignment method that is based on
dynamic programming (30). Instead of producing a pairwise alignment, this
algorithm outputs a pattern common to the two sequences. The pattern is basic-

155
INGE JONASSEN

Figure 4 Example of a pattern-driven discovery method applied to a set of four sequences:

ACGTT, TACGTA, GCACGT, and TTACGTAA, The solution space consists of all words over the
DNA alphabet (no wildcards or group characters), and patterns are sought that match all four
sequences. Patterns matching all four are boxed, and only these are extended. The longest
pattern found to match all four sequences is ACGT, The 'e' at the top is the empty pattern
which matches any sequence.

ally of the PROSITE type, but the set of possible amino acid groups is limited and
given by an Amino Acid Class Covering (AACC) hierarchy. The two sequences
aligned are now replaced by their common pattern, and the procedure is repeated
until there remains only one pattern matching all of the input sequences. For an
example, see Figure 4. Note that this approach is very analogous to the pro-
gressive multiple alignment methods used for example by Thompson et al. (31)
and Taylor (32).

4 The Pratt programs

The Pratt programs take as input a set of (un-aligned) sequences and finds pat-
terns matching at least a (user-defined) minimum number of the sequences. The
patterns are of the type used in PROSITE and can include both character groups
and flexible length wildcards. For instance, Pratt is able to automatically
rediscover the pattern C-x(2,4)-C-x(3)-[LIVMFYWC]-x(8)-H-x(3,5)-H from the set of
over 300 classical zinc-finger-containing protein sequences in SWISS-PROT in
less than a minute on a modern workstation. No prior information about the
pattern is given to Pratt and the sequences are input unaligned.
The user can input constraints on the patterns to be considered, effectively
defining the solution space to be used (cf. three-step approach). The user also
chooses how many of the input sequences a pattern should match, in order to
be output. We will call any pattern matching at least this number of sequences,

156
 METHODS FOR DISCOVERING CONSERVED PATTERNS IN PROTEIN SEQUENCES

a conserved pattern. Pratt uses a two-step search for finding conserved patterns
from the chosen solution space having maximum fitness. The fitness is norm-
ally defined as the information content of the pattern (see below).
In the following sections we give a practical guide to how Pratt should be
used and then some details about the algorithms used in Pratt. For more detailed
technical descriptions, see the original papers (15, 33) and the Pratt home page
on the World Wide Web: http://www.ii.uib.no/~inge/Pratt.html.

4.1 Using Pratt

When using Pratt from a command line environment, it is started using the
command
pratt <format> file [options]
where format is one of fasta or swissprot and file is a file containing the
sequences in the indicated format. Optionally, one can specify options on the
command-line. If no options are given, a menu appears. Figure 5 shows the menu
of Pratt when run on a file MUTT containing 26 unaligned sequences. The menu
can be used to choose values for a number of parameters and also to obtain help.
The parameters fall within a few main classes:
(a) Pattern conservation. Using options CM (respectively, C%) one can set the
minimum number (respectively, percentage of the input sequences) of
sequences a pattern should match.
(b) Pattern restrictions. A number of parameters are used to constrain the
patterns to be considered. For example PL can be used to set the maximum
length of a pattern, PN to set the maximum number of non-wildcard sym-
bols, PX to set the maximum length of a wildcard region. For constraining
the flexibility to be allowed in wildcard regions, FN is used to set the maxi-
mum number of flexible wildcards and FL to set the maximum flexibility of a

Figure 5 Pratt menu. The options are explained in the text.

157
INGE JONASSEN

wildcard. Note that the PX constraint also applies to patterns to be found

during the initial search. For some examples, see Table 2. Also, one can choose
which pattern symbols should be used during initial pattern search and
during pattern refinement using options BI (BF) and BN.
(c) Pattern scoring. By default patterns are scored by their information con-
tent. Optionally one can use a scoring function derived from the Minimum
Description Length (MDL) principle. Patterns can also be ranked by their
positive predictive value (PPV)—in this case the name of a file containing a
sequence database in flat file format must be given using option SF.
(d) Search parameters. By default, the shortest sequences will be used for
deriving a pattern graph (see below). Optionally this can be generated instead
from a special query sequence or from an alignment, in which case only
patterns that match the query sequence respectively are consistent with the
alignment, will be considered in the search.
(e) Output format. Filename for output can be chosen using option OF, format
of patterns using option OP, if pattern matches should be shown (option OA),
etc.
(f) Help and control. Help can be obtained by typing in option H, the search
started by using option X, or abandoned using option Q.
All parameters that can be set using the menu can alternatively be set from the
command line by adding '-<menu option> <value>' to the command. For
instance
pratt fasta seqs -cm 20
tells Pratt to analyse the sequences in the file seqs (fasta format) to find
patterns matching at least 20 of the sequences. When command line options are
used, the menu will not appear unless the option -menu is used.
For example, if the file c2h2 contains the sequences (in fasta format) of the
proteins in the classical c2h2 zinc finger family, these can be analysed using the
command:
pratt fasta c2h2 -px 15

Table 2 The table shows some example patterns, and for each example, the minimum
values to be used for some Pratt parameters if the pattern is to be discovered. For example,
in order to discover the bottom-most pattern, one needs to increase the value of the PX
parameter to at least 12

Minimum parameter values

Pattern PL PN PX FL FN
C-x(3)-C 5 2 3 0 0
C-x(2,5)-C 7 2 5 3 1
C-C-D-E-x(7)-C 12 5 7 0 0
C-x(2,4)-C-x(12)-H-x(3,5)-H 25 4 12 2 2
Default values 50 50 5 2 2

158
 METHODS FOR DISCOVERING CONSERVED PATTERNS IN PROTEIN SEQUENCES

using the option -px 15 to allow long wildcards (if we want to re-discover the
known motif, we need to set PX to at least 12 since during the initial pattern
search, a spacing of 12 is needed between the last conserved cysteine and the
first histidine). Also, if one wants to find all patterns matching a minimum 90%
of these sequences, one can use the command
pratt fasta c2h2 -px 15 -c% 90

4.2 Pratt: Internal search methods

The search for conserved patterns is done in two phases:
(a) Initial pattern search. Search for patterns having only single character
elements and wildcards (possibly of flexible length). For example, in this
phase the pattern Pratt can discover that the pattern A-x(4) -D-X-E is con-
served. Optionally, group characters can be allowed also in this step, at the
cost of increased computing time.
(b) Pattern refinement. Take each of the best patterns from phase 1, collect
the matching sequence segments and check if wildcard positions can be
replaced by group characters so that the pattern remains conserved. Also, the
pattern may be extended to the right (but not to the left) in this phase. For
example, the pattern above can now be refined to
A-x-[KER]-x(2)-D-[ILV]-E-x(4)-[KR].

4.2.1 Initial pattern search

In the first version of Pratt (15), in the first phase a search tree containing all
patterns to be considered, is explored. The label of the root node is the empty
pattern. A node having the pattern P as label has children with labels P-x(i,j)-A,
that is P extended with a wildcard region (empty if i=j=0) and a single character
A, for all allowable values of i, j, and A.
The search starts at the root of the tree. At each step of the search a node in
the tree having label P is considered. It is assumed that P is conserved. Then, all
children of P are generated, and for each of these it is checked whether the
corresponding pattern is conserved (whether it matches the minimum number
of sequences). If a pattern is conserved, this is recursively analysed using the
same procedure. If no extension of P is conserved, then, depending on the score
of P, it is included in the list of the best patterns that are subsequently input to
the refinement algorithm.
In the second version of Pratt (33), instead of considering the full tree of all
patterns in the solution space, a pattern graph is used to define the set of
patterns to be considered. The pattern graph is a node- and edge-labelled directed
acyclic graph. The nodes are labelled with non-wildcard pattern elements and
the edges have as labels wildcard lengths. A path in the graph defines a pattern
and from this pattern, more generalized versions will be generated. An example
of a pattern graph is shown in Figure 6.

159
INGE JONASSEN

Figure 6 Example of a pattern graph. The paths in the graph define the patterns A-B-x(0,2)-
x(3,3)-D, A-B-x(0,2)-C, A-B, B-x(0,2)-C-x(3,3)-D, B-x(0,2)-C, C-x(3,3)-D, A-x(l,3)-C-x(3,3)-D, A-
x(l,3)-C.

The initial search explores all patterns that can be derived from paths in the
pattern graph and that are contained in the class of patterns defined by the user.
The search is focused on finding only the highest scoring patterns. Branch-and-
bound techniques are used to avoid considering parts of the search space that
cannot possibly contain patterns with higher scores than already identified
patterns.
Also, heuristics have been implemented that effectively avoids exploring
search paths unlikely to produce patterns scoring higher than patterns already
found. The user can adjust the greediness of the search. Setting the E parameter
to zero gives non-heuristic search (guaranteed to find highest-scoring patterns),
setting E to 1 gives the same guarantee in cases where no flexibility is allowed in
wildcard regions, and E values above 1 gives increasingly greedy search. The
default value is 3. The more greedy the search, the faster it will be, and the more
likely Pratt is to not find the highest scoring patterns. Experiments have shown
that E = 3 gives a good compromise between speed and accuracy for protein
sequences, while for DNA a lower value should be used (for instance, E = 1.5).

4.2.2 Pattern refinement

During refinement, each position in the fixed-length wildcard regions (exclud-
ing regions x(i, j) where j > i) is analysed and it is checked whether replacing the
wildcard with an allowed group character (the allowed groups are given as a list)
so that the resulting pattern remains conserved. It might be that there are several
wildcard positions that can be replaced by group characters, so that replacing
any one of these gives a conserved pattern, but if all are replaced simultane-
ously, the resulting pattern will not be conserved. As there are exponentially
many (in the number of positions that can be replaced) subsets of replacements
that could be done, it is computationally expensive to consider all. Therefore, in
the second version of Pratt (33), a heuristic refinement algorithm is used. The
degree of greediness is adjusted using the same E-parameter as for the initial
search.

4.2.3 Block data structure

To find all matches to each pattern quickly, a special data structure initially
proposed by Neuwald and Green (14) is used. Assume that the input sequences
are S = { SI, S2,..., Sn}. For some fixed parameter w (typically 50), let B be the set
of all w-segments (consecutive substrings of length w) from all of the sequences
in S. Also construct w-segments in the end of the sequences by appending to

160
 METHODS FOR DISCOVERING CONSERVED PATTERNS IN PROTEIN SEQUENCES

Figure 7 Example of the set B of w-segments made for a set of sequences. The B segments
are used in the block data structure.

each sequence w-1 dummy symbols '-'. For an example of the w-segments made
for a set of sequences, see Figure 7.
Now, for each amino acid symbol a, and for each i between 1 and w, construct
the set bi,a that is the set of all w-segments having character a in position i. These
sets can be used to quickly find the set of w-segments matching any pattern
considered by Pratt not having length exceeding w. For instance, the set of
segments matching A-x(2) -B is b1,A nb4,B. In the recursive search, the set of
segments matching P is used together with the block data structure to find
the segments matching each extension P-x(i,j)-A of P. For a more detailed
description, see Jonassen et al. (15).

4.3 Scoring patterns

Pratt can score discovered patterns using different fitness functions. By default
the patterns are evaluated by their information content (15). This measure de-
pends only on the pattern itself and is only appropriate for ranking patterns
matching the same number of sequences. The information content of a pattern
is a measure of the information gained about an unknown sequence when one
is told that the sequence matches the pattern. It increases with the number of
single character elements (any one single character contributing the same) and
with the number of group character elements (more ambiguous contributing
less). A penalty is subtracted for flexible wildcard regions, more specifically x(i,j)
is penalized by c(j-i) where c is a parameter whose default value is 0.5.
As an alternative, Pratt can score the discovered patterns using the minimum
description length (MDL) principle based fitness measure described by Brazma
et al. (17). This assigns a score to a pattern that depends on the pattern's infor-
mation content and on how many sequences it matches. The fitness measure
definition contains some parameters that can be used to slant the optimum
towards strong patterns matching few sequences or towards weaker patterns
matching many sequences. Brazma et al (17) used the measure in a method for
simultaneously finding subfamilies and patterns in a set of unaligned (and un-
labelled) sequences. It was required that each pattern matches all sequences in
one of the subfamilies. The method uses Pratt to find patterns matching differ-
ent sized subsets of sequences. Next it selects in a greedy fashion a collection of
the patterns that cover the input sequence and has a high fitness value.

161
INGE JONASSEN

If the aim is to find patterns to be used for classification, Pratt can evaluate
the discovered patterns by their positive predictive value (PPV, see Section 2.4.2,
Classification). It is assumed that the sequences under analysis are all in the
SWISS-PROT database (20). The number of false positives for each pattern is
found by matching the patterns against the SWISS-PROT database that must be
locally available in flat file format.

5 Structure motifs
Finding recurring patterns (motifs) in protein structures help to better under-
stand the rules underlying the formation of protein structures. Since structure is
better conserved during evolution than sequence, structural similarities can also
help to identify remote evolutionary relationships. Structure motifs can help in
approaching the structure prediction problem and in assigning function to pro-
teins. Structure motifs can represent common structural features at different
levels. For example, they can represent packing of secondary structure elements,
local packing of residues, and atom coordinates of binding atoms in active site
(or ligand binding) residues. Structure motifs describing functional sites in
proteins, have been developed by, for example, Wallace et al. (34). They call their
motifs 'templates' and suggest that they can be used for finding functional
sites in proteins. They have also developed a database PROCAT of such templates
(35) which allows the user to search for 3-D enzyme active sites in a protein
structure.
In order to find recurring patterns in protein structures one can use methods
for the comparison of protein structures. A number of such methods have been
developed, most of them for comparing pairs of structures, but also some for
multiple structure comparison. The methods differ in what similarities they are
able to find. Some represent the structures as composed of secondary structure
elements (alpha helices, beta strands, and loops) and have provided methods for
finding patterns of conserved patterns at this level. Other methods find patterns
of residues (or atoms) that have similar configurations in space. Brown et al. (36)
gives a survey of a large number of different methods focusing on their way of
representing similarities. An important difference between methods is whether
they require matched elements to be in the same order along the proteins'
primary structure. A number of different methods have been used, including
extensions of the dynamic programming algorithms used for pairwise sequence
(37), use of graph-theoretic methods (38), and methods from computer vision
(39).
Also, some more direct methods for structure motif discovery have been sug-
gested. Here we will describe the SPratt program, a more detailed description of
which can be found in (40).

5.1 The SPratt program

The idea behind the SPratt program was to use the Pratt method developed for
sequence motif discovery, to discover structure motifs. This can be done by

162
 METHODS FOR DISCOVERING CONSERVED PATTERNS IN PROTEIN SEQUENCES

encoding structural features in the form of strings and input these to Pratt
producing patterns common to the strings. Next one needs to check if the
patterns found in the structure description strings correspond to similarities
between the structures. The encoding of structural properties that we adopted
was one described by Karlin and Zhu (41). In their method, they make one string
per residue in each structure. The strings contain information about the spatial
neighbourhood of each residue. Karlin and Zhu describe alternative methods for
making these strings, and we chose one of them to be used in SPratt.

Overview of the method:

1. For each residue a, we make two neighbour strings Ca and Na. The string Ca
starts with a followed by all the residues C-terminal to a, whose spatial
distance to a is below a user-chosen threshold dmax. The residues are ordered
in N-to-C chain order. Analogously, the string Na starts with a and contains
in C-to-N order the residues preceding a in the chain which are spatially
close to a (distance below dmax).
2. Run Pratt twice, once on the complete set of C and once on the N strings.
Search for patterns whose matches start from the beginning of the matched
residue strings and that match residue strings from at least the minimum
number of structures chosen by the user.
3. For each pattern, consider the neighbourhood string matches and retrieve
the substructures (list of residues) corresponding to each such match. Com-
pare the spatial geometry of the substructures by calculating the root mean
square deviation (RMSD) when superposing each pair of substructures.
4. Output the patterns for which all pairwise RMSD values are below some
upper limit, and rank them by I/R where I is the information content of the
neighbourhood string pattern and R is the maximum RMSD for any pair of
matching substructures.
The method differs from other structure comparison methods in that it con-
siders information from all structures simultaneously in step 2. Other methods
perform a number of pairwise structure comparisons and combine the results to
find motifs shared by all or most of the structures under study. By utilizing in-
formation from all structures simultaneously, SPratt avoids considering patterns
common to pairs of structures but not shared by the others. On the other hand,
when Pratt is used to analyse the neighbourhood strings, many patterns can be
found which later prove not to reflect similar substructures (i.e. they do not
superpose very well). It might be advantageous to use additional structural in-
formation in this step to avoid considering such patterns. We explore extending
the encoding of the neighbourhoods to include information about secondary
structure and restrict Pratt to match only residues from the same sort of
secondary structure (alpha-helix, beta-strand, or loop).
The patterns found by SPratt are evaluated using a very simple function, which
basically rewards patterns containing more residues and imposing stricter re-
strictions on the amino acids allowed for each residue, and penalizes patterns

163
INGE JONASSEN

for which the occurrences do not superpose very well using a measure of RMSD.
However, the RMSD value for superposing the coordinates of a small number of
residues, for example 4 is not very informative. The significance of the patterns
found by Pratt, can be further assessed using the structure alignment program
SAP (42). We have done this by rewarding alignment of residues in agreement
with the motif, and in this Way checking whether the matching of the few
residues described by the motif can be extended to an alignment of larger parts
of the structures.

6 Examples
In (15) we described the application of the first version of Pratt to the analysis of
some protein families in PROSITE. For example, we analysed the Snake toxin
family (PS00272 in PROSITE) containing 164 sequences of average length 64. We
retrieved the sequences from the SWISS-PROT database and input them (un-
aligned) to the Pratt program. Using default parameters (which requires patterns
to match all the sequences), no patterns were found. However, using Pratt to
discover patterns matching at least 155 out of the 164 sequences, we got the
pattern G-C-x(l,3)-C-P-x(8,10)-C-C-x(2)-[PDEN]. This pattern turned out not to
match any sequences in SWISS-PROT apart from the family members and was
since included in the PROSITE database as the pattern for this family.
Using the SPratt program we analysed a set of cupredoxin protein structures
(40). The proteins were selected from the cupredoxins super-family in SCOP (43)
so that all pairwise sequence similarities were 30% or less. The 10 structures
were input to the SPratt program and it was instructed to search for patterns
containing single residue elements and also allowing the match-set [MLQ] (since
the methionine ligand-binding residue is known to be substituted by L or Q. in
some proteins). SPratt used two minutes and identified three patterns all match-
ing around the copper binding sites. The substructure occurrences of the
pattern superpose with very low RMSD values (0.7 A or less). We also used the
motif identified by SPratt to guide the structure alignment program SAP using
the pairs of equivalenced residues as extra constraints on the alignment. This
resulted in a greatly improved alignment with RMSD of 1.56 A over 63 pairs of
residues (as compared to the alignment with RMSD of 5.1 A over 26 residues
when no SAP was run without any motif information).

7 Conclusions
Patterns and pattern discovery tools can help in the analysis of protein families,
i.e. sets of proteins believed to share structural and/or functional properties. The
most well conserved parts of the sequences or structures can be identified and it
can be analysed whether the conserved patterns are statistically significant and
therefore likely to have biological importance. Furthermore, the patterns can be
used to identify additional related proteins. Patterns can describe protein prop-

164
 METHODS FOR DISCOVERING CONSERVED PATTERNS IN PROTEIN SEQUENCES

erties at sequence level (sequence patterns) or at structure level (structure

patterns). There exist a number of databases of protein families that give for
each family one or several sequence patterns that can be used to identify more
family members. We described in some detail the PROSITE, BLOCKS, Identify,
PRINTS, and the Pfam databases. Structure pattern databases are starting to appear,
for instance PROCAT is a database of 3-D enzyme active site templates (35).
For sequence patterns we made a distinction between deterministic and prob-
abilistic pattern. Regular expression type patterns fall into the first class and
profiles and HMMs fall into the second class. The different types of patterns used
each have their strengths and weaknesses. For example, regular expression
patterns are easily interpreted, but provide less expressive power than profiles
or HMMs.
A number of methods have been developed for the automatic discovery of
conserved patterns in protein sequences. We summarized a framework com-
prising a three-step approach that can be used to better understand the myriad
of different methods. One sequence motif discovery program, Pratt, was de-
scribed in some more detail, focusing on the more practical aspects of how it
can be used to find conserved regular expression type patterns in unaligned
protein sequences.
Structure motifs can describe properties at different levels, e.g. at atom group
or residue level or at secondary structure level. Most structure comparison
methods are for comparing pairs of structures, but some of these have been ex-
tended to the comparison of several by combining the results of a set of pairwise
comparisons. We described in some detail a tool SPratt (Structure Pratt) which is
able to find patterns of locally packed residues whose spatial positions are
similar in a set of protein structures. This method utilizes information from all
structures simultaneously in the search for conserved motifs.
Motifs provide a powerful classification tool for determining structure and
function of proteins coming out of the numerous genome projects. Also,
methods for finding conserved patterns in structures and sequences can be used
to extract information from the large amounts of biomolecular data. In this
chapter we have only considered patterns and motifs in proteins, but pattern
discovery methods can also be used in other applications. For example, recently,
a pattern discovery method was used to find putative regulatory elements in sets
of gene upstream regions from the Yeast genome (44).

References
1. Bairoch, A., Bucher, P., and Herman, K. (1996). Nucleic Acids Res., 24,189.
2. Dayhoff, M. O., Schwartz, R. M., and Orcutt, B. C. (1978). In Atlas of protein sequence and
structure (ed. M. O. Dayhoff), Vol. 5, Suppl. 3, p. 345. National Biomedical Research
Foundation, Washington DC.
3. Henikoff, S. and Henikoff, J. G. (1992). Proc. Natl. Acad. Sci. USA, 89,10915.
4. Needleman, S. B. and Wunsch, C. D. (1970).J. Mol Biol., 48, 443.
5. Altschul, S. F., Gish, W., Miller, W., Myers, E. W., and Lipman, D. J. (1990).J. Mol. Biol.,
215, 403.

165
INGE JONASSEN

6.
Lipman, D. J. and Pearson, W. R. (1985). Science, 277, 1435.
7.
Brazma, A., Jonassen, I., Eidhammer, I., and Gilbert, D. (1998). J. Comp. Biol., 5, 279.
8.
Pietrokovski, S., Henikoff, J. G., and Henikoff, S. (1996). Nucleic Acids Res., 24, 197.
9.
Attwood, T. K., Beck, M. E., Bleasby, A. J., Debtyarenko, K., and Smith, D. J. P. (1996).
Nucleic Adds Res., 24,182.
10. Gribskov, M., McLachland, A. D., and Eisenberg, D. (1987). Proc. Natl. Acad. Sri. USA, 84,
4355.
11. Krogh, A., Brown, M., Miah, I. S., Sjoelander, K., and Haussler, D. (1994). J. Mol. Biol.,
235,1501.
12. Brown, M., Hughey, R., Krogh, A., Mian, I. S., Sjoelander, K., and Haussler, D. (1993). In
Proc. 1st Int. Conf. On Intell. Systems for Mol. Biol, pp. 47-53. AAAI Press.
13. Altschul, S. F., Carroll, R. J., and Lipman, D. J. (1989). J. Mol. Biol., 207, 309.
14. Neuwald, A. F. and Green, P. (1994). J. Mol. Biol., 239, 698.
15. Jonassen, I., Collins, J. F., and Higgins, D. G. (1995). Protein Sci., 4,1587.
16. Rissanen, J. (1978). Automotica-J.IFAC, 14, 465.
17. Brazma, A., Jonassen, I., Ukkonen, E., and Vilo, J. (1996). In Proc. 4th Int. Conf. On Intell.
Systems for Mol Biol, pp. 34-43. AAAI Press.
18. Jonassen, I., Helgesen, C, and Higgins, D. (1996). Reports in Informatics, report no.
116. Dept. of Informatics, Univ. of Bergen, Norway.
19. Sternberg, M. J. E. (1991). Nature, 349, 111.
20. Bairoch, A. and Boeckmann, P. (1992). Nucleic Acids Res., 20, 2019.
21. Lathrop, R., Webster, T., Smith, R., Winston, P., and Smith, T. (1993). In Artificial
intelligence and molecular biology (ed. L Hunter), pp. 211-58. AAAI Press/The MIT Press.
22. Tatusov, R. L, Altschul, S. F., and Koonin, E. V. (1994). Proc. Notl. Acad. Sri. USA, 91,12091.
23. Bucher, P.and Bairoch, A. (1994). In Proc of 2nd Int. Conf. On Intell. Systems for Mol. Biol.,
pp. 53-61. AAAI Press.
24. Sonnhammer, E. L., Eddy, S. R., Birney, E., Bateman, A., and Durbin, R. (1996). Nucleic
Acids Res., 26, 320.
25. Nevill-Manning, C. G., Wu, T. D., and Brutlag, D. L. (1998). Proc. Natl. Acad. Sri. USA, 95,
5865.
26. Saqi, M. A. S. and Sayle, R. (1994). Comput. Appl. Biosci., 10, 545.
27. Sayle, R. A. and Milner-White, E. J. (1995). Trends Biochem. Sci., 20, 374.
28. Smith, H. O., Annau, T. M., and Chandrasegaran, S. (1990). Proc. Natl. Acad. Sri. USA, 87,
826.
29. Smith, R. F. and Smith, T. F. (1990). Proc. Natl. Acad. Sci. USA, 87,118.
30. Smith, T. F. and Waterman, M. S. (1981). J. Mol. Biol., 147,195.
31. Thompson, J. D., Higgins, D. G., and Gibson, T. J. (1994). Nucleic Acids Res., 22, 4673.
32. Taylor, W. R. (1988). J. Mol. Evol., 28, 161.
33. Jonassen, I. (1997). Comput. Appl. Biosci., 13, 509.
34. Wallace, A. C., Borkakoti, N., and Thornton, J. M. (1997). Protein Sci., 6, 2308.
35. http://www.biochem.ucl.ac.uk/bsm/PROCAT/PROCAT.html
36. Brown, N. P., Orengo, C. A., and Taylor, W. R. (1996). Comput. Chem., 20, 359.
37. Taylor, W. R. and Orengo, C. A. (1989). J. Mol. Biol., 208,1.
38. Artymiuk, P. J., Porrette, A. R., Grindley, H. M., Rice, D. W., and Willett, P. (1994).
]. Mol. Biol, 243, 327.
39. Nussinov, R. and Wolfson, H. J. (1991). Proc. Natl. Acad. Sri. USA, 88,10495.
40. Jonassen, I., Eidhmmer, I., and Taylor, W. R. (1999). Proteins: Struct. Funct. Genet, 34, 206.
41. Karlin, S. and Zhu, Z.-Y. (1996). Proc. Natl. Acad. Sri. USA, 93, 8344.
42. Taylor, W. R. (1999). Prot. Sci., 8, 654
43. Murzin, A. G., Brenner, S. E., Hubbard, T., and Chothia, C. (1995).J. Mol. Biol., 247, 536.
44. Brazma, A., Jonassen, I., Vilo, J., and Ukkonen, E. (1998). Genome Res., 8,1202.

166
 Chapter 8
Comparison of protein
sequences and practical
database searching
Golan Yona and Steven E. Brenner*
Department of Structural Biology, Stanford University, USA.
* Department of Plant and Microbial Biology, University of California,
Berkeley, USA

1 Introduction
During the last three decades a considerable effort has been made to develop
algorithms that compare sequences of biological macromolecules (proteins, DNA).
The purpose of such algorithms is to detect evolutionary, and thus structural
and functional, relations among sequences. Successful sequence comparison
would allow us to infer the biological properties of new sequences from data
accumulated on related genes. For example, a similarity between a translated
nucleotide sequence and a known protein sequence suggests a homologous
coding region in the corresponding nucleotide sequence. Significant sequence
similarity among proteins may imply that the proteins share the same second-
ary and tertiary structures, and have close biological functions. The prediction of
unknown protein structures is often based on the study of known structures of
homologous proteins.
Today, the routine procedure for analysis of a new protein sequence almost
always starts with a comparison of the sequence to hand with the sequences in
one or more of the main sequence databases. A new sequence is analysed by
extrapolating the properties of its 'Neighbours' in a database search. Such
methods have been applied during the last three decades with much success and
have helped to identify the biological function of many protein sequences, as
well as to reveal many distant and interesting relationships between protein
families. Actually, more sequences have been putatively characterized by data-
base searches than by any other single technology.
Detecting homology may often help in determining the function of new pro-
teins. By definition, homologous proteins have evolved from the same ancestor
protein. The degree of sequence conservation varies among protein families.
Yet, homologous proteins almost always have the same fold (1-3). Although the
common evolutionary origin of two proteins is almost never directly observed,

167
GOLAN YONA AND STEVEN BRENNER

we can deduce homology, with a high statistical confidence, given that the
sequence similarity is significant.
In principle, similarity does not necessarily imply homology (similarity may
be quantified whereas homology is a relation that either holds or does not hold).
Therefore, similarity should be used carefully in attempting to deduce homology.
The deduction of biological function out of sequence similarity is not straight-
forward, and sequence comparison procedures may lead to false conclusions
when applied simple-mindedly. Today sequence comparison algorithms are
accompanied with statistical estimates which provide a measure of statistical
significance of the observed sequence similarities. These estimates can further
help in assessing the significance of the similarity, and in many cases can lead to
deduction of homology. The confidence in the deduction clearly depends on the
level of statistical significance. In this view, database searches should be treated
as experiments analogous to wet-lab characterization. Their use deserves the
same care both in the design of the experiment and in the interpretation of
results.
Planning a good experiment requires understanding of the methods being
applied. Fundamentally, database searches are a simple operation: a query
sequence is aligned with each of the sequences (called targets) in a database. A
score is computed from each alignment, and the query/target pairs with the best
scores are then reported to the user. Statistics are used to help improve the
ability to interpret these scores and distinguish true relations between proteins
from chance similarities. A more detailed description of this process, the
sequence-comparison algorithms, the scoring schemes, and the statistics of
sequence alignments is given next.

2 Alignment of sequences
During evolution, sequences have changed by insertions, deletions, and muta-
tions. These evolutionary events may be traced today by applying algorithms for
sequence alignment. Suppose that a DNA sequence a has evolved to the
sequence b through substitutions, insertions and deletions. This transformation
can be represented by an alignment where a is written above b with the com-
mon (conserved) bases aligned appropriately. For example, say that a = ACTTGA
and b is obtained by substituting the second base from C to G, inserting an A
between the second and the third bases, and by deleting the fifth base (G). The
corresponding alignment will be:
a = A C - T T G A
b = A G A T T - A
score =1 0- 1 1 1-1 1
We usually do not actually know which sequence evolved from the other.
Therefore the events are not directional and insertion of A in b might have been
a deletion of A in a.
In a typical application we are given two related sequences and we wish to

168
 COMPARISON OF PROTEIN SEQUENCES AND PRACTICAL DATABASE SEARCHING

recover the evolutionary events that transformed one to the other. The goal of
sequence alignment is to find the correct alignment that encodes the true series
of evolutionary events that have occurred. The alignment can be assigned a
score which accounts for the number of identities (a match of two identical
letters), the number of substitutions (a match of two different letters), and the
number of gaps (insertions/deletions). For example, in the alignment above, a
score of 1 was given for each identity, a score of 0 was given for each sub-
stitution, and a negative score of -1 was given for each gap. Overall, the align-
ment scored 2, which is the sum of all pair scores and gap scores. In general, the
scores for identities and substitutions which are used to score the alignment are
called the scoring matrix, and the scores for gaps are called gap penalties.
Altogether they are called the scoring scheme (see Section 4.5 for details). With
high (positive) scores for identities, and low (or negative) scores for substitutions
and gaps, the basic strategy towards tracing the correct alignment seeks the
alignment which scores best. In the following sections we describe in detail the
common algorithms for sequence comparison. The discussion focuses on the
comparison of protein sequences, but it holds for DNA sequences as well.

2.1 Rigorous alignment algorithms

There are several different alignment algorithms which have become a standard
tool for biologists. The rigorous algorithms use dynamic programming to find
the optimal alignment.

2.1.1 Global alignment

The first to propose a dynamic programming algorithm for comparison of
macromolecules, were Needleman and Wunsch (4). Their algorithm performs a
global alignment of the sequences; i.e. an alignment where all letters of a and
b are accounted for. This type of alignment is appropriate when similarity is
expected along the whole or most of the sequence.
Formally, let s(ai,bj) be the similarity score of ai,bj (the scoring matrix) and let
a > 0 be the penalty for deleting/inserting of one amino acid. The score of an
alignment with Nij matches of ai and bj and Ngap insertions/deletions is defined as:

In sequence evolution, an insertion or deletion of a segment (several adjacent

amino acids) usually occurs as a single event. That is, the opening of the gap is
the significant event. Therefore, most computational models assign a penalty
for a gap of length k that is smaller than the sum of k independent gaps of length
I, by charging large penalty for opening a gap, and a smaller penalty for each
extension (affine or linear gap penalty). If the penalty for gap of length k is a(k),
and Nk-gap, is the number of gaps of length k in a given alignment, then the score
of this alignment is defined in this case as:

169
GOLAN YONA AND STEVEN BRENNER

The global similarity of sequences a and b is defined as the largest score of

any alignment of sequences a and b, i.e.

In principle, the number of possible alignments is exponentially large, what

makes it impossible to perform a direct search. However, a dynamic program-
ming algorithm makes it possible to find the optimal alignment without check-
ing all possible alignments, but only a very small portion of the search space. In
brief, the idea is that every subalignment in the optimal alignment should be
optimal as well (otherwise it would be possible to improve the overall alignment
by improving the subalignments, in contrast with its definition as optimal align-
ment). Since any optimal subalignment (say, of the substring a1, a2 ... ai,with the
substring bi, b 2 . . . bj) can end only in one of following three ways:

every possible subalignment is calculated only once, and in constant time1, out
of its optimal subalignments.
Formally speaking, denote by Sij the score of the best alignment of the sub-
string ai a2... ai with the substring b1 b2 • • •bj,i.e.

Assume that the gap penalty is constant and equals a. Then, after an initial-
ization step

(where n and m are the lengths of the sequences a and b respectively) define Si,j
recursively

Therefore, the score S(a,b) can be calculated recursively. Since the subalignment
for each i and j has to be calculated, the time complexity of this algorithm is
proportional to the product of the lengths of the sequences compared (a quad-
ratic time complexity). In practice, the scores are stored in a two-dimensional
array of size (n+1). (m+1). The initialization set the values at row zero and
column zero and the computation proceeds row by row so that the value of each
matrix cell is calculated from entries which were already calculated (see Figure 1).

2.1.2 Local alignment

In many cases the similarity of two sequences is limited to a specific motif or
domain, the detection of which may yield valuable structural and functional
insights, while outside of this motif/domain the sequences may be essentially

1
This is true with linear gap functions. With non-linear gap penalties, the calculation of this
optimal subalignment may need up to i+j+1 operations.

170
 COMPARISON OF PROTEIN SEQUENCES AND PRACTICAL DATABASE SEARCHING

Figure 1 Calculating the global similarity score. The score of the (i,j) entry in the matrix is
calculated from three matrix cells: the one on the left, the one on the top, and the one
located at the top left corner of the current cell. In case of a non-constant gap penalty we
need also to check all the cells in the same row and all the cells in the same column (along
the dashed lines).

unrelated. In such cases global alignment may not be the appropriate tool. In
the search for an optimal global alignment, local similarities may be masked by
long unrelated regions. Consequently, the score of such an alignment can be as
low as for totally unrelated sequences. Moreover, the algorithm may even mis-
align the common region. Therefore, usually it is better to compare sequences
locally. A local alignment of a and b is defined as an alignment between a
substring of a and a substring of b. The local similarity of sequences a and b is
defined as the maximal score over all possible local alignments.
The algorithm which finds the best local alignment is based on a minor mod-
ification of the dynamic programming algorithm for global alignment. Specific-
ally, whenever the score of the optimal subalignment of two subsequences
becomes negative, the score is set to zero, meaning that the corresponding
subsequences should not be aligned. Following the notations of the previous
section, Si,j is now defined

In the literature, this algorithm is often called the Smith-Waterman (SW)

algorithm, after those who introduced this modification (5).
There is a lot of literature regarding dynamic programming algorithms in
general (6), and for sequence comparison specifically (7-9). The interested reader
is referred to these books for more details on the algorithmic aspects of this
method, as well as its computational aspects.

2,2 Heuristic algorithms for sequence comparison

In a typical application new protein sequence is compared with all sequences in
the database (library sequences), in search of related proteins. Because of its
quadratic time complexity, the dynamic programming algorithms may not be

171
GOLAN YONA AND STEVEN BRENNER

suitable for this purpose. For example, the comparison of a sequence, of average
length of 350 amino acids, against a typical database (like SWISSPROT (10), with
more than 80000 sequences), may take few CPU hours on a standard PC of
nowadays (Pentium-Ill).
Several algorithms have been developed to speed up the alignment procedure.
The two main algorithms are FASTA (11) and BLAST (12). These are heuristic
algorithms which are not guaranteed to find the optimal alignment. However,
they proved to be very effective for sequence comparison, and they are signifi-
cantly faster than the rigorous dynamic programming algorithm.2

2.2.1 BLAST (Basic Local Alignment Search Tool)

BLAST compares two sequences and seeks all pairs of similar segments, whose
similarity score exceeds a certain threshold. These pairs of segments are called
'high scoring segment pairs' (HSPs). A segment is always a contiguous sub-
sequence of one of the two sequences. Segment pair is a pair of segments of the
same length, one from each sequence. Hence the alignment of the segments is
without gaps. The score of the match is simply the sum of matches of the amino
acids (defined by a scoring matrix) along the segment pair. The segment pair
with the highest score is called the 'maximum segment pair' (MSP).
To identify the HSPs (and particularly, the MSP), the algorithm starts by
locating 'seeds' of similarity among the query sequence and the database
sequence that score at least T, and then extends them in both directions until
the maximum possible score for the extension is reached. The changes in the
threshold T permit a tradeoff between speed and sensitivity. A higher value of T
yields greater speed, but also an increased probability of missing weak similarities.
Finally, multiple MSP regions are combined. For each consistent combination,
its probability is calculated using the Poisson or sum statistics (14) and the most
significant hits (lowest probability) are reported.
The algorithm is an outgrowth of the statistical theory for local alignments
without gaps (see Section 3). This theory gives a framework for assessing the
probability that a given similarity between two protein sequences (i.e. the MSP)
could have emerged by chance. If the probability is very low, then the similarity
is statistically significant and the algorithm reports the similarity along with
its statistical significance. Though the algorithm may miss complex similari-
ties which include gaps, the statistical theory of alignments without gaps
provided a reliable and efficient way of distinguishing true homologies from

2
In the last few years, biotechnology companies such as Compugen and Paracel, have developed
special purpose hardware that accelerates the dynamic programming algorithm (13). This special-
purpose hardware has again made the dynamic programming algorithm competitive with
FASTA and BLAST, both in speed and in simplicity of use. However, meanwhile, FASTA and
BLAST have become standard in this field and are being used extensively by biologists all over
the world. Both algorithms are fast, effective, and do not require the purchase of additional hard-
ware. BLAST has an additional advantage, as it may reveal similarities which are missed by the
dynamic programming algorithm, for example when two similar regions are separated by a long
dissimilar region.

172
 COMPARISON OF PROTEIN SEQUENCES AND PRACTICAL DATABASE SEARCHING

chance similarities, thus making this algorithm an important tool for molecular
biologists.
Current improvements of BLAST allow gapped alignments, by using dynamic
programming to extend a central seed in both directions (15). This is com-
plemented by PSI-BLAST, an iterative version of BLAST, with a position-specific
score matrix (see Section 4.5) that is generated from significant alignments found
in round i and used in round i+1. The latter may better detect weak similarities
that are missed in database searches with a simple sequence query.

2.2.2 FASTA
FASTA is another heuristic that performs a fast sequence comparison. The
algorithm starts by creating a hash table of all k-tuples (a string of length k) in
the query sequence (usually, k = 1 or 2 for protein sequences, where k = 1 gives
higher sensitivity). This table stores the k-tuples in a way which enables fast
accession, and restoration of each k-tuple. Then, when scanning a library
sequence, each k-tuple of the library sequence is looked up in the hash table,
and if it is found (this means k-tuple identity) it is marked. At a second stage, the
ten regions with the highest density of identities are rescanned. Common k-
tuples which are on the same diagonal (same offset in both sequences), and not
very far apart (the exact parameters are set heuristically), are joined to form a
region (a gapless local alignment, or HSP in BLAST terminology). The regions are
scored to account for the matches as well as the mismatches, and the best region
is reported (its score is termed 'initial score' or 'initl'). Then, the algorithm tries
to join nearby high scoring regions, even if they are not on the same diagonal
(the corresponding score being termed 'initn score'). Finally, a bounded dynamic
programming is run in a band around the best region, to obtain the 'optimized
score'. If the sequences are related then the optimized score is usually much
higher than the initial score.

3 Probability and statistics of sequence alignments

In the evolution of protein sequences, not all regions mutate at the same rate.
Regions which are essential for the structure and function of proteins, are more
conserved. Therefore, significant sequence similarity of two proteins may reflect
a close biological function or a common evolutionary origin. The algorithms
that were described in the previous section can be used to identify such simi-
larities. However, on any two input protein sequences, even if totally unrelated,
the algorithms almost always find some similarity. For unrelated sequences this
similarity is essentially random. As the length of the sequences compared in-
crease, this random similarity may increase as well. Therefore, in order to assess
the significance of a similarity score it is important to know what score to
expect simply by chance.
Naturally, we would like to identify those similarities which are genuine, and
biologically meaningful. In the view of the last paragraph, the raw similarity
score may not be appropriate for this purpose. However, when the sequence

173
GOLAN YONA AND STEVEN BRENNER

similarity is statistically significant we can deduce, with high confidence level,

that the sequences are related.3 The reverse implication is not always true. We
encounter many examples of low sequence similarity despite functional and
structural similarity (16-18).
Though statistically significant similarity is neither necessary nor sufficient
for a biological relationship, it may give us a good indication of such relation-
ship. When comparing a new sequence against the database, in search of close
relatives, this is extremely important, as we are interested in reporting only sig-
nificant hits, and sorting the results according to statistical significance seems
reasonable.
To estimate the statistical significance of similarity scores, a statistical theory
should be developed. A great effort was made in the last two decades to establish
such statistical theory. Currently, there is no complete theory, though some
important results were obtained. These results have very practical implications
and are very useful for estimating the statistical significance of similarity scores.
The statistical significance of similarity scores for 'real' sequences is defined by the
probability that the same score would have been obtained for random sequences.
The statistical results concern the similarity scores of random sequences, when
the similarity scores are defined by ungapped alignments. However, these results
have created a framework for assessing the statistical significance of various
similarity scores, including gapped sequence alignments, and recently, even
structural alignments (19).
Readers who are primarily interested in practical applications (rather than
the statistical basis) of the methods might like to proceed to Section 4.

3.1 Statistics of global alignment

Though the distribution of global similarity scores of random sequences has not
been characterized yet, some important properties of this distribution were
partly determined. The main characteristic of this distribution is the linear
growth (or decline, depends on the mean of the scoring matrix) with the
sequence length, i.e. the expected global similarity score grows linearly with
the sequence length. However, the growth rate has not been determined.
The statistical significance of a similarity score obtained for 'real' sequences,
which exceeds the expected score by a certain amount, is estimated by the prob-
ability that the similarity score of random sequences would exceed the expected
mean by the same amount. However, since the distribution of scores is un-
known, the available estimates give only a rough bound on that probability. The
variance of the global similarity score has not been determined either, and the
best results give only an upper bound.
In practice, it is possible to approximate the distribution empirically by
shuffling the sequences and comparing the shuffled sequences. By repeating

3
Two exceptions are segments with unusual amino acid composition, and similarity that is due
to convergent evolution.

174
 COMPARISON OF PROTEIN SEQUENCES AND PRACTICAL DATABASE SEARCHING

this procedure many times it is possible to estimate the mean and the variance
of the distribution, and a reasonable measure of statistical significance (e.g. by
means of the z-score) can be obtained. Formally, denote by S the global simi-
larity score. Let ul and o2 be the mean and the variance of the distribution of
scores. Then, the z-score associated with the score S is defined as

This score measures how many units of standard deviation apart the score S is
from the mean of the distribution. The larger it is, the more significant is the
score S.

3.2 Statistics of local alignment without gaps

The statistics of ungapped similarity scores has been studied extensively since
the early 1990s. The exclusion of gaps allowed a rigorous mathematical treat-
ment, and several important results were obtained. Karlin and Altschul (21) have
shown that for two random sequences of length n and m, the score of the best
ungapped local alignment (the MSP score in BLAST jargon) is centred around

where X is a parameter that depends on the overall background distribution

of amino acids in the database, and the scoring matrix. That is, the local simi-
larity score grows logarithmically with the length of the sequences, and with
the size of the search space (n-m).
This result in itself is still not enough to obtain a measure of statistical signi-
ficance for local similarity scores. This can be done only once a concentration of
measure result is obtained or the distribution of similarity scores is defined.
Indeed, one of the most important results in this field is the characterization of
the distribution of local similarity scores without gaps. This distribution was
shown to follow the extreme value distribution (20-22).
Formally, as the sum of many random variables is distributed normally,
then the maximum of many random variables is distributed as an extreme
value distribution (23). This distribution is characterized by two parameters:
the index value u and the decay constant X. (for u = 0 and A, = 1, the distribution is
plotted in Figure 2). The distribution is not symmetrical. It is positive definite and
unimodal with one peak at u. Practically, the score of the best local alignment
(the MSP score) is the maximum of the scores of many independent alignments,
which explains the observed distribution. Specifically, S, the local similarity
score of two random sequences of length n and m, is distributed as an extreme
value distribution and

where u = (In Kmn)/X, and K is a constant that can be estimated from the back-
ground distribution and the scoring matrix (Karlin & Altschul 1990).

175
GOLAN YONA AND STEVEN BRENNER

Figure 2 Probability density function for the extreme value distribution with u = 0 and X = 1.

For a large x we can use the approximate 1 - exp(-e -x ) ~ e - x .Therefore, for a

large x,

This result helps to calculate the probability that a given MSP score could have
been obtained by chance. The score will be statistically significant at the 1% level
if S > x0 where x0 is determined by the equation Kmne -x *° = 0.01. In general, a
pairwise alignment with score S has a p-value of p where p = Kmne~*'s. I.e., there
is a probability p that this score could have happened by chance.
The probability p, that a similarity score S could have been obtained simply
by chance from the comparison of two random sequences, should be adjusted
when multiple comparisons are performed. One example of this is when a
sequence is compared with each of the sequences in a database with D
sequences. Denote by p-match a match between two sequences that has a p-
value < o (i.e. its score > S). The probability P of observing at least one p-match
(i.e. at least one 'success'), in a database search follows the Poisson distribution

and for Dp < 0.1 is well approximated by

Since not all library sequences have the same probability of sharing a similar
region with the query sequence, D should be replaced with the effective size of
the database. If the query sequence is of length n, and the (pairwise) alignment
of interest involves a libaiy segment of length m, and the database has a total of
N amino acids, then D should be replaced with N/m. Thus,

so the effective size of the search space is Nn (intuitively, this is the number of
possible starting positions of a match).

176
 COMPARISON OF PROTEIN SEQUENCES AND PRACTICAL DATABASE SEARCHING

It is very common to use the expectation value (e-value) as a measure of

statistical significance. The expectation value of the Poisson distribution is given
by

and as discussed above, D should be replaced with N/m. Hence

This is the expected number of distinct matches (segment pairs) that would
obtain a score > S by chance in a database search, with a database of size N
(amino acids) and composition P (the background distribution of amino acids).
The higher it is, the match is less significant. For example, if £ = 0.01, then the
expected number of random hits with a score > S is 0.01. In other words, we
may expect a random hit with that score only once in 100 independent searches.
If E = 10, then we should expect 10 hits with a score > S by chance, in a single
database search. This means that such a hit is not significant. (Note that
E = P for P< 0.1.)
Finally, by setting a value for £ and solving the equation above for S, it is
possible to define a threshold score, above which hits are reported. This is the
score above which the number of hits that are expected to occur at random is
< E. Therefore, we can deduce that a match with this score or above reflects true
biological relationship, but we should expect up to £ errors per search. The
specific value of £ affects both the sensitivity of a search (the number of true
relationships detected) and its selectivity (the number of errors). A lower value
of E would decrease the error rate. However, it would decrease the sensitivity as
well. A reasonable choice for E is between 0.1 and 0.001.

3.3 Statistics of local alignment with gaps

Though local alignments without gaps may detect most similarities between
related proteins, and give a good estimation of the similarity of the two
sequences, it is clear that gaps in local alignments are crucial in order to obtain
the correct alignment, and for a more accurate measure of similarity. However,
no precise model has been proposed yet to explain gaps in alignments. Moreover,
introducing gaps in alignments greatly complicates their mathematical tract-
ability. Rigorous results have been obtained only for local alignments without
gaps.
Recent studies suggest that the score of local gapped alignments can be
characterized in the same manner as the score of local ungapped alignments: As
was mentioned in the previous section, the local ungapped similarity score grows
logarithmically with the sequence's length and the size of the search space.
Arratia and Waterman (24) have shown that for a range of substitution matrices
and gap penalties, local gapped similarity scores have the same asymptotic
characteristic. Furthermore, empirical studies (25, 26) strongly suggest that local
gapped similarity scores are distributed according to the extreme value dis-
tribution, though some correction factors may apply (27). Based on empirical

177
GOLAN YONA AND STEVEN BRENNER

observations, Pearson (28) has derived statistical estimates for local alignment
with gaps, using the extreme value distribution for scores obtained from a
database search. A database search provides tens of thousands of scores from
sequences which are unrelated to the query sequence, and therefore are effect-
ively random. As discussed above, these scores are thus expected to follow the
extreme-value distribution. This is true as long as the gap penalties are not too
low. Otherwise the alignments shift from local to global and the extreme value
distribution no longer apply.
Since the logarithmic growth in the sequence length holds in this case, scores
are corrected first for the expected effect of sequence length. The correction is
done by calculating the regression line S = a + b .In n for the scores obtained in
a database search, after removing very high scoring sequences (probably related
sequences). The process is repeated as many as five times. The regression line
and the average variance of the normalized scores are used to define the z-
score:

and the distribution of z-scores is approximated by the extreme value distribution

where c1 and c2 are constants, and the expectation value is defined as before by
£(z-score> x) = N • p where N is the number of sequences in the database (the
number of tests).
This empirical approach has the advantage of internal calibration of the
accuracy of the estimates, and has proved to be very accurate in estimating the
statistical significance of gapped similarity scores (28). (See also refs. 18 and 29.)

4 Practical database searching

4.1 Types of comparison
To formulate the database search 'experiment', it is first necessary to decide
what types of sequences will be compared: DNA, Protein, or DNA as Protein. The
algorithms described above may be applied to the comparison of protein
sequences as well as to DNA sequences (coding or non-coding regions). However,
the comparison of protein sequences has proven to be a much more effective
tool (3). Though the evolutionary events occur at the DNA level, the main
genetic pressure is on the protein sequence. Moreover, mutations at the DNA
level do not necessarily change the encoded amino acid due to the redundancy
of the genetic code. Mutations often result in conservative substitutions at the
protein level, namely, replacement of an amino acid by another amino acid with
similar biochemical properties. Such changes tend to have only a minor effect
on the protein's functionality. Therefore, if the sequence under consideration
either is a protein or codes for a protein, then it is almost always the case that

178
 COMPARISON OF PROTEIN SEQUENCES AND PRACTICAL DATABASE SEARCHING

Table 1 Comparison programs and types of comparison

Programs Query DB Comparison Common Use

blastn, fasta, ssearch DNA DNA DNA-level Seek identical DNA sequences,
and splicing patterns
blastp, fasta, ssearch Protein Protein Protein-level Seek homologous proteins
blastx, fastx DNA Protein Protein-level Query new DNA to find genes and seek
homologous proteins
tblastn, tfasta, tfastx Protein DNA Protein-level Search for genes in un-annotated DNA
tblastx DNA DNA Protein-level Discover gene structure

the search should take place at the protein level, as proteins allow one to detect
far more distant homologies than DNA. Another aspect is that in DNA com-
parisons, there is noise from comparisons of non-coding frames (though this
latter issue still arises in DNA as Protein searches). DNA versus DNA comparison
is typically only used to find identical regions of sequence in a database. One
would do such a search to discover whether another group has sequenced or
studied a gene, and to learn where it is expressed or where splice junctions occur.
In short, protein-level searches are valuable for detecting evolutionarily related
genes, while DNA searches are best for locating nearly identical regions of
sequence (see Table 1 for available comparison programs and the corresponding
types of comparison).
4.2 Databases
Next, it is necessary to select a database to search against. There are several
commonly used databases (e.g. GenBank, SwissProt, ESTs, etc.). For homology
searches, it is best to use a comprehensive collection of all known proteins. Two
such databases are available. One is the nr database at the NCBI website
(http://www.ncbi.nlm.nih.gov/). The nr (which stands for non-redundant) protein
database combines data from several sources (GenPept, SwissProt, PIR, RPF, and
PDB) removes the redundant identical sequences, and yields a collection with
nearly all known proteins. The second nr database is available at the ExPASy
website in Switzerland (http://www.expasy.ch/). Both databases are frequently
updated, to incorporate as many sequences as possible. Obviously, a search will
not identify a sequence that has not been included in the database, and since
databases are growing so rapidly, it is essential to use a current database.
The main sources of these non-redundant databases are the SwissProt database
and the TrEMBL database (10), the PIR database (30), and the GenPept database
(31). The SwissProt database is maintained at the ExPASy centre in Switzerland.
This is a non-redundant highly annotated database which offers a lot of valuable
biological information on almost all of its entries (more than 86000 in the latest
release, June 2000). Such information may include for example the description
of the function of a protein, its domain structure, post-translational modifica-
tions, etc. This database is supplemented by TrEMBL, which is a collection of
all the translations of EMBL nucleotide sequence entries not yet integrated in
SwissProt. For most of these entries some biological information is available,

179
GOLAN YONA AND STEVEN BRENNER

Table 2 Sequence databases. Number of entries is updated to June 2000.

Protein Number Availability Description

Database Entries
nr (ExPasy 385002 ftp://www.expasy.ch/databases/ consists of SwissProt, TrEMBL
sp_tr_nrdb/
nr(NCBI) 508388 ftp://ncbi.nlm.nih.gov/blast/db/ consists of GenPept, SwissProt,
PIR, RPF, PDB
SwissProt 86337 http://www.expasy.ch/sprot/ non-redundant, high level of
sprot-top.html annotation
TrEmbl 298665 http://www.expasy.ch/sprot/ non-redundant, computer
sprot-top.html annotated
PIR 180605 http://www-nbrf.georgetown.edu/ non-redundant, annotated, family
pirwww/pirhome.shtml classification
GenPept 544510 http://www.ncbi.nlm.nih.gov/ translation of DNA sequences in
Entrez/protein.html GenBank
PDB 12426 http://www.rcsb.org/pdb/index. repository of all known 3D
html structures
Genomes http://www.ncbi.nlm.nih.gov/ protein sequences sorted by
Entrez/Genome/org.html organism

DNA Number Availability Description

Database Entries
GenBank 5691170 http://www.ncbi.nlm.nih.gov/ DNA sequence
Entrez/nucleotide.html
EMBL 5865742 http://www.ebi.ac.uk.embl/ DNA sequence
DDBJ 5962608 http://www.ddbj.nig.ac.jp/ DNA sequence

usually based on sequence analysis carried by the ExPASy team. PIR is another
database that offers a lot of biological information on entries through an exten-
sive annotation as well as classification to families and superfamilies and links
to alignments with other family members. GenPept is a database that contains
all translations of DNA sequences in the GenBank database.
Several specialized databases are also available, all of which overlap with the
composite non-redundant databases. For example, if one is interested in search-
ing for proteins of known structure, it is best to just search the smaller PDB
database. Other specialized databases are available for each of the fully sequenced
genomes, as well as for subsets of protein families (such as protein kinases
or immunoglobulins), etc. See Table 2 for a list of the main databases (see also
Chapters 9 and 10).
One may also wish to search DNA databases at the protein level. Programs
can do so automatically by first translating the DNA in all six reading frames and
then making comparisons with each of these conceptual translations. The nr
DNA database (containing most known DNA sequence except GSS, EST, STS, or
HTGS sequences) is useful to search when hunting new genes; the identified
genes in this database would already be in the protein nr database. Searches
against the GSS, EST, STS, and HTGS databases can find new homologous genes,
and are especially useful to learn about expression data or genome map location.

180
 COMPARISON OF PROTEIN SEQUENCES AND PRACTICAL DATABASE SEARCHING

Table 3 Availability of sequence comparison programs

Program FTP site Run over the Web

ssearch ftp://ftp.virginia.edu/pub/fasta/ http://www2.ebi.ac.uk/bic_sw/
http://genome.dkfz-heidelberg.de/genweb/
http://sgbcd.weizmann.ac.il/genweb/
http://www.ch.embnet.org/software/FDF_form.
html
fasta ftp://ftp.virginia.edu/pub/fasta/ http://www2.ebi.ac.uk/fasta3/
blastp ftp://ncbi.nlm.nih.gov/blast/ http://www.ncbi.nlm.nih.gov/BLAST/
http://www2.ebi.ac.uk/blastall/
http://www.ch.embnet.org/software/
BottomBLAST.html

4.3 Algorithms
The choice of the comparison algorithm should be based on the desired com-
parison type, the available computational resources, and the goals of the search.
All standard comparison algorithms can be run over the Web and can be down-
loaded from the FTP site to run locally (see Table 3). The rigorous Smith-
Waterman algorithm is available, as well as the FASTA program, within the
FASTA package. This algorithm is more sensitive than the others, but it is also
much slower. The FASTA program is faster, and with the parameter ktup set to 1,
is almost as sensitive as the Smith-Waterman algorithm (32, 18). The fastest
algorithm is BLAST, the newest versions of which support gapped alignments
(15) and provide a reliable, sensitive and fast option (the older versions are slower,
detect fewer homologs, and have problems with some statistics). Iterative pro-
grams like PSI-BLAST require extreme care in their operation, as they can provide
very misleading results; however, they have the potential to find more homologs
than purely pairwise methods.

4.4 Filtering
The statistics for database searches assumes that unrelated sequences look
essentially random with respect to each other. Specifically, the theoretical re-
sults that were obtained for the statistics of local alignments without gaps (see
Section 3.2) are subject to the restriction that the amino acid composition of the
two sequences that are compared are not too dissimilar (20). Assuming that both
sequences are drawn from the background distribution, the amino acid
composition of both should resemble the background distribution. Without this
restriction the statistical estimates overestimates the probability of similarity
scores, and indeed, this is observed in protein sequences with unusual com-
positions (18, 29). The most common exceptions are long runs of a small number of
different residues (such as a poly-alanine tract). Such regions of a sequence may
spuriously obtain extremely high match scores. For this reason, it is recom-
mended to filter out these regions using programs such as SEG (33). The NCBI

181
GOLAN YONA AND STEVEN BRENNER

BLAST server will automatically remove such sections in proteins, replacing

them with X, if default filtering is selected. DNA sequences will be similarly
masked by DUST. Though these programs automatically remove the majority of
problematic matches, some problems invariably slip through; moreover, valid
hits may be missed due to masking of part of the sequence. Therefore, it may be
helpful to try using different masking parameters.
Other sorts of filtering are also often desirable; for example, iterative searches
are prone to contamination by regions of proteins that resemble coiled-coils or
transmembrane helices. Here, one protein that is similar only because it has the
general characteristics may match initially. The profile then emphasizes these
inappropriate characteristics, eventually causing many spurious hits. Heavily
cysteine rich proteins can also obtain anomalous high scores. If these charac-
teristics are not filtered, then it is necessary to carefully review the alignment
results to ensure that they have not led to incorrect matches.

4.5 Scoring matrices and gap penalties

The next step is to choose the set of parameters for the sequence comparison
algorithm. Namely, the scoring matrix and the gap parameters. The default
matrices offered with the comparison algorithm (e.g. BLOSUM62 with BLAST,
BLOSUM50 with FASTA) are a safe choice. However, it may be fruitful to check
other matrices as well. Several different approaches were taken to derive re-
liable and effective scoring matrices. The most effective matrices are those that
are based on actual frequencies of mutations that are observed in closely related
proteins. These matrices reflect the biochemical properties of the amino acids,
which influence the probability of mutual substitution (exchange occur more
frequently among amino acids that share certain properties), and amino acids
with similar properties have high pairwise score. Matrices which are based on
sequence alignments include the family of PAM matrices (34) (and their im-
provement by ref. 35), the BLOSUM matrices (36), and Gonnet matrix (37). Other
matrices, which proved to be very effective for protein sequence comparison,
are those that are based on structural principles and aligned structures (38, 39).
The two most extensively used families of scoring matrices are the PAM
matrices and the BLOSUM matrices. A detailed description of these matrices is
given in the next two sections.
4.5.1 The PAM family of scoring matrices
PAM matrices were proposed by Dayhoff et al. in 1978 based on observations of
hundreds of alignments of closely related proteins. The frequencies of sub-
stitution of each pair of amino acids were extracted from alignments of proteins
of small evolutionary distance, below 1% divergence, i.e. at most one mutation
per 100 amino acids, on average. These frequencies, normalized to account for
the frequencies of random occurrences of single amino acids, resulted in the
PAM-1 probability transition matrix. The PAM-1 matrix reflects an amount of
evolutionary change that yields on average one mutation per 100 amino acids.
Accordingly, it is suitable for comparison of proteins which have diverged by 1%

182
 COMPARISON OF PROTEIN SEQUENCES AND PRACTICAL DATABASE SEARCHING

or less. The acronym PAM stands for Percent of Accepted Mutations (and hence
the distance is in percentages) or for Point Accepted Mutations (and hence the
distance in number of mutations per 100 amino acids).
The PAM-1 matrix is then extrapolated to yield the family of PAM-k matrices.
Each PAM-k matrix is obtained from PAM-1 by k consecutive multiplication, and
is suitable for comparison of sequences which have diverged k%, or are k
evolutionary units apart. For example, PAM-250 = (PAM-1)250 reflects the fre-
quencies of mutations for proteins which have diverged 250% (250 mutations
per 100 amino acids). The actual scoring matrices that are used by search pro-
grams are derived from the transition probability matrices and the background
probabilities. The score of each pair s(a,b) is defined as the logarithm of the
likelihood ratio of the transition probability Mab (mutation) versus the prob-
ability of a random occurrence of the amino acid b in the second sequence, i.e.,
s(a,b) = logMab/pb.
The PAM matrices were later refined by Jones et al. (35) based on much larger
data set. The significant differences were detected for substitions that were
hardly observed in the original data set of (34).
The PAM-250 matrix. The PAM-250 matrix is one of the most extensively
used matrices in this field. This matrix corresponds to a divergence of 250
mutations per 100 amino acids. Naturally one may ask whether it makes sense
to compare sequences which have diverged this much. Surprising as it may
seem, when calculating the probability that a sequence remains unchanged
after 250 PAMs (this is given by the sum EapaMaa where pa is the probability of a
random occurrence of amino acid s and Maa is the diagonal entry in the PAM-250
matrix that corrresponds to the amino acid a) the outcome is that such
sequences are expected to share about 20% of their amino acids. For reference,
note that the expected percentage of identity in a random match is 100Eap2a,
and for a typical distribution of amino acids (in a large ensemble of protein
sequences), we should expect less than 6% identies.

4.5.2 The BLOSUM family of scoring matrices

Unlike PAM matrices, which are extrapolated from a single matrix PAM-1, the
BLOSUM series of matrices was constructed by direct observation of sequence
alignments of related proteins, at different levels of sequence divergence. The
matrices are based on 'blocks'—a collection of multiple alignments of similar
segments without gaps (40), each block representing a conserved region of a
protein family. These blocks provide a list of (accepted) substitutions, and a log-
odds scoring matrix can be defined based on the observed relative frequency
of aligned pairs of amino acids qab, and the expected probability of pairs eab
estimated from the population of all observed pairs

To reduce the bias in the amino acid pair frequencies caused by multiple counts
from closely related sequences, segments in a block with at least x% identity are

183
GOLAN YONA AND STEVEN BRENNER

clustered and pairs are counted between clusters, i.e., pairs are counted only be-
tween segments less than x% identical. When counting pairs frequencies between
clusters, the contributions of all segments within a cluster are averaged, so that
each cluster is weighted as a single sequence. Varying the percentage of identity
x within clusters results in a family of matrices BLOSUM-x, where x ranges from
30 to 100. For example, BLOSUM-62 is based on pairs that were counted only
between segments less than 62% identical.

4.5.3 Choosing the scoring matrix

When comparing two sequences, the most effective matrix to use is the one
which corresponds to the evolutionary distance between them (41). However,
we usually do not know this distance. Therefore, it is recommended to use
several scoring matrices which cover a range of evolutionary distances, for
example PAM-40, PAM-120, and PAM-250. In general, low PAM matrices are well
suited to finding short but strong similarities, while high PAM matrices are best
for finding long regions of weak similarity.
Exhaustive evaluations have been carried out to compare the performance of
different scoring matrices (42, 32). These studies show that log-odds matrices
derived directly from alignments of highly conserved regions of proteins (such as
BLOSUM matrices or the Overington matrix, which is based on structural align-
ment (39)) outperform extrapolated log-odds matrices based on an evolutionary
model, such as PAM matrices. Moreover, the accuracy of alignments based on
extrapolated matrices decreases as the evolutionary distance increases. This sug-
gests that extrapolation cannot accurately model distant relationships, and that
the PAM evolutionary model is inadequate. BLOSUM matrices were shown to be
more effective in detecting homologous proteins. Specifically, BLOSUM-62 and
BLOSUM-50 gave superior performance in detecting weak homologies. These
matrices offer good overall performance in searching the databases. The best
hybrid of matrices for searching in different evolutionary ranges is either
BLOSUM 45/62/100 or BLOSUM 45/100 plus the Overington matrix.

4.5.4 Gap penalties

There is no mathematical model to explain the evolution of gaps. Practical
considerations (the need for a simple mathematical model, time complexity)
have led to the broad use of linear gap functions, where the penalty for a gap of
length k is given by a(k) = a0 + k, a1. Usually a large penalty is charged for
opening a gap (a0), and a smaller penalty is charged for each extension (o 1 ).
Gonnet et al. (37) have proposed a model for gaps that is based on gaps
occurring in pairwise alignments of related proteins. The model suggests an
exponentially decreasing gap penalty function. However, a linear penalty func-
tion has the advantage of better time complexity, and in most cases the results
are satisfactory. Therefore the use of linear gap functions is very common.
The gap parameters that are used as default in the standard comparison pro-
grams are usually optimized based on extensive evaluations (32), and it is rarely
beneficial to change these from their defaults.

184
 COMPARISON OF PROTEIN SEQUENCES AND PRACTICAL DATABASE SEARCHING

4.5.5 Position dependent scores

In many proteins, mutations are not equally probable along the sequence. Some
regions are functionally/structurally important and consequently, the effect of
mutation in these regions can be drastic. They may create a non-functional pro-
tein or even prevent the molecule from folding into its native structure. Such
mutations are unlikely to survive, and therefore these regions tend to be more
evolutionary conserved than other, less constrained regions (e.g. loops) which
can significantly diverge.
Accordingly, it may be appropriate to use position-dependent scores for mis-
matches and gaps. The incorporation of information about structural preferences
can lead to alignments that are more accurate biologically. If a protein's structure
is known, the secondary structure should be taken into account. In the absence
of such data, general structural criteria, such as the propensities of amino acid
for occurring in secondary structures versus loops can be taken into account. For
example, the probability of opening a gap in existing secondary structure can be
decreased, while the probability for opening/inserting a gap in loop regions can
be increased.
Usually position-specific scoring matrices, or profiles, are not tailored to a
specific sequence. Rather, they are built to utilize the information in a group of
related sequences, and provide representations of protein families and domains.
These representations are capable of detecting subtle similarities between dis-
tantly related proteins. Without going into detail, profiles are usually obtained
by applying algorithms for multiple alignment (i.e. a combined alignment of
several proteins) to align a group of related sequences. The frequency of each
amino acid at each position along the multiple alignment is then calculated.
These counts are normalized and transformed to probabilities, so that a prob-
ability distribution over amino acids is associated with each position. Finally, the
scoring matrix is defined based on these probability distributions as well as on
the similarities of pairs of amino acids (taken from a standard scoring matrix).
For example, the score for aligning the amino acid a at position i of the profile is
given by

where s(a,b) is the similarity of amino acids a and b according to some scoring
matrix. For a review on algorithms for multiple alignment and profile techniques,
see refs 7-9, 43, 44.

4.6 Command line parameters

The command line parameters of the search programs are generally divided into
three groups. The first group is the set of parameters which specify the input
and output filenames, and the database name. These are the only mandatory
parameters. All other parameters are optional and are set default values other-
wise. For example, the basic command line for SSEARCH, FASTA, BLAST, and
gapped BLAST are:

185
GOLAN YONA AND STEVEN BRENNER

ssearch -Q, query-file -O out-file database

fasta -Q, query-file -O out-file database
blastp database query-file
blastpgp -i query-file -o out-file -d database
The second set of parameters affects the comparison algorithm. This set in-
cludes the scoring matrix and the gap penalties and the parameters used to
control the sensitivity of the search. By altering the later, it is possible to make
the program run slower and be more sensitive, or to run faster at the cost of
missing more homologs. BLAST has few such parameters. Currently, it is very
rare for users to alter these options from the defaults. The FASTA program has
one such parameter that a user will often want to set, called letup. Searches with
ktup = 1 are slower, but are more sensitive than BLAST;ktup = 2 is faster but less
effective.

Table 4 Parameters for sequence comparison programs. PSI-BLAST and gapped BLAST are
executed by the same program (blastpgp). The default mode is a simple gapped BLAST (i.e.,
the parameter j is set to 1).

Program Parameter Use

search -Q filename query file
-0 filename output file
-E evalue evalue threshold (only hits with evalue below this threshold are
reported)
-d number maximal number of alignments displayed
-H suppresses histogram of scores
fasta -Q filename query file
-0 filename output file
-E evalue evalue threshold (only hits with evalue below this threshold are
reported)
-d number maximal number of alignments displayed
-H suppresses histogram of scores
ktup number controls sensitivity (can be either 1 or 2 for proteins and up to 4 for
DNA)
blastp E=evalue evalue threshold (only hits with evalue below this threshold are
reported)
V=number maximal number of hits reported
B=number maximal number of alignments displayed
H=1 display histogram of scores
blastpgp -d databasethe database searched
-i filenamequery file
-o filenameout file
-e evalue evalue threshold (only hits with evalue below this threshold are
reported)
-v number maximal number of hits reported
-b number maximal number of alignments displayed
-j number maximal number of iterations (PSI-BLAST)
-C filename saves a checkpoint profile in a file after each iteration (PSI-BLAST)
-R filename reads the initial profile from a file (PSI-BLAST)
-h evalue evalue threshold for inclusion in a profile (PSI-BLAST)

186
 COMPARISON OF PROTEIN SEQUENCES AND PRACTICAL DATABASE SEARCHING

Finally, there is a third set of parameters which controls the output of the
program, e.g. how many results are reported, and how many alignments are
displayed. The number of hits reported is often controlled by the e-value
parameter (see Section 3.2). For example, by default, the BLAST programs will
report only matches with an e-value up to 10 (this parameter also affects the
sensitivity of the method, in an indirect manner). The total number of matches
is limited to the best 500, and detailed information with the alignment is
provided for up to 100 pairs. To retrieve more matches, these numbers can be
altered (see Table 4).

5 Interpretation of results
Interpretation of the results of a sequence database search involves first eval-
uating the matches, to determine whether they are significant and therefore
imply homology. The most effective way of doing so is through use of the
statistical scores (the e-values). The e-values are more useful than the raw or bit
scores, and they are far more powerful than percentage identity (which is best
not even considered unless the identity is very high) (18). Fortunately, the e-
values from FASTA, SSEARCH, and gapped BLAST seem to be accurate and are
therefore easy to interpret (18, 29).
The e-value (or expectation-value) of a match should measure the expected
number of sequences in the database which would achieve a given score. There-
fore, in the average database search, one expects to find ten random matches
with e-value score of 10; obviously, such matches are not significant. However,
lacking better matches, sequences with these scores may provide hints of func-
tion or suggest new experiments. Scores below 0.01 would occur by chance only
very rarely, and are therefore likely to indicate homology, unless biased in some
way. Scores of near le-50 are now seen frequently, and these offer extremely
high confidence that the query protein is evolutionarily related to the matched
target in the database.
Inferring function from the homologous matched sequences is a process still
fraught with difficulty. If the score is extremely good and the alignment covers
the whole of both proteins, then there is a good chance that they will share the
same or a related function. However, is dangerous to place too much trust in the
query having the same function as the matched protein: functions do diverge,
and organismal or cellular roles may alter even when biochemical function is
unchanged. Moreover, a significant fraction of functional annotations in data-
bases are wrong (45), so one needs to be suspicious. There are other complexi-
ties; for example, if only a portion of the proteins align, they may share a
domain which only contributes an aspect of the overall function. It is often the
case that all of the highest-scoring hits align to one region of the query, and
matches to other regions need to be sought much lower in the score ranking.
For this reason, it is necessary to consider carefully the overlap between the
query and each of the targets.
Database search methods are also limited because most homologous sequences

187
GOLAN YONA AND STEVEN BRENNER

have diverged too far to be detected by pairwise sequence comparison methods

(16-18). Thus, failure to find a significant match does not necessarily indicate
that no homologs exists in the database. In such cases more sophisticated
methods must be applied. For example, iterative search programs such as the
profile based PSI-BLAST program (15) or the HMM based SAM-T98 (46) are ad-
vanced and sensitive search tools. However, these programs should be carefully
used as they can lead to false positives by diverging from the original query
sequence, and creating a profile that represents unrelated sequences.
The most powerful tools today are those that incorporate information from a
group of related sequences. This strategy has led to the compilation of databases
of protein families and domains. These databases have become an important
tool in the analysis of newly discovered protein sequences. They usually offer
biologically valuable information about domains and the domain structure of
proteins, through multiple alignments and schematic representations of
proteins, and can help to detect weak relationships between remote homologs.
Such methods are described in chapters 3, 4, and 5. However, family databases
are limited because they typically contain less than half of the proteins in
sequence databases. Moreover, many families have not yet been characterized,
while others are currently too sparse to yield reliable models. For this reason,
database searches are crucial in the analysis of newly sequenced genes that have
no clear homologs with known families, and by integrating the information
obtained from a database search one may discover clues about the function of
the new gene.

6 Conclusion
One should neither have excessive faith in the results of a database search, nor
should they be blithely disregarded. The standard search programs such as
FASTA, gapped BLAST and SSEARCH are well-tested and reliable indicators of
sequence similarity, and their underlying principles are straightforward. These
programs and their parameters have been optimized for the hundreds of
thousands of runs every day. If one is careful about posing the database search
experiment and interprets the results with care, sequence comparison methods
can be trusted to rapidly and easily provide an incomparable wealth of bio-
logical information.

References
1. Sander, C. and Schneider, R. (1991). Database of homology-derived protein structures
and the structural meaning of sequence alignment. Proteins, 9, 56.
2. Hilbert, M., Bohm, G., andjaenicke, R. (1993). Structural relationships of homologous
proteins as a fundamental principle in homology modeling. Proteins, 17, 138.
3. Pearson, W. R. (1996). Effective protein sequence comparison. In Methods in enzymology
(ed. R. F. Doolittle), Vol. 266, pp. 227-58. Academic Press.
4. Needleman, S. B. and Wunsch, C. D. (1970). A general method applicable to the search
for similarities in the amino acid sequence of two proteins. J. Mol. Biol., 48, 443.

188
 COMPARISON OF PROTEIN SEQUENCES AND PRACTICAL DATABASE SEARCHING

5. Smith, T. F. and Waterman, M. S. (1981). Comparison of Biosequences. Adv. App. Math.,

2, 482.
6. Cormen, T. H., Leiserson, C. E., and Rivest, R. L. (1990). Introduction to algorithms. MIT
Press/McGraw-Hill Book Company.
7. Waterman, M. S. (1995). Introduction to computational biology. Chapman & Hall,
London.
8. Setubal, J. C. and Meidanis, J. (1996). Introduction to computational molecular biology. PWS
Publishing Co., Boston.
9. Gusfield, D. (1997). Algorithms on strings, trees, and sequences: computer science and
computational biology. Cambridge University Press, Cambridge.
10. Bairoch, A. and Apweiler, R. (1999). The SWISS-PROT protein sequence data bank and
its supplement TrEMBL in 1999. Nucleic Acids Res., 27, 49.
11. Pearson, W. R. and Lipman, D. J. (1988). Improved tools for biological sequence
comparison. Proc. Natl. Acad. Sri. USA, 85, 2444.
12. Altschul, S. F., Carrol, R. J., and Lipman, D. J. (1990). Basic local alignment search tool.
J. Mol. Biol., 215, 403.
13. Compugen LTD. BIOCCELERATOR Manual. http://www.compugen.co.il
14. Altschul, S. F., Boguski, M. S., Gish, W. G., and Wootton, J. C. (1994). Issues in
searching molecular sequence databases. Nature Genet., 6, 119.
15. Altschul, S. F., Madden, T. L, Schaffer, A. A., Zhang, J., Zhang, Z., et al. (1997). Gapped
BLAST and PSI-BLAST: a new generation of protein database search programs. Nucleic
Arias Res., 25, 3389.
16. Murzin, A. G. (1993). OB(oligonucleotide/oligosaccharide binding)-fold: common
structural and functional solution for non-homologous sequences. EMBO J., 12:3, 861.
17. Pearson, W. R. (1997). Identifying distantly related protein sequences. Comp. Appl.
Biosci., 13:4, 325.
18. Brenner, S. E., Chothia, C., and Hubbard, T. J. P. (1998). Assessing sequence
comparison methods with reliable structurally identified distant evolutionary
relationships. Proc. Notl. Acad. Sri. USA, 95, 6073.
19. Levitt, M. and Gerstein, M. (1998). A unified statistical framework for sequence
comparison and structure Comparison. Proc. Natl. Acad. Sri. USA, 95, 5913.
20. Karlin, S. and Altschul, S. F. (1990). Methods for assessing the statistical significance of
molecular sequence features by using general scoring schemes. Proc. Natl Acad. Sri.
USA, 87, 2264.
21. Dembo, A. and Karlin, S. (1991). Strong limit theorems of empirical functionals for
large exceedances of partial sums of i.i.d variables. Ann. Prob., 19, 1737.
22. Dembo, A., Karlin, S., and Zeitouni, O. (1994). Limit distribution of maximal non-
aligned two-sequence segmental score. Ann. Prob., 22, 2022.
23. Gumbel, E. J. (1958). Statistics of extremes. Columbia University Press, New York.
24. Arratia, R. and Waterman, M. S. (1994). A phase transition for the score in matching
random sequences allowing depletions. Ann. Appl. Prob., 4, 200-225.
25. Smith, T. F., Waterman, M. S., and Burks, C. (1985). The statistical distribution of
nucleic acid similarities. Nucleic Acids Res., 13, 645.
26. Waterman, M. S. and Vingron, M. (1994). Rapid and accurate estimates of statistical
significance for sequence data base searches. Proc. Notl. Acad. Sci. USA, 91, 4625.
27. Altschul, S. F. and Gish, W. (1996). Local alignment statistics. In Methods in enzymology
(ed. R. F. Doolittle). Vol. 266, pp. 460-80. Academic Press.
28. Pearson, W. R. (1998). Empirical statistical estimates for sequence similarity searches.
J. Mol. Biol., 276, 71.
29. Yona, G., Linial, N., and Linial, M. (1999). ProtoMap: Automatic classification of
protein sequences, a hierarchy of protein families, and local maps of the protein
space. Proteins, 37, 360.

189
GOLAN YONA AND STEVEN BRENNER

30. George, D. G., Barker, W. C., Mewes, H. W., Pfeiffer, F., and Tsugita, A. (1996). The PIR-
International protein sequence database. Nucleic Adds Res., 24, 17.
31. Benson, D. A., Boguski, M. S., Lipman, D. J., Ostell, J., Ouellette, B. F., Rapp, B. A., et al.
(1999). GenBank. Nudeic Acids Res., 27, 12.
32. Pearson, W. R.' (1995). Comparison of methods for searching protein sequence
databases. Protein Sri., 4, 1145.
33. Wootton, J. C. and Federhen, S. (1993). Statistics of local complexity in amino acid
sequences and sequence databases. Comput. Chem., 17, 149.
34. Dayhoff, M. 0., Schwartz, R. M., and Orcutt, B. C. (1978). A model of evolutionary
change in Proteins. In Atlas of protein sequence and structure (ed. M. Dayhoff), Vol. 5,
Suppl. 3, pp 345-52. National biomedical research foundation, Silver Spring, MD.
35. Jones, D. T., Taylor, W. R., and Thornton, J. M. (1992). The rapid generation of
mutation data matrices from protein sequences. Comput. Appl. Biosci., 8:3, 275.
36. Henikoff, S. and Henikoff, J. G. (1992). Amino acid substitution matrices from protein
blocks. Proc. Natl Acad. Sri. USA, 89, 10915.
37. Gonnet, G. H., Cohen, M. A., and Benner, S. A. (1992). Exhaustive matching of the
entire protein sequence database. Science, 256, 1443.
38. Risler, J. L., Delorme, M. O., Delacroix, H., and Henaut, A. (1988). Amino acid
substitutions in structurally related proteins. A pattern recognition approach.
Determination of a new and efficient scoring matrix. J. Mol. Biol., 204, 1019.
39. Johnson, M. S. and Overington, J. P. (1993). A structural basis for sequence
comparisons. An evaluation of scoring methodologies. J. Mol. Bid., 233, 716.
40. Henikoff, S. and Henikoff, J. G. (1991). Automated assembly of protein blocks for
database searching. Nudeic Acids Res., 19, 6565.
41. Altschul, S. F. (1991). Amino acid substitution matrices from an information theoretic
perspective. J. Mol. Biol., 219, 555.
42. Henikoff, S. and Henikoff, J. G. (1993). Performance evaluation of amino acid
substitution matrices. Proteins, 17, 49.
43. Gribskov, M. and Veretnik, S. (1996). Identification of sequence patterns with profile
analysis. In Methods in enzymology (ed. R. F. Doolittle). Vol. 266, pp. 198-211. Academic
Press.
44. Taylor, W. R. (1996). Multiple protein sequence alignment: algorithms and gap
insertion. In Methods in enzymology (ed. R. F. Doolittle). Vol. 266, 343-67. Academic
Press.
45. Brenner, S. E. (1999). Errors in genome annotation. Trends Genet, 15, 132.
46. Karplus, K., Barrett, C., and Hughey, R. (1998). Hidden markov models for detecting
remote protein homologies. Bioinformatics, 14:10, 846.

190
 Chapter 9
Networking for the biologist
R. A. Harper
EMBL-European Bioinformatics Institute, Wellcome Trust Genome Campus,
Cambridge, UK.

Riding along in my automobile my baby beside me at the wheel

Cruising and playing the radio with no particular place to go

Chuck Berry

1 Introduction
Every research worker would like to have the tools on hand to make his job
quicker and more efficient, and with the advent of the World Wide Web many
of the tasks associated with molecular biology have become freely available
online. In the past when a scientist wanted to know something about a par-
ticular subject then the first option was to talk to colleagues in the laboratory
and ask for their advice. If that was not sufficient then it was off to the library to
scan abstracts or the latest journals for the relevant information.
However times are changing and so are working habits. Why ask questions
from people in your laboratory when you can ask the same question on the
Bionet newsgroups http://www.bio.net from research workers all over the world?
Why thumb through textbooks for references when you can type in keywords
to an Internet search engine such as Lycos or Alta Vista and get a satisfactory
answer in no time at all? But often you find that the major search engines index
everything on the Web, which makes it difficult to find exactly what you want.
So often it is more profitable to use search engines that are totally dedicated to
biology.
In Europe you could use BiowURLd http://search.ebi.ac.uk:8888/compassl or Bio-
Hunt http://www.expasy.ch/BioHunt, which deal exclusively with biology-related
subjects. Another comprehensive listing exists at the Virtual Library in the
BioSciences division. http://www.vlib.org/Biosciences.html, and from China there is
the NEE-HOW project, http://biology.neehow.org which is an invaluable resource
for research workers from the Pacific rim.
In the USA one of the original and best lists of Biological resources, put to-
gether by Keith Robison can be found at Harvard http://golgi.harvard.edu/biopages.list
and of course there is the ever popular Pedro's BioMolecular Research Tools at

191
R. A. HARPER

http://www.public.iastate.edu/~pedro/research_tools.html. If you are looking specific-

ally for software related to bioinformatics, then there is the BioCatalogue at
http://www.ebi.ac.uk/biocat/biocat.html or if you are looking for an obscure database
then there is DBcat http://www.infobiogen.fr/services/dbcat from Infobiogen the
EMBnet node in France.
There are also a few good newsletters, which deal specifically with what is
happening in the world of bioinformatics. EMBnet produces a quarterly news-
letter, which gives an update of the latest developments at the different EMBnet
nodes throughout Europe. The EMBnet News can be found at the URL
http://www.ebi.ac.uk/embnet.news/embnetmenu.html.
The EBI has its own industry programme and they produce a newsletter called
the Bioinformer: http://bioinformer.ebi.ac.uk/newsletter/. One special feature within
this newsletter is the BioEvents Calendar, http://bioinformer.ebi.ac.uk/Events/ that
allows people to advertise workshops, conferences, or symposiums. In the USA
there are two major newsletters associated with bioinformatics. The NCBI news-
letter is at http://www.ncbi.nlm.nih.gov/Web/Newsltr/indeK.html and the National
Centre for Genomic Research, the NCGR newsletter, is at http://www.ncgr.org/
ncgr/ncgr_newsletter.html
The focus of this article is to help research workers avoid the World Wide
Wait while using World Wide Web.

2 The changing face of networking

In the early 1990s academic research workers had the networks all to them-
selves. Today however, the demography of those using the networks and their
reasons for using the networks have completely changed. The competition for
bandwidth is fierce between the commercial and academic sector.
The Georgia Institute of Technology (http://www.gvu.gatech.edu/user_surveys/)
has been conducting user surveys on the use of the Internet since 1994 (see
Figure 1). Over a four-year period there have been many radical changes in
attitude towards the use and abuse of the Web. The most recent surveys show
that when it comes to using the WWW the two main activities that people
engage in are collecting personal information, and using the Web purely for
entertainment. The academic no doubt will be distressed that work and educa-
tion only occupy equal third place. Academics are no longer the only people
using the internet and they may feel that their research work suffers because of
the 'info-tourists' on the web. Gone are the days when the only people on the
network were scientists with Unix boxes. More and more people are coming
online from home and the humble PC seems to have cornered the market (see
Figure 2).
In the past scientists relied on centralized systems, with systems administra-
tors installing and maintaining programmes. Nowadays since the installation of
many programmes on PC's has been fully automated, scientists are doing it for
themselves. This means that the scientist needs to be aware of the trends that
are driving the internet forward. Applications will be written for platforms that

192
 NETWORKING FOR THE BIOLOGIST

Figure 1 Primary uses of the Web. (Copyright 1994-1998 Georgia Tech Research
Corporation. All rights Reserved. Source: GVU's WWW User Survey
www.gvu.gatech.edu/user_surveys)

Figure 2 Primary places of WWW access. (Copyright 1994-1998 Georgia Tech Research
Corporation. All rights Reserved. Source: GVU's WWW User Survey
www.gvu.gatech.edu/user_surveys)

are being used the most. If the scientist insists that they can get by with their
VT100 terminal and a text based Lynx browser very soon they will be unable to
browse sites that are visually rich or rely on Java scripts or corba interfaces. It is
clear from the latest survey results that the most used widely used computing
platform is Windows 95 (Figure 3). No doubt this is partly due to the popularity of
Microsoft Internet Explorer which comes bundled with the operating system.
The browser wars between Netscape and Microsoft have already led to legal
battles in the American courts.

193
R. A. HARPER

Figure 3 Primary computing platform. (Copyright 1994-1998 Georgia Tech Research

Corporation. All rights Reserved. Source: GVU's WWW User Survey
www.gvu.gatech.edu/user_surveys.)

2.1 Networking in Europe

When it comes to networking not all countries are created equal. The EMBnet
organization has developed a service called 'Network Performance monitoring
in EMBnet'. This project has monitored the efficiency of networking throughout
Europe between the EMBnet nodes. The URL that gives the results from this
project is http://www.cmbi.kun.nl/Ping/.

Figure 4 Average round trip times times in ms from EmbNet node CAOS/CAMM in the
Netherlands to the EBI in the UK.

194
 NETWORKING FOR THE BIOLOGIST

In 1995, a Network Usage and Quality Advisory Group of the Dutch Network
organisation SURFnet, defined 'an upper RTT (Round Trip Time) limit of 125 msec.
without packet loss' as a minimum QOS (Quality of Service) level for interactive on-
line work. The RTT values from the Dutch Embnet node to the EBI can be
represented by the following graph.
This shows that the RTT from the Netherlands to the EBI in the UK has
consistently been below the recommended time of 125 msec, which means
scientists from the Netherlands, should have no difficulties in contacting the EBI
web server. It is interesting to note that the results from October 1998 show that
of the thirty-two nodes monitored, twenty-two have a RTT of less than 125 msec.
This must surely be good news for networking within Europe (Table 1).
It is essential that research workers learn to use the services provided for
them within their own countries. Penalties are always paid when you network
across international borders. It would seem that the more borders you cross, the
less efficient the network becomes. However, networking within your own
country is more efficient because more often than not a basic infrastructure
already exists between the major universities. In the mind of the molecular
biologist, however, Mecca is either at the NCBI or EBI and that is the direction
they religiously point their browsers to, only to suffer frustration when they
cannot get their work done due to poor bandwidth and increased traffic directed
towards these sites. For this reason EMBnet tries to co-ordinate their activities so
that all the EMBnet nodes provide easy access for database query and retrieval.
Many of the EMBnet nodes use a mirror package to update their databases on a
daily basis, via remote ftp from the databases stored on the EBI anonymous ftp
server at ftp://ftp.ebi.ac.uk/pub/databases.
The major databases such as EMBL or SWISS-Prot are then indexed at the
EMBnet nodes and can be queried with the SRS package. SRS which was de-
veloped at EMBL Heidelberg by Thure Etzold, has been adopted by many of the
EMBnet nodes throughout Europe and also abroad. SRS is also unique in that it
is able to index very many different databases. A list of all EMBNET sites that use
SRS is given in Table 2.

2.2 The way we were . . . e-mail servers for sequence

retrieval
Networking for the biologist has a very short history and many of the services
developed in the eighties are still in use today. Indeed for people with very bad
network connections the use of E-mail servers is still the preferred method of
obtaining sequences or running homology similarity searches such as FASTA or
Blast. The main depositories for sequence data are are found in the UK at the
European Bioinformatics Institute (EBI), and at the National Centre for Bio-
technology Information (NCBI) in the United States. In addition these two
institutes collaborate with the DDJBB in Japan.
Both the EBI and NBCI run e-mail servers that will allow you to retrieve
sequences via e-mail. To obtain information on how to run the e-mail server at

195
R. A. HARPER

Table 1 Table of EMBNET Internet PING results from 1-30 September 1998. These give
network response times from the CAOS/CAMM centre in Nijmegen (The Netherlands) to all
of the European EMBNET sites. The main figures show the average, minimum, and maximum
RTT times.

Node % Loss Avg RTT Min RTT Max RTT

(ms) (ms) (ms)
nl_caoscamm 0 1 1 30
uk_ucl 0 32 20 334
de_embl 98 33 26 86
be_ben 3 34 19 567
uk_hgmp 0 35 22 290
uk_ebi 1 36 22 251
uk_sanger 1 37 23 394
uk_seqnet 2 42 30 289
ch_expasy 5 42 28 328
ch_isrec 3 43 28 258
de_dkfz 5 46 31 336
fi_CSC 2 50 43 247
no_bio 0 52 45 156
se_bmc 0 56 44 564
fr_infobiogen 1 56 31 249
dk_biobase 5 59 49 654
at_biocenter 11 64 49 279
es_cnb 3 65 44 415
fr_genethon 1 65 35 851
de_mips 7 67 41 1091
ie_incbi 4 74 39 558
hu_abc 6 87 50 542
it_icgeb 4 132 58 2299
it_cnr 8 176 82 1214
us_ncbi 3 181 106 2804
gr_imbb 15 188 76 808
il_inn 95 275 149 521
pt_pen 3 372 103 3724
pl_ibb 3 588 560 876
au_angis 15 669 433 1828
za_sanbi 9 816 687 2084
cn_peking 34 904 742 3160

EBI you simply send a e-mail message to [email protected] and include in the
main body of the message the word help and full instructions will be sent via e-
mail on how to operate the service.
A similar method for sequence retrieval is employed by the NCBI and the e--
mail query system utilizes the Entrez retrieval system that they have developed

196
 NETWORKING FOR THE BIOLOGIST

Table 2 A List of the EMBNET nodes and some other sites around the world which support
SRS

WEHI, Melbourne, Australia

Vienna Biocenter EMBnet Node, Vienna, Austria
Belgian EMBnet Node (BEN), Brussels, Belgium
DBBM-IOC, Fiocruz, Rio de Janeiro, Brazil
CBR-NRC, Halifax, Canada The Genome Mine, Base4 Bioinformatics, Canada
CBI EMBnet Node, University of Beijing, China
CSC, Otaniemi, Espoo, Finland
INFOBIOGEN, Villejuif, France Institut Pasteur, Paris, France
LBMRPM INRA/CNRS, Auzeville, Toulouse, France
DKFZ, Heidelberg, Germany
EMBL, Heidelberg, Germany
GBF, Braunschweig, Germany
MIPS-MPG/GSF, Martinsried/Munich, Germany
Bioinformatics Centre, University of Pune, India
INCBI EMBnet Node, Dublin, Ireland
Weizmann Institute BCD, Rehovot, Israel
CNR EMBnet Node, Bari, Italy
CRISCEB, Second University of Naples, Italy
IVR, Kyoto University, Japan
Biotek EMBnet Node, Oslo, Norway
IBB-PAS EMBnet Node, Warsaw, Poland
IGC EMBnet Node, Oeiras, Portugal
SRCG, Novosibirsk, Siberia, Russia
BIC, National University Hospital, Singapore
CNB EMBnet Node, Madrid, Spain
Biomedical Centre (BMC), Uppsala, Sweden
ExPASy, Geneva, Switzerland
CAOS/CAMM Center, Nijmegen, The Netherlands
RIGEB-MRC, Gebze, Kocaeli, Turkey
Adlib, CAB International, Wallingford, UK
EMBL-EBI, Hinxton, Cambridge, UK
HGMP-RC, Hinxton, Cambridge, UK
MBDC Oxford, Oxford University, UK
SEQNET EMBnet Node, Daresbury, UK
Sanger Centre, Hinxton, Cambridge, UK
lUBio, Indiana University, USA

for their website. Many people would argue that getting sequence via e-mail is
old-fashioned technology. It is primitive in that it only delivers simple ascii-
formatted text. However the e-mail query server at the NCBI is clever enough to
be able to return the sequence to you in a variety of different formats including
GenBank, FASTA, or Html.

197
R. A. HARPER

Using [email protected] for the retrieval of sequences

and software
To request:
• Specific help on the sequence databases • The sequence called PIP03XX (nucleotide)
such as EMBL or SWISS-PROT • The sequence called WAP_MOUSE
• General help on software (protein)
• The sequence with accession number • The sequence submission form
X03392 (nucleotide)
You would write the following commands directly into the body of an e-mail message:
HELP NUC
HELP PROT
HELP SOFTWARE
GET NUC:PIP03XX
GET NUC:X03392
GET PROT:WAP_MOUSE
GET DOC:DATASUB.TXT
and then mail the commands to the e-mail address [email protected] You would then
receive the results back in your mailbox via e-mail.

It is often more convenient to shoot off a query by e-mail and get an answer
within a few minutes than it is to struggle with tiying to access a website
that has bandwidth problems. The address for the NCBI e-mail server is at
[email protected]. To receive full instructions on how the server works just
send an e-mail message to [email protected] and in the main body of the
message type the word help. I have often found that people who have used an e-
mail server generally have a better understanding of databases and sequence
retrieval than those who have only used a WWW interface.

Using [email protected] for sequence retrieval

Examples:
DBn
UID U30150
Will search the nucleotide database for an entry whose accession number is U30150. Since no
DOPT line is present, the record will be displayed the record in the default GenBank format.

DB n
UID U30150,U30153
DOPT f

198
 NETWORKING FOR THE BIOLOGIST

Will search the nucleotide database for entries whose accession numbers are U30150 and
U30153, and display them in FASTA format.

DB m
UID 88055872
DOPT r
HTML
Will search the MEDLINE database for the record with MEDLINE UID 88055872 and dis-
play it in MEDLINE Report format. Send the results in HTML format for viewing through
a WWW browser.

DB p
UlDsp|P11598|
DOPT m
Will search the protein database, using a FASTA formatted UID, to retrieve the entry whose
Swiss-Prot accession number is P11598, and display the MEDLINE links for that protein
record as document summaries.

2.3 Similarity searches via e-mail

The two most popular e-mail servers dealing with similarity searches are Blast
from the NCBI. and FASTA from EBI. For help regarding these e-mail servers you
can send an e-mail message to either [email protected] or [email protected] and
complete instructions on how to formulate an e-mail message to be processed
by these servers will be returned to you via e-mail. Again it should be stressed
that once you understand how to compose an e-mail message to submit a Blast
query via E-mail, then you can be more discriminating when you are asked to
repeat the procedure via the WWW. As it is most people just opt for the default
parameters and never experiment with different options.

Blast similarity search e-mail server at NCBI

To submit a Blast similarity search at the NCBI a e-mail message should be composed as
follows.
From: [email protected] Tue Jul 28 21:36:38 1998
Date: 28 Jul 1998 21:29:02-EDT
To: [email protected]
Subject:
PROGRAM blastn
DATALIB month
EXPECT 0.75

199
R. A, HARPER

BEGIN
>XYZ012 mygene XYZ
tgcttggctgaggagccataggacgagagcttcctggtgaagtgtgtttcttgaaatcat
The actual search request begins with the mandatory parameter 'PROGRAM' in the first
column followed by the value 'blastn' (the name of the program) for searching nucleic
acids. The next line contains the mandatory search parameter 'DATALIB' with the value
'month' for the newest nucleic acid sequences. The third line contains an optional
EXPECT parameter and the value desired for it. The fourth line contains the mandatory
'BEGIN' directive, followed by the query sequence in FASTA/Pearson format. Each line of
information must be less than 80 characters in length. Once the e-mail message has been
sent it will be processed automatically at the NCBI and the results returned to your e-mail
address once they have been computed.

The BLAST algorithm was developed by the National Center for Biotech-
nology Information at the National Library of Medicine. The BLAST family of
programs employs this algorithm to compare an ammo acid query sequence
against a protein sequence database or a nucleotide query sequence against a
nudeotide sequence database, as well as other combinations of protein and
nucleic acid. If you use BLAST as a tool in your published research, the following
reference should be cited:

Altschul. S. F., Madden, T. L., Schaffer, A. A., Zhang, J., Zhang, Z., Miller, W.,
and Lipman, D. J. (1997).Gapped BLAST and PSI-BLAST: a new generation of
protein database search programs. Nucleic Acids Res., Sept. 1, 25(17), 3389.

It used to be that the NCBI exclusively provided access to BLAST but in recent
years you can now run BLAST searches from many different sites around the
world, which is a clear indication that this programme has become a very
popular method for doing homology searches. The fact that it appears in so
many places may be due to the fact that it is available for free from the NCBI
anonymous ftp server at ftp:///ftp.ncbi.nlm.nih.gov/blast/.
Historically the EBI has always provided hornology searches through FASTA.
The following reference should be cited when you have used FASTA:

Pearson, W. R. and Lipman, D. J. (1988). Improved tools for biological

sequence analysis. Proc. Natl. Acad. Sri. USA, 85, 2444.

Web-based FASTA applications can be found at the F.B1 http://www.ebi.ac.uk/fasta3

and at DDJB http://www.ddbj.nig.ac.jp/E-mail/homology.html

200
 NETWORKING FOR THE BIOLOGIST

FASTA similarity search e-mail server at EBI

EXAMPLE OF A SIMPLE SUBMISSION
PATH [email protected]
TITLE My Sequence
LIB swall
SEQ
MMFSGFNADYEASSSRCSSASPAGDSLSYYHSPADSFSSMGSPVNAOJDFC
TDLAVSSANHPIVrAISTSPDLQWLVQPALVSSVAPSQTRAPHPFGVPA
END
The PATH is the e-mail address of the person to whom the results should be sent. TITLE is
anything that you want to appear on the subject line of the returned mail. LIB is the data-
base that you want to search against. The sequence itself should be enclosed between
SEQ and END.

2.4 Speed solutions for similarity searches

In recent years there has been an increase in the use of specialized hardware for
doing similarity searches. Four companies in particular have pioneered this
approach, and the turn around time for running a search against the whole of
Swiss-Prot has been reduced to around 10 seconds using the Smith-Waterman
algorithm.

2.4.1 Time Logic

Time Logic (http://www.timelogic.com) from the USA has introduced DeCypher Bio-
informatics Accelerators and they have implemented a number of algorithms
namely, Gapped BLAST 2 (includes entire heuristic search suite: blastn, blastp,
blastx, tblastn, tblastx) PSI-BLAST. Affine Smith-Waterman, FrameSearch,
ProfileSearch, ProfileScan, and ClustalW with graphical rendition of dendro-
gram (Java applet). The WWW query interface for a Smith-Waterman similarity
search is shown inFigure5 and some results from that search are given in Figure 6.
Timelogic also have a Blast search running on their hardware at the NCGR at the
URL:
http://seqsim.nogr.org/newBlast.html

2.4.2 Compugen
Compugen have succeeded in introducing the Biocellorator to many pharma-
ceutical companies to aid them in their search for new and novel drugs. The EBI
has a biocellorator, which is online and is available for public use. At the EBI
there are two different interfaces to this service. The one provided by Com-
pugen called GeneWeb and a simple custom interface developed at the EBI at
http://www.ebi.ac.uk/bic_sw. The interface to the BIC-SW at the EBI is very compact

201
R. A. HARPER

Figure 5 DeCypher interface for Smith-Waterman similarity searches.

and easy to use. Most people just accept the default settings, paste in their query
sequences and run the program. The interface page is shown in Figure 7 and
some results are shown in Figure 8.
If you are from the Mediterranean area then perhaps it would be more
convenient to try the GeneWeb interface from the Weizmann Institute in
Israel, which is also open to the public for unregistered users. The URL is
(http://sgbcd.weizmann.ac.11:80/cgi-bin/gcnweb/main.cgi)

2.4.3 Paracel FDF

At the Swiss EMBnet node you can find the Paracel Fast Data finder (FDF), which
is designed to help bioinformatics departments dramatically increase the rate at
which they can find high-scoring potential gcnomic targets. Paracel claim that
GeneMatcher is the first commercially available genetic data analysis system to
use custom ASIC technology that can analyse similarities or differences in DNA
or protein sequences up to 1000 times faster than traditional computer systems.

202
 NETWORKING FOR THE BIOLOGIST

Figure 6 Results from Decypher Smith-Waterman similarity search.

Competition to discover novel genes is of great interest to pharmaceutical

companies because if it is possible to identify just one critical target gene then
this can result in an application for a patent on a product. Therefore any method
that combines speed with sensitivity is a very valuable tool in the hands of the
research worker. The main search interface and some results are shown in
Figures 9 and 10.

3 Sequence retrieval via the WWW

If you are in a country with a poor Internet connection then working with E-
mail servers for the retrieval of sequences is often the best option. However,
there are many excellent servers in different parts of the world and they should
not be ignored, even if you do live on the other side of the planet. Your geo-
graphical location should be the first consideration when accessing a remote
site. It is best to access a site that is in close proximity. Two of the most popular
services for sequence retrieval arc Entrez from the NCBI and SRS from the EBI.
However there are other options available and if you are in the Pacific rim area
then it might be worthwhile to look at the services offered by DDBJ in Japan
http://www.ddbj.nig.ac.jp/searches-e.html or the Maestro service from the National
Centre for Genomic Research (NCGR) http://www.ncgr.org/gsdb/maestro/index.html
on the West Coast of the USA.

203
R. A. HARPER

Figure 7 BIC-SW interface for performing Smith-Waterman homology searches.

Figure 8 Results from BIC-SW similarity search at the E8L

204
 NETWORKING FOR THE BIOLOGIST

Figure 9 Paracel PDF interface at the Swiss EMBnet node.

3.1 Entrez from the NCBI

The NCBI is the only place in the world where you will find the Entrez service
and it concentrates on a few databases namely, nucleotide sequences, amino
acid sequences, 3-D structures, Genomes. Taxonomy, and Literature-PubMed
http://www.ncbi.nlm.nih.gov/Entrez/. One of its strengths is that it provides access to
PubMed and this is a key factor in its popularity and success. Effective August 3,
1998, NLM implemented a system enhancement that dramatically increases the
speed of the system. This redesign of the way PubMed stored and retrieved infor-
mation will improve users search time—a search that previously took approxi-
mately 18 seconds to run in PubMed now runs under 2 seconds.
In Entrez you select the database you wish to query, for example the protein
database and then you are allowed to string a number of keywords together, like
'Rhizobium Ausubel nodulation1 and those entries that meet the criteria will be
displayed. An example is shown in Figure 11.

3.2 SRS from the EBI

SRS http://srs,ebi.ac.uk/ is a very powerful tool for querying databases and it would
seem to be the preferred querying system within Europe. You select a database

205
R. A. HARPER

Figure 10 The first few lines of some search results from an FDF search at the Swiss
EMBnet node.

Figure 11 An example query using Entrez and the NCBI.

206
 NETWORKING FOR THE BIOLOGIST

Figure 12 An example query using SRS at the EBI,

and fill in your search criteria as keywords. For example in Figure 12, we see a
sample query using the fields Organism (Rhizobium), authors (Ausubel), Seq-
Length (a range 500:700) and the date (a range l-Jan-1998:30-Dec-1990),
SRS will then display two hits in Swiss-Prot for that particular year with a
sequence range between 500 and 700 (Figure 13). It should also be noted that SRS
also gives the possibility to launch an application such as BLAST or FASTA tor
any of the sequences that you care to select. You may also select different views
of a sequence. For example the FASTA format, which then allows you to launch
a multiple sequence alignment using ClustalW, directly from within SRS. This

Figure 13 The results page from the query in Figure 12.

207
R. A. HARPER

Figure 14 The results from launching clustalw as an application on the results in Figure 13.

method is a great time saver since there is no need to cut and paste your
sequence into a separate ClustalW application (Figure 14).

4 Submitting sequences
Not only does the research worker want to query, retrieve and analyse sequences.
occasionally they also want to submit their own sequences to the databanks be
it GenBank or EMBL. The three major organizations that collect sequence in-
formation work in collaboration with each other so that sequences entered into
GenBank are transferred daily by FTP to both EBI and DDBJ (and vice versa) in an
attempt to keep the major databases synchronized.
At any given time the three institutes are continually swapping data so it is a
false idea to believe that any one database is more current than the other. All
three institutes have online methods of submitting sequence data through the
Web. The NCBI were the first to come online with BANKIT. The EBI then
followed with WEB1N and the Japanese at DDBJ have Sakuara,
It should also be noted that the NCBI developed a stand-alone programme
for MAC'S, PC's, and Unix called SEQUIN that allows the end-user to enter their
data from a personal computer and to send the submission via e-mail or to
simply post the disk to the appropriate institute where it is then uploaded into
the database. Sequin is strongly recommended if you have bulk submissions to
make,

208
 NETWORKING FOR THE BIOLOGIST

4.1 Bankit at NCBI

Bankit is convenient for quick submission of sequence data to the NCBI. BankIt
allows you to enter sequence information into a form, edit as necessary, and add
biological annotation (e.g. coding regions, mRNA features). BankIt transforms
your data into GenBank format for you to review and when your record is com-
pleted, it can be submitted directly to GenBank. You have the option of adding
information by using text boxes to describe in your own words the source of the
sequence and its biological features. The entry screen from Bankit is shown in
Figure 15. The GenBank annotation staff reviews the submitted textual infor-
mation, incorporates it into the appropriate structured fields, and returns the
record by e-mail for your review.

4.2 Sequin from NCBI

Sequin is stand-alone program for the MAC, PC/Windows and UNIX. Sequin is an
interactive, graphically oriented program based on screen forms and controlled
vocabularies that guide you through the process of entering your sequence and
providing biological and bibliographic annotation. Sequin is designed to simplify
the sequence submission process and to provide graphical viewing and editing
options. This program is optimal for submitting multiple sequences, mutation
studies, phylogenetic sets, population sets, and segmented sets. It incorporates

figure 15 The Welcoming page of the Bankit service from the NCBI.

209
R. A. HARPER

Figure 16 The welcoming screen of the Sequin data submission program from the NCBI.

robust error checking and accommodates very long sequences and complex
annotations.
Although Sequin has been implemented by the NCBI, the opening screen
allows you to select which database you would like to submit you sequence to be
it GenBank, EMBL, or DDBJ. Usually when a sequence is submitted there may be
a process whereby the submitter has to be in contact with the annotators of the
sequence by telephone to clarify certain details. Therefore it is wise to choose a
submission centre in your geographical region if you want to avoid long dis-
tance telephone calls. A screen capture of Sequin is shown in Figure 16,
Once you have completed the submission depending on which database you
have selected at the beginning you will be prompted to send an e-mail to gb-
[email protected] for NCBI, [email protected] for EMBL or [email protected]
for DDBJ. Sequin runs on Macintosh, PC/Windows, and UNIX computers. The
program itself, along with its on-line help documentation, is available by anony-
mous FTP from the EBI (UK) at ftp://ftp.ebi.ac.uk/pub/software/sequin/ or from the
NCBI (USA) at ftp://ncbi.nlm.nih.gov/sequin/ A useful FAQ to help you if you run
into problems during submission can be found at. http://www.ebi.ac.uk/~sterk/
sqndocs/index.html

4.3 Webin from EBI

The EBI WWW tool (WebIn) guides the user through a sequence of WWW forms
allowing the user to submit sequence data and descriptive information in an
interactive and easy way (see Figure 17). All the information required to create a
database entry will be collected during this process:

210
 NETWORKING FOR THE BIOLOGIST

(a) Submitter information.

(b) Release date information,
(c) Sequence data, description, and source information.
(d) Reference citation information.
(e) Feature information (e.g. coding regions, regulatory signals etc.).
Data submissions are usually processed within two working days of receipt
and the authors are sent notification of their accession number(s). Authors will
be asked whether their submitted data can be made available to the public
immediately or whether they should be withheld until an author-specified date.
Data are never withheld after publication.
Once a database entry has been created from a submission, a copy is sent to
the submitter for their reference and for comments or corrections. However, it
often happens that the entry is correct when it is created but, with the passage
of time, becomes out of date. The authors may make corrections to the sequence
itself, or may discover new features of the sequence. Since such findings are
often not published, the only way to keep entries correct and up to date is if the
authors communicate their new findings to the database. At the EBI this can be
done by completing an update form available from the Anonymous FTP, site
FTP.EBI.AC.UK in the file: pub/databases/embl/release/update.doc or via the
WWW at the URL http://www.ebi.ac.uk/ebi_docs/update.html.
A new service that has been instituted at EBI is scanning for vectors before

Figure 17 A page from the Webin service of the EBI.

211
R. A. HARPER

submitting your sequence. You are able to check your sequence data prior to
submission for potential vector contamination by running a BLASTN search
against EMVEC, a vector database containing information on more than 2000
vectors from the EMBL/GenBank/DDBJ Database SYN(thetic) division. The results
will list sequences producing significant alignments and associated information
like vector name, score, alignment, etc. The EB1 suggests that you remove vector
contamination from your sequence data before submitting to the database.

4.4 Sakura from DDJB

SAKURA is a web-based DNA data submission system for DDBJ. The URL for
SAKURA is http://sakura.ddbj.nig.ac.jp which can be accessed from the DDBJ
Home Page (http://www.ddbj.nig.ac.jp). You can select either the English or
Japanese version. However, data input must be done in English only, regardless
of language version selected . SAKURA allows you to save your document before
completion and submit multiple sequences sequentially (see Figure 18).

5 Conclusions
Historically there has been a collaboration between EBI, NCBI, and DDBJ, These
three sites are still the only places that have the infrastructure set up to handle

Figure 18 A page from the SAKURA DNA submission system from DDBJ.

212
 NETWORKING FOR THE BIOLOGIST

the submission of nucleotide sequences to the databases, be they EMBL or

Genbank or DDBJ. For this reason they are also looked upon as the only places
where you can do queries and retrieval, or perform homology searches, or
multiple sequence alignments. This is no longer true and with the advent of
EMBnet, many of the national nodes are able to supply services that are not
offered by the major centres. These three major centres have a policy of making
all of their databases publicly available, and when distributed network of
databases exists in many different parts of the globe then it can only be for the
benefit of molecular biologists worldwide.

References
General references to articles about biological services on the internet.
1. Aldhous, P. (1993). Managing the genome data deluge. Science, 262, 502.
2. Altschul, S. et al. (NCBI) (1994). Issues in searching molecular sequence databases.
Nature Genet., 6(Feb), 119.
3. Appel, R. D., Sanchez, J.-C., Bairoch, A., Golaz, O., Ravier, F., Pasquali, C., et al. (1996).
The Swiss-2DPAGE database of two-dimensional polyacrylamide gel electrophoresis,
its status in 1995. Nucleic Acids Res., 24(1), 180.
4. Ashburner, M. and Goodman, N. (1997). Informatics—genome and genetic databases.
Curr. Opin. Genet. Dev., 7, 750.
5. Bairoch, A., Bucher, P., and Hofman, K. (1996). The Prosite Database, its status in 1995.
Nucleic Acids Res., 24(1), 189.
6. Bairoch, A. and Apweiler, R. (1996). The Swiss-Prot Protein sequence data bank and its
new supplement Trembl. Nucleic Acids Res., 24(1), 21.
7. Bairoch, A. (1996). The ENZYME Data Bank in 1995. Nucleic Acids Res., 24(1), 221.
8. Bairoch, A. (1991). SEQANALREF: a sequence analysis bibliographic reference
databank. Camput. Appl. Biosci., 7(2), 268.
9. Bleasby, A., Griffiths, P., Harper, R., Hines, D., Hoover, K., Kristofferson, D., et al.
(1992).Electronic communications and the new biology. Nucleic Acids Res., 20(16),
4127.
10. Coulson, A. (1994). High performance searching of biosequence databases. Trends
Biotechnol, 12, 76.
11. Fuchs, R. (1994). Sequence analysis by electronic mail: a tool for accessing Internet
e-mail servers. Comput. Appl. Biosci., 10(4), 413.
12. Gershon, D. (1997). Bioinformatics in the post-genomic age. Careers and recruitment
article. Nature, 389, 417.
13. Gershon, D. (1995). The boom in bioinformatics (employment review). Nature, 375,
262.
14. Harper, R. (EBI). (1995). World Wide Web resources for the biologist. Trends Genet.,
11(6), 223.
15. Holm, L. and Sander, C. (1996). The FSSP database: fold Classification based on
structure-structure alignment of proteins. Nucleic Acids Res., 24(1), 206.
16. Marshall, E. (1996). Hot property: biologists who compute. Science, 272, 1730.
17. O'Donnell, C. (1994). Obtaining software via INTERNET. Methods Mol. Biol., 24, 345.
18. Peitsch, M. C., Wells, T. N., Stampf, D. R., and Sussman, J. L. (1995). The Swiss-
SDImage collection and PDB-Browser on the World-Wide Web. Trends Biochem. Sci.,
20(2), 82.
19. Pietrokovski, S., Henikoff, J. G., and Henikoff, S. (1996). The Blocks database a system
for protein classification. Nucleic Acids Res., 24(1), 197.

213
R. A. HARPER

20. Roberts, R. J. and Macelis, D. (1996). REBASE—restriction enzymes and methylases.

Nucleic Acids Res,, 24(1), 223.
21. Rodriguez-Tome, P., Stoehr, P., Cameron, G. N., and Flores, T. P. (1996). The European
Bioinformatics Institute (EBI). Nucleic Acids Res., 24(1), 6.
22. Smith, T. F. (1990). The history of the genetic sequence databases. Genomics, 6(4), 701.
23. Stoehr, P. J. and Omond, R. A. (1989). The EMBL Network File Server. Nucleic Acids Res.,
17(16), 6763.
24. Williams, G. W. and Gibbs, G. P. (1990). Automatic updating of the EMBL database via
EMBNet. Comput. Appl. Biosci., 6 (2), 122.

214
 Chapter 10
SRS—Access to molecular
biological databanks and
integrated data analysis tools
D. P. Kreil and T. Etzold
EMBL-European Bioinformatics Institute, Wellcome Trust Genome Campus,
Cambridge, UK.

1 Introduction
This first section gives an introduction to SRS. Section 2 (A user's primer) is a
tutorial that demonstrates basic tasks: simple database queries, exploiting links
between databases, exploration of results, and launching analysis tools. Section 3
(Advanced tools and concepts) builds on the skills imparted by the tutorial, and
shows how to refine queries, create custom views on data, and use distributed
SRS resources. Section 4 (SRS server side) introduces aspects of using a local SRS
installation, and outlines how to take advantage of one. Section 5 (Where to
turn for help) suggests where to turn to, if this chapter does not address a par-
ticular question or problem.

1.1 SRS fills a critical need

Everyone who has faced the problem of finding particular information (e.g.
experimental results) in the printed literature knows to value computer-read-
able storage of data. Besides allowing advanced methods of data retrieval and
visualization, computerized storage also facilitates systematic combination of
data from multiple sources, comparative studies, and methodical application of
analysis tools to large sets of data. The rapid growth of available molecular
biological databases now gives us access to an unprecedented amount of infor-
mation by computer.
The increasing specialization of research and the resulting fragmentation of
molecular biology is clearly reflected in the choice of databases, which are many
and varied in content type and form. The wide differences that can be observed
in the database formats and technologies employed are partly caused by the
different nature of the results gathered, and cannot solely be blamed on an in-
difference on the part of the database designers towards efforts at standardization

215
D. P. KREIL AND T. ETZOLD

While adhering to present standards is certainly to be advocated, it would often

significantly impede the progress of individual research groups. Also, many data-
bases, and consequently the respective data formats, have grown over history.
The large number of different databases alone is a serious problem for knowing
where to turn for specific information. DBCAT, a manually maintained catalogue
of databases, currently lists around 400 databases relevant to biology (1). The
variation in database formats and technologies that need to be dealt with yet
adds significantly to an already complex task of fruitfully accessing the available
data. The technical difficulties encountered particularly obstruct more sophisti-
cated leverage of data, such as systematic application of analysis tools, or bring-
ing data from different areas together. Even though many databases directly or
indirectly reference data stored elsewhere, these links are difficult to exploit,
owing to the large differences between individual database implementations.
While a standardization of formats, or at least an agreed upon interface for data-
base interconnectivity, would greatly alleviate these problems, none of the
attempts so far has achieved the hoped-for degree of acceptance (2).
SRS has evolved to overcome many critical technical difficulties in database
access, and the integration of databases, and analysis tools. Already in earlier
versions SRS was considered a promising approach to the problems it set out to
solve—respected as 'the paragon of connectivity' (3). The current version SRS-5.1
now offers an automatically maintained catalogue-database of over 350 different
databases and their documentation to help users identify databases of interest
and locate an appropriate server (4). Users can access all these databases through
the uniform SRS query interface on the Web. The interface is simple to use, yet
allows complex queries, including the use of logical operators, combining fields
from multiple databases, and following implicit and explicit links between
databases. Results can be browsed in various views, combining data from several
databases. Database entries can be passed to analysis tools such as sequence simi-
larity search programs, restriction map analysis, or phylogenetic algorithms.
The results generated can be used in further queries or analyses.

1.2 History, philosophy, and future of SRS

The SRS system started out as a Sequence Retrieval System (5) that employed
sophisticated parsing and indexing of database text files (a parser is a computer
program that breaks down text into recognized strings of characters for further
analysis). Plain text is the lingua franca of data exchange. Whether a researcher
has entered experimental results into a spreadsheet, or a database centre main-
tains a repository of complex data structures using an advanced database
management system, everyone can provide a dump of their data in the form of
plain text files. SRS does not convert these files, but rather leaves the original
data in place and reads items directly from there, preserving the original con-
text when the complete entry is viewed. The parsed items are used for display of
selected details of a database record, and for the construction of index files that
allow efficient data retrieval (6) and searchable links (7) between database

216
 SRS—ACCESS TO MOLECULAR BIOLOGICAL DATABANKS

entries. SRS was first applied to databases with information on protein or nucleo-
tide sequences. It has since developed into a much more general data-retrieval
tool. It is used, for example, for bibliographical databases, hierarchically struc-
tured databases like taxonomies, or clinical data on mutations.
Besides giving researchers the freedom to store information the way they
want to, an approach centred on the use of plain text files benefits from a
medium that is portable across different computer architectures and is usually
easy to read for humans. In the integration of distributed resources, this can be
quite helpful in resolving conflicts that may be triggered by the individually
evolving components.
SRS parsers, definitions of database structures and interfaces to analysis tools,
and a considerable amount of the SRS core functionality itself are written in
Icarus, a language especially designed for that task. Icarus is used to define de-
scriptive data structures (meta-data) which are extensively used to represent SRS
concepts like database structures. Icarus code in parsers is interpreted. These
two design features allow rapid development: using meta-data reduces the
amount of code to be written, and interpreted parsers can be modified and
tested fast—it often takes just a few hours to integrate a new database starting
from scratch. Not only has this approach proven to scale very well with the
number of databases to integrate, the flexible combination of a recursive break-
down of structure and the powerful in-place processing of parsed data that is
characteristic of SRS parsers handles even very complex database formats
elegantly. Developers who have used an interpreted language before will have
little difficulty adapting to Icarus. We think this is reflected by its ready
acceptance by SRS server maintainers around the world who have added well
over hundred databases to the system.
Icarus is evolving towards a general-purpose object-oriented programming
language with special support for recursive lazy parsing and definition of meta-
data structures. We expect that it will play a central role in scripting and large-
scale data analysis in the future of SRS.
SRS integrates and interfaces to databases and analysis tools of other re-
searchers, and it is also often used as an engine for database access in other
systems like 0PM (8), and BioKleisli (9)—'wrap and be wrapped'! Communi-
cation with an SRS system can be through the World Wide Web (the most
popular form of access), from the Unix command line, from within an Icarus
script, or using the C application programming interface (API). Current new
developments include prototypes of a Corba server, and of a Perl API. Adding
support for other language API's (e.g. for Java via JNI) would be straightforward.

2 A user's primer
This section will introduce core SRS functionality with simple step-by-step
instructions. It concludes with an overview of the screens covered in the primer.
We suggest that new users follow the examples given below. The concepts are
much more readily understood with some hands-on experience.

217
Figure 1 (a) The SRS Home Page at EMBLEBI, http: //srs.ebi.ac.uk/. (b) The 'Top Page': Select the databases to query.
 SRS—ACCESS TO MOLECULAR BIOLOGICAL DATABANKS

2.1 A simple query

SRS is a very powerful and complex system. Nevertheless, the basic features are
easy to use, and its interface also caters for the not so experienced user. Protocol 7
is a guide to performing a simple query and should be accessible for absolute
beginners.

Performing a simple SRS query

1 Connect to an SRS server, e.g. http://srs.ebi.ac.uk/ at EMBL-EBI (Figure la).
2 Press the 'Start' button to begin a new SRS session.a
3 Select one or severalb databases to query, e.g. SWISSPROT. Database names are
displayed in groups. If necessary, expand and collapse groups by clicking on the '+'
and '-' buttons respectively (Figure 1b).
4 Press the 'Continue' button to get the Query Form (Figure 2).
5 Select the field to query in the drop-down box, e.g. 'Description'.
6 Enter the term to search for, e.g. 'Tetracycline' (without the quotes). By default,
the wildcard '*' will be appended, thus also matching 'Tetracyclines' etc. Deselect
this option for a verbatim search.c
7 Press the 'Do Query' button to submit your query.
a
SRS will store your work. If you plan to come back to it. add the 'resume' link from this page
to your bookmarks. At EMBL-EBI, an SRS work session is deleted when it has not been used for
two consecutive days. Other public sites may have similar mechanisms in place.
b
When several databases are selected, only fields present in all the selected databases are
available in the Query Form.
c
Avoid using wildcards when not needed since this considerably speeds up the search.

Figure 2 An SRS query form showing an example of a simple query.

219
D. P. KREIL AND T. ETZOLD

If the list of results is long, the user can page through the list using the links
at the bottom of the screen ('go to entries in chunk [. ..]').Every item in the list
of results has an identifier composed of database name and entry-i.d., such as
'SWISSPROT:LPTR_BACST'. They are hyper-linked to the respective complete
entries.
To submit another query of the same database(s), press the 'Query Form'
button and continue with Protocol 1. step 5. For a new query, press the 'Top Page'
button and go to Protocol 1, step 3, SeeFigure6 for more options,

2.2 Exploiting links between databases

In contrast to many other systems, SRS allows queries that search so-called links.
The meaning of links between entries depends on the databases involved, yet it
is usually evident. For example, an entry E in SWISS-PROT is linked to those
EMBL entries which contain nucleotide sequences that belong to the protein
described by the entry E. Similarly, a SWISS-PROT entry E is Linked to those PDB
entries which hold a 3D-structure for the protein described by the entry E. EMBL
entries are linked to updates in EMBLNEW that supersede the original entries in
EMBL. The example given in the next protocol assumes that the reader has
followed Protocol 1.

Applying a link query to selected entries

Selecting a set of entries
1 Start from a page showing the results of a query, e.g. after following Protocol 1.
Performing a simple SRS query.
2 Mark the tick boxes next to the entries you wish to include or exclude. Use the
radio button to choose to operate on the 'selected' or on 'all but selected' entries.
For example, chose 'selected' and mark entry 'SWISSPROT:TCRB_BACSU+.

Performing a link query

1 Press the 'Link' button.
2 Select the database(s) to link to [cL Protocol 1. step 3], e.g. EMBL.
3 Press the 'Continue' button to ask for the list of results.

Protocol 2 shows how to apply a link query to a selected set of entries. The
example query retrieves an EMBL entry that contains the nucleotide sequence of
the gene that encodes the protein described by 'SWISSPROT:TCRB_BACSU'. In
addition, entries that hold larger stretches of genomic DNA containing this gene
are retrieved (such as the complete genome of Bacillus subtilis).

220
 SRS—ACCESS TO MOLECULAR BIOLOGICAL DATABANKS

2.3 Using Views to explore query results

There are several pre-defined Views, which present various aspects of a list of
results. Besides the Views that display just the entry-id's, or the complete
entries, the selection of Views available is database dependent. Protocol 3 shows
how to display a set of selected entries in a specific View. The example given in
the next protocol assumes that the reader has followed Protocol 1.

Displaying selected entries with one of the pre-defined

views
Selecting a set of entries
1 Start from a page showing the results of a query, e.g. after following Protocol 1.
2 Mark the tick boxes next to the entries you wish to include or exclude. Use the radio
button to choose to operate on the 'selected' or on 'all but selected' entries. For
example, chose 'selected' and mark the entries 'SWISSPROT:LPTR_BACST',
'SWISSPROT:LPTR_BACSU', 'SWISSPROT:TCR1_ECOLI', and 'SWISSPROT:
TCR3_ECOLI'.

Changing the View

1 Select a View from the drop-down menu next to the 'view' button, e.g. 'Sequence
Simple'.
2 Optionally change the number of entries to be shown per screen (. . . entries in chunks
of. . .').
3 Press the 'view' button.

Figure 3 Showing a selection of entries in the View 'SequenceSimple'.

221
Figure 4 More examples of Views showing the same selection of entries. The screen on the right (b) employs a Java Applet to display
various local protein properties as controlled by the user.
 SRS—ACCESS TO MOLECULAR BIOLOGICAL DATABANKS

Some pre-defined Views present connected fields from multiple databases by

linking the displayed entries to other databases and integrating the results. See
below for an example.

2.4 Launching analysis tools

Besides browsing results of database field and link queries, external application
programs can be used to analyse data. Typical of such applications are database
searches by sequence similarity, e.g. BLAST (10), construction of multiple
sequence alignments, e.g. CLUSTALW (11), restriction map analysis, and tools
predicting various sequence properties (trans-membrane regions, protein second-
ary structure, etc).

Launching an external application program for selected

entries
Selecting a set of entries
1 Start from a page showing the results of a query, e.g. after a query for 'uroplakin'
in the 'Description' field of SWISS-PROT (cf. Protocol 1).
2 Mark the tick boxes next to the entries you wish to include or exclude. Use the
radio button to choose to operate on the 'selected' or on 'all but selected' entries.
For example, chose 'selected' and mark the entry 'SWISSPROT:UPKB_MUSVI'.
Request an application launch
1 Select a tool from the drop-down menu next to the 'launch' button, e.g. 'SW', the
Smith & Waterman hardware accelerated search at EMBL-EBI.
2 Press the 'launch' button to get a page of application options.

Adjusting launch parameters and starting the analysis

\ Check and/or amend your original selection of input to the application (for
example, by selecting only a fragment of the sequence to be searched for),
2 Adjust the application parameters according to your data and your preferences.
Hyper-links offer additional help that is displayed in a separate window to aid data
entry (see Figure 5a).
3 Consider turning off the 'automatically display result' feature because some
browsers do not correctly support it.
4 Press the 'Continue' button.
5 If the 'automatically display result' feature was disabled, wait until the next page
has completely loaded; this may take a while. Then follow the 'Display Results' link
at the top of the screen.

223
Figure 5 The application launch page of the 'SW' (Smith-Waterman search) application program. Here users can set application
parameters. Clicking on a hyper-finked parameter prompts will open a separate window with related help text. This window is re-used to
reduce screen clutter. (b) Results of an 'SW' application launch as seen with the view 'SwMofeFamilies'.
 SRS—ACCESS TO MOLECULAR BIOLOGICAL DATABANKS

Follow Protocol 4 to launch an external application for selected entries. The

results of the application launch as shown in Figure 5b are an example of a more
sophisticated pre-defined View. (See Protocol 3 on how to choose a pre-defined
View.) For each sequence hit by the search, a normalized score (Z-score), the raw
score, and the number of gaps in the alignment of the query sequence with the
hit sequence are displayed. Then each hit is linked to the database searched
(SWISS-PROT, in this case) to display the full-length description and the sequence
length from that record. Further links to PROSITE and PFAM give information
on membership to well classified families. In our example, the search returned
related Uroplakin 1A/B sequences as top hits. Examination of the match data
together with the family classifications strongly suggests that the longer rank-
ing sequences are not distantly related to the query, but rather scored because
they all have a trans-membrane-4 domain. Also, note how the data in PROSITE
and PFAM augment each other.

2.5 Overview
Here we give a flow-chart like overview of the interface screens covered by the
protocols of this section (Figure 6). Users may abort operations and return to the
Top Page or a new Query Form by pressing the respective buttons shown at the
top of most screens.

3 Advanced tools and concepts

In this section we introduce advanced methods of working with SRS. Users are
advised to become familiar with these as soon as they have mastered the basics,
as the complexity of the molecular biological data and their storage demands
fairly sophisticated data handling skills.

3.1 Refining queries

Due to the nature of molecular biological databases, even simple questions often
require a complex query. Here we explain several techniques that help the user
go beyond the first simple queries.

3.1.1 How the system works

As in many cases, it helps to know a little about what is happening under the
hood. In a nutshell, SRS parsers read text from a file into entries. These are then
parsed into fields, which usually are further decomposed or processed. In par-
ticular, it is the parser's job to extract terms to be written into index files. These
index files make fast queries of database fields possible. Obviously, they also
determine what fields are available for queries, and what terms a query can be
matched against.

3.1.2 The index browser

Consider searching for human SWISS-PROT entries. SWISS-PROT has an 'organ-
ism' field which can be queried for 'human'—or should it be 'homo sapiens', or

225
D. P. KREIL AND T. ETZOLD

Figure 6 An overview of the pages discussed so far. The thick arrows show the path of
usual progression as described in the protocols. There may be an intermediate page before
the list of results after an application launch.

'homo sapiens sapiens', or a combination of all these terms? The index browser is
the easiest way to inspect the terms a query can be matched against. Always
consult the index browser when working with an unfamiliar database or data-
base field, and when in doubt. Protocol 5 shows how to do that.
Looking at the returned values, one sees the need to restrict the search to
'human' without a trailing wildcard '*' (to exclude, e.g. the human viruses).
Checking the index for terms matching 'homo*', one respectively finds the same
number of entries for 'homo' and 'homo sapiens' as for 'human'. Queries using
logical operators (see below) can be used to check the suspicion that the three
terms are equivalent and that it suffices to use only one of them.

226
 SRS—ACCESS TO MOLECULAR BIOLOGICAL DATABANKS

Browsing the index for a database field

1 In the query form, select a database field from the drop-down box, e.g. 'Organism'.
2 Press the 'Info' button next to the selected field name to get the respective Index
Browser.
3 Modify the query mask, e.g. enter 'human*' (without the quotes),
4 Press the 'List values' button.

Typical use of the index browser includes:

(a) Checking whether single words or phrases are available from the index for
searching. Most free-text fields (such as 'Description') are usually indexed
using single words, while fields with a more standardized vocabulary often
store longer phrases in the index (e.g. the 'Keywords' field of SWISS-PROT).
When a field is indexed using single words, logical operators must be used to
combine words. Instead of asking for 'tetracycline represser', the query must
request 'tetracycline' and 'represser' (see below).
(b) Learning about the terms used in fields that have a controlled vocabulary
(e.g. the SWISS-PROT Feature Key. 'FtKey').
(c) Learning about special standardized formats used to represent data in indices
of particular fields (see the 'Citation' of SWISS-PROT, as an example).
(d) Looking for alternative spellings or words with typing errors using wildcards
or a regular expression. SRS regular expressions arc delimited by forward
slashes, e.g. '/p[ 0—9 ] | /', matching 'p' followed by one or more digits.
The index browser also shows documentation for the database field, and
status information on the respective index (Figure 7).

3.1.3 Complex queries with logical operators

Logical operators are a powerful extension to queries, whether one needs to
refine or broaden a search, or it is necessary to combine words of a phrase for
searching a field with a single word index. SRS supports the following logical
operators:
• 'a AND b', which requires both the terms a and b to be present—in queries: V;
• 'a OR b', which requires at least one of the terms a and b to be present—in
queries:' |':
• 'a BUTNOT b', which requires the term a to be present and b to be absent—in
queries: ' I'.
The short forms of the logical operators ('&' '|', and '!') can be used when typing
in the query form. Please note that parentheses cannot be used for grouping,

227
D. P. KREIL AND T. ETZOLD

Figure 7 The index browser page for the database field 'Organism'.

Figure 8 This query retrieves all entries on tetracycline that are relevant to antibiotic
resistance but are neither trans-membrane proteins nor peptides. Note that one can directly
search for the phrase 'antibiotic resistance' in the 'Keyword' index. To search for
entries described by 'resistance protein', one has to ask for each word separately
(namely, 'resistance & protein') because the index for 'Description' only contains single
words (cf. Index Browser, above).

and that expressions are evaluated strictly left to right (all operators have the
same precedence). However, the drop-down menu 'Combine searches with' can
be used to select a logical operator that will be applied to join the lines of the
query form, figure 8 shows a query that retrieves all entries on tetracycline that
are relevant to antibiotic resistance but are neither trans-membrane proteins
nor peptides.

3.1.4 The Query Manager

Often it is helpful or even necessary to break complex operations into smaller.
more manageable tasks. SRS stores user queries and makes them available for
later reuse through the Query Manager, Selected query results may be anno-
tated, saved to disk, deleted, displayed in particular Views, or combined using
logical operators.
Given the previous example, consider the wish to exclude those entries with
an uncertain sequence. Following the next few steps as outlined below will
accomplish just that. To avoid editing the original query, which one might want

228
 SRS—ACCESS TO MOLECULAR BIOLOGICAL DATABANKS

Figure 9 A query to retrieve entries described as 'tetracycline' and with a sequence about
which there is disagreement in the available literature.

Figure 10 The Query Manager page offers previous results for reuse and allows for more
complex operations.

to keep for reference, run a new query as shown in Figure 9. Go to the Query
Manager page, select the last two entries, and finally choose 'BUTNOT' as
operator and click on the 'combine' button (Figure 10).
Advanced users will learn to appreciate the option to enter more complex
query expressions directly. The previous result can also be obtained by entering
'[SWISSPROT-Description: tetracycline ! peptide] & [SWISSPROT-
Keywords: antibiotic resistance 1 transmembrane] ! ( [SWISSPROT-
FtKey: conflict ] > parent)' (without the quotes) into the window next to
the 'expression' button. Press the 'expression' button to request the evaluation
of the query. This example demonstrates several features:
(a) Simple queries take the form '[DatabaseName-FieldName:QueryString]'.
Logical operators may be used with the same restrictions as in the Query
Form. For brevity, the abbreviation of the field name, as displayed in the
Index Browser, can be used, e.g. ' [ SWISSPROT-des: tetracyc* ]' to query
the 'Description' field.

229
D. P. KREIL AND T. ETZOLD

(b) Queries can be combined using logical operators. Here, parentheses may be
used for grouping.
(c) The syntax for a link query is 'A < B' or 'A > B', where A and B can each be a
database name or the name of a set of entries (such as those automatically
assigned, e.g. 'Q2'). The first expression retrieves those entries of A that are
linked to B, while the second one retrieves the entries of B that are linked to
A (as suggested by the points of the 'arrows'). Thus, 'A > B' and 'B < A' are
equivalent expressions.
(d) Special links: Some entries can have sub-entries (such as the SWISS-PROT
Features). These are indexed and can be queried just like the main entries.
Linking with the special pre-defined 'parent' provides a mechanism to access
the entries that contain the retrieved sub-entries. There are other special
links for moving up and down in hierarchical databases and to link hits of
a search application to the originally searched database (. . .searchdb').
SRS Links can always be searched in both directions; they are bi-directional. If a
specialized database explicitly refers to a common (e.g. repository) database, one
can thus also request linked entries of the specialized database starting from an
entry in the common database. The designers of the common database do not
need to know even of the existence of the specialized database. SRS also finds
entries that are linked via an intermediary. For example, a SWISS-PROT entry
that has links to both EMBL and PDB creates (indirect) links between the respect-
ive EMBL and PDB entries. Consequentially, indirect links can involve any
number of intermediate linking steps. SRS uses the shortest way through the
net of databases when resolving a link. SRS server administrators can set up
penalties reflecting the desirability of using a particular inter-database link.
Users can chose a different path through the network by explicitly requesting a
series of links (. . . A > B > ... > C').

3.2 Creating custom Views

The simplest way to have query results displayed in a custom View is specifying
the fields of interest in the Query Form. There are two conceptually different
View types that can be requested.
(a) List Views usually try to conserve as much as possible of the entry as for-
matted in the original database text file. List Views thus typically act as a
filter, displaying only those lines of the entry that have information pertain-
ing to the selected fields. Often, the text is pretty-printed and contains hyper-
links where appropriate. It is not, however, altered in content. List Views are
best employed when information in a database entry needs to be read in its
original context. This is typically the case for fields that implicitly or ex-
plicitly refer to information elsewhere in the entry. Figure 11 shows an
excerpt of an entry that would considerably lose in clarity when displayed
out of context as fields in a table.
(b) Table Views give a better overview by collecting only requested data for

230
 SRS—ACCESS TO MOLECULAR BIOLOGICAL DATABANKS

display. Values are extracted from the original entry format of the database
text file, and the data are usually re-formatted for easy readability and con-
cise display. Table Views are ideal for summaries, excerpts, and also when
data from different databases needs to be combined. All the Views shown in
Figure 3, Figure 4, and Figure 5 are Table Views.
The View Manager allows the creation of more complicated Views, such as the
one show in Figure 5b. Also, Views can be deleted. Named custom Views can be
created from scratch, or derived by editing existing Views. Named custom Views
are then available just like the pre-defined Views. Users can specify the type of
the View (List View or Table View), and whether the abbreviated forms or the
complete names of the displayed database fields are to be shown in Table Views.
Views are defined for a set of databases (the root-libraries). Views also may in-
clude fields that are only present in some of the chosen root-libraries. Many
elaborate Views link to additional databases that are to be specified (the leaves of
the View). When all the involved databases have been selected, the user chooses
the respective database fields that should be shown, and the formats in which to
display data objects (e.g. sequences, alignments) of particular fields.
Advanced options allow the following:
(a) 'Display only number of linked entries': The query that links a displayed
entry to a leaf database may result in a set of several entries. For summaries,
the size of this set (instead of the individual entries) can be displayed; this
includes a hyper-link to the set of entries for inspection of details.
(b) 'Use view to display entries': In List Views, the entries of each leaf-database
can be shown in a specified View.

Figure 11 For this List View, the fields 'AccNumber' (accession number), 'Description',
'CitationNo', 'Authors', 'Citation', 'Sequence', and 'FtDescription' (feature sub-entry
description) have been selected for display. The 'Swiss' (SWISS-PROT) format has been
chosen for the sequence field.

231
D. P. KREIL AND T. ETZOLD

(c) 'Use query instead of link': This allows complex operations, such as per-
forming a link following a specified path through the net of databases, or
excluding certain entries. The set 'entry' holds the entry to be displayed and
can be used in the query expression.

3.3 SRS world wide: using DATABANKS

The public network of SRS servers currently provides the scientific community
with access to ~350 different databases. The 40 sites located in 26 countries
offer a total of ~ 1300 databank copies.
To facilitate identification of databases of interest and choosing an appro-
priate server, a new component of SRS version 5.1 generates DATABANKS, a
database of databanks, by traversing the public SRS servers around the world (4),
Automated nightly compilation of the data guarantees that the catalogue is up
to date.
SRS provides a framework for database documentation in form of the 'data-
bank information page', which has & standardized layout for easy reference and
includes a general description, references and internet links, as well as detailed
documentation of database fields. Each entry in DATABANKS contains a copy of
the SRS databank information page as provided by the server it was collected
from, and it concludes with an overview of alternate sites. A typical entry is
shown in Figure 12. If a stable connection to a particular site could not be
established, the site is moved to the end of the list of alternatives. In these cases,
data from previous runs are used as backup. A record of when the backup was
originally retrieved indicates whether it might be out of date.
DATABANKS is typically searched by databank name or description. Protocol
6 is a guide to solving common questions using DATABANKS. More generally,
any field present in the databank information pages, as well as site and server
characteristics, can be queried.

Search SRS world wide

Query DATABANKS
1 Connect to an SRS server that offers DATABANKS, for example http://srs.ebi.ac.uk/ at
EMBL-EBI.
2 Select 'SRS World Wide' from the SRS home page. This leads to a list of known
public SRS servers.
3 Select 'Search' to get to the Query Form for DATABANKS.
Identification of databases of Interest
1 Enter a query into the form. (See Figure 13 for an example.)
2 Restrict the search to a subset of DATABANKS that includes only one site from each
group of alternativesa as shown in Figure 13.b

232
 SRS—ACCESS TO MOLECULAR BIOLOGICAL DATABANKS

3 Press the 'Continue' button to submit your query.

Choosing an appropriate server
1 Go back to the DATABANKS Query Form.
2 Enter your query into the form, e.g. ask for databanks named 'PIRALN' (without the
quotes).
3 Press the 'Continue' button to submit your query.
a
Currently, this representative site is chosen as the site that has the most extensive databank
information page,
b
This requests the 'selection flag' by including' [ DATABANKS-SelFag:*]' in your query.

Consider a user who is interested in databases offering sequence alignments,

which hold information on well-characterized protein domains or families and
can be used for functional assignments or phylogenetic examinations. Selecting
the field 'Description' in the query form and asking for 'sequence' and 'align',
yields a list of approximately 60 databank copies. By restricting the search as in-
structed by Protocol 6, part B; however, the query only fetches a more man-
ageable list of 15 representative databanks. In addition to general databases of
protein domains or families such as PFAM(12), PRINTS (13), or PIRALN (14), a user
will also find specialized databases, such as HOVERGEN (vertebrates, 15), AMmtDB
(vertebrate mitochondria, 16), RDP (ribosomes, 17), FSSP and HSSP (protein struc-
ture, 18 and 19), or TRAKSFAC (transcription factors, 20), The user can now browse
the descriptions of the databases retrieved and refine or broaden the search.
Having decided on a particular database, e.g. PIRALN, one usually finds that
more than one server maintains a copy of the database, and the list of results
shows alternative sites (see Figure 14). The number of indexed entries and the
release number (where assigned by the server maintainers) help users to choose
a nearby site that has a current version of the database.
Both the list of results and the overview of alternate sites compiled for each
entry of DATABANKS provide direct links for remote queries. These lead to the
respective query forms of the databases at the remote sites as specified.

3.4 Interfacing with SRS over the network

This section demonstrates with a few examples how to access SRS directly over a
network connection. Currently, that means using the Web to request specific
parts of the SRS Web interface. The Web interface is rendered by the program
'wgetz'. which sites may provide in different locations, e.g. 'http://srs.ebi .
ac.uk/srs5bin/cgi-bin/wgetz' for EMBL-EBI, An easy way of getting the
path to wgetz is looking at the 'Databanks' hyper-link on the SRS home page.
To request the launch of wgetz, append the character '?' and any parameters
that you wish to supply. If you specify more than one parameter, separate them

233
D. P. KREIL AND T. ETZOLD

Figure 12 A typical DATABANKS entry. The entry contains a copy of the respective remote
SRS databank information page, which includes a description, references and links, as well
as detailed documentation of database fields and indices. It concludes with a listing of
alternative sites that offer ENZYME. Direct links to these sites and the remote query forms
for ENZYME are provided. For uses in the network vicinity of a particular DATABANKS server,
the relative response times compiled by that server give a clue to the net distances to other
sites ('N/A 1 indicates problems connecting at the specified time).

234
 SRS—ACCESS TO MOLECULAR BIOLOGICAL DATABANKS

Figure 13 Query for databanks that have a description containing the terms 'sequence'
and 'align'. The second line of the query form requests that the results be restricted to
one representative databank for each group of alternatives.

Figure 14 The results of a query for databanks named 'PIRALN', The number of indexed
entries and the release number (where assigned by the server maintainers) help users to
choose a nearby server that offers a current version of the appropriate database.

with '| ' instead of with spaces. To include literal spaces in your parameters,
replace them with ' % 2 0'.
There are various ways in which entries can be displayed, for example:
(a) As plain lists of entries: Specify your query as you would in the expression
box of the Query Manager.
(b) As individual entries: Specify the '-e' (entry) switch together with an SRS
query, e.g..../wgetz?-e+ [SWISSPROT-ID:LPTR_BACST].
(c) As lists of entries that can be browsed: Specify the '-sl' (sequence list)
switch with an SRS queiy. This yields a page that looks like the list of results
in the standard SRS Web interface.

235
D. P. KREIL AND T. ETZOLD

For each of the above, a particular View may be requested, e.g. '-view +
SequenceSimple'.
Some SRS functions can be accessed directly as entry points into the system,
e.g. requesting the query form for a specified set of databases. The switch '-l'
(libraries) sets the databases to operate on, the '-fun' (function) switch selects
the operation requested: e.g. '.../wgetz?-fun+PageQueryForm+-l+SWISS
PROT%20SWISSNEW'.
Coming back to a previous session, or to maintain data across several wgetz
calls, a user context needs to be specified with the '-id' switch. The id of a
session is part of most of the links it displays; it is easily seen in the 'resume'
link on the Top Page (cf. Figure 6). Chapter 3 of the on-line manual contains
further information on linking to SRS servers using the Web.
Future versions of SRS will also support other mechanisms of remote access
to SRS servers. We have already developed a prototype Corba server. Generally,
SRS is able to serve structured data and methods that operate on them to any
object oriented or procedural environment.

4 SRS server side

SRS may be installed locally on your site. This section outlines how to take
advantage of a local SRS system.

4.1 User's point of view

In addition to accessing their SRS system over the Web, local users can run
queries from the command line, use scripting to automate repetitive tasks or
perform elaborate analyses, and modify parsers (if the administrators of the
system permit that).
To activate a particular installation of SRS on your site, use '. prep_srs. sh'
from Bourne, Korn or POSIX shells, or'source prep_srs' from C shells. If the
sourced files are not in your path, they are provided in the 'etc' sub-directory of
the place where SRS was installed (the SRS root directory). After this initial-
ization, one can use the 'getz' program to query the SRS system; e.g. 'getz
[SWiSSPROT-Des:tetracycline] -f des ' will print those lines of the
matching entries that hold the Description field. Use Chapter 4 of the on-line
manual as a reference and for more examples.
While 'getz' is quite popular among users and is frequently used within
shell scripts or Perl programs, it is inefficient for repeated fine-grained access to
data. The only way to avoid a huge number of 'getz' invocations is to request
many fields in one query, and users often end up parsing the resulting 'getz'
output—what a waste considering that SRS has already parsed the entry into
fields! A better solution is integrating a programming language with SRS:
(a) The entire SRS functionality can be accessed through the C application pro-
gramming interface (API). See Chapter 12 of the on-line manual for instruc-
tions on linking your C code with the SRS library, and several examples.

236
 SRS—ACCESS TO MOLECULAR BIOLOGICAL DATABANKS

More examples can be found in the 'demo' sub-directory of the SRS installa-
tion. This is the interface to use for the most extensive access to SRS features.
However, only the functions introduced in the examples are guaranteed to
be supported in future versions—other functions may change as we improve
and extend the system.
(b) More and more, Icarus allows access to SRS functionality. Already now, SRS
queries and access to token tables is supported. Token tables hold the results
of the parsing process. In future versions, application launching and all other
SRS functions and data will be accessible through Icarus. The example shown
in Figure 15a prints the entry-ids and the molecular weights of all matches to
the query 'SWISSPROT-des:Tetracycline*'. The example shown in Figure
15b takes an SRS query string on the command line and dumps the 'fields'
token table. See Chapters 5-7 of the on-line manual to learn more about
Icarus.
(c) Future versions of SRS will offer native language interfaces to other general
purpose languages like C++ or Java, and to popular scripting languages like
Perl (for which a prototype has already been developed)
Sometimes users find they struggle to extract particular data through a query.
Often a simple modification or extension of the parsers involved helps a lot.
Clearly, many users do not want to deal with the parsers themselves, or they
even may not have permission to do so. Still, they need to know when it pays to
ask for help. Typical problems that can be solved include:

Figure 15 Example for using Icarus as a cripting language (for SRS version 5.1 or higher).
The program displayed on the left prints the entry-ids and the molecular weights of all
matches to the query 'SWlSSPROT-des:Tetracycline*'. The program shown on the right takes
an SRS query string as its argument on the command line and prints the parser's 'fields'
token table: For each token, it shows both the token code (a label, which can be used to
determine how a token is further processed), and the token string. The token object (which
is used, e.g. to hold the sequence object) could have been accessed using Stoken.obj.
$Query returns a $Set object. See Chapter 7 of the on-line manual for a reference of Icarus
classes and chapters 5 and 6 to learn more about Icarus syntax and functions.

237
D. P. KREIL AND T. ETZOLD

(a) Extracting data from a field: If you find yourself trying to query certain
parts of a field regularly, that extraction process should be delegated to the
parser. In a parser, this extraction can also be much more sophisticated.
(b) Querying phrases: The parser controls which terms can be matched by a
query (see Sections 3.1.1 and 3.1.2). Sometimes the way a parser breaks a field
into indexed terms is not well suited for answering a particular question.
Changing the parser or adding a new field that takes particular requirements
into account solves the problem. Consider a comment field, for which the
parser extracts single words for indexing; the query 'not & known' will not
only find the phrase 'not known', but also retrieve instances that contain the
two words out of context. To address this problem, the parser might be
changed to watch, e. g., for words preceded by 'not', and write the two word
phrase to the index.
(c) Similarly, queries that use logical operators to combine data from several fields
are sometimes not sufficient if the necessary context has been lost in the
indexing process. Changing the parser is again the only satisfactory solution.

4.2 Administrator's point of view

Within the scope of this chapter, we can only give some help in getting an SRS
system up and running. Each of the following topics, however, should be of
further interest to SRS server administrators:
(a) The automation of tasks in server, database, and application maintenance.
(b) The design and implementation of new modules for the integration of data-
bases and application programs. This must include a discussion of lazy, forced,
and explicit parsing, and how to best make use of this layered approach.
(c) A manual of Icarus that introduces new users, and serves as a reference to
experienced programmers. Icarus is an efficient, expressive scripting lan-
guage that has been particularly designed for the purpose of describing data-
structures and the rapid development of highly flexible parsers. Yet Icarus
has evolved towards a general-purpose object-oriented programming language,
which we expect to play a central role in scripting and large-scale data
analysis in the future of SRS.
Clearly, these issues are all rather complex and deserve a dedicated discussion
in themselves.

4.2.1 How to get your own SRS system

Download SRS from ftp://ftp.ebi.ac.uk/pub/software/unix/srs/.
the current version is in file 'srsS.l.tar.gz' and is freely available to the
public. Please note that future versions of SRS will continue to be freely
available for academic non-commercial use. Future versions of SRS and services
are provided by LION Bioscience Ltd (http: //www.lionbio.co.uk/).
To install SRS in a place of your choice, unpack the distribution file (e.g.
'gzip —cd srs5. l.tar.gz | tar —xf -'). This creates a directory for SRS,

238
 SRS—ACCESS TO MOLECULAR BIOLOGICAL DATABANKS

which we call the SRS root directory, and which we will print as '. . .' for the rest
of this chapter. From this directory, first run './srsinstall all'. To also
install the SRS web interface, run './srsinstall www' next. At completion,
this prints two lines that have to be inserted into the 'srm.conf file of your
web server. Ask your system administrator to do this if necessary. If there is no
web server installed on your site, save the lines for later reference. You then
need to have a Web server installed. We now normally use Apache (see
http://www.apache.org/).

4.2.2 Basic configuration steps

SRS obviously conies without any databases or applications. However, a large
number of database modules come with the distribution. These modules allow
SRS to index and access the respective databases. The files that constitute a
module are stored in '.../icarus/db/'. For each database, respectively, there is:
(a) A 'module-name. is'-file that holds the parser.
(b) A 'module-name. i'-file defines the database fields and links that can be queried.
(c) A 'module-name. it'-file that contains optional database documentation, which
is used to construct the databank information page.
The databank information pages usually contain fields that report database
sources, indicating where the required files can be downloaded. This can directly
be read in the 'module-name. it' -files. More conveniently, this information can
be queried for in the Database of Databanks as introduced in Section 3.3. This
also gives access to the hundreds of modules developed worldwide, of which
only some are included with the standard distribution. If a database module is
offered by remote servers only, download the files that constitute the respective
database module by following the hyper-links provided in the 'SRS Description'
field of the DATABANKS entry.
Edit '.../icarus/db/srsdb.i' to reflect the databases available at your
installation:
(a) There needs to be a 'file:module-name.i' command to include the appro-
priate field definitions.
(b) Each database needs a unique id number: Edit the list 'libIds' accordingly.
When adding a new database, specify the database as defined by '$Library'
in the 'module-name. i'-file.
(c) The locations of the database files are declared in the 'libs' list of '$LibLoc'
statements. SRS uses this list to determine which databases are available.
Therefore, remove statements for databases not kept at your site, or turn
them into comment lines, which start with a hash ('#'). Edit the other
directories as appropriate. Again, when adding a new database, specify the
database as defined by $Library in the 'module-name, i'-file.
To commit your changes, run 'srssection'. In preparation for this or any other
SRS program, initialize your environment: Users of Bourne, Korn or POSIX shells

239
D. P. KREIL AND T. ETZOLD

execute '. . . ./etc/prep_srs.sh', while C shell users 'source . . ./etc/

prep_srs'. (This sets the path for executables, the environment variable
'$SRSROOT' and several others.)
All of the above also applies to modules for application programs, only they
are not yet included in DATABANKS.

4.2.3 Prepare your databases for access through SRS

After preparing the appropriate database text files and configuring SRS accord-
ingly, a set of indices has to be created for each database. These indices allow
fast queries of database fields and links.
The program 'srscheck' generates the script 'srsupdate' and shows which
indices need to be created or updated. Use the '-l' (library) switch to check the
indices of a particular database only. Run srsupdate to actually build the
required indices. For small databases, this is a matter of minutes. Indexing large
databases like EMBL may take several hours, though. At successful completion,
your databases are ready for SRS queries.

5 Where to turn to for help

If you experience difficulties, do not despair!
(a) If you have installation problems, please read carefully through the README
file that comes with the distribution.
(b) Try to make good use of the on-line manuals. While it is sometimes difficult
to find particular bits of information there, they are quite extensive, and
should certainly be consulted first. As the individual sections can be quite
large, using the text search feature of your browser can be quite helpful. Users
will want to focus on Chapters 1 and 2, which deal with the Web interface
and the Query language, respectively. Administrators should start with
Chapters 4, 10, and 11. Chapters 10 and 11 contain a lot of information on
installation and set-up, and Chapter 4 is a reference of SRS programs avail-
able from the command line. Use the to search within chapters, and do not
be fooled by the section headings. Please note that the chapter numbers refer
to the on-line manual of SRS-5.1.
(c) Try asking colleagues for help. There is a strong community of SRS
aficionados out there that will be happy to help (that is, if you have looked
through the on-line manual first).
(d) Send your question to news://localnews/bionet.software.srs, a
newsgroup dedicated to SRS related questions.
(e) If you discover a bug, or have a problem of very technical nature, report it by
e-mail to srsdev@ebi .ac.uk.
If you need more help configuring and customising your system, you may want
to learn about SRS workshops and consultancy available from LION Bioscience
Ltd ([email protected], uk).

240
 SRS—ACCESS TO MOLECULAR BIOLOGICAL DATABANKS

Acknowledgements
Many people have contributed to the wealth of SRS database and application
modules that is publicly available now, and we are indebted to them all!
We wish to warmly thank Rob Falla for his help and the fruitful late-night
discussions. Also, we gratefully thank Mark Wooding who thoroughly examined
the final draft of this chapter. Any errors were certainly introduced afterwards!

References
1. Discala, C., et al. (1998). DBCAT, the public catalog of databases. INFOBIOGEN,
Villejuif, France;contact [email protected].
2. Frishman, D., Heumann, K., Lesk, A., and Mewes, H.-W. (1998). Bioinformatics, 14, 551.
3. Brenner, S. E. (1995). Science, 268, 622.
4. Kreil, D. P. and Etzold, T. (1998). Trends Biochem. Set., 24, 155.
5. Etzold, T., Ulyanov, A, and Argos, P. (1996). In Methods in enzymology (ed. R. F. Doolittle).
Vol. 266, p. 114. Academic Press.
6. Etzold, T. and Argos, P. (1993). Comput. Appl. Biosci., 9, 49.
7. Etzold, T. and Argos, P. (1993). Comput. Appl. Biosci., 9, 59.
8. Markowitz, V. M. and Ritter, O. (1995). J. Comput. Biol., 2, 537.
9. Davidson, S. B., Overton, C., Tannen, V., and Wong, L. (1997). Int. J. Digit. Libr., 1, 36.
10. Altschul, S. F., et al. (1990). J. Mol. Biol., 215, 403.
11. Thompson, J. D., Higgins, D. G., and Gibson, T. J. (1994). Nucleic Acids Res., 22, 4673.
12. Sonnhammer, E. L, et al. (1998). Nucleic Acids Res., 26, 320.
13. Attwood, T. K., Beck, M. E., Bleasby, A. J., and Parry-Smith, D. J. (1994). Nucleic Acids Res.,
24, 182.
14. Barker, W. C., et al. (1998). Nucleic Acids Res., 26, 27.
15. Duret, L, Mouchiroud, D., and Gouy, M. (1994). Nucleic Acids Res., 22, 2360.
16. Lanave, C., et al. (1999). Nucleic Acids Res., 27, 134.
17. Maidak, B. L, et al (1997). Nucleic Acids Res., 25, 109.
18. Holm, L, et al. (1992). Protein Set., 1, 1691.
19. Sander, C. and Schneider, R. (1991). Proteins, 9, 56.
20. Heinemeyer, T., et al. (1998). Nucleic Acids Res., 26, 362.

241
This page intentionally left blank
 List of suppliers

Anderman and Co. Ltd., 145 London Road, Bio 101 Inc., PO Box 2284, La Jolla, CA
Kingston-upon-Thames, Surrey KT2 6NH, 92038-2284, USA. Tel: 001 760 598 7299
UK. Fax: 001 760 598 0116
Tel: 0181 541 0035 Fax: 0181 541 0623 URL: http://www.biol01.com

Beckman Coulter (UK) Ltd., Oakley Court, Bio-Rad Laboratories Ltd., Bio-Rad House,
Kingsmead Business Park, London Road, Maylands Avenue, Hemel Hempstead,
High Wycombe, Buckinghamshire HP11 Hertfordshire HP2 7TD, UK.
1JU, UK. Tel: 0181 328 2000 Fax: 0181 328 2550
Tel: 01494 441181 URL: http://www.bio-rad.com
Fax: 01494 447558 Bio-Rad Laboratories Ltd., Division
URL: http://www.beckman.com Headquarters, 1000 Alfred Noble Drive,
Beckman Coulter Inc., 4300 N Harbor Hercules, CA 94547, USA.
Boulevard, PO Box 3100, Fullerton, CA Tel: 001 510 724 7000
92834-3100, USA. Fax: 001 510 741 5817
Tel: 001 714 871 4848 URL: http://www.bio-rad.com
Fax: 001 714 773 8283
URL: http://www.beckman.com CP Instrument Co. Ltd., PO Box 22, Bishop
Stortford, Hertfordshire CM23 3DX, UK.
Becton Dickinson and Co., 21 Between Tel: 01279 757711 Fax: 01279 755785
Towns Road, Cowley, Oxford 0X4 3LY, UK. URL: http://www.cpinstrument.co.uk
Tel: 01865 748844
Fax: 01865 781627 Dupont (UK) Ltd., Industrial Products
URL: http://www.bd.com Division, Wedgwood Way, Stevenage,
Becton Dickinson and Co., 1 Becton Drive, Hertfordshire SG1 4QN, UK.
Franklin Lakes, NJ 07417-1883, USA. Tel: 01438 734000
Tel: 001 201 847 6800 Fax: 01438 734382
URL: http://www.bd.com URL: http://www.dupont.com
Dupont Co. (Biotechnology Systems
Bio 101 Inc., c/o Anachem Ltd., Anachem Division), PO Box 80024, Wilmington, DE
House, 20 Charles Street, Luton, 19880-002, USA.
Bedfordshire LU2 0EB, UK. Tel: 001 302 774 1000
Tel: 01582 456666 Fax: 01582 391768 Fax: 001 302 774 7321
URL: http://www.anachem.co.uk URL: http://www.dupont.com

243
LIST OF SUPPLIERS

Eastman Chemical Co., 100 North Eastman Invltrogen Corp., 1600 Faraday Avenue,
Road, PO Box 511, Kingsport, TN 37662- Carlsbad, CA 92008, USA.
5075, USA. Tel: 001 760 603 7200
Tel: 001 423 229 2000 Fax: 001 760 603 7201
URL: http://www.eastman.com URL: http://www.invitrogen.com
Invitrogen BV, PO Box 2312, 9704 CH
Fisher Scientific UK Ltd., Bishop Meadow Groningen, The Netherlands.
Road, Loughborough, Leicestershire LE11 Tel: 00800 5345 5345
5RG, UK. Fax: 00800 7890 7890
Tel: 01509 231166 URL: http://www.invitrogen.com
Fax: 01509 231893
URL: http://www.fisher.co.uk Life Technologies Ltd., PO Box 35, Free
Fisher Scientific, Fisher Research, 2761 Fountain Drive, Incsinnan Business Park,
Walnut Avenue, Tustin, CA 92780, USA. Paisley PA4 9RF, UK.
Tel: 001 714 669 4600 Tel: 0800 269210
Fax: 001 714 669 1613 Fax: 0800 838380
URL: http://www.fishersci.com URL: http://www.lifetech.com
Life Technologies Inc., 9800 Medical Center
Fluka, PO Box 2060, Milwaukee, WI 53201, Drive, Rockville, MD 20850, USA.
USA. Tel: 001 301 610 8000
Tel: 001 414 273 5013 URL: http://www.lifetech.com
Fax: 001 414 2734979
URL: http://www.signia-aldrich.com Sharp & Dohme, Research Laboratories,
Fluka Chemical Co. Ltd., PO Box 260, CH- Neuroscience Research Centre, Terlings
9471, Buchs, Switzerland. Park, Harlow, Essex CM20 2QR, UK.
Tel: 0041 81 745 2828 URL: http://www.msd-nrc.co.uk
Fax: 0041 81 756 5449 MSD Sharp and Dohme GmbH, Lindenplatz
URL: http://www.sigma-aldrich.com 1, D-85540, Haar, Germany.
URL: http://www.msd-deutschland.com
Hybald Ltd., Action Court, Ashford Road,
Ashford, Middlesex TW15 1XB, UK. Mllllpore (UK) Ltd., The Boulevard,
Tel: 01784 425000 Blackmoor Lane, Watford, Hertfordshire
Fax: 01784 248085 WD1 8YW, UK.
URL: http://www.hybaid.com Tel: 01923 816375
Hybaid US, 8 East Forge Parkway, Franklin, Fax: 01923 818297
MA 02038, USA. URL: http://www.milupore.com/local/UK.htm
Tel: 001 508 541 6918 Millipore Corp., 80 Ashby Road, Bedford,
Fax: 001 508 541 3041 MA 01730, USA.
URL: http://www.hybaid.com Tel: 001 800 645 5476
Fax: 001 800 645 5439
HyClone Laboratories, 1725 South HyClone URL: http://www.millipore.com
Road, Logan, UT 84321, USA.
Tel: 001 435 753 4584 New England Blolabs, 32 Tozer Road,
Fax: 001 435 753 4589 Beverley, MA 01915-5510, USA.
URL: http://www.hyclone.com Tel: 001 978 927 5054

244
 LIST OF SUPPLIERS

Nikon Inc., 1300 Walt Whitman Road, Promega Corp., 2800 Woods Hollow Road,
Melville, NY 11747-3064, USA. Madison, WI 53711-5399, USA.
Tel: 001 516 547 4200 Tel: 001 608 274 4330
Fax: 001 516 547 0299 Fax: 001 608 277 2516
URL: http://www.nikonusa.com URL: http://www.promega.com
Nikon Corp., Fuji Building, 2-3, 3-chome,
Marunouchi, Chiyoda-ku, Tokyo 100, Japan. Qlagen UK Ltd., Boundary Court, Gatwick
Tel: 00813 3214 5311 Road, Crawley, West Sussex RH10 2AX, UK.
Fax: 00813 3201 5856 Tel: 01293 422911 Fax: 01293 422922
URL: http://www.nikon.co.jp/main/index_e.htm URL: http://www.qiagen.com
Qiagen Inc., 28159 Avenue Stanford,
Nycomed Amersham pic, Amersham Place, Valencia, CA 91355, USA.
Little Chalfont, Buckinghamshire HP7 9NA, Tel: 001 800 426 8157
UK. Fax: 001 800 718 2056
Tel: 01494 544000 Fax: 01494 542266 URL: http://www.qiagen.com
URL: http://www.amersham.co.uk
Nycomed Amersham, 101 Carnegie Center, Roche Diagnostics Ltd., Bell Lane, Lewes,
Princeton, NJ 08540, USA. East Sussex BN7 1LG, UK.
Tel: 001 609 514 6000 Tel: 01273 484644 Fax: 01273 480266
URL: http://www.amersham.co.uk URL: http://www.roche.com
Roche Diagnostics Corp., 9115 Hague Road,
PerkIn Elmer Ltd., Post Office Lane, PO Box 50457, Indianapolis, IN 46256, USA.
Beaconsfield, Buckinghamshire HP9 1QA, Tel: 001 317 845 2358
UK. Fax: 001 317 576 2126
Tel: 01494 676161 URL: http://www.roche.com
URL: http://www.perkin-elmer.com Roche Diagnostics GmbH, Sandhoferstrasse
116, 68305 Mannheim, Germany.
Pharmacia Biotech (Blochrom) Ltd., Unit 22,
Tel: 0049 621 759 4747
Cambridge Science Park, Milton Road,
Fax: 0049 621 759 4002
Cambridge CB4 OFJ, UK.
URL: http://www.roche.com
Tel: 01223 423723
Fax: 01223 420164 Schlelcher and Schuell Inc., Keene, NH
URL: http://www.biochrom.co.uk 03431A, USA.
Pharmacia and Upjohn Ltd., Davy Avenue, Tel: 001 603 357 2398
Knowlhill, Milton Keynes, Buckinghamshire
MK5 8PH, UK. Shandon Scientific Ltd., 93-96 Chadwick
Tel: 01908 661101 Road, Astmoor, Runcorn, Cheshire WA7
Fax: 01908 690091 1PR, UK.
URL: http://www.eu.pnu.com Tel: 01928 566611
URL: http://www.shandon.com
Promega UK Ltd., Delta House, Chilworth
Research Centre, Southampton SO16 7NS, Sigma-Aidrich Co. Ltd., The Old Brickyard,
UK. New Road, Gillingham, Dorset XP8 4XT, UK.
Tel: 0800 378994 Tel: 01747 822211
Fax: 0800 181037 Fax: 01747 823779
URL: http://www.promega.com URL: http://www.sigma-aldrich.com

245
LIST OF SUPPLIERS

Sigma-Aldrich Co. Ltd., Fancy Road, Poole, Stratagene Europe, Gebouw California,
Dorset BH12 4QH, UK. Hogehilweg 15, 1101 CB Amsterdam
Tel: 01202 722114 Zuidoost, The Netherlands.
Fax: 01202 715460 Tel: 00800 9100 9100
URL: http://www.sigma-aldrich.com URL: http://www.stratagene.com
Sigma Chemical Co., PO Box 14508,
St Louis, MO 63178, USA. United States Biochemical, PO Box 22400,
Tel: 001 314 771 5765 Cleveland, OH 44122, USA.
Fax: 001 314 771 5757 Tel: 001 216 464 9277
URL: http://www.sigma-aldrich.corn

Stratagene Inc., 11011 North Torrey Pines

Road, La Jolla, CA 92037, USA.
Tel: 001 858 535 5400
URL: http://www.stratagene.com

246
 Index

Brackets denotes figures Compugen computer company expectation value (e-value) 177,
201 187
1D-3D profile 5 consensus sequence 145, [146] extreme value distribution
175-6, [176]

alignment statistics DALI structure comparison

FASTA sequence search
background distribution 175 program 39
program 173, 185-7, [186],
significance 173 DATABANKS sequence retrieval
201, 207
program 232-3, [234], [235]
fold recognition 3, [4]
DBCAT database catalogue 216
DDBJ (web site) 195, 208
Bankit sequence submission DEJAVU structure comparison
program 209, [209] program 40 gap-BLAST sequence search
Bayesian methods in HMMs deterministic pattern 146 program 107, 109
87-8 DIAIIGN multiple sequence GA_FTT structure comparison
BioCatalogue (web site) 192 alignment program 64 program 36
BiowURLd (web site) 191 Dirchlet mixtures 84, 90, 147 genetic algorithm in structure
BLAST sequence search DOMO sequence alignment comparison 36
program 172-3, 185-7, database 106 GENFIT structure comparison
[186], 199-200, 207, 223 double dynamic programming program 42-6
BLOCKS sequence alignment method 27, 39-40 GenTHREADER 8
database 81, 106, 147, 152 DSC structure prediction global sequence alignment
BLOSUM substitution matrix method 127 169-70
55, 183-4 DSSP secondray structure GOR structure prediction
program 119 method 121
dynamic programming in graph-matching in structure
structure comparison 20, comparison 26, 28
ASP experiment 9 26-8
clique detection in structure dynamic programming in
comparison 28 threading 5, 6
hidden Markov models (HMM)
CLUSTAL multiple sequence
77, 144
alignment program 61,
HMMER2 sequence alignment
[62], 207, 223
program 79, 82-4
COBBLER sequence search EBI (web site) 195, 205
program 108 EMBL sequence database 195,
codon based alignment 68 220
coiled-coil structure prediction EMBnet 195-7 Internet 191
137 Entrez sequence retrieval iterated sequence search
comparative genomics 51 program 203, 205, [206] methods 99-100, 109

247
INDEX

jackknife test 119 phylogenetic trees 98 root mean square deviation

JPRED structure prediction position-specific scoring matrix (RMSD) 26, 29-30, 163
method 129 (PSSM) 101, 185
PRALIGN multiple sequence
alignment program 65 Sakura sequence submission
LAMA multiple sequence Pratt sequence pattern program 212, [212]
alignment program 110 discovery program SARF2 structure comparison
local sequence alignment 156-162, [157] program 42-5
170-171 Pred2ary structure prediction Seqin sequence submission
low-complexity sequence program 124 program 209, [210]
regions 53 PREDATOR structure prediction sequence alignment 51, 168,
method 126-7 [171]
PRINTS sequence alignment algorithm 169-172, [171],
MACAW multiple sequence database 81, 105, 147, 181
alignment program 65, 98, 152 gap penalty 56, 169-70,
101 probabilistic pattern 146 184
Maestro sequence retrieval probability value (p-value) 177 global/local 54, [54]
program 203 176 matrices 54-5, 182-4
MAST sequence search program PROCLASS sequence alignment sequence databank searching
108 database 106-7 56-7, 103, 108, 167, 178
minimum description length ProDom sequence alignment sequence homology 53
148 database 106 sequence motif 143
motif 143 profile-HMMs 86-8, 146-7, 1 sequence retrieval 195, 203,
multiple alignment 88 [204]
databases 72 progressive multiple alignment SMART sequence alignment
editing 69-72 61-3 program 81
of sequences 57-60, 95-6 Prosite SPratt structure pattern
profile 77-9, 86-7, 101-2, database 105, 145-7, 151-2, discovery program 162-4
146-7, 185 [152] SRS (Sequence Retrieval System)
multiple structure comparison sequence pattern 97, 144-6 195, 205, [207], 215, [218]
41 sequence profiles 80 SSAP (SAP) structure comparison
protein comparison program 40, 164
features 19 SSEARCH sequence search
NCBI (web site) 192, 195, 199 relationships 20 program 185-7, [186]
NEE-HOW project 191 protein family databases 105 SSPRED structure prediction
neural network method 123 protein secondary structure method 129
NNSSP structure prediction 113-4, 116 structure comparison
method 125 prediction 113 accuracy 29
from multiple alignment databases 46
131-2, [133] structure environment 5, 6
PAM (Dayhoff) substitution protein sequence structure prediction accuracy
matrix 55, 182-3 databases 105, 179-81, 215 118
Paracel computer company Logo 72, 97, [97] structure superposition 17
202, [205] motif 93, 143 multiple 21
patten pattern 143 of globins 24
databases 150 pseudocounts 103 using quaternions 23
classification 149 protein structure superfold 2
discovery 154, [156] comparison of 15 SWISS-PROT 220-3, [222]
sensitivity 150 patterns in 143, 162-4
specificity 150 topology [115], 116
Pedro's BioMolecular Research psi-BLAST sequence search threading 1, 5, 138
Tools (web site) 191 program 81-2, 107-9, 188 accuracy 8-9
Pfam protein family database genome data 7, [8]
80, 105, 153 protein domains 11
PHD structure prediction regular expression 146 Time Logic computer company
program 7, 123 residue pair potentials 6 201, [202]

248
 INDEX

topological equivalence 25 VERTAA structure comparison World Wide Web (WWW) 191
TOPS topology program 116-7 program 37, [38]
transmembrane prediction of
structure 133-6 z-score 178
topology 136 Webin sequence submission
program 210, [211]

249