Lecture Notes in Bioinformatics 2666

Edited by S. Istrail, P. Pevzner, and M. Waterman

Editorial Board: A. Apostolico S. Brunak M. Gelfand
T. Lengauer S. Miyano G. Myers M.-F. Sagot D. Sankoff
R. Shamir T. Speed M. Vingron W. Wong

Subseries of Lecture Notes in Computer Science

3
Berlin
Heidelberg
New York
Hong Kong
London
Milan
Paris
Tokyo
Concettina Guerra Sorin Istrail (Eds.)

Mathematical Methods
for Protein Structure
Analysis and Design

C.I.M.E. Summer School

Martina Franca, Italy, July 9-15, 2000
Advanced Lectures

13
Series Editors
Sorin Istrail, Celera Genomics, Applied Biosystems, Rockville, MD, USA
Pavel Pevzner, University of California, San Diego, CA, USA
Michael Waterman, University of Southern California, Los Angeles, CA, USA

Volume Editors
Concettina Guerra
Università degli Studi di Padova
Dipartimento di Ingegneria dell’Informazione
via Gradenigo 6a, 35131 Padova, Italy
E-mail: guerra@dei.unipd.it
Sorin Istrail
Celera Genomics, Applied Biosystems
45 West Gude Drive, Rockville, MD 20850, USA
E-mail: sorin.istrail@celera.com

Cataloging-in-Publication Data applied for

A catalog record for this book is available from the Library of Congress.

Bibliographic information published by Die Deutsche Bibliothek.

Die Deutsche Bibliothek lists this publication in the Deutsche Nationalbibliografie;
detailed bibliographic data is available in the Internet at <http://dnb.ddb.de>.

CR Subject Classification (1998): J.3, F.2, H.2, G.2, I.3.5, I.4

ISSN 0302-9743
ISBN 3-540-40104-0 Springer-Verlag Berlin Heidelberg New York

This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting,
reproduction on microfilms or in any other way, and storage in data banks. Duplication of this publication
or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965,
in its current version, and permission for use must always be obtained from Springer-Verlag. Violations are
liable for prosecution under the German Copyright Law.
Springer-Verlag Berlin Heidelberg New York
a member of BertelsmannSpringer Science+Business Media GmbH

http://www.springer.de

© Springer-Verlag Berlin Heidelberg 2003

Printed in Germany
Typesetting: Camera-ready by author, data conversion by Boller Mediendesign
Printed on acid-free paper SPIN: 10927403 06/3142 543210
Preface

The papers collected in this volume reproduce contributions by leading schol-

ars to an international school and workshop which was organized and held with
the goal of taking a snapshot of a discipline under tumultuous growth. Indeed,
the area of protein folding, docking and alignment is developing in response
to needs for a mix of heterogeneous expertise spanning biology, chemistry,
mathematics, computer science, and statistics, among others.
Some of the problems encountered in this area are not only important for
the scientific challenges they pose, but also for the opportunities they disclose
in terms of medical and industrial exploitation. A typical example is offered by
protein-drug interaction (docking), a problem posing daunting computational
problems at the crossroads of geometry, physics and chemistry, and, at the
same time, a problem with unimaginable implications for the pharmacopoeia
of the future.
The school focused on problems posed by the study of the mechanisms be-
hind protein folding, and explored different ways of attacking these problems
under objective evaluations of the methods. Together with a relatively small
core of consolidated knowledge and tools, important reflections were brought
to this effort by studies in a multitude of directions and approaches. It is
obviously impossible to predict which, if any, among these techniques will
prove completely successful, but it is precisely the implicit dialectic among
them that best conveys the current flavor of the field. Such unique diversity
and richness inspired the format of the meeting, and also explains the slight
departure of the present volume from the typical format in this series: the
exposition of the current sediment is complemented here by a selection of
qualified specialized contributions.
The topics covered in this volume pinpoint major issues arising in the de-
velopment and analysis of models, algorithms and software tools centered on
the structure of proteins, all of which play crucial roles in structural genomics
and proteomics. The study of 3D conformations and relationships among pro-
teins is motivated by the belief that the spatial structure, more than the
primary sequence, dictates the function of a protein. The largest repository of
VI Preface

3D protein structures is the Protein Data Bank (PDB), currently containing

about 17,000 proteins. The PDB has experienced a sustained growth and is
expected to continue to grow at an increasing pace in the near future. The
available structures are classified into a relatively small number of families
and folds, according to their three-dimensional conformation. While the num-
ber of proteins will continue to grow, it is widely believed that the number
of new folds will remain relatively stable. Structural comparisons involving
these structures are at the core of docking and the classification of proteins
and sub-aggregates, and motif searches in sequence and protein databases,
and ultimately they contribute to understanding the mechanics of folding in
living organisms.
The first three chapters of this volume contain material that was pre-
sented at the school. The chapter entitled “Protein Structure Comparison:
Algorithms and Applications,” by G. Lancia and S. Istrail, focuses on the
algorithmic aspects and applications of the problem of structure comparison.
Structure similarity scoring schemes used in pairwise structure comparison
are discussed with respect to the ability to capture the biological relevance
of the chemical and physical constraints involved in molecular recognition.
Particular attention is paid to the measures based on contact map similarity.
The chapter “Spatial Pattern Detection in Structural Bioinformatics,” by
H.J. Wolfson, discusses the task of protein structural comparison as well as the
prediction of protein-protein, protein-DNA or protein-drug interaction (dock-
ing). Different protein shape representations are used in biological pattern
discovery. The paper discusses the shape representations best suited to each
computational task, then outlines some rigid and flexible protein structural
alignment algorithms, and discusses the issues of rigid bound versus unbound
and flexible docking.
The chapter “Geometric Methods for Protein Structure Comparison,” by
C. Ferrari and C. Guerra, discusses, from a theoretical point of view, geometric
solutions to the problem of finding correspondences between sets of geometric
features, such as points or segments. After reviewing existing methods for the
estimation of rigid transformations under different metrics, the paper focuses
on the use of the secondary structures of proteins for fast retrieval of similarity.
It also deals with the integration of strategies using different levels of protein
representations, from atomic to secondary structure level.
The chapter “Identifying Flat Regions and Slabs in Protein Structures,”
by M.E. Bock and C. Guerra, presents geometric approaches to the extraction
of planar surfaces, which is motivated by the problem of identifying packing
regions in proteins.
The two contributions, “OPTIMA: a New Score Function for the Detection
of Remote Homologs,” by M. Kann and R.A. Goldstein, and “A Comparison of
Methods for Assessing the Structural Similarity of Proteins,” by D.C. Adams
and G.J.P. Naylor, deal with the problem of protein comparison, focusing on
different similarity functions for sequence and structure comparison.
 Preface VII

The next three papers, “Prediction of Protein Secondary Structure at High

Accuracy Using a Combination of Many Neural Networks,” by C. Lundegaard,
T.N. Petersen, M. Nielsen, H. Bohr, J. Bohr, S. Brunak, G. Gippert and O.
Lund, “Self-consistent Knowledge-Based Approach to Protein Design,” by A.
Rossi, C. Micheletti, F. Seno, A. Maritan, and “Learning Effective Amino-
Acid Interactions,” by F. Seno, C. Micheletti, A. Maritan and J.R. Banavar,
discuss techniques and criteria for protein folding and design.
The paper “Protein structure from solid-state NMR,” by J.R. Quine and
T.A. Cross, presents a mathematical analysis for solid-state nuclear magnetic
resonance (NMR). Finally, the contribution “Protein-like Properties of Simple
Models,” by Y.-H. Sanejouand and G. Trinquier, focuses on properties relevant
to the sequence-structure relationships.
The school was attended by 56 participants from 10 countries. Lectures
were given by Prof. Ken Dill, University of California (USA), Prof. Arthur
Lesk, University of Cambridge Clinical School (UK), Prof. Michael Levitt,
Stanford University School of Medicine (USA), Prof. John Moult, University
of Maryland (USA), and Prof. Haim Wolfson, Tel Aviv University (Israel)
Invited talks at the workshop were given by Prof. Mary Ellen Bock, Purdue
University (USA) and Dr. Andrea Califano, IBM Yorktown (USA).

Concettina Guerra
Sorin Istrail
Contents

Protein Structure Comparison: Algorithms and Applications

Giuseppe Lancia, Sorin Istrail . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3 Applications of Structure Comparisons . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
4 Software and Algorithms for Structure Comparison . . . . . . . . . . . . . . . . 11
5 Problems Based on Contact Map Representations . . . . . . . . . . . . . . . . . 20
6 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Spatial Pattern Detection in Structural Bionformatics
Haim J. Wolfson . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2 Protein Shape Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3 Protein Structural Alignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4 Protein-Protein Docking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
Geometric Methods for Protein Structure Comparison
Carlo Ferrari, Concettina Guerra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
2 Protein Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3 Structural Comparison: Problem Formulation . . . . . . . . . . . . . . . . . . . . . 62
4 Representation of Rigid Transformations . . . . . . . . . . . . . . . . . . . . . . . . . 63
5 Determination of 3D Rigid Transformations . . . . . . . . . . . . . . . . . . . . . . . 68
6 Geometric Pattern Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
7 Indexing Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
8 Graph-Theoretic Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
9 Integration of Methods for Protein Comparison Using Different
Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
X Contents

10 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
11 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Identifying Flat Regions and Slabs in Protein Structures
Mary Ellen Bock, Concettina Guerra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
2 A Geometric Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
3 An Improved Geometric Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4 Hough Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5 Performances of the Two Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
6 Plane Detection in Proteins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
7 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
OPTIMA: A New Score Function for the Detection of Remote
Homologs
Maricel Kann, Richard A. Goldstein . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
A Comparison of Methods for Assessing the Structural
Similarity of Proteins
Dean C. Adams, Gavin J.P. Naylor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
2 The DALI Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
3 The Root Mean Square Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
4 Geometric Morphometrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5 Comparison of Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
Prediction of Protein Secondary Structure at High Accuracy
Using a Combination of Many Neural Networks
Claus Lundegaard, Thomas Nordahl Petersen, Morten Nielsen,
Henrik Bohr, Jacob Bohr, Søren Brunak, Garry Gippert, Ole Lund . . . . . 117
1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
 Contents XI

Self-consistent Knowledge-Based Approach to Protein Design

Andrea Rossi, Cristian Micheletti, Flavio Seno, Amos Maritan . . . . . . . . . 123
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
2 The Design Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
3 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
Protein Structure from Solid-State NMR
John R. Quine, Timothy A. Cross . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
1 Discrete Curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
2 Tensors and NMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
3 Structure from Orientational Constraints . . . . . . . . . . . . . . . . . . . . . . . . 134
4 Acknowledgment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
Learning Effective Amino-Acid Interactions
Flavio Seno, Cristian Micheletti, Amos Maritan, Jayanth R. Banavar . . 139
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
2 Models and Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
Proteinlike Properties of Simple Models
Yves-Henri Sanejouand, Georges Trinquier . . . . . . . . . . . . . . . . . . . . . . . . . . 147
1 The 3x3x3 Cubic Lattice Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
2 N-Soft-Spheres Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152

List of Participants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155

Protein Structure Comparison: Algorithms and
Applications

Giuseppe Lancia1 and Sorin Istrail2

1
Dipartimento Elettronica e Informatica, Via Gradenigo 6/A, Padova, Italia
lancia@dei.unipd.it
2
Celera Genomics, 45 West Gude Dr., Rockville, MD
sorin.istrail@celera.com

1 Introduction
A protein is a complex molecule for which a simple linear structure, given by
the sequence of its aminoacids, determines a unique, often very beautiful, three
dimensional shape. Such shape (3D structure) is perhaps the most important
of all protein’s features, since it determines completely how the protein func-
tions and interacts with other molecules. Most biological mechanisms at the
protein level are based on shape-complementarity, so that proteins present
particular concavities and convexities that allow them to bind to each other
and form complex structures, such as skin, hair and tendon. For this reason,
for instance, the drug design problem consists primarily in the discovery of
ad hoc peptides whose 3D shape allows them to “dock” onto some specific
proteins and enzymes, to inhibit or enhance their function.
The analysis and development of models, algorithms and software tools
based on 3D structures are heceforth very important fields of modern Struc-
tural Genomics and Proteomics. In the past few years, many tools have
emerged which allow, among other things, the comparison and clustering of
3D structures. In this survey we cannot possibly recollect all of them, and even
if we tried, our work would soon be incomplete, as the field is dynamically
changing, with new tools and databases being introduced constantly. At the
heart of all of them, there is the Brookhaven Protein Data Bank, the largest
world’s repository of 3D protein structures.
Two typical problems related with protein structure are Fold Prediction
and Fold Comparison. The former problem consists in trying to determine the
3D structure of a protein from its aminoacid sequence alone. It is a problem
of paramount importance for its many implications to, e.g., medicine, genetic
engineering, protein classification and evolutionary studies. We will not discuss

C. Guerra, S. Istrail (Eds.): Protein Structure Analysis and Design, LNBI 2666, pp. 1-33, 2003.
 Springer-Verlag Berlin Heidelberg 2003
2 Giuseppe Lancia and Sorin Istrail

the Fold Prediction problem here, if not marginally, but we will instead focus
on algorithms and applications for protein fold (structure) comparison. The
following are some of the reasons why the Fold Comparison problem is also
extremely important:
• For determining function. The function of a new protein can be determined
by comparing its structure to some known ones. That is, given a set of pro-
teins whose fold has already been determined and whose fuction is known,
if a new one has a fold highly similar to a known one, then its function will
similar as well. This type of problems imply the design of search algorithm
for 3D databases, where a match must be based on structure similarity.
Analogous problems have already been studied in Computational Geom-
etry and Computer Vision, where a geometric form or object has to be
identified by comparing it to a set of known ones.
• For clustering. Given a set of proteins and their structures, we may want
to cluster them in families based on structure similarity. Furthermore, we
may want to identify a consensus structure for each family. In this case,
we would have to solve a multiple structure alignment problem.
• For assessment of fold Predictions. The Model Assessment Problem is the
following: Given a set of “tentative” folds for a protein, and a “correct”
one (determined experimentally), which of the guesses is the closest to the
true? This is, e.g., the problem faced by the CASP (Critical Assessment
of Structure Prediction) jurors, in a biannual competition where many re-
search groups try to predict protein structure from sequence. The large
number of predictions submitted (more than 10,000) makes the design of
sound algorithms for structure comparison a compelling need. In particu-
lar, such algorithms are at the base of CAFASP, a recent Fully Automated
CASP variant.
The problem of Structure Similarity/Alignment determination is in a way
analogous to the Sequence Alignment problem, but the analogy is only su-
perficial and it breaks down when it comes to their complexity. There is a
dramatic difference between the complexity of sequence alignment and struc-
ture alignment. As opposed to the protein sequence alignment, where we are
certain that there is a unique alignment to a common ancestor sequence, in
structure comparison the notion of a common ancestor does not exist. Simi-
larity in folding structure is due to a different balance in folding forces, and
there is not necessarily a one-to-one correspondence between positions in both
proteins. In fact, for two homologous proteins that are distantly related, it is
possible for the structural alignment to be entirely different from the correct
evolutionary alignment [14].
By their nature, three-dimensional computational problems are inherently
more complex than the similar one-dimensional ones for which we have more
effective solutions. The mathematics that can provide rigorous support in
understanding models for structure prediction and analysis is almost nonex-
istent, as the problems are a blend of continuous, geometric- and combinato-
 Protein Structure Comparison: Algorithms and Applications 3

rial, discrete-mathematics. Not surprisingly, various simplified versions of the

structure comparison problems were shown NP-complete [17, 18].
The similarity of two structures can be accessed on whole proteins or
on local domains (e.g., in a clearly defined multi–domain target). This is
analogous to global vs local sequence alignments, and different techniques
and similarity scores should be used for the two cases.
In this article we focus on the algorithmical aspects and applications of
the problem of structure comparison. We will pay particular attention to the
measures based on contact map similarity. For a more general treatment of
different algorithms and measures, the reader is referred to a good survey by
Lemmen and Lengauer [30].
An inherent difficulty in pairwise structure comparison is that it requires
a structure similarity scoring scheme that captures biological relevance of the
chemical and physical steric constraints involved in molecular recognition. In
this respect, contact maps seem to have many advantages over other represen-
tations. The information of contacts is relatively simple but complete, and it
provides a reliable first–approximation to the overwelming complexity of the
problem.
Among the classical measures (non contact map–based) for structure com-
parison, there is the RMSD (Root Mean Square Deviation) which tries to
determine an optimal rigid superpositioning of a set of residues of one struc-
ture into a predetermined set of corresponding residues of another. The RMSD
measure has many recognized flaws. Most notably, it is a global measure which
can be a very poor indicator of the quality of a model when only parts of the
model are well predicted. In fact, the wrongly predicted regions produce such a
large RMSD that it is impossible to see if the model contains ”well predicted”
parts at all. Despite its drawbacks, the RMSD measure is widely used, and
many other measures are based on it. We will see, e.g., the algorithms Max-
Sub and GDT, used in CAFASP, the algorithms MNYFIT and STAMP, used
by the database HOMSTRAD, 3dSEARCH, used in the database SCOP, the
server DALI, and others.
With some notable exception, (e.g. 3dSEARCH) of algorithms whose com-
plexity and properties were analized rigorously and in detail, these algorithms
are usually heuristic procedures using some sort of simple local search to op-
timize a certain objective function. These heuristics are typically based on
reasonable assumptions on the properties that the optimal solution should
have in order to heavily prune the search space. From a theoretical com-
puter scientist’s perspective, the analysis and foundation of these algorithms,
if present at all, are often unsatisfactoy, and we remain in doubt that different,
more sophisticated approaches (e.g. Branch–and–Bound, or Metaheuristics ad
hoc) may be able to compute solutions with a better value of the particular
objective function.
Every optimization model set up for a problem in molecular biology is
only an arbitrary representation of the natural process, and hence it is ar-
guable that its optimal solution is in fact the “right” one. This argument is
4 Giuseppe Lancia and Sorin Istrail

usually employed to justify the use of simple heuristics instead of any rigor-
ous techniques or provable property of the model. However, after a heuristic
optimization, the so–called “global optimum” found is then used to draw
conclusions of biological meaning or support some theory on the biological
process. Hence, it is important for the quality of the solution to be as good
as possible, which justifies the use of sophisticated optimization techniques.
It is auspicable therefore an ever increasing cooperation of the community of
theoretical computer scientists, applied discrete mathematicians and molec-
ular biologists to address these problems starting from their computational
complexity, devising exact and approximation algorithms, studying lower and
upper bounds, and designing effective heuristics of many sorts.
The remainder of the paper is organized as follows. We start with a short
account on 3D structures, their properties and how they are determined. The
rest of the article is divided roughly in two parts. The first part focuses on ap-
plications of structure alignment problems, while the second part is centered
on algorithms. This second part is in turn divided in two, first talking about
measures which are not based on contact maps, and then focusing on contact
map problems. Among the applications of structure comparison, we describe
some databases based on structure similarity (PDB, SCOP, HOMSTRAD,
SLoop, CAMPASS, FSSP) and the problem of protein folding prediction as-
sessment, relevant, e.g., in the CASP and CAFASP competitions. We then
describe some of the main algorithms for structure comparison, and in partic-
ular those used by some of the applications previously mentioned. We survey
RMSD and its variants, DALI, MaxSub, Lgscore, GDT, Geometric Hashing
(3dSEARCH), MNYFIT and STAMP. In the final part of the article, we talk
about novel algorithms and problems based on the contact map representa-
tion of structures. We review the literature and some of our work on exact
and heuristic algorithms for the maximum contact map overlap problem. The
chapter is closed by a section on possible generalizations and directions for
future work.

2 Preliminaries
A protein consists of a chain of aminoacids, of length varying from about 50 to
3,000 and more. The chemical structure of a single aminoacid is comprised of
a carbon atom (called Cα ) connected to a carboxyl group and an amine group,
a hydrogen atom and a part depending on the specific aminoacid, called the
residue. The amine group of one aminoacid is linked to the carboxyl group of
the next one, giving rise to the linear chain. The backbone of the protein is
the sequence of Cα atoms, to which the rest of the atoms are attached.
The chemical properties and forces between the aminoacids are such that,
whenever the protein is left in its natural environment, it folds to a spe-
cific 3-dimensional structure, called its native, which minimizes the total free
energy. Several experiments have confirmed that two proteins with the same
 Protein Structure Comparison: Algorithms and Applications 5

aminoacid sequence have the same 3D structure under natural conditions, and
that, if a folded protein is artificially stretched to a chain and then released,
it will fold again to the same native.
The 3D structure of a protein is fully specified by the set of (x, y, z)–
coordinates of its atoms, with respect to an arbitrary origin. An alternative,
weaker description is the set of distances between either all atoms or only
the Cα s. The main procedures used to to determine the 3D structure of a
protein are X-ray crystallography and Nuclear Magnetic Resonance. In X–ray
crystallography the 3D structure of a molecule is determined by X–ray diffrac-
tion from crystals. This technique requires the molecule to be first crystalized,
at perfect quality; then, its diffraction pattern, produced by X–irradiation is
studied. This involves the analysis of thousands of spots, each with a posi-
tion and an intensity. Obtaining good crystals and studying the phases of the
waves forming each spot are very complex problems. For these reasons, X–ray
crystallography is a very long and delicate process, which may take several
months to complete. The result is the set of spacial coordinates of each atom
in the structure. For a good description of this technique, see [47] or [5].
Nuclear Magnetic Resonance (NMR) is performed in an aqueous solution,
where the molecules tumble and vibrate from termal motion. NMR detects
the chemical shift of atomic nuclei with nonnull spin. In order to get an ad-
equate resolution, the molecule must tumble rapidly, which typically limits
the size of the proteins to which this technique is applicable. The result of
NMR is a set of (estimates of) pairwise distances between the atoms and
hence it yields a collection of structures (namely all those compatible with
the observed atom–to–atom distances) rather than a single one. According to
some studies, the results of NMR are “not as detailed and accurate as that
obtained crystallographically” [8].
For a protein we distinguish several levels of structure. At a first level, we
have its primary structure, given by the monodimensional chain of aminoacids.
Subsequent structures depend on the protein fold. The secondary structure
describes the protein as a chain of structural elements, the most important of
which are α–helices and β–sheets. The tertiary structure is the full description
of the actual 3D fold, while the quaternary structure describes the interaction
of several copies of the same folded protein. Correspondingly, when comparing
two proteins, one may use algorithms that highlight differences and similarities
at each of these levels. Algorithms for the comparison of primary structures
are known as sequence alignment methods and are beyond the scope of this
paper. Here, we will focus mainly on algorithms based on the secondary and
tertiary structures.
The input to a generic structure alignment algorithm contains the de-
scription of the n–ary structures of two proteins (e.g., for n = 3, the set of 3D
atomic coordinates of two different molecules). Loosely speaking, the goal of
the algorithm is to find a geometrical transformation (typically rigid, made
of rotation and translation) which maps a “large” number of elements of the
first molecule to corresponding elements of the second. Such algorithms can
6 Giuseppe Lancia and Sorin Istrail

be of two types, i.e. sequence order dependent or sequence order independent.

In the sequence dependent situation, the mapping must take into account the
“identity” of the atoms, i.e. an atom of one structure can only be mapped
to the same atom in the other. This way the problem is reduced to a 3D
curve matching problem, which is essentially a monodimensional case, and so
computationally easier. This approach can be useful for finding motifs pre-
serving the sequence order. A sequence order independent algorithm does not
exploit the order of atoms, that is, treats each atom as an anonymous “bead”
indistinguishable from the others. This is a truly 3D task, which can detect
non–sequential motifs and binding sites. It can be used to search structural
databases with only partial information and it is robust to insertion and dele-
tions. Consequently, it is also more challenging computationally.
In the remainder of this survey we will discuss algorithms of both types.
The most important sequence order dependent similarity measure is certainly
the Root Mean Square Deviation (RMSD). This is the problem of rigid super-
position of n points (ai ) and n
points (bi ), by means of a rotation R and a
translation t which minimize i |Rai + t − bi |2 , and can be solved in O(n)
time [55] by Linear Algebra techniques. As already mentioned, this measure
is not flawless: in fact, the RMSD for a pair of structures almost identical,
except for a single, small region of dissimilarity, can be very high.

3 Applications of Structure Comparisons

3.1 Databases Organized on Structure

The Protein Data Bank (PDB, [3]) is the most comprehensive single worldwide
repository for the processing and distribution of 3D biological macromolecular
structure data. As soon as a new 3D structure has been determined, it can
be deposited in the PDB, this way making it readily available to the scientific
comunity. Currently, the 3D structures of proteins in the PDB are determined
mainly by X-ray crystallography (over 80%) and Nuclear Magnetic Resonance
(about 16%). The remaining structures are determined by other methods, such
as theoretical modeling, which are not experimental in nature, but rather
based on computation and/or similarity to known proteins.
In this archive, one can find the primary and secondary structure infor-
mation, atomic coordinates, crystallographic and NMR experimental data, all
annotated with the relevant bibliographic citations, in a suitable text file for-
mat. Several viewers are available which interpret this format to give a 3D
graphic rendering of the protein.
 Protein Structure Comparison: Algorithms and Applications 7

Fig. 1. The growth of the PDB.

The amount of information in the PDB has increased rapidly over the
past years, passing from the roughly one thousand structures of 1990 to over
14,000 of today (see Figure 3.13 ). In order for this information to be usable and
accessible, it must be organized into databases which can be queried, e.g., for
global/local 3D similarity and/or complementarity, or for finding evolutionary
related homologs. A number of such databases has been created in the last
decade, each with its own special features. Common to all of them is the
use of structure comparison algorithms as the basic tool for clustering and
information retrieval. In the remainder of the section we survey a few of the
most known, i.e. SCOP, FSSP, HOMSTRAD, SLoop and CAMPASS.

SCOP
In the SCOP (Structural Classification Of Proteins, [42]) database, all pro-
teins of known structure are related to each other according to their structural
similarity and evolutionary relationships. The database was created by both
manual inspection and the use of algorithms and automated methods. It con-
tains detailed information on all known protein folds, and on all structurally
close relatives to a given protein. The SCOP database is also a starting point
for the construction of more specific databases, of which we will mention a
few later on in the survey.
As of 2000, the database contained information on some 12,000 entries from
the PDB, with relative folds, families and superfamilies statistics. A hidden
Markov Model for SCOP superfamilies has recently been proposed by Gough
et al [19]. One of the programs used in the partially automatic classification
process in SCOP is 3dSEARCH [51], a procedure based on Geometric Hashing,
which is a technique originally developed in the field of Computer Vision. We
will shortly describe the technique and the algorithm later in this paper.
3
Figure from http://www.rcsb.org/pdb/holdings.html
8 Giuseppe Lancia and Sorin Istrail

HOMSTRAD

The HOMologous STRucture Alignment Database (HOMSTRAD, [39]) is a

curated database of structure-based alignments for homologous protein fami-
lies. It originated from a small database of seven family alignments [44], which
has been gradually extended and updated over the years. This database pro-
vides structural alignments in various formats, annotated and displayed by
using the program JOY [38] which was developed to highlight structural fea-
tures. These alignments have been carefully examined by manual editing and
are organized in homologous families, i.e., either having a common evolution-
ary origin, or a high percentage of sequence identity (most of the families
have on average more than 30 percent identity). The classification scheme
adopted by HOMSTRAD is a combination of many others, among which
those adopted by SCOP. The sequence alignment programs used to help in
the classification are BLAST and FUGUE. In HOMSTRAD, the known pro-
tein structures are clustered in evolutionary related families, together with a
sequence representative of each familiy. These representatives are computed
on the basis of common 3D features by using structure alignment programs
such as MNYFIT and STAMP, described later in this article. The database
also provides superimposed structures and links to other databases and align-
ment/comparison programs. The use of HOMSTRAD and JOY is suitable
for comparative modelling and accurate structure alignments, while SCOP is
more suited for the hierarchical classification of protein structures.

SLoop

The basis of the SLoop database [7] is in the work of Donate et al. [11], who de-
scribed a classification of protein loops and loop regions. SLoop is a database
of super secondary fragments, in which proteins are clustered according to
the similarity of their secondary structures, in particular the lenght and type
of loop regions. The clustering of structures derives form a two–stage pro-
cess. First, the program SSTRUC is used to identify the loop regions, and
each loop is separated into groups according to the length and type of the
bounding secondary structures. Then, the loops within each group are pair-
wise superimposed, the relative similarities are computed and then used in a
classical clustering scheme.

CAMPASS

The CAMbridge database of Protein Alignments organised as Structural Su-

perfamilies (CAMPASS, [53]) database is a collection of structure–based se-
quence alignments of protein(domain)s belonging to a set of superfamilies.
Currently, it is restricted to proteins sharing the same topology, and an arbi-
trary cut-off of 25% sequence identity has been used to eliminate homologous
entries within a superfamily. The superfamilies have been chosen according
 Protein Structure Comparison: Algorithms and Applications 9

to the SCOP database and the literature. Individual domains within multi-
domains proteins have been considered separately. Altough this database is
founded on primary structure similarity, it fits within the scope of this survey
since for the distantly related members of each superfamily, simple multiple
sequence alignment procedures are not appropriate. Hence, the alignment of
superfamily members is based on the conservation of structural features like
the presence of secondary structures, hydrogen bonding and solvent accessi-
bility. For this type of alignments, the program COMPARER [49] has been
used, which takes into account structural information. In the present version,
the database consists of 52 superfamilies for which structure-based sequence
alignments are available.

FSSP

The Fold classification based on Structure–Structure alignment of Proteins

(FSSP, [26]) database provides all–against–all structure comparison of all cur-
rent PDB structures with more than 30 residues. The alignment and clusters
are updated continuously via the DALI search engine (described later in this
article). The classification is fully automatic. As of this writing, the database
contains some 2,700 sequence families, out of a population of roughly 25,000
protein structures.
The clustering of all proteins in the PDB is done by first identifying a
set of “representative” folds (which are highly dissimilar) and partitioning all
the remaining proteins into homologs of the structures in the representative
set (where a sequence is considered a homolog of another if they have at
least 25% sequence identity). Then, an all–againts–all comparison is performed
on the sequences in the representative set, this way inducing a clustering of
all structures and, implicitely, their homologs. The FSSP entries include the
resulting alignments as well as structure alignments of the representatives and
their homologs.

3.2 Fold Prediction and Assessment

The problem of protein fold prediction is perceived as one of the core ques-
tions, albeit extremely complex, for the molecular biology and gene therapy
of the 21st century. The amount of time and work required to determine a
protein structure experimentally makes the design of faster yet highly reli-
able methods a compelling need. One of the possibilities which is extensively
studied is the creation of algorithms capable of computing the final fold of
a protein from its aminoacid sequence. This is a challenging computational
problem and the prize for its solution includes, beyond the obvious scientific
and medical implications, also relevant economical interests. For this reasons,
some companies among which IBM, are devoting a good deal of research into
this problem. In 1999 IBM announced a $100 million research project, named
Blue gene [2], for the development of a system capable of more than one
10 Giuseppe Lancia and Sorin Istrail

petaflop (1015 operations per second) which will tackle the problem of protein
folding by simulating the actual physical phenomenon in a massive, parallel,
computation.
The computational problem of protein fold prediction is beyond the scope
of this paper. For a good introduction to the topic see, e.g., [34]. What is
relevant for our purposes is the assessment of the prediction quality, i.e., how
can we judge if a prediction is “close enough” to a model? Clearly, we need
tools of structure comparison which, given a model determined experimentally,
and a prediction of the former, are capable to come up with a score (possibly,
a real number in the range [0, 1], with 0 meaning completely wrong, 1 meaning
completely right).
We now describe some of the situations where this problem occurs and
how it is currently solved.

CASP and CAFASP

The Critical Assessment of techniques for protein Structure Prediction (CASP

[40, 41]) is an international bi–annual exercise in which various research groups
try to predict the 3D structure of some proteins from their aminoacid se-
quences. The experiment consists of three phases: First the prediction targets
are collected from the experimental community; then the predictions are col-
lected from the modeling comunity; finally, the assessment and discussion of
the results takes place. Depending on the method used to predict, there are
different categories, i.e. Comparative Modeling, Fold Recognition or Thread-
ing, Docking and ab initio Folding. The rules of the exercise do not forbid
human intervention in making the predictions. In fact, the most successfull
teams so far are those who can annoverate some of the best experts in the field.
Similarly, the assessment of the results is not done automatically, but, again,
human expertise is required. Thus, for each prediction category, a group of
evaluators is asked to compile a ranking of the teams. Some automatic meth-
ods are employed as well, with the caveat that measures such as RMSD and
other rigid–body superposition methods have been found unsuited for the
recognition of well predicted substructures or domains. Since it has been so
far unclear even if a single, objective, quantitative assessment based on just
one measure can exist, in the CASP experiments a large number of scores
have been employed, but the final word is left to the expert evaluation and
human visualization. Some of the assessment techniques are described in [35].
The large number of models submitted (11136 in CASP4) underlines the ever
increasing importance of automatic measures.
CAFASP [12] is a Fully Automated CASP, in which human intervention is
forbidden. In this competition, held in parallel with CASP, both predictions
and evaluations must be performed by unsupervised programs. Although it is
widely believed that human-expert predictions and assessments can be signif-
icantly more accurate than their automatic counterparts, one of the goals of
CAFASP is to push the research for reducing the gap between the two. The
 Protein Structure Comparison: Algorithms and Applications 11

main algorithm used in CAFASP for scoring the predictions is MaxSub, de-
scribed later in this paper. This algorithm was partially validated by scoring
some of the CASP3 Fold-Recognition (FR) models and comparing the rank-
ings obtained to the human–expert assessment carried out by CASP jurors.
Even if some differences were observed, the top six predicting groups ranked
by MaxSub were the same as those ranked by CASP3. On the other hand,
this algorithm had very little success in comparing hard FR targets to their
models, because it fails to detect local structural similarities. In the words of
CAFASP organizers [13], even for a target with a known fold, the fact that
MaxSub scored a prediction with zero does not necessarily imply that the pre-
diction is totally wrong. (...) Because of the relatively low sensitivity of our
automated evaluation, a “human expert” is required to learn more about the
prediction capabilities of the servers.

Live Bench

The Live Bench Project [6] is an effort aimed at the continuous benchmark-
ing of protein structure prediction servers. This project implements an au-
tomatic large–scale assessment of the performance of publicly available fold–
recognition/prediction servers, such as PDB–Blast, GenTHREADER, FFAS,
T98-lib, 3D-PSSM and INBGU. The goal is to provide a consistent, automa-
tized framework in which such servers can be evaluated and compared. Every
week, all new PDB protein structures are submitted to all participating fold
recognition servers. For instance, in the period October 1999–April 2000, 125
targets were used for comparisons, divided into 30 “easy” structures and 95
“hard” ones. The results of the predictions are compared using similar algo-
rithms as in CAFASP, i.e., MaxSub and Lgscore. Note that, differently from
CASP and CAFASP, these “predictions” are in fact done after the struc-
tures are made known and available on the PDB, and hence the reliability of
LiveBench is based on the assumption that the evaluated servers do not use
any hidden feature directing the prediction towards the correct answer. One
of the main results of the LiveBench project was to recognize that the servers
largely disagree in many cases (they were able to produce structurally similar
models for only one half of the targets, and for only one third of the targets
such models were significantly accurate) but that a “combined consensus”
prediction, in which the results of all servers are considered, would increase
the percentage of correct assignments by about 50%.

4 Software and Algorithms for Structure Comparison

In this section we review some of the existing algorithms for protein struc-
ture comparison. We start with the RMSD measure, which is at the basis of
other algorithms described, such as MaxSub, GDT, 3dSEARCH and MNY-
FIT. The section includes both sequence order dependent and sequence order
12 Giuseppe Lancia and Sorin Istrail

independent algorithms. In the latter we find DALI, Lgscore, 3dSEARCH and

STAMP.
A general comment which applies to almost all of the algorithms described
is a lack of rigorousness from a Theoretical Computer Science point of view.
That is, these algorithms are usually heuristic procedures using some sort of
local search to optimize a certain objective function. The algorithms that we
surveyed are not “approximate” as intended in Combinatorial Optimization
(i.e., there is no guarantee that the solution found will be within a certain
factor of the optimum, or of a bound to the optimum, which is never stud-
ied), but are mostly fairly simple heuristics which use “common sense” to
prune the search space. That is, they exploit some reasonable assumptions
on the properties that the optimal solution should have (which, however, are
not proved to always hold) to discard partial solutions from further consid-
erations. This way of proceeding makes certainly sense for hard problems
such as these, and in fact, the ideas on which Branch-and-Bound is based
are somewhat similar. However, Branch-and-Bound is an exact and rigorous
method, which guarantees an optimal solution or, if stopped before reaching
one, returns the maximum error percentage in the best solution found. We
think that Branch-and-Bound should also be considered, at least to tackle the
smaller instances, as a viable option for these problems. Also, in the realm
of local search heuristic procedures, there are several sophisticated paradigms
that have proved effective for many hard combinatorial problems. In partic-
ular, Genetic Algorithms, Tabu Search with its variants, Simulated Anneal-
ing, Randomized Algorithms and various Metaheuristics. Given the relatively
small effort required in implementing such searches (most of these procedures
are available as general purpose libraries and the user has to implement only
one or two core modules) we think that more work should be devoted into in-
vestigating which of these techniques is more suitable for a particular problem
at hand.

RMSD

Given n points (a1 , . . . , an ), with ai = (ai1 , ai2 , ai3 ) and n points (b1 , . . . , bn ),
a “rigid body superposition”, is a composition of a rotation and a traslation,



that takes the points ai into a i and makes the sum of differences |a i − bi | as
small as possible.
More specifically, let ∆ be defined by

∆2 = min |Rai + t − bi |,
R,t
i

where R is a rotation matrix (i.e., det R = 1) and

√ t a translation vector. The
Root Mean Square Deviation of a and b is ∆/ n.
The optimal traslation t is determined easily, since it can be proved that
it must take the center of mass of the ai points into the center of mass of the
 Protein Structure Comparison: Algorithms and Applications 13

bi s. To determine the

n optimal R one must consider the correlation matrix A
defined by Aij = h=1 ahi bhj , as it can be shown that the optimal rotation
maximizes Tr(RA). One way to maximize Tr(RA), exploits the fact that the
problem is three dimensional. Represent the matrix R as

R = 1 + sin θM + (1 − cos θ)M 2

where
 
0 n −m
M =  −n 0 l 
m −l 1
and θ represent the rotation angle around an axis in the direction of the
unit vector u = (l, m, n). It then follows that

Tr(RA) = Tr(A) + Tr(M A) sin θ + (Tr(A) − Tr(M 2 A)) cos θ

and hence


max Tr(RA) = Tr(A) + (Tr(M A))2 + (Tr(A) − Tr(M 2 A))2 .
θ

So, for a fixed axis of rotation, the calculation of the optimal angle θ is
immediate [36].
Alternative ways of determining the optimal rotation matrix R, based on
the computation of the eigenvalues of AT A, or on the quaternion method, are
described in [31].
As previously mentioned the RMSD is widely used as a sequence depen-
dent measure of 3D similarity, i.e., to find an optimal mapping of points of a
structure into a predetermined set of corresponding points in another struc-
ture. Despite some known flaws, the RMSD measure is at the hearth of many
other measures. The main problem with RMSD is that it is more sensitive to
small regions of local differences than to large regions of similarity. In fact,
small distances between well-matched atoms have a much lesser impact on
the RMSD than do the very large distances between poorly matched Cα s.
Another issue with RMSD is that it is not designed to be equivalent be-
tween different targets. For instance, the quality of a model with an RMSD
of a 4 Å for a 40 residues long target is not the same as that of a model of 4
Å RMSD over 400 residues.
The problem of local different regions affecting the total score is common
to other global measures, but for the RMSD value the effect is larger than for
other scores, e.g., the contact map similarity. In this situation, a solution is
typically to first calculate the similarity for segments of the protein and then
define a normalized score based on the number of residues in the segments.
However, the right relationship between the length of the segment and the
14 Giuseppe Lancia and Sorin Istrail

value must be defined, and this is not a trivial problem. There are two con-
flicting issues: we are interested in large segments that are highly similar, but
the similarity and the length of the segments are inversely proportional.
Similar problems are faced by “trimming” methods which, given a current
alignment, re-define iteratively a core of pairs of residues that are matched
within a small distance, work on these only, and then try to extend the align-
ment. The threshold and the degree of trimming are to some extent arbitrary,
and this choice affects the final outcome.

DALI

In DALI [25], the structural alignment of two proteins is done indirectly, not
by comparing the actual structures, but their distance matrices. The distance
matrix of a folded protein P is the matrix DP = (dP ij ) of the Euclidean dis-
tances between all pairs (i, j) of its residues. The matrix provides a 2D repre-
sentation of a 3D structure, and contains enough information for retrieving the
actual structure, except for overall chirality [22]. The idea underlying DALI is
that if two structures are similar, then their distance matrices must be similar
too. An analogous idea is used to compare structures via their contact maps
and is described later in this article. In order to find similarities between two
distance matrices, DA and DB , DALI uses a heuristic trick, and looks for all
6 × 6 submatrices of consecutive rows and columns in DA and DB (that is, it
maps iA , iA + 1, . . . , iA + 5 into iB , iB + 1, . . . , iB + 5, and j A , j A + 1, . . . , j A + 5
into j B , j B + 1, . . . , j B + 5) to find patterns of similarity. This is done in the
first step of the algorithm. The set {iA , . . . , iA + 5, j A , . . . , j A + 5} with its
counterpart in B, is called a contact pattern. The second step attempts at
merging the contact patterns into larger, consistent, alignments. An align-
ment of two proteins A and B is a one-to-one function M of some residues
MA (the matched ones) of A into some residues of B. DALI tries to determine
the alignment that maximizes the global objective function

S(M ) = φ(i, j, M (i), M (j)).
i,j∈MA

This objective function, which looks only at matched residues, depends on

the particular φ used, which in turn depends on the Cα –Cα distances dA ij and
B
dM(i),M(j) . The two main forms for φ used by DALI are a rigid and an elastic
ij − dM(i),M(j) |,
score. The rigid score is defined by φ(i, j, M (i), M (j)) := θ − |dA B

where θ = 1.5Å is the zero level of similarity. The more complex elastic score
uses relative, other than absolute deviations, and an envelope function to
weigh down the contribution of long distance matched pairs.
Some heuristic rules are employed to speed up the search. For instance,
only a subset of all contact patterns is considered. Also, overlapping contact
patterns (they could possibly overlap by 11 out of 12 residues) are suppressed
by partitioning the protein in nonoverlapping structural elements and merging
 Protein Structure Comparison: Algorithms and Applications 15

repetitive segments. Further rules to prune the search involve the suppression
of pairs of patterns for which the sum of distances of one is not within a given
interval of the same value in the other. Other rules of the same nature are
employed which we do not describe here. The final optimization is done via a
Monte Carlo algorithm, which is basically the local search strategy known as
Simulated Annealing. In this search a neighborhood of a solution is explored
for a good move, leading to a new solution. A move which takes into a solution
worse than the current one can be accepted, but with probability inverse to the
degradation in the solution quality. The temperature of the system is also used
to favour or make more difficult such non–improving moves. DALI uses two
basic moves, named expansion and trimming ones. Expansion moves try to ex-
tend a current partial alignment by adding to it a contact pattern compatible
with one of the currently matched residues (i.e., containing the same match),
and then possibly removing some matches that have become noncompatible.
Trimming moves simply remove from a current alignment any subalignment
of 4 elements that gives negative contribution to the total similarity score.
These moves are alternated as one trimming cycle every five expansion cycles.
DALI was used to carry out an all–against–all structure comparison for
225 representative protein structures from the PDB, providing the basis for
classification in the FSSP database. Also, the DALI server can be used to
submit the coordinates of a query protein structure which is then compared
against the PDB.

MaxSub

MaxSub [50] is an algorithm explicitely developed to be used for the automatic

assessment of protein structure similarity (in particular within the CAFASP
exercise). Therefore it was written with the goals of (a) being simple and
objective, (b) producing one single number (score) that measures the amount
of similarity between two structures and (c) performing similarly to human-
expert evaluations. These goals were partly met, and the measure was partially
validated when, on a particular category and subset of predictions, it ranked
in the top the same groups that were ranked in the top by the CASP human
evaluators. A prediction can be evaluated in different aspects. For example,
in the Homology Modeling category one may want to score the accuracy of
the loops or the side chain packing. Or, in the Fold Recognition category, one
may want to evaluate wether a compatible fold was recognized, regardless of
the quality of the alignment obtained. MaxSub, in particular, focuses on the
quality of the alignment of the models to the target. It is a sequence–dependent
assessment, i.e., only corresponding residues are compared. The result is a
single scalar in the range of 0 (completely wrong) to 1 (perfect model). The
scalar is a normalization of the size of the largest ”well–predicted” subset,
according to a variation of a formula devised by Levitt and Gerstein [32].
MaxSub is also similar to GDT, a method developed by Zemla et al. [59] for
being used in CASP3, which attempts to find a large set of residues which
16 Giuseppe Lancia and Sorin Istrail

can be superimposed over the experimental structure within a certain error.

GDT is briefly described later in this survey.
The input to MaxSub are the 3D coordinates A = {a1 , . . . , an } and B =
{b1 , . . . , bn } of the Cα atoms of a prediction and a target. A solution is a
subset M ⊂ {1, . . . , n} (called a match) such that the pairs (aj , bj ) for j ∈ M
can be superimposed ”well-enough” (i.e., below a given threshold d) by a
transformation T (rotation and translation). The objective is to maximize
the size of M , i.e., find the largest subset of residues which superimpose well
upon their correspondents.
Given M , the value of the best superposition can be easily found, (analo-
gously to the RMSD determination previously described), by using the trans-
formation TM which minimizes, over all T ,


j∈M ||aj − T (bj )||
RM S(M ) =
|M |
where || · || is the Euclidean norm.
The optimization of M is performed by a simple heuristic algorithm. This
algorithm is based on the reasonable (although not rigorous) assumption that
a good match must contain at least L ≥ 4 consecutive pairs, i.e. {(ai , bi ), i =
j, j + 1, . . . , j + L − 1} for some j. So the idea is, for a fixed L, to look, in time
O(n), for all matches of L consecutive elements, and try to extend each of
them. The extension of a match M is also done heuristically, in 4 iterations,
where at each iteration k the pairs (ai , bi ) for which ||ai − TM (bi )|| < kd/4
are added to M .
Let M ∗ be the best match found by the above procedure. In order to
return a single, normalized score, a final value S is computed as

 1
S(M ∗ ) = /n,
i∈M ∗
1 + ( ddi )2

where di = ||ai − TM ∗ (bi )||. In our opinion, there are several non rigorous
steps that should be addressed in evaluating MaxSub and its reliability. For
instance, although the basic objective optimized by the algorithm is maximize
|M | such that di ≤ d for all i ∈ M , the optimization of S (which
is the score
1
eventually returned) calls ultimately for maximize over all M , i∈M d2 +d i
2

which is a different objective with possibly a different optimum. Also, there

are many arbitrary parameters, such as L, the number 4 of iterations in the
extension phase, and the threshold d. Some experiments have shown that the
algorithm does not depend on the choice of d too much, while the dependance
on the other parameters is not described with the same accuracy. As far as
how close the solution is to the optimum, a comparison to the best of a random
set of 70,000 matches was used. Given the astronomical number of possible
solutions, it is arguable if such method can be used to conclude anything
about the actual performance of the algorithm.
 Protein Structure Comparison: Algorithms and Applications 17

Lgscore

The measure Lgscore [9] is used in LiveBench and CAFASP for the automatic
assessment of fold recognition problems. This measure is statistically based,
and relies on the following formula, by Levitt and Gerstein [32], for the simi-
larity of two structures, after a superposition in which M denotes the aligned
residues:

 1 Ng
Sstr (M ) = K 2
− ,
1 + (di /d0 ) 2
i∈M

where K and d0 are constants (usually set to 10 and 5Å), di is the distance
between the aligned pairs i of Cα atoms, and Ng is the number of gaps in the
alignment. The P–value of a similarity is the likelihood that such a similarity
could be achieved by chance. Levitt and Gerstein showed how to compute the
P–values for a distribution of Sstr depending on the length of the alignment.
Lgscore is the negative log of the P–value for the most significant subpart of
the alignment. In order to find such most significant segment, two heuristic
algorithms are used, i.e. “top–down” and “bottom–up”. The top–down ap-
proach consists of a loop in which (1) a superposition is done of all residues
that exist in the current model and the target; (2) the P–value of this super-
position is stored and the residues that are furthest apart are deleted in the
model and the target. The loop is repeated as long as there are at least still
25 residues. The alignment with the best P–value is returned. The bottom–up
approach essentially tries to match a fragment i, . . . , i+j of j ≥ 25 consecutive
residues in the model and the target, for different values of i and j and re-
turns the best P–valued fragment. None of the two approaches dominates the
other, although the authors report that the bottom–up algorithm found the
best subset in most cases. The arbitrary value of 25 residues as a threshold for
fragment length is justified in [9] since “short segments are given unrealistic
good P–values”.

GDT

GDT (Global Distance Test, [59]) is an algorithm for identifying large sets of
residues (not necessarily continuous) in a prediction and a target which do
not deviate more than a given cutoff value. The algorithm is heuristic, and
similar to Lgscore. It consists of a loop, started from the alignment between
model and target of the longest continuous fragments within the cutoff value,
found by enumeration, and all three–residue segments. At each iteration, an
RMS transfom is computed for the given alignment, and the pairs of residues
whose distance is greater than the threshold are removed. The loop is ended
as soon as no residues are removed by one iteration.
18 Giuseppe Lancia and Sorin Istrail

Geometric Hashing and 3dSEARCH

Comparing protein structures and, more generally, querying databases of 3D

objects, such as the PDB, can be regarded as special cases of Object Recog-
nition problems in Computer Vision. Therefore, algorithms originally devel-
oped in the field of Computer Vision have now found a suitable application
in Structural Genomics as well. Geometric hashing [28] is an example of such
a technique. Geometric Hashing can be used to match a 3D object (a pro-
tein structure) against one or more similar objects, called “models” (e.g. a
database organized on structure). The key idea is to represent each model
under different systems of coordinates, called reference frames, one for each
triplet of non collinear points of each model. In any such system, all points
of the model (which are vectors of reals) can be used as keys (i.e., indices
in a hash, or look–up, table) to store the information about the model they
belong to. In a preprocessing phase, all models are processed this way, and
an auxiliary, highly redundant, hash table is created. In the querying phase,
executed each time a target is to be matched against the models, the follow-
ing steps are taken. For each reference frame of the target, the points of the
target (in the coordinate system defined by the reference frame) are used to
access the hash table and hence retrieve some of the models. These models
define a match list for the particular reference frame. A match list gives a set
of alignments of the points in the target and in each of the models retrieved.
The match lists are then merged in order to find larger alignments. The final
step is to compute an RMSD transformation for each alignment obtained, so
as to find the model most similar to the target.
The program 3dSEARCH [51], used by the database SCOP, implements
the strategy of geometric hashing on the secondary structure representation
of each protein.

MNYFIT

The program MNYFIT [54], which is used by the database HOMSTRAD,

belongs to a suite of programs called COMPOSER, for homology modeling
of protein structures. With COMPOSER one can work out the construction
of a predicted 3D structure based on one or more known homologous ones.
MNYFIT is a core module within this software collection, and can handle up to
10 molecules of about 500 residues each. It performs the superposition of two
or more related protein structures, individues structurally equivalent residues
and their location, and with this information is able to define a common
framework for a set (family) of proteins.
The main algorithm used by MNYFIT is the least squares fitting procedure
of McLaughlan [36] for rigid body superposition, akin to the RSMD procedure
discussed above.
The superpostion is performed initially using at least three atoms from
the Cα backbones of each molecule, which occupy topologically equivalent
 Protein Structure Comparison: Algorithms and Applications 19

positions in all the molecules to be fitted. For two or more superimposed

structures, a threshold is used to define which of the Cα atoms are consid-
ered structurally equivalent (i.e. those whose distance in the superposition
is smaller than the threshold). Then, the alignment is iteratively updated
with the double objective of increasing the number of structurally equivalent
atoms while at the same time keeping a low Root Mean Square Deviation
between the equivalenced atoms in the superimposed structures. When more
than two molecules are superimposed, the objective is to determine a con-
sensus structure (called “framework”), which should be as close as possible
to all involved structures, and which represents the average positions of the
structurally equivalent atoms of all superimposed molecules. The framework
returned from the final step of the procedure is used by COMPOSER for
further model–building operations.

STAMP
One of the programs used by the database HOMSTRAD for the detection
of homologous protein families is STAMP [48], a program for the alignment
of protein sequences based on their 3D structures. The core of this program
contains an implementation the basic Smith and Waterman dynamic pro-
gramming procedure for sequence alignment [52], but using suitable similarity
scores which express the probability of residue–residue structural equivalence.
These scores are computed according to the equation by Argos and Rossmann
[1]

d2ij s2ij
Pij = exp exp ,
−2E12 −2E22
where dij is the distance between the Cα atoms of residues i and j, and
sij measures the local main chain conformation. The Smith and Waterman
procedure is embedded in a loop as follows: at each iteration, the output of
the dynamic program is a residue–residue equivalence (the sequence align-
ment). This list of equivalences is used to compute a best fit transformation
minimizing the RMSD, via the least square method of McLaughlan. The new
set of coordinates and residue–residue distances, obtained under this trans-
formation, are then used to recompute the similarity score values, which are
then used for another round of the Smith Waterman procedure. The loop is
repeated iteratively until the alignment becomes stable.
The method has proved experimentally effective, allowing the generation
of tertiary structure based multiple protein sequence alignments for a vari-
ety of protein structural families. However, the method is only effective for
proteins which share a good deal of global topological similarity, while fails if
applied to, e.g., proteins with common secondary structures but with different
connectivity, orientation or organization.
STAMP allows for the specification of a minimum number of equivalent
residues to be matched in two structures, the reversals of strand directions, the
20 Giuseppe Lancia and Sorin Istrail

swapping of sequence segments and more. The output contains two measures
of alignment confidence: a “structural similarity score”, which can also be used
to measure the functional and evolutionary relationship, and an “individual
reside accuracy” which is intended to measure the quality of the topological
equivalence of the pairs of aligned residues.

5 Problems Based on Contact Map Representations

A contact map [33, 21] is a binary version of the distance matrix representation
of a protein structure. More specifically, the contact map of a folded protein
of n residues is a 0-1, n × n matrix C, whose 1–elements correspond to pairs
of amino acids in three–dimensional “contact”. A contact can be defined in
many ways. Typically [37], one considers Cij = 1 when the distance of two
heavy atoms, one from the i–th aminoacid and one from the j–th aminoacid of
the protein, is smaller than a given threshold (e.g., 5Å). The framework of the
contact map representation of proteins is very appealing, since this intuitive
and fairly simple representation is already complex enough to capture the
most important properties of the folding phenomenon. It has been shown
that it is relatively easy to go from a map to a set of possible structures to
which it may correspond [21, 56]. This result has opened the possibility of
using contact maps to predict protein structure from sequence, by predicting
contact maps from sequence instead. Vendruscolo and collaborators, among
others, have looked at the problem of devising an energy function based on
contacts, which should be minimized by the protein’s native state contact map
[57, 45]. For this purpose, they have set up a system of linear inequalities, with
20 × 20 variables Cab for all pairs of aminoacids a and b, which represent the
weight to give to a contact between the aminoacids a and b. The inequalities
are built as follows. Given the contact map of a correct structure r, there
is an inequality for any alternative structure w over the same sequence of
aminoacids, imposing that the energy of r is lower than that of w. Alternative
structures are obtained, e.g., by threading through other known folds. The
results are that “a simple pairwise contact energy function is not capable of
assigning the lowest energy to an experimentally determined structure” [45],
but by using corrective factors, such as a distinction between contacts on the
surface or in the core, and simple distance–dependant interaction weights, one
can achieve contact potentials which are in fact often stabilized by the native
contact maps. The use of energy function minimization to predict contact
maps is just one possible way. To the best of our knowledge, very little success
has been met so far in the contact map prediction problem. It is possible that
the research on this question will be boosted by the fact that the competition
CAFAPS has recently introduced the new “contacts” prediction category.
The statistics of contact maps have been studied as well, and it has been
shown that the number of contact maps corresponding to the possible config-
urations of a polypeptide chain of n residues, represented by a self–avoiding
 Protein Structure Comparison: Algorithms and Applications 21

G1

G2

Fig. 2. An alignment of value 5

walk in the d–dimensional lattice, grows exponentially with n for all d ≥ 2

[58].

5.1 Using Contact Maps for Structure Comparison

Besides their use for protein fold prediction, contact maps can be exploited
to compare 3D structures. The basic idea is fairly obvious: if two structures
are similar, we should expect their contact maps to be similar as well, and
conversely. Hence, we can use an indirect method for structure comparison,
i.e., contact map comparison instead. In our previous work [29] we have de-
signed an exact algorithm based on an Integer Programming formulation of
this problem, which we will now review.
We can regard the contact map of a protein p as the adjacency matrix of a
graph Gp . Each residue is a node of Gp , and there is an edge between two nodes
if the the corresponding residues are in contact. The Contact Map Overlap
(CMO) problem, calls for determining a sequence–independent alignment of
some residues in the first protein (nodes in G1 ) with residues of the second
protein (nodes in G2 ) which highlights the largest set of common contacts as
follows: The value of the alignment is the number of contacts (i.e., edges) in
the first map whose endpoints are aligned with residues that are also in contact
in the second map. This value is called the overlap for the two proteins, and
the optimization problem is to find the maximum overlap. Figure 5.1 shows
two contact maps and a feasible alignment.
This measure was introduced in [15], and its optimization was proved
NP-hard in [18], thus justifying the use of sophisticated heuristics or Branch–
and–Bound methods.
Some of the most powerful algorithms for finding exact solutions of com-
binatorial optimization problems are based on Integer Programming (IP) [43].
The IP approach consists in formulating a problem as the maximization of a
linear function of some integer variables and then solving it via Branch–and–
Bound. The upper bound comes from the Linear Programming (LP) relax-
ation, in which the variables are not restricted to be integer, and is polyno-
22 Giuseppe Lancia and Sorin Istrail

mially solvable. When the LP–relaxation value is close to the value over the
integers, then the bound, and hence the pruning of the search space, is effec-
tive. In order to obtain good bounds, the formulation is often reinforced by
the use of additional constraints, called cuts (from which the approach name,
Branch–and–Cut): these are constraints that do not eliminate any feasible in-
teger solution, but make the space of fractional solutions smaller, this way
decreasing the value of the LP bound. In many cases a good IP formulation
requires an exponential number of constraints and/or cuts. This would make
its practical solution impossible, unless there is a way to include all of them
only implicitely. This way exists, and works as follows. Given an LP fractional
solution x∗ and an inequality ax ≤ b, we say that the inequality is violated
by x∗ if ax∗ > b. If we have an exponential number of inequalities, we can
solve the LP with only a (small) subset of them, obtain a solution x∗ and
then check if any of the (exponentially many) inequalities that were left out
is violated by x∗ . If not, the current solution is optimal with respect to the
full formulation, otherwise, we can add the violated inequality to the current
LP and iterate the process. The check for a violated inequality is called sepa-
ration and is carried out by a separation algorithm. By a fundamental result
of Grötschel, Lovász and Schrijver [20], the existence of a polynomial–time
separation algorithm is a necessary and sufficient condition for solving the
whole exponential–sized LP relaxation in polynomial time.
The CMO problem can be reduced to a (very large) Maximum Indepen-
dent Set (MIS) problem on a suitable graph. An independent set is a set of
vertices such that there is no edge between any two of them. The MIS is a
classic problem in combinatorial optimization which has a definition nice and
simple, but is one of the toughest to solve exactly. However, by exploiting the
particular characteristics of the graphs derived from the CMO problem, we
can in fact solve the MIS on graphs of 10,000 nodes and more.
The natural formulation of the MIS as an IP problem has
a binary variable
xv for each vertex v, with the objective of maximizing v xv , subject to
xu + xv ≤ 1 for all edges {u, v}. This formulation gives a very weak bound,
but

it can be strengthened by the use of clique inequalities cuts, such as
v∈Q xv ≤ 1, which say that any clique Q can have at most one node in
common with any independent set. The addition of these constraints can lead
to tight formulations. This is exactly the case for the CMO problem. In [29]
we formulated the CMO problem as an IP and solved it by Branch–and–
Cut, where the cuts used are mainly clique–inequalities. Although there is an
exponential number of different clique inequalities, we can characterize them
completely and separate over them in fast (O(n2 )) polynomial time. Note that
finding cliques in a graph is in general a difficult problem, but in this case
we can solve it effectively since it can be shown that the underlying graph
is perfect. The following section gives some details on the formulations and
results from [29]. Further details and full proofs can be found in the original
paper.
 Protein Structure Comparison: Algorithms and Applications 23

IP Formulation
We can phrase the CMO problem in graph–theoretic language as follows:
We are given two undirected graphs G1 = (V1 , E1 ) and G2 = (V2 , E2 ), with
ni = |Vi | for i = 1, 2. A total order is defined on V1 = {a1 < . . . < an1 } and
V2 = {b1 < . . . < bn2 }. It is customary to draw such a graph with the vertices
arranged increasingly on a line. We denote an edge by an ordered pair (i, j),
with a tail in the left endpoint and a head in the right endpoint.
A non–crossing alignment of V1 in V2 is defined by any two subsets of the
same size k, {i1 , . . . , ik } ⊆ V1 and {u1 , . . . , uk } ⊆ V2 , where i1 < i2 . . . < ik
and similarly for the uh ’s. In this alignment, uh is aligned with ih for 1 ≤
h ≤ k. Two edges (contacts) (i, j) ∈ E1 and (u, v) ∈ E2 are shared by the
alignment if there are l, t ≤ k s.t. i = il , j = it , u = ul and v = ut (see
Figure 5.1). Each pair of shared edges contributes a sharing to the objective
function. The problem consists in finding the non–crossing alignment which
maximizes the number of sharings.
An alignment corresponds to a set of lines connecting nodes of V1 and V2
in the usual drawing with V1 drawn on the top and V2 on the bottom. We
denote such a line for i ∈ V1 and j ∈ V2 by [i, j]. We say that two lines cross
if their intersection is a point. The sharings (e1 , f1 ) and (e2 , f2 ) can be both
achieved by an alignment if and only if they are compatible, i.e. no two of the
lines betweens the tails of e1 and f1 , the tails of e2 and f2 , the heads of e1
and f1 and the heads of e2 and f2 cross. A set of sharings is feasible if the
sharings are all mutually compatible, otherwise it is infeasible. Similarly we
define a feasible and infeasible set of lines. If we draw the lines connecting the
endpoints of an infeasible set of sharings, we have an infeasible set of lines.
We denote by yef a binary variable for e ∈ E1 and f ∈ E2 , which is 1 iff
the edges e and f are a sharing in a feasible solution. The objective function
of CMO is

max yef , (1)
e∈E1 ,f ∈E2

and the constraints are

ye1 f1 + ye2 f2 ≤ 1, (2)

for all e1 , e2 ∈ E1 , f1 , f2 ∈ E2 s.t. (e1 , f1 ) and (e2 , f2 ) are not compatible.
It can be shown that the formulation made of constraints (2) gives a very
weak LP bound, and also contains too many constraints. To strengthen the
bound, we use a new set of variables and strong cuts. We introduce a new
set of binary variables xiu for i ∈ V1 and u ∈ V2 , and constraints forcing the
nonzero x variables to represent a non–crossing alignment. We then bound
the y variables by means of the x variables, so that the edges (i, j) and (u, v)
can be shared only if i is mapped to u and j to v. For i ∈ V1 (and analogously
for i ∈ V2 ), let δ + (i) = {j ∈ i + 1, . . . , n1 : (i, j) ∈ E1 } and δ − (i) = {j ∈
1, . . . , i − 1 : (j, i) ∈ E1 }. Then we have the following constraints:
24 Giuseppe Lancia and Sorin Istrail
 
y(i,j)(u,v) ≤ xiu , y(j,i)(u,v) ≤ xiv (3)
j∈δ + (i) j∈δ − (i)

for all i ∈ V1 , (u, v) ∈ E2 , and analogous constraints for i ∈ V2 and

(u, v) ∈ E1 . We call these activation constraints. Finally, the noncrossing
constraints are of the form:

xiu + xjv ≤ 1 (4)

for all 1 ≤ i ≤ j ≤ n1 , 1 ≤ v ≤ u ≤ n2 s.t. i  = j ∨u = v.
Our IP formulation for the max CMO problem is given by (1), (3), and (4),
where x and y are all binary variables. The cuts in the x variables are described
next.
First, we make clear the connection to the independent set problem. We
define two graphs Gx and Gy . In Gx there is a node Niu for each line [i, u]
with i ∈ V1 and u ∈ V2 and two nodes Niu and Njv are connected by an edge
iff [i, u] and [j, v] cross. Similarly, in Gy there is a node Nef for each e ∈ E1
and f ∈ E2 and two nodes Nef and Ne
f
are connected by an edge iff the
sharings (e, f ) and (e , f  ) are not compatible. Then, a selection of x variables
feasible for all noncrossing constraints corresponds to an independent set in
Gx and a feasible set of sharings is an independent set in Gy . The maximum
independent set in Gy is the optimal solution to CMO. All cuts valid for the
independent set problem can be applied to the x and y variables, and, most
notably, the clique inequalities.
The separation of the clique inequalities works as follows. Given a frac-
tional solution x∗ , we look for the maximum weight clique in Gx , where x∗ij
is the weight of Nij . If the maximum weight clique weighs less than 1, then
there are no violated clique inequalities, otherwise, we can add at least one
such cut to the current LP.
Finding maximum weight cliques is a polynomial problem when the un-
derlying graph is perfect. We can prove that

Theorem 5.1. The graph Gx is perfect.

There are algorithms for finding a max weighted clique in a weakly trian-
gulated graph (as Gx can be shown to be) of time O(n5 ), due to Hayward,
Hoang, Maffray [23] and Raghunathan [46]. However, for this specific graph
we can do better and find max weighted cliques in time O(n2 ), thus making
a huge difference in the practical solution of the problem.
We start by chacterizing all cliques in Gx , i.e. sets of alignment lines which
are all mutually crossing. We define the following notion of a triangle: this is
a set of lines with one common endpoint and a second endpoint in a range of
consecutive nodes, like T (i, j|u) := {[i, u], [i + 1, u], . . . , [j − 1, u], [j, u]} where
i ≤ j ∈ V1 and u ∈ V2 , and T (i|j, u) := {[i, j], [i, j + 1], . . . , [i, u − 1], [i, u]}
where

i ∈ V1 and j ≤ u ∈ V2 . For S a set of lines, by x(S) we denote the value
[i,j]∈S xij .
 Protein Structure Comparison: Algorithms and Applications 25
a1 an a1 an
1 1

b1 bn c1 cn
2 2

Fig. 3. Left: A zigzag path P (bold) and the set T (P ). Right: Same path after
flipping V 2.

Call a1 , an1 , b1 , and bn2 the set of terminal nodes. Consider a path P which
passes through all the terminal nodes, and alternates nodes of V1 and V2 in a
zig–zag fashion: That is, we can orient the path so that a1 is the first of the
nodes of V1 visited by the path, and if ak has been visited by the path, then
all of the nodes in V1 visited after ak are to its right. Similarly, bn2 is the first
of the nodes of V2 visited by the path, and if bh has been visited by the path,
then all of the nodes in V2 visited after bh are to its left. Note that any such
path must start and end at a terminal node (see figure 5.1, Left), and must
always include the lines [a1 , bn2 ] and [an1 , b1 ]. For each node of degree two in
P , a triangle is defined by considering the set of lines incident on the node
and contained within the two lines of the path. Let TA (P ) (TB (P )) be the set
of triangles defined by P with tip in the nodes of V1 (V2 ) having degree two
in P . We define T (P ) := TA (P ) ∪ TB (P ). The following theorem characterizes
the cliques of Gx .

Theorem 5.2. A set Q of lines is a maximal clique in Gx if and only if there

exists a zigzag path P such that Q = T (P ).

The inequalities x(T (P )) ≤ 1 for all zigzag paths P are therefore the
strongest clique cuts for this particular independent set problem. We now
show that to find the most violated such inequality in time O(n2 ). To make
the following argument easier, we rename the nodes of V2 as {c1 , . . . , cn2 }, so
that the leftmost node c1 is bn2 and the rightmost, cn2 , is b1 (i.e., we flip the
nodes of V2 with respect to the usual drawing). A zigzag path P now looks as
a path which goes from left to right both in V1 and V2 . We call such a path a
leftright path (see figure 5.1, Right).
With respect to the new drawing of W , orient each line [a, c] in the two
possible ways and, given a real vector x∗ , define the length for each arc (a, c) ∈
V1 ×V2 and (c, a) ∈ V2 ×V1 as follows: l(a, c) = x∗ (T (a|1, c))−x∗ (T (1, a|c)) and
l(c, a) = x∗ (T (1, a|c)) − x∗ (T (a|1, c)). The lengths of four special arcs are de-
fined separately, as l(a1 , c1 ) = 0, l(c1 , a1 ) = 0, l(an1 , cn2 ) = x∗ (T (an1 |1, cn2 ))
26 Giuseppe Lancia and Sorin Istrail

and l(cn2 , an1 ) = x∗ (T (1, an1 |cn2 )). Now, consider a leftright path P starting
with either the arc (a1 , c1 ) or (c1 , a1 ) and ending with either the arc (an1 , cn2 )
or (cn2 , an1 ). Call l(P ) the standard length of this path, i.e. the sum of arcs
lengths. We then have the following lemma.

Lemma 5.1. For a leftright path P , l(P ) = x∗ (T (P )).

Hence, to find the max–weight clique in Gx , we just need to find the

longest leftright path. This can be computed effectively, by using Dynamic
Programming. Call V (i, j) the length of a longest leftright path starting at
ai and using nodes of V2 only within cj , cj+1 , . . . , cn2 . Also, call V (i, j) the
length of a longest zigzag path starting at cj and using nodes of V1 only within
ai , ai+1 , . . . , an1 . Then we have the following recurrences:

V (i, j) = max{l(ai , cj ) + V (i + 1, j), V (i, j + 1)},

V (i, j) = max{l(cj , ai ) + V (i, j + 1), V (i + 1, j)}.

These recurrences can be then solved backwards, starting at (n1 , n2 ), in
time O(n2 ).

Genetic Algorithms and Local Search

In a Branch–and–Cut algorithm, pruning of the search space happens when-

ever the upper bound ub to the optimal solution of a subroblem is smaller
than a lower bound lb to the global optimum. As described before, ub is the
value of the LP–relaxation. It is important then to have good (i.e., large) lower
bounds as well. Any feasible solution gives a lower bound lb to the optimum.
In our work we have developed heuristics of two types for generating good
feasible solutions, i.e. Genetic Algorithms and Steepest Ascent Local Search.
A generic Genetic Algorithm (GA, [24, 16]) mimics an evolutionary pro-
cess with a population of individuals that, by the process of mutation and
recombination, gradually improve over time. For optimization problems, an
individual is a candidate solution, mutation is a slight perturbation of the
solution parameters, and recombination is a “merging” of two candidate so-
lutions. The use of GAs for Structural Genomics problems is not new, e.g.,
GAs were used for protein fold prediction [27, 10].
Our application of the GA encodes the solutions as sets of alignment edges,
associating a residue in one contact map graph with a residue in the other. A
mutation will slightly shift one edge, and randomly add new edges in any avail-
able space, while recombination will pick edges out of two candidate solutions
and create a new candidate solution using those edges.
Figure 5.1 shows two mutations. The first mutation shifts the alignment
edges on the circled nodes to the right by one position, causing the right-most
edge to be removed and a new edge to be be inserted. The second mutation
shifts the circled edges to the left by one position, causing the left-most shifted
 Protein Structure Comparison: Algorithms and Applications 27

(a) Initial State

(b) After the first mutation

(c) After the second mutation

Fig. 4. The mutation operator. The top alignment shows the state before any
mutations, the middle alignment shows the state after the first mutation but before
the second mutation, the bottom alignment shows the state after both mutations.
Dotted lines are edges that have been removed by a mutation, dashed lines are edges
that have been added after a mutation is performed.

edge to be removed and a new edge to be randomly inserted on the right end
– exactly one of the dashed lines is inserted.
The recombination operator is used to create a new candidate solutions
(the child) from existing solutions. This is done by randomly selecting two
existing solutions (the parents) by a standard GA method, i.e., randomly, but
biased towards good solutions. A set of contiguous edges is randomly selected
from one of the parents and is copied directly to the child. Next, all edges
from the second parent that do not cross edges in the child are copied to the
child. Finally, new edges are added in any available positions, exactly as was
done with mutation.
Our Steepest Ascent Local Search heuristic algorithms follow the standard
approach: Let s be a feasible current solution. The neighborhood of s is the set
of solutions which can be obtained by applying a move to s. If all solutions
in the neighborhood are no better than the current solution s, s is a local
optimum and the search is terminated. Otherwise, the move that results in the
best solution value is applied to s, and the search continues. Since converging
to a local optimum is very fast, the search can be repeated many times, each
time starting from a random feasible solution.
A feasible contact map alignment solution is identified by a pair of lists
of the same size, (A, B), where A = (a1 < . . . < ak ) are nodes from G1 and
B = (b1 < . . . < bk ) are nodes from G2 . The alignment maps residue ai in the
first graph to residue bi in the second graph (see Figure 5.1(a)).
Our two local search algorithms differ only in the definition of the moves
that generate a neighborhood. The first algorithm uses moves that add a
single specific line to the solution, removing any lines that cross the new line.
28 Giuseppe Lancia and Sorin Istrail
a1 a2 a3 a4 a5

b1 b2 b3 b4 b5
(a) Initial State (b) M(7,4)

(c) M+(7,4) (d) M-(6,3)

Fig. 5. Moves for the two algorithms. (a) Starting solution s. (b) A move from the
first algorithm applied to s. (c) An increasing move applied to s. (d) A decreasing
move applied to s.

Formally, the move M (a, b) is defined for all a ∈ G1 − A and b ∈ G2 − B and

will add the line [a, b], removing all lines [aj , bj ] such that a < aj b > bj
or a > aj b < bj (see Figure 5.1(b)). This moves allows for big “jumps” in
the solution space, by introducing very skewed lines and by removing many
lines at once, and is suitable for instances in which the contact maps are not
similar.
The second algorithm uses two types of moves, a decreasing move and an
increasing move. The decreasing move, M − (a, b), defined for a ∈ A and b ∈ B,
simply removes a from A and b from B. Figure 5.1(d) shows the decreasing
move M − (6, 3), which removes a4 and b2 from their respective lists. The
increasing move, M + (a, b), where a and b are as defined for M (a, b) in the
first algorithm, adds a to A and b to B (see Figure 5.1(c)). These moves do
not introduce very skewed lines easily and so are suited for similar proteins,
in which good solutions are made of many parallel lines.

Computational Experiments

Our program has been implemented and run on some proteins from the PDB.
This is the first time that exact solutions have been found for real instances of
this problem. We have run our procedure on a set of 269 proteins, with sizes
ranging from 64 to 72 residues and 80 to 140 contact each. The set was chosen
to contain both a large number of similar proteins, as well as a large number
of dissimilar proteins. An all-against-all computation would have resulted in
36046 alignments; we selected a subset of 597 alignments, so that there would
be an equal number of similar pairs and dissimilar pairs.
In order to perform such a massive computation on a standard, single–
processor Pentium PC, we have limited the time devoted to each individual
instance to one hour (details can be found in [29]). Therefore, some prob-
lems have not been solved to optimality. However, even within the time limit
 Protein Structure Comparison: Algorithms and Applications 29

Fold Family Residues Seq. Sim. RMSD Proteins

1 Flavodoxin-like CheY-related 124 15-30% < 3Å 1b00, 1dbw, 1nat, 1ntr,
1qmp, 1rnl, 3cah, 4tmy
2 Cupredoxins Plastocyanin/ 99 35-90% < 2Å 1baw, 1byo, 1kdi, 1nin,
azurin-like 1pla, 2b3i, 2pcy, 2plt
3 TIM beta/ Triosephosphate 250 30-90% < 2Å 1amk, 1aw2, 1b9b, 1btm,
alpha-barrel isomerase 1hti, 1tmh, 1tre, 1tri,
1ydv, 3ypi, 8tim
4 Ferratin-like Ferritin 170 7-70% < 4Å 1b71, 1bcf, 1dps, 1fha
1ier, 1rcd

Fig. 6. The Skolnick set.

set, we have been able to solve 55 problems optimally and for 421 problems
(70 percent) the gap between the best solution found and the current upper
bound was less than or equal to 5, thereby providing a strong certificate of
near–optimality. The results also show the effectiveness of our lower bound-
ing heuristic procedures, and in particular, of our genetic algorithm. The GA
heuristic turned out to be superior to the others, finding 52 of the 55 optimal
solutions.
In a second test, we used our programs to cluster proteins according to
their Contact Map Overlap. I.e., given a set of proteins, we compute a nor-
malized score based on their best CMO alignment, and check if pairs with
high score are actually coming from the same family.

5.2 The Skolnick Clustering Test

A test set was suggested to us by Jeffrey Skolnick. The set contains 33 proteins
classified by SCOP into four families: CheY-related, Plasto–cyanin/azurin-
like, Triosephosphate isomerase and Ferratin (see Figure 5.1). Since this pro-
teins are relatively large (beyond the capability of Branch–and–Cut exact so-
lution, the problem is NP–hard after all), we used the heuristics that worked
well for providing lower bounds in the Branch–and–Cut algorithm, i.e. GA
and Local Search.
We applied the heuristics to all 528 contact map pairs, and clustered the
proteinss based on the following similarity score:
cij
sij =
min{mi , mj }
where mi and mj are the number of contacts in proteins i and j, and cij
is the number of shared contacts in the solution found. Note that 0 ≤ sij ≤ 1.
A score of 1.0 would indicate that the smaller contact map is completely
contained in the larger contact map. We found that 409 alignments with
score between 0.0 and 0.313 were almost exclusively between contact maps
in different families; 1.3% of these were alignments within the same family.
The remaining 119 alignments, with score between 0.314 and 0.999, were all
30 Giuseppe Lancia and Sorin Istrail

between contact maps in the same family. Hence, we validated the CMO score
for clustering with 98.7% accuracy (1.3% false negatives).

5.3 Conclusions

In order to promote the use of contact map overlap for structure comparison,
more work should be devoted into designing effective algorithms for proteins of
larger size than our current limit. Armed with such an algorithm, one could
run an all against all alignment of the PDB structures, and cluster them
according to their contact map similarity.
Our Branch–and–Bound and heuristic algorithms can be extended to new
contact map–based similarity measures. For instance, one can consider the
introduction of weights we on the contacts e = {i1 , i2 }. These weights can be
based on the residues appearing at positions i1 and i2 , or on the distance be-
tween i1 and i2 in the folded protein. This would allow to model a situation in
which some contacts are “more important” to preserve than others. Similarly,
penalties uij can be used to weigh each residue–residue xij alignment. Finally,
the objective function could be made parametric in the number of residues
considered, so that the measure can be used, e.g., for local alignments. The
problem would then read: for a given k find the best contact map alignment
which maps exactly k residues of the first proteins in k of the second.
These and similar measures will be the object of our future research.

6 Acknowledgements
Support for Lancia was provided in part by the Italian Ministry of Univer-
sity and Research under the National Project “Bioinformatics and Genomics
Research”, and by the Research Program of the University of Padova.

References
1. P. Argos and M. Rossmann. Exploring structural homology of proteins. J. Mol.
Biol. 105 (1976) 75–95
2. F. Allen et al. Blue Gene: A vision for protein science using a petaflop super-
computer. IBM System Journal 40(2) (2001) 310-321
3. H.M.Berman, J.Westbrook, Z.Feng, G.Gilliland, T.N.Bhat, H.Weissig,
I.N.Shindyalov, P.E.Bourne, The Protein Data Bank. Nucl. Ac. Res. 28
(2000) 235–242
4. T. L. Blundell. Structure-based drug design. Nature 384 (1996) 23–26
5. C. Branden and J. Tooze, Introduction to Protein Structure. Garland, 1999
6. J. M. Bujnicki, A. Elofsson, D. Fischer and L. Rychlewski. LiveBench-1: Contin-
uous benchmarking of protein structure prediction servers. Prot. Sc. 10 (2001)
352–361
 Protein Structure Comparison: Algorithms and Applications 31

7. D. F. Burke, C. M. Deane, and T. L. Blundell. Browsing the SLoop database of

structurally classified loops connecting elements of protein secondary structure.
Bioinformatics 16(6) (2000) 513–519
8. T. E. Creighton, Proteins: Structures and Molecular Properties. Freeman, New
York, 1993
9. S. Cristobal, A. Zemla, D. Fischer, L. Rychlewski and Arne Elofsson. How can
the accuracy of protein models be measured? (2000) submitted
10. J. Devillers (ed), Genetic Algorithms in Molecular Modeling, Academic Press,
London, 1996
11. L. E. Donate, S. D. Rufino, L. H. Canard, and T. L. Blundell. Conformational
analysis and clustering of short and medium size loops connecting regular sec-
ondary structures: a database for modeling and prediction. Protein Sci. 5(12)
(1996) 2600–2616
12. D. Fischer et al.. CAFASP-1: Critical Assessment of Fully Automated Structure
Prediction Methods. Proteins Suppl 3 (1999) 209–217
13. http://www.cs.bgu.ac.il/~ dlfisher/CAFASP2
14. A. Godzik. The structural alignment between two proteins: Is there a unique
answer?. Prot. Sc. 5 (1996) 1325–1338
15. A. Godzik, J. Skolnick and A. Kolinski, A topology fingerprint approach to
inverse protein folding problem. J. Mol. Biol. 227 (1992) 227–238
16. D. E. Goldberg. Genetic Algorithms in Search, Optimization and Machine
Learning, Addison-Wesley, 1989
17. D. Goldman. PhD Thesis, U.C. Berkeley, 2000
18. D. Goldman, S. Istrail and C. Papadimitriou, Algorithmic Aspects of Protein
Structure Similarity. Proc. of the 40th IEEE Symposium on Foundations of Com-
puter Science, 512–522, 1999.
19. J. G. Gough, C. Chothia C., K. Karplus, C. Barrett and R. Hughey. Optimal
Hidden Markov Models for all sequences of known structure. Currents in Com-
putational Molecular Biology (Miyano S., Shamir R., Toshihisa T. eds.), Univ.
Acad. Press Inc. Tokyo, 2000.
20. M. Grötschel, L. Lovász and A. Schrijver. The Ellipsoid Method and its Conse-
quences in Combinatorial Optimization. Combinatorica 1 (1981) 169–197
21. T. F. Havel, G. M. Crippen and I. D. Kuntz, Biopolymers 18 (1979), 73
22. T. F. Havel, I. D. Kuntz and G. M. Crippen. The theory and practice of distance
geometry. Bull. Math. Biol. 45 (1983) 665–720
23. R. B. Hayward, C. Hoang and F. Maffray. Optimizing Weakly Triangulated
Graphs. Graphs and Combinatorics 5 (1987) 339–349
24. J. H. Holland. Adaptation in Natural and Artificial Systems. MIT Press, 1992
25. L. Holm and C. Sander. Protein Structure Comparison by Alignment of Distance
Matrices. J. Mol. Biol. 233 (1993) 123–138
26. L. Holm and C. Sander. Mapping the protein universe. Science 273 (1996) 595–
602
27. M. Khimasia and P. Coveney, Protein Structure prediction as a hard optimiza-
tion problem: the genetic algorithm approach, 1997
28. Y. Lamdan, J. T. Schwartz and H. J. Wolfson. Object Recognition by Affine
Invariant Matching. Proc. of the IEEE Conference on Computer Vision and
Pattern Recognition, 335–344, 1988
29. G. Lancia, R. Carr, B. Walenz and S. Istrail, 101 Optimal PDB Structure Align-
ments: A Branch-and-Cut Algorithm for the Maximum Contact Map Overlap
32 Giuseppe Lancia and Sorin Istrail

Problem. Proc. of the 5th ACM REsearch in COMputational Biology, 193–202,

2001
30. C. Lemmen and T. Lengauer, Computational methods for the structural align-
ment of molecules. Journal of Computer–Aided Molecular Design 14 (2000) 215–
232
31. A. M. Lesk. Computational Molecular Biology. In: Encyclopedia of Computer
Science and Technology (A. Kent, J. Williams, C. M. Hall and R. Kent eds.) 31
(1994) 101–165
32. M. Levitt and M. Gerstein. A Unified Statistical Framework for Sequence Com-
parison and Structure Comparison. Proc. Natl. Acad. Sc. 95 (1998) 5913–5920
33. S. Lifson and C. Sander, Nature 282 (1979), 109
34. G. M. Maggiora, B. Mao, K. C. Chou and S. L. Narasimhan. Theoretical and
Empirical Approaches to Protein–Structure Prediction and Analysis. In: Meth-
ods of Biochemical Analysis 35 (1991) 1–60
35. A. Marchler–Bauer and S. H. Bryant. Comparison of Prediction Quality in the
Three CASPS. Proteins Suppl 3 (1999) 218–225
36. McLaughlan, Acta Crystallogr. (1979)
37. L. Mirny and E. Domany. Protein fold recognition and dynamics in the space
of contact maps. Proteins 26 (1996) 391–410
38. K. Mizuguchi, C. M. Deane, T. L. Blundell, M. S. Johnson and J. P. Overington.
JOY: protein sequence-structure representation and analysis. Bioinformatics 14
(1998) 617-623
39. K. Mizuguchi, C. M. Deane, T. L. Blundell, and J. P. Overington. HOMSTRAD:
a database of protein structure alignments for homologous families. Protein Sci.
7(11) (1998) 2469–2471
40. J. Moult, T. Hubbard, S. Bryant, K. Fidelis, J. Pedersen and Predictors. Critical
Assessment of Methods of Proteins Structure Prediction (CASP): Round II.
Proteins Suppl 1 (1997) dedicated issue
41. J. Moult, T. Hubbard, K. Fidelis, and J. Pedersen. Critical Assessment of Meth-
ods of Protein Structure Prediction (CASP): Round III. Proteins Suppl 3 (1999)
2–6
42. A. G. Murzin, S. E. Brenner, T. Hubbard and C. Chothia. SCOP: a structural
classification of proteins database for the investigation of sequences and struc-
tures. J. Mol. Biol. 247 (1995) 536–540
43. G. L. Nemhauser and L. Wolsey. Integer and Combinatorial Optimization, John
Wiley and Sons, 1988
44. J. P. Overington, M.S. Johnson, A. Sali, and T.L. Blundell, Tertiary structural
constraints on protein evolutionary diversity; Templates, key residues and struc-
ture prediction. Proc. Roy. Soc. Lond. B 241 (1990) 132–145
45. K. Park, M. Vendruscolo and E. Domany, Toward an Energy Function for the
Contact Map Representation of Proteins. PROTEINS: Structure, Function and
Genetics 40 (2000) 237–248
46. A. Raghunathan. Algorithms for Weakly Triangulated Graphs, U.C. Berkeley
Tech. Rep., CSD-89-503 (1989)
47. G. Rhodes, Crystallography Made Crystal Clear. Academic Press, 2nd Ed., 1999
48. R. B. Russell, G. J. Barton. Multiple protein sequence alignment from tertiary
structure comparison. PROTEINS: Struct. Funct. Genet. 14 (1992) 309–323
49. A. Sali and T. L. Blundell. Definition of general topological equivalence in pro-
tein structures. A procedure involving comparison of properties and relationships
 Protein Structure Comparison: Algorithms and Applications 33

through simulated annealing and dynamic programming. J. Mol. Biol. 212(2)

(1990) 403-428
50. N. Siew, A. Elofsson, L. Rychlewski and D. Fischer, MaxSub: An Automated
Measure for the Assessment of Protein Structure Prediction Quality. Bioinfor-
matics 16 (2000) 776–785
51. A. P. Singh and D. L. Brutlag 1998. Protein Structure Alignment: A comparison
of methods. Bioinformatics (2000), Submitted.
52. T. F. Smith and M. S. Waterman. Identification of common molecular subse-
quences. J. Mol. Biol. 147 (1981) 195–197
53. R. Sowdhamini, D. F. Burke, J. F. Huang, K. Mizuguchi, H. A. Nagarajaram,
N. Srinivasan, R. E. Steward, and T. L. Blundell. CAMPASS: a database of
structurally aligned protein superfamilies. Structure 6(9) (1998) 1087–1094
54. M. J. Sutcliffe, I. Haneef, D. Carney and T. L. Blundell. Prot. Eng. 1 (1987)
377–384
55. S. Umeyama. Least–Squares Estimation of Transformation Parameters Between
Two Point Patterns. IEEE Transactions on Pattern Analysis and Machine In-
telligence, PAMI-13(4) (1991) 376–386
56. M. Vendruscolo, E. Kussell and E. Domany, Recovery of protein structure from
contact maps. Fold. Des. 2 (1997) 295–306
57. M. Vendruscolo, R. Najmanovic and E. Domany, Protein Folding in Contact
Map Space. Phys. Rev. Lett. 82(3) (1999) 656–659
58. M. Vendruscolo, B. Subramanian, I. Kanter, E. Domany and J. Lebowitz, Sta-
tistical Properties of Contact Maps. Phys. Rev. E 59 (1999) 977–984
59. A. Zemla, C. Venclovas, J. Moult and K. Fidelis. Processing and Analysis of
CASP3 Protein Structure Predictions. Proteins Suppl 3 (1999) 22–29
Spatial Pattern Detection in Structural
Bionformatics

Haim J. Wolfson

School of Computer Science, Tel Aviv University, Tel Aviv 69978, Israel
wolfson@post.tau.ac.il

1 Introduction

Geometric pattern detection and recognition appears as a major task in var-

ious fields such as Computer Vision ([27]), Biometrics ([35]), Medical Image
Processing ([29]), Classification of Anatomical Form ([7]) and more. In the
last decade a new major application of 3D geometric pattern discovery has
emerged in the rapidly developing field of Bioinformatics ([60]), which is deal-
ing with the development of algorithms for Molecular Biology applications.
In its early stages Bioinformatics dealt almost exclusively with pattern
discovery in DNA and protein sequences ([58]), which resulted in the devel-
opment of efficient algorithms for biological sequence and multiple sequence
alignment ([28]). These algorithms have become an indispensable tool in the
analysis and interpretation of novel genomic data supplied by the Human
Genome project ([18]). One of the key applications of pattern discovery is in
the elucidation of the function of novel proteins based on their global sequence
similarity to proteins of known function. Another key application is the detec-
tion of short sequential patterns (motifs), which are responsible for a similar
specific function in otherwise different proteins (e.g. the calcium binding EF
motif described in chapter 2 of [8]).
Elucidation of protein function is one of the key tasks of Molecular Biology.
Proteins are most versatile molecules, which are involved in major processes
of living organisms, such as catalysis, metabolism, signaling, transport, regu-
lation, detection and destruction of foreign invaders by the immune system,
assembly of new proteins and more. To quote A.M. Lesk ([44]) ”In the drama
of life on a molecular scale, proteins are where the action is”. Under natural
conditions, which depend on temperature and the solvent, protein molecules
possess (almost) unique stable three-dimensional structures, which are defined
by their amino-acid sequence. This structural information is significantly more

C. Guerra, S. Istrail (Eds.): Protein Structure Analysis and Design, LNBI 2666, pp. 35-56, 2003.
 Springer-Verlag Berlin Heidelberg 2003
36 Haim J. Wolfson

informative for the deduction of protein function, than the sequence informa-
tion alone. Since protein structure is more conserved (during evolution) than
its sequence, we can detect proteins with less than 25% sequence similarity,
yet with roughly similar structure and function. It should be noted, though,
that sometimes functionally non related proteins might fold into similar stable
structures. Among the 15,000 structures in the Protein Data Bank ([4], circa
March 2002) there are only about 700 structurally different single chain pro-
tein folds. Since proteins function by association, the spatial arrangement of
certain residues on the protein molecular surface is crucial for their function.
The books by Lesk ([44]) and Branden & Tooze ([8]) give an excellent in-depth
introduction into protein structure, while lively descriptions and illustrations
can be found in the popular book of Goodsell ([26]).
Most of the protein structures known today have been detected by X-ray
crystallography and NMR based techniques. These methods are time con-
suming and cannot be applied to all proteins due to physical constraints.
The most challenging task is, naturally, to computationally detect the fold
of novel proteins from basic principles. The crucial role of protein structure
analysis in the elucidation of protein function has triggered the initiation of
the Structural Genomics ([9]) project as a natural follow up of the Human
Genome project. The aim of this project is to detect (mainly, by X-ray crys-
tallography) the structures of representatives from each cluster of sequentially
homologous proteins and, consequently, to model the structures of the other
proteins in the respective clusters based on sequence similarity to the detected
structural template. This project is expected to supply numerous novel protein
structures for subsequent structural and functional analysis, which requires a
computational infrastructure similar to the one that was developed for protein
sequence analysis. This algorithmic infrastructure, which gained the name of
Structural Bioinformatics, receives as input the 3D shapes of proteins and,
thus, from the mathematical viewpoint belongs to the discipline of Geomet-
ric Computing, which includes Computational Geometry, Computer Vision,
Robotics, Computer Graphics, Medical Imaging and more. In particular, the
structural analog of the classical protein sequence alignment tool, which is
protein structure alignment, is conceptually similar to the classic problems of
Object Recognition in Computer Vision ([51]). Both are problems of spatial
pattern detection.
In our discussion we shall distinguish between pattern recognition and pat-
tern detection or pattern discovery. While the term pattern recognition will be
used for the detection of an a-priori known pattern/template in a database
of shapes, the term pattern detection will apply for the detection of a-priori
unknown (structural) patterns, which appear in several shapes (or recurrently
appear in a single shape) of the database.
In Structural Bioinformatics spatial pattern discovery dominates the task
of protein structural comparison and the task of the prediction of protein-
protein, protein-DNA or protein-drug interaction (docking). In the general
case the individual molecular structures also posses internal degrees of free-
 Spatial Pattern Detection in Structural Bionformatics 37

dom, so we are faced with pattern detection in flexible spatial structures. In

this paper we shall shortly outline the major computational issues one is fac-
ing in the above mentioned tasks and survey some of the work done at the
Tel Aviv University Structural Bioinformatics group ([72]) on these topics.
We shall first discuss the shape representation issue of protein structures and
what are the relevant shape representations for the different computational
tasks, then we shall outline some rigid and flexible protein structural align-
ment algorithms, and discuss the issues of rigid bound versus unbound and
flexible docking.

2 Protein Shape Representation

Different protein shape representations are used in biological pattern discov-
ery. The choice of the proper representation depends (first of all) on the bio-
logical task at hand and (to a lesser extent) on the algorithmic methodology.
Like any molecule a protein can be viewed as a set of its individual atoms,
where each atom type is modelled as a ball of a given radius. Since the lengths
of these radii are restricted to a relatively small interval and there are also
stringent restrictions on the minimal distance between the centers of these
atomic balls, such molecular shapes can be very efficiently handled by vari-
ous Computational Geometry algorithms dealing with intersection and loca-
tion queries ([30]). The full volumetric representation is important in docking
studies, where one is fitting a pair of molecules, as if they were separate pieces
of a 3D jigsaw puzzle. There one tries, first, to detect complementary spatial
patterns on the surfaces of the shapes ([14]), and in a later stage has to verify
that the proposed fit does not cause intersection of the interiors of the two
volumes.
A protein can also be viewed as a folded 3-D curve of amino acids, the,
so called, polypeptide chain ([44]). Each amino acid is built from a central
Cα atom to which are attached a hydrogen atom, carboxyl and amino groups
and a specific side chain residue, which differs from one amino acid to the
other. Usually, the proteins function via exposed surface residues, while the
internal residues supply the structural scaffold. Thus, if our aim is to detect
similar spatial patterns in the protein folds, we can often reduce a protein
representation to the set of its Cα atom centers, thus representing each amino
acid by a single 3D point. If the order of the amino-acids on the polypeptide
chain is relevant to the task at hand, we further reduce a protein to a 3D curve
sampled at the centers of its Cα atoms. This is, essentially, a one-dimensional
structure. Whether one should use the non-connected 3D point constellations
or could consider a protein as a curve depends on the biological problem that
we have to address. If one is looking for similar functional patterns of residues
appearing on the protein surface, then the unconnected 3D representation is
appropriate, while for evolutionary classification of protein folds the 3D curve
representation might be a natural choice.
38 Haim J. Wolfson

A much coarser representation of the polypeptide chain can be achieved by

viewing it as an ensemble of secondary structure elements. Each element is a
regular structure of consecutive amino acids, which are formed in the protein
interior. The most common secondary structures are the α-helices which have
a helical structure and β-sheets, consisting of β-strands ([44]). Both the α-
helix axes and the β-strands can be modelled as roughly linear segments of
10-15 amino acids. Thus, one can reduce by an order of magnitude the size of
the point set representation by representing a protein as a set of line segments.
Here, as well, one might consider either ordered or unordered sets, depending
on the problem at hand ([38, 1]). Of course, one could consider mixed sets of
points and segments.
A major issue in protein shape representation is the modelling of flexibility,
or internal degrees of freedom. While the native protein structure is assumed
to be rigid, under different physical conditions or in association with other
molecules, the protein might change its 3D shape. One has to consider both
major changes, like domain movements ([25]), or minor frequently occurring
changes in the shape and location of the side chains lining the outer molecular
surface. This problem is especially acute in pattern discovery tasks associated
with the prediction of protein-protein interaction (docking). The tasks then
resemble articulated object recognition in computer vision ([71]) or matching
of non-rigid human organs in medical image processing.
To summarize, depending on the biological task at hand, proteins might
have very different shape representations, which directly affect both the size
of the input and the algorithmic complexity of handling these shapes. Pattern
detection in flexible shapes is obviously much more complex than in rigid
ones. Viewing a protein as a 3D curve, which is in essence a one-dimensional
entity, enables application of significantly more efficient algorithms than the
handling of a protein as a disconnected set of 3D points, and reduction to
secondary structures reduces the input size by an order of magnitude. Yet,
for the detection of functional patterns on protein surfaces and for docking
prediction, one needs to consider inherently 3D shape representations.

3 Protein Structural Alignment

3.1 Alignment of Rigid Structures

Assuming that proteins are rigid structures, the structural alignment of a

pair of proteins is defined by a three dimensional rigid transformation (rota-
tion and translation) which optimally superimposes their shapes. Naturally,
the meaning of optimal shape superimposition depends on the protein shape
representation. For comprehensive reviews on protein structural alignment al-
gorithms see [20, 42]. As was mentioned in section 2 proteins may undergo
conformational changes. These are generated by rotational movements around
 Spatial Pattern Detection in Structural Bionformatics 39

covalent bonds. Hinge and shear movements of protein domains have been ob-
served as a combination of such rotations ([25]). Thus, a more complicated
task is to discover structural similarity among molecules modulo internal de-
grees of freedom as well.
There are several applications where structural alignment plays a key role.
Among these applications are the classification of the proteins universe by
shape similarity ([47, 32, 14]), and the detection of three dimensional struc-
tural patterns, so called, structural motifs, which imply similar protein
function. Structural alignment has also key applications to Computer-Assisted
Drug Design, where one is trying to detect or design drugs which snugly fit
some functional active site of a protein (receptor) molecular surface. Detection
of structural similarity between the active site of a protein and a molecular
surface patch of a novel protein implies that the new protein is a candidate
receptor for the same drug. Likewise, if one detects other drugs having a cer-
tain structural similarity of functional groups to the original one, these drugs
become candidate inhibitors for the same receptor.
Let us formulate the rigid structural alignment task as a particular instance
of pattern detection, which is often called partial matching:
Given two rigid sets of features in the 3-dimensional space and the group
of rigid 3D transformations, the partial matching task is to recover a transfor-
mation (or several transformations), which superimpose a large enough subset
of the first structure onto the second one.
Note that in partial matching not all the features of either structure will
have a matched counterpart. Moreover, one does not know in advance which
of the features will have a counterpart. One should also note that in biological
applications two features (e.g. atomic centers) are considered superimposed if
the distance between them (in some metric) is less than a predefined thresh-
old. Another practical observation is that we are not necessarily looking for
the maximally matching subsets. Due to inaccuracies and fuzziness of corre-
spondence the ”correct solution” might not achieve the maximal score of a
given algorithm. Moreover, even non-maximal matching substructures might
imply similarity in biological function. Thus, one is looking for large match-
ing subsets (and not only for the maximal one) each of which can represent a
correct solution.
In order to tackle the partial matching task, one has to address two prob-
lems - feature correspondence and corresponding feature superposition. The
correspondence problem is the difficult one. Once a correspondence hypoth-
esis is established, there are well known efficient methods to find the rigid
transformation which superimposes them with minimal least squares distance
among the matching feature points (see, e.g. [36, 33, 59]).
We have tackled the partial matching problems in Structural Biology by
methods which were originally developed for the object recognition task in
Computer Vision and Robotics [40, 71]. A major goal of computer vision
systems is to efficiently perform model-based object recognition in cluttered
scenes [5, 12]. Specifically, given a database of familiar objects (models) and a
40 Haim J. Wolfson

newly observed scene, the task is to detect all the appearances of the models
in the scene as well as the transformation between their pose in the model
database and the scene. The objects appearing in the scene may be partially
occluded and additional objects not appearing in the database may clutter
the scene. Objects are usually represented by feature sets, such as interest
points, line segments, surface patches etc.
One can notice that there is a striking conceptual similarity between the
two tasks. Actually, the model-based object recognition task, as formulated
above, is analogous to the structural alignment of a newly discovered molec-
ular structure against a database of such structures. The analog of the model
database is the molecular structural database, and the analog of the newly ob-
served scene is the new molecular structure. Partial occlusion and additional
scene clutter correspond to the fact that we are looking for previously unde-
fined structural patterns. This analogy is even more direct, when in Computer
Vision one is dealing with data acquired by a 3-D sensor, such as a range sen-
sor. In essence, both in Computer Vision and in Structural Biology we are
faced with the task of spatial pattern detection or partial matching . Guided
by this conceptual similarity we have introduced Computer Vision based ob-
ject recognition/matching techniques into Structural Bioinformatics [51].
Due to the need to compare databases of protein and DNA sequences,
algorithms for character string matching have been extensively applied in
Molecular Biology (see e.g. [58, 28]). Most of the methods are based on the
dynamic programming approach, although hashing techniques have been ap-
plied as well to speed up processing [46]. Consequently, dynamic programming
is deeply rooted in computational molecular biology and there have been sig-
nificant attempts to tackle the three-dimensional structural matching task
using this technique [63].
In order to apply the dynamic programming method, one has to exploit
the order of the amino acid residues on the polypeptide chain. Thus, as was
mentioned in section 2, from a purely geometric viewpoint the problem tack-
led is not that of 3D point set set partial matching, but of 3D curve partial
matching. This is a significantly easier problem, since a curve in any dimen-
sional space is just a one dimensional entity. The location of a point on a
curve is fully defined by only one parameter, which is the arc-length ([17]).
We shall see below that the sequence order constraint allows us to tackle the
flexible alignment task without prior knowledge of hinge positions in the pro-
teins ([61]). However, sequence order dependent matching algorithms cannot
detect geometric patterns which do not depend on such an order, especially,
molecular surface motifs.
Let us very shortly outline the sequence order independent protein struc-
tural alignment technique ([51, 2]). To facilitate the exposition we explain our
method for structural alignment of Cα atom sets. We consider rigid objects,
which are constellations of points. Each such point is located at the center of
a Cα atom and may have a label (e.g. a residue name, or a residue type).
 Spatial Pattern Detection in Structural Bionformatics 41

The technique we present is geared towards the efficient comparison of a

target molecule to a database of known molecules. The pairwise structural
comparison problem is just a particular case, where the database consists of
one model. To simplify the exposition we describe our method for the pairwise
comparison of point sets.
The three major steps of our approach are: detection of seed matches; clus-
tering of seed matches; extension and verification of the candidate matching
hypotheses.
1) Detection of seed matches. This is the heart of the algorithm. The goal
is to establish localized structural alignments, which we nickname seed
matches. The rationale of the localized approach is twofold - first, there
is a biological justification for a local search, since the amino acids of bio-
logically relevant structural patterns (e.g. active sites) are predominantly
in spatial proximity, and, second, localization reduces the complexity of
the computationally intensive step of the algorithm.
The Geometric Hashing Technique [40, 51, 2] is applied to generate these
seed matches. The localization is achieved by limiting the Geometric Hash-
ing only to points with pairwise distance below a certain predefined thresh-
old.
Seed matches which receive a relatively high Geometric Hashing score are
retained for further processing. Such a seed match is represented by a list
of matching pairs of atoms (match-list) and by a 3-D rigid transformation
(rotation and translation). The number of matching pairs should be above
a threshold (minimal match-list size). Each pair in the list specifies a
correspondence between an atom from one structure and an atom from
the other structure. The transformation represents the 3-D rotation and
translation, which superimposes the atoms of the first structure onto the
corresponding atoms of the second structure.
Note that this step may produce several candidate seed matches which
induce (almost) similar transformations, obtained from different localized
subsets. Thus, the next step clusters all these matching atoms together.
2) Clustering of seed matches. In the second step seed matches represent-
ing almost identical transformations are detected, and their match-lists
are merged. A rotation and translation of one set onto the other giving
the minimal least squares deviation of the matching points is computed
[36, 59]) .
3) Extension and verification of match hypotheses. The relatively re-
liable correspondence lists of the seed matches obtained after the cluster-
ing procedure are extended to contain additional matching pairs. First, the
structures are superimposed according to the transformation determined
by the seed match of the cluster. Then, pairs of atoms that are ”close”
enough after the superimposition become prospective additional matching
pairs. Since proteins are usually quite dense in space, each atom from one
structure may have several ”close” neighbors in the other structure. To
42 Haim J. Wolfson

choose the appropriate pairs a heuristic iterative matching algorithm is

applied, which minimizes the sum of the distances between all the matched
pairs.
Finally, the best extended matches are reported. The quality of the match
is determined mainly by the number of the matching pairs of atoms and,
in some extent, by the least squares distance between these matching atoms
sets. Naturally, one would prefer to incorporate in the final ranking biological
information, which could assist in the detection of the biologically meaningful
alignments, even if they are not large in size.
The algorithm was applied to the detection of a structurally non redundant
representation of the Protein Data Bank ([14]), as well as to the creation of
a structurally non redundant dataset of protein-protein interfaces ([65]). The
algorithm can be accessed via the WWW at ([72]).

3.2 Structural Alignment of Flexible Proteins

In the previous discussion we treated proteins as rigid shapes. However, pro-

teins do possess internal rotational degrees of freedom. In particular, as clas-
sified by [25], large proteins, which are built of several domains, often have
different conformations, which can be modelled as hinge or shear movements
of respective domains. These movements are important for protein function,
such as biomolecular recognition/docking. A flexible structural comparison
algorithm would enable us to perform the same searches as a rigid one with
the added strength of being able to detect partial matches, even if one of the
matching substructures has undergone conformational changes. This ability is
especially important in the search of drug molecules against a database, since
the vast majority of molecules have many potential conformations.

Flexible Alignment with Pre-defined Hinge Locations

For the sake of clarity let us discuss the following problem: one molecule is
rigid and the other is composed of two rigid parts (domains) which are joined
by a hinge. A hinge in this case is a rotational joint with full 3-D rotational
freedom. In practice, rotations are only around specific bonds, but we allow
this more general 3-D rotation model since it implicitly allows approximation
of several consecutive (or, nearby) bond rotations, as well. The fact that one
molecule is rigid can be assumed without loss of generality, while the intro-
duction of several hinges in the other molecule is a straightforward extension
of the algorithm that we present. Our flexible partial matching method is
based on the ideas that we have introduced for articulated (flexible) object
recognition in Computer Vision [71, 3]. It exploits an associative memory in-
dexing approach reminiscent of the one used in Geometric Hashing, except
that this time we have the additional task of handling the internal flexibility.
Of course, one can apply the rigid matching method to the individual rigid
 Spatial Pattern Detection in Structural Bionformatics 43

parts, and then check whether there is a pair of high scoring hypotheses, one
for each part, which are spatially consistent with the single flexible object
input. However, our aim is to handle the information obtained from all the
molecule rigid parts in an integrated manner, so that the whole is more than
just the sum of its parts. We achieve this goal by accumulating evidence on
the position of the hinge location of the model molecule relative to the target
molecule. Since the hinge belongs to both rigid parts and its location is not
influenced by the internal rotation, both parts contribute evidence which is
integrated in our algorithm.
Below we describe the algorithm in more detail. Assume that we have a
database of flexible model molecules, each consisting of two rigid parts joined
by a hinge at a known position. Given a target molecule, we want to find
a large partial match of the target with some model molecule allowing both
external rotation and translation and internal rotation at the hinge. Thus, we
are trying to detect a flexible spatial pattern. The molecules are represented
by their Cα backbone atoms.
We first preprocess the database of the flexible molecules, and then per-
form the pattern discovery (recognition) versus the target. For each database
(model) molecule the following preprocessing is done:
(a) The molecule is represented by its Cα backbone atoms, as interest fea-
tures.
(b) The (known) hinge location is chosen as the origin of a 3-D reference
which is denoted as the ‘hinge frame’. The orientation of this frame is set
arbitrarily.
(c) Each ordered non-collinear triplet of atoms belonging to a single rigid part
and satisfying both proximity and non-degeneracy constraints is defined
as a ‘feature group’. An unambiguous 3D reference frame is defined for the
triangle. For example, given the non-collinear triplet of points p0 , p1 , p2 ,
one can set the origin of the reference frame at p0 . To set the reference
set axes, let us define two vectors v1 = p1 − p0 ; v2 = p2 − p0 . Define v1
as the direction of the x-axis, the cross product v1 ⊗ v2 as the direction
of the z-axis, and the unit vector of the y-axis as the cross product of the
unit vectors of the x and z axes.
(d) The (rotation and translation invariant) triple of triangle side lengths
serves as an address to a (hash) table, where one stores the information
on the model molecules, rigid part, feature group and, especially, on the
transformation between the feature group reference frame and the hinge
frame.
In the pattern discovery/recognition stage we repeat a similar proce-
dure for the rigid target molecule exploiting the information accumulated in
the associative memory:
44 Haim J. Wolfson

(a) The target is represented by its Cα backbone atoms, as interest features.

(b) Non-collinear triplets satisfying the proximity and non-degeneracy con-
straints are tagged as feature groups. A 3D reference frame is defined for
each feature group as above. The triangle side lengths invariants serve as
an address to the table and the transformations stored there are applied
to the feature group reference frame. This transforms the reference frame
to a new frame, which is a candidate ‘hinge frame’. Votes are accumulated
for the location of candidate hinge frames.
(c) After completion of stage (b) we consider pairs (model molecule, hinge
location), which have received a high vote. (In an ideal situation, such a
candidate solution should exhibit two clusters of internal rotations repre-
senting the groups of the different molecule parts.)
(d) High scoring solutions can be further evaluated by different physico-che–
mical criteria as well as criteria requiring participation of both molecule
parts in the detected pattern etc. Finally, match-lists are derived and a
best least-squares match for each rigid part is computed.
In this algorithm we have exploited the fact that both parts incorporate
the same hinge by locating the global reference frame of the flexible model-
molecule at the hinge. In such a way, both parts contribute votes to a reference
frame at the same location, although at different orientations.
Although here, our algorithm is described for molecules with a single hinge,
it is extended to multiple hinges by the following enhancement. Instead of
having one ‘hinge frame’ one can define multiple ‘hinge frames’, each of them
centered at a different hinge. In the preprocessing stage, for each feature group
on a single part, one should encode the transformations from its frame to all
the ‘hinge frames’, which are located on that part. Thus, e.g., on a part with
two hinges, two transformations will be stored for each feature group, while on
a part with one hinge only, one transformation will be stored. The recognition
phase will remain as described above, except that each target feature group
votes for as many frames as the number of different transformations stored in
its table entry.
We have implemented variants of this technique for flexible protein and
small molecule structural alignment ([69] and the FlexMol algorithm of [61]),
and protein-ligand docking [55, 57].

Simultaneous Flexible Alignment and Hinge Location Detection

The previous exposition referred to the case where we have a-priori knowledge
of the candidate hinge positions on one of the molecules, or enumerate these
positions, if it is computationally feasible. To overcome this handicap, we
have recently developed a novel algorithm, which automatically detects both
the candidate hinge positions on one of the molecules and the largest aligned
flexible spatial patterns. The algorithm, FlexProt ([61]) accepts as input two
molecules, assuming that the first is rigid and the second is flexible (the sit-
uation is symmetric from the mathematical standpoint), and outputs lists of
 Spatial Pattern Detection in Structural Bionformatics 45

best alignments ranked according to the number of detected rigid parts, the
size of the overall alignment, the size of the individual rigid parts, the RMSD
of the alignment and some additional criteria. Thus, the algorithm overcomes
the requirement of prior partition of one of the molecules to rigid parts. Yet,
in order to accomplish this task efficiently, it exploits locally the amino acid
sequence order and is not fully sequence order independent as the algorithm
described above.
The goal of the FlexProt algorithm is to divide the two protein molecules
into a minimal number of separate consecutive fragments of maximal size, such
that the fragments which are matched will be almost congruent (-congruent).
Two rigid fragments are -congruent, if they have the same number of Cα
atoms and there exists a 3-D rotation and translation which superimposes the
corresponding atoms with an RMSD less than some pre-defined threshold .
The arrangement of the matching fragments should be consistent with their
order on the protein chain. Flexible regions are located between the rigid
matching fragments. A trivial way to achieve a flexible alignment of maximal
size is to allow flexibility between each pair of neighboring amino acids (Cα
atoms), thus aligning completely the two molecules, if they are of the same
length. Thus, to avoid such absurd ”optimal” solutions, our goal is to minimize
the number of flexible regions or rigid fragments. Clearly, the two goals of
maximal matching size and minimal number of flexible regions are conflicting
and some compromise heuristic should be applied.
The input to the algorithm are two protein molecules M1 and M2 , each
being represented by the sequence of its Cα atom coordinates. Assume that
molecule M1 has undergone hinge bending movements at several locations
along its backbone. Further assume that between the flexible joint regions
there are fragments without a significant structural change. The resulting
hinge-bent molecule is denoted M2 . Under our assumptions there exists a set
of rigid fragments of M2 that are -congruent to the corresponding set of frag-
ments of M1 . The model presented applies not only to different conformations
of a given molecule, but to the general case of flexible motif detection between
two molecules with different sequences.
The algorithm has the following major steps:
1. Detect all the large enough -congruent rigid fragment pairs, one from
each molecule. This is done by aligning each atom pair (one atom from
each molecule) and extending the alignment to the left and to the right,
along the backbone chain, until the RMSD of the superposition of the two
fragments gets larger than . The procedure is efficient, since (based on
the calculations of the previous steps) one can compute the RMSD of each
extension in O(1) time ([36, 59]). Thus, the procedure is linear in the size
of the matched Cα atom pairs. As a by-product we also get the rotation
and translation, which best aligns the pair of fragments. At the end of this
step, we have recovered all the possible -congruent rigid fragment pairs
together with the transformations, which superimpose them.
46 Haim J. Wolfson

2. Construct long enough flexible alignments by concatenating matching

rigid fragment pairs in a way that is consistent with the amino acid se-
quence order, and does not create too large gaps between consecutive frag-
ments. This is done by a graph-theoretic method reminiscent of the one
applied in the FASTA algorithm for sequence alignment ([28]). A graph
is built with the congruent fragment pairs as its vertices. A directed edge
joins a pair of vertices, if the fragments they represent might be consecu-
tive in the final alignment. Weights are assigned to the edges, rewarding
long matching fragments and penalizing big gaps as well as large discrep-
ancies in the size of the gaps for the pair of proteins. A virtual (source)
vertex, which is connected with a zero weight edge to each original vertex
in the graph, is added and a Single-Source Shortest Paths algorithm ([16])
is applied to the graph starting from the virtual vertex. This algorithm
detects the shortest paths from the virtual vertex to each other vertex
in the graph. The number of vertices in each such path represents the
number of rigid fragments in a potential flexible alignment. For each such
number, the candidate solutions are sorted according to the total size and
the RMSD of the fragment alignment.
3. For the leading candidate solutions the rotation and translation for each
fragment pair is recovered, and consecutive fragment pairs are joined into
epsilon-congruent regions, if their transformations are close enough. This
may happen, if between the fragments, one had insertions/deletions which
did not affect the overall 3D structure of the local region. Thus, the dis-
covered flexible pattern, appearing in both proteins is represented by a
number of epsilon-congruent region pairs, each consisting of one or of
several epsilon-congruent fragment pairs.
The theoretical complexity of the algorithm is bounded by O(n4 ) (where n is
defined as the size of the larger molecule), however in practice its performance
is much faster. We have done numerous experiments with FlexProt, which can
be accessed via the WWW at [72]. The method proved to be robust and highly
efficient comparing a pair of flexible structures of about 300 amino acids each
in average time of 7 seconds on a standard desktop PC (400MHz, 256 Mb
RAM). For further reference see [61].

4 Protein-Protein Docking
Protein-protein interactions play a major role in all biological processes. In
addition, protein-drug docking is a major tool in computer assisted drug de-
sign. The binding affinity of the molecules is affected mainly by electrostatic,
hydrophobic and van der Waals interactions. Since these non-covalent inter-
actions are weak and act at short distances, in order to be effective the in-
teracting molecules have to be very close to each other. As a result shape
complementarity of the interacting molecules becomes a necessary condition
 Spatial Pattern Detection in Structural Bionformatics 47

for docking. Thus, the majority of the docking methods are searching, first, for
complementary spatial patterns on the molecular surfaces of the interacting
molecules ([43, 31]). The spatial pattern detection problem in docking is very
challenging, since molecules usually undergo conformational changes upon as-
sociation. This is aggravated by the fact that the residues on the molecular
surface, which take part in docking, are more flexible than the residues in the
protein interior.
When evaluating docking algorithms, one should distinguish between the,
so called, ‘bound’ and ‘unbound’ cases. In the ’bound’ case we are given the
co-crystallized complex of two molecules. The complex is artificially separated
by randomly rotating and translating one of the molecules. Now, the goal be-
comes to reconstruct the original complex. No conformational changes are
involved in the ‘bound’ case. Success in bound docking examples is a natural
pre-requisite for any docking algorithm, yet such a success does not ensure ad-
equate performance in ‘real life’ cases, which are ‘unbound’. In the ’unbound’
situation we are given two molecules in their native conformations. The goal
of the algorithm is to predict the ‘correct’ structure of the complex. In the
few tens of cases, where we do have an independently resolved structure of
the complex, one can verify the quality of the algorithm’s prediction. Most of
the docking algorithms encounter difficulties with the ‘unbound’ case ([31]).
As was mentioned above, significant geometric surface complementarity
is usually a prerequisite for successful docking. We shall discuss mainly this
geometric part of the docking methodology, where the task is to detect large
enough patches on the molecular surfaces of the docked molecules, which are
of complementary shape. Assuming that the molecules do not undergo large
conformational changes, we are faced again with the 3D rigid inexact partial
matching task.
Some of the algorithms approach this task by direct enumeration of the
six dimensional rigid transformation space, to detect a translation and ro-
tation, which best superimposes one molecule onto the surface of the other.
Most of these methods ([70, 66, 67, 68, 24, 10, 11]) use brute force search of
the 3 rotational parameters and the Fast Fourier Transform technique ([37])
for fast enumeration of the translations. Such algorithms are computationally
expensive. Other algorithms define discrete interest features on the molecular
surfaces and apply a partial shape matching algorithm on these features in a
way similar to protein structural alignment algorithms. The pioneering dock-
ing algorithm in this direction was suggested by Kuntz et al. [39], where the
problem was reduced to the detection of large enough cliques in the, so called,
docking graph. We have applied the Geometric Hashing technique ([21]) for
rigid docking as well as a variant of Geometric Hashing and Pose Cluster-
ing ([48, 49]), which proved to be relatively robust even for unbound docking
([50]). The flexible pattern detection method described in section 3.2 was ap-
plied to the docking of flexible molecules ([56, 57]). In protein-drug docking,
one is often looking for complementary patterns of hydrogen bond donors
versus hydrogen bond acceptors ( [52, 53]).
48 Haim J. Wolfson

Most of the geometric docking algorithms can be roughly divided into the
following major steps :
1. Molecular Surface Representation - a popular representation is that
of the solvent accessible surface as calculated by Connolly ([14, 13]).
Sparser discrete interest features, such a points and associated normals
([15, 45, 49]) are extracted in this stage.
2. Focusing on candidate binding (active) sites - in order to signifi-
cantly reduce the number of false positives and reduce computation time
it is desirable to focus a-priori on the approximate areas of the molec-
ular surface, where binding is likely to appear. Such candidate binding
sites are usually detected by biological and shape criteria. An excellent
example of ’biologically’ defined binding regions are the complementarity
defining regions(CDRs) in antibodies ([44]). An example of a shape cri-
terion is the binding of drugs and small ligands in the large cavities of a
receptor. There one might restrict the receptor surface to be explored to
such cavities ([39]).
3. Complementary spatial pattern detection - this is the heart of the
geometric docking algorithm and usually applies a partial shape matching
algorithm on a set of discrete features, such as surface points with asso-
ciated normals, which should align in roughly opposite directions. The
output of this step is a set of candidate rigid transformations, which dock
one molecule to the other.
4. Geometric Complementarity Scoring and Ranking - since mole–
cules cannot penetrate into each other, candidate transformations from
the previous step are discarded if they cause a significant penetration.
Minor penetrations are allowed to reflect conformational changes of the
molecular surface upon docking. In this step one can also calculate the size
of the detected complementary surface (size of binding site) and score the
hypotheses according to this size. (The fact that candidate solutions can
be discarded due to penetration supplies a powerful false positives filter,
which does not exist in the structural alignment algorithms described in
section 3.)
5. Biological Scoring and Re-ranking- in this step one would like to
accept the high enough scoring hypotheses of the previous step and re-rank
them according to a free-energy function ([10]), which could discriminate
between the biologically valid hypotheses and decoy complexes, which
exhibit only geometric complementarity. To date this step seems to be the
real bottleneck in the performance of unbound rigid docking algorithms
([31]).
We have recently developed an efficient rigid docking algorithm, which per-
forms relatively well in the unbound docking task ([19]). The geometric part
of the algorithm is based on methodology developed for Computer Vision
applications and it extends the ideas presented in ([22, 50]). The algorithm
partitions the molecular surface of the molecules into convex, concave and
 Spatial Pattern Detection in Structural Bionformatics 49

flat local patches of almost equal size. Patches with higher probability of be-
longing to the binding site are considered, and complementary configurations
of pairs of interest points with associated normals are detected. Alignment
of such pairs induce rigid transformations, which are subsequently tested for
shape penetration and scored by geometric complementarity. The use of sur-
face patches reduces the number of potential docking solutions, while still (in
most tested cases) retaining the correct transformation. The algorithm can
treat receptors and ligands of variable sizes. It succeeds in docking of large
proteins (antibody with antigen) and small drug molecules. The running times
of the algorithm are on the order of seconds for small drug molecules and
minutes for large proteins. The algorithm was tested on most of the known
benchmark complexes (see e.g. [11]) with satisfactory results. Figure 4 illus-
trates a challenging antibody-antigen docking example. We briefly sketch the
outline of this algorithm, which is presented in [19].

Fig. 1. Unbound docking of the Antibody Fab 5G9 with Tissue factor (PDB codes
1FGN,1BOY). The antibody is depicted by ribbons and the CDRs by atomic balls.
The antigen is depicted as a bright backbone at the top of the figure. The dark
(antigen) backbone represents the best solution obtained by our program (which
was ranked 8’th by its score) superimposed on the complex with RMSD 2.27Å.

The algorithm follows the basic steps outlined above :

Molecular Surface Representation

We compute both a dense and sparse representation of the molecular surfaces

for both each molecules. The dense surface representation is calculated using
50 Haim J. Wolfson

the MS program [14, 13], which outputs, concave, convex and saddle patches
that are sampled with high density. In addition, the sparse surface representa-
tion of [45], which retains for each patch one point with its associated normal
is calculated.

Segmentation of the Molecular Surface by Convexity

The input to this step is the sparse set of critical points. The goal is to di-
vide the surface to patches of almost equal area. Each patch is of homogenous
shape, namely, we segment the surface into convexities, concavities and flats.
In order to decide the shape type in the vicinity of a single critical surface
point we apply the shape function used in [15, 48]. A sphere is centered at each
critical point and the shape function measures the ratio of the sphere volume,
which is occupied by the molecule interior. The radius of the shape function
sphere is selected according to the molecule size. We use 6Å for proteins and
3Å for small ligands. According to the histogram of the shape function values
for the molecule, two cut-values are calculated, which split the critical points
into three equal size sets of knobs (convex), flats and holes (concave). Using
graph-theoretic techniques followed by a split-and-merge algorithm the molec-
ular surface is divided into connected, almost equal size patches of concave,
convex and flat points.

Detection of Potential Active/Binding Sites

A docking algorithm should successfully detect the binding sites of both

molecules and their correct alignment. Knowledge of the binding site of at
least one molecule greatly reduces the space of possible docking interactions,
reducing both the run-time of the algorithm and the expected number of false
positives. There are major differences in the interactions of different types of
molecules. We develop filters for every type of interaction and focus only on
the patches that were selected by the filter.
• Protease-inhibitor interactions: Hot Spot Filtering. A number of experi-
mental and computational studies ([6, 34]) have shown that protein-protein
interfaces have conserved polar and aromatic ‘hot-spot’ residues. We have
used the results of these studies to select patches that have high probability
of belonging to an active site. The other patches are discarded.
• Antibody-Antigen interactions: Detection of CDRs. It is well known ([44,
8]) that antibodies bind to antigens through their hypervariable (HV) re-
gions, also called complementarity-determining regions (CDRs). The three
heavy-chain and three light-chain CDR regions are located on the loops
that connect the β strands of the variable domains. The CDRs are de-
tected by aligning the sequence of the antibody to a consensus sequence of
a library of antibodies. We restrict our docking algorithm to the patches
of the CDR regions.
 Spatial Pattern Detection in Structural Bionformatics 51

Matching of Complementary Patches

Given the patches of the pair of molecules we compute hypothetical docking

transformations based on local geometric complementarity. We try to match
convex patches with concave patches and flat patches with flat patches. We use
two techniques - single patch versus single patch matching and pair of patches
versus pair of patches matching. Single Patch Matching is used for docking of
small ligands, like drugs or peptides, while Pairwise Patch Matching is used for
protein-protein docking. Pairs of critical points with their associated normals
are used as basic feature sets for matching by applying a combination of the
Geometric Hashing ([40] and the Pose Clustering algorithm ([62]) techniques.
Each such feature has a rotation and translation invariant shape signature,
which includes the distance between the pair of points and the three angles
that the normals and the distance segment define. Only features with almost
similar shape signatures are aligned and the resulting transformation is stored.
The method is similar to the one applied in [48]. The motivation in pairwise
patch matching is that molecules interacting with big enough contact area
must have more than one patch in their interface. Therefore matching two
patches simultaneously will result in numerically more stable transformations.
We consider only neighboring patches and choose the two points of the basic
feature set from separate patches.

Pose Clustering.

Matching of local features may lead to multiple instances of similar pose.

Therefore clustering is applied to reduce the number of potential solutions.
We employ two clustering techniques. The first is clustering by the 6D trans-
formation parameters, which is coarse, yet very fast, and thus applied first.
The second is RMSD based clustering (similar to the one used in [54]), which
is more exact but also much slower.

Steric Clash Detection and Ranking by Size of the

Complementary Spatial Patterns

For each candidate transformation from the previous step, we superimpose the
molecules according to the transformation. Now, we have to detect and discard
those transformations, which cause unacceptable steric clashes (penetrations),
and rank the rest of the solutions according to their shape complementarity.
The detection of steric clashes and the scoring of shape complementarity re-
quires to know the distance of the molecular surface points of one molecule
from the molecular surface of the other. In general, points of one molecule,
which penetrate deeply into the interior of the other discard the transforma-
tion, moderately penetrating points are penalized, and points, which are close
to the molecular surface of the other molecule receive a positive score (binding
site area). In order to accomplish this task efficiently, we compute in advance
52 Haim J. Wolfson

the distance transform of the first (stationary) molecule and calculate the
score of each transformation using this distance transform. To further speed
up the penetration detection and geometric scoring calculation, we construct
a multi-resolution data structure for the transformed molecule. Only trans-
formations that have not been rejected and received a large enough score at
the lower resolutions, are re-scored at a higher resolution.

Refinement of the Candidate Transformations

The transformation from the matching step was computed based on aligning
pairs points/ normals. Since the interface includes a much higher number of
points, the transformation can be refined to improve geometric complemen-
tarity. Similar to the structural alignment algorithms described in section 3,
for each candidate transformation a new extended match-list is compiled and
the rotation and translation giving minimal RMSD for that match-list is cal-
culated. This can be done in several iterations.

5 Summary
We have presented several applications of spatial pattern discovery algorithms
to major tasks in Structural Bioinformatics, such as protein structural align-
ment and protein-protein docking. The exposition concentrated on algorithms
developed by the interdisciplinary Structural Bioinformatics Group of Tel
Aviv University and was not intended to review the enormous amount of
work on this topic, which is done elsewhere.
Among the challenging tasks ahead of us is the development of efficient
multiple structural alignment algorithms ([41]) and flexible multiple struc-
tural alignment algorithms. In docking, one should find better ways to handle
surface side chain flexibility, and, especially, detect new ways for biological
re-ranking of the geometrically ranked solutions. Our experience shows, that
even in unbound rigid docking, we almost always detect a solution, which is
very close to the ‘correct’ one, among the few hundred top ranked geome–tric
solutions. This implies that one needs an efficient and robust biologically based
score, which could successfully re-rank these top few hundred geome–trically
derived hypotheses.
Successful docking algorithms might have direct implications on auto-
mated protein folding as well. If one could derive from the protein sequence
partial structures of building blocks ([64]), then folding could be reduced to
the docking of these building blocks subject to some proximity constraints.
Work in this direction is underway.
 Spatial Pattern Detection in Structural Bionformatics 53

Acknowledgments

I thank my long time colleague Ruth Nussinov and our joint graduate stu-
dents, in particular, Dina Duhovny and Maxim Shatsky for valuable discus-
sions and input. This research has been supported in part by the Israel Science
Foundation administered by the Israel Academy of Sciences - Center of Ex-
cellence in ”Geometric Computing”, and by the Tel Aviv University Research
Foundation.

References
1. V. Alesker, R. Nussinov, and H. J. Wolfson: Identification of non-topological
motifs in protein structures. Protein Engineering. 9(12) (1996) 1103–1119
2. O. Bachar, D. Fischer, R. Nussinov, and H. J. Wolfson: A computer vision
based technique for 3-d sequence independent structural comparison. Protein
Engineering. 6(3) (1993) 279–288
3. A. Beinglass and H. J. Wolfson: Articulated object recognition, or, how to gen-
eralize the generalized Hough transform. In Proc. of the 1991 IEEE Computer
Society Conf. on Computer Vision and Pattern Recognition (1991) 461–466
4. H. M. Berman, J. Westbrook, Z. Feng, G. Gilliland, T. N. Bhat, H. Weissig, I. N.
Shindyalov, and P. E. Bourne: The protein data bank. Nucleic Acids Research.
28 (2000) 235–242
5. P. J. Besl and R. C. Jain: Three-dimensional object recognition. ACM Comput-
ing Surveys. 17(1) (1985) 75–154
6. A. A. Bogan and K. S. Thorn: Anatomy of hot spots in protein interfaces. J.
Mol. Biol. 280 (1998) 1–9
7. F. L. Bookstein: Morphometric Tools for Landmark Data: Geometry and Biol-
ogy. Cambridge Univ. Press, Cambridge, UK 1991
8. C. Branden and G. Tooze: Introduction to Protein Structure. Garland Publish-
ing Inc., New York 1991
9. S. K. Burley: An Overview of Structural Genomics. Nature Struct. Bio. (Struc-
tural Genomics Suppl.) (2000) 932–934
10. J. C. Camacho, D. W. Gatchell, S. R. Kimura, and S. Vajda: Scoring docked
conformations generated by rigid body protein protein docking. PROTEINS:
Structure, Function and Genetics. 40 (2000) 525–537
11. R. Chen and Z Weng: Docking unbound proteins using shape complementarity,
desolvation, and electrostatics. ROTEINS: Structure, Function and Genetics.
47 (2002) 281–294
12. R. T. Chin and C. R. Dyer: Model-based recognition in robot vision. ACM
Computing Surveys. 18(1) (1986) 67–108
13. M.L. Connolly: Analytical molecular surface calculation. J. Appl. Cryst. 16
(1983) 548–558
14. M.L. Connolly: Solvent-accessible surfaces of proteins and nucleic acids. Science.
221 (1983) 709–713
15. M.L. Connolly: Shape complementarity at the hemoglobin α1 β1 subunit inter-
face. Biopolymers. 25 (1986) 1229–1247
54 Haim J. Wolfson

16. T. H. Cormen, C. E. Leiserson, and R. L. Rivest: Introduction to Algorithms.

The MIT Press, 1990
17. M. P. DoCarmo: Differential Geometry of Curves and Surfaces. Prentice-Hall,
Englewood Cliffs, N.J. 1992
18. U. S. DOE: Human genome project information.
http://www.ornl.gov/hgmis/publicat/publications.html
19. D. Duhovny, R. Nussinov, and H.J. Wolfson: Efficient unbound docking of rigid
molecules. submitted
20. I. Eidhammer, I. Jonassen, and W. Taylor: Structure comparison and structure
patterns. J. Comp. Biol. 7(5) (2001) 685–716
21. D. Fischer, S. L. Lin, H.J. Wolfson, and R. Nussinov: A geometry-based suite
of molecular docking processes. J. Mol. Biol. 248 (1995) 459–477
22. D. Fischer, R. Norel, R. Nussinov, and H. J. Wolfson: 3-d docking of protein
molecules. In Fourth Symposium on Combinatorial Pattern Matching. Lecture
Notes in Computer Science 684 (1993) 20–34
23. D. Fischer, R. Tsai, R. Nussinov, and H.J. Wolfson: A 3-d sequence-independent
representation of the protein databank. Protein Engineering. 8(10) (1995) 981–
997
24. H. A. Gabb, R. M. Jackson, and J. E. Sternberg: Modelling protein docking
using shape complementarity, electrostatics, and biochemical information. J.
Mol. Biol. 272 (1997) 106–120
25. M. Gerstein, A. M. Lesk, and C. Chothia: Structural mechanisms for domain
movements in proteins. Biochemistry. 33(22) (1994) 6739–6749
26. D. S. Goodsell: Our Molecular Nature : the body’s motors, machines, and mes-
sages. Springer-Verlag, New York 1996
27. W. E. L. Grimson: Object Recognition by Computer. MIT Press, 1990
28. D. Gusfield: Algorithms on Strings, Trees, and Sequences : Computer Science
and Computational Biology. Cambridge University Press, 1997
29. J. V. Hajnal, D. L. G. Hill, and D. J. Hawkes: Medical Image Registration. CRC
Press, Boca Raton 2001
30. D. Halperin and M. H. Overmars: Spheres, molecules and hidden surface re-
moval. Computational Geometry: Theory and Applications. 11(2) (1998) 83–
102
31. I. Halperin, B. Ma, H. J. Wolfson, and R. Nussinov: Principles of Docking: An
Overview of Search Algorithms and A Guide to Scoring Functions. PROTEINS:
Structure, Function and Genetics. 47 (2002) 409–443
32. L. Holm and C. Sander: Searching protein structure databases has come of age.
PROTEINS: Structure, Function and Genetics. 19 (1994) 165–173
33. B. K. P. Horn: Closed-form solution of absolute orientation using unit quater-
nions. J. Opt. Soc. Amer. A. 4(4) (1987) 629–642
34. Z. Hu, B. Ma, H.J Wolfson, and R. Nussinov: Conservation of polar residues as
hot spots at protein–protein interfaces. PROTEINS: Structure, Function and
Genetics. 39 (2000) 331–342
35. A.K. Jain, R. Bolle, and S. Pankanti(ed’s): Biometrics: Personal Identification
in Networked Society. Kluwer Acad. Pub., Boston 1999
36. W. Kabsch: A solution for the best rotation to relate to sets of vectors. Acta
Crystallogr. A32 (1976) 922–923
37. E. Katchalski-Katzir, I. Shariv, M. Eisenstein, A.A. Friesem, C. Aflalo, and
I. A. Vakser: Molecular Surface Recognition: Determination of Geometric Fit
 Spatial Pattern Detection in Structural Bionformatics 55

between Protein and their Ligands by Correlation Techniques. Proc. Natl. Acad.
Sci. USA. 89 (1992) 2195–2199
38. I. Koch, T. Lengauer, and E. Wanke: An algorithm for finding maximal common
subtopologies in a set of protein structures. J. Comp. Biol. 3 (1996) 289–306
39. I.D. Kuntz, J.M. Blaney, S.J. Oatley, R. Langridge, and T.E. Ferrin: A geometric
approach to macromolecule-ligand interactions. J. Mol. Biol. 161 (1982) 269–
288
40. Y. Lamdan and H. J. Wolfson: Geometric Hashing: A general and efficient
model-based recognition scheme. In Proceedings of the IEEE Int. Conf. on Com-
puter Vision. (1998) 238–249
41. N. Leibowitz, Z.Y. Fligelman, R. Nussinov, and H. J. Wolfson: Multiple struc-
tural alignment and core detection by geometric hashing. In Int. Conf. on Intell.
Systems for Mol. Bio. (1999) 169–177
42. C. Lemmen and T. Lengauer: Computational methods for the structural align-
ment of molecules. J. Computer-Aided Molecular Design. 14(3) (2000) 215–232
43. T. Lengauer and M. Rarey: Computational methods for biomolecular docking.
Current Opinion in Structural Biology. 6 (1996) 402–406
44. A. M. Lesk: Introduction to Protein Architecture : the structural biology of pro-
teins. Oxford Univ. Press, 2001
45. S. L. Lin, R. Nussinov, D. Fischer, and H. J. Wolfson: Molecular surface rep-
resentation by sparse critical points. PROTEINS: Structure, Function and Ge-
netics. 18 (1994) 94–101
46. D. J. Lipman and W. R. Pearson: Rapid and Sensitive protein Similarity
Searches. Science. 227 (1985) 1435–1441
47. A. G. Murzin, S. E. Brenner, T. Hubbard, and C. Chothia: SCOP: a struc-
tural classification of proteins database for the investigation of sequences and
structures. J. Mol. Biol. 247 (1995) 536–540
48. R. Norel, S. L. Lin, H. J. Wolfson, and R. Nussinov: Shape complementarity at
protein-protein interfaces. Biopolymers. 34 (1994) 933–940
49. R. Norel, S. L. Lin, H. J. Wolfson, and R. Nussinov: Molecular surface com-
plementarity at protein-protein interfaces: The critical role played by surface
normals at well placed, sparse points in docking. J. Mol. Biol. 252 (1995) 263–
273
50. R. Norel, D. Petrey, H. J. Wolfson, and R. Nussinov: Examination of shape com-
plementarity in docking of unbound proteins. PROTEINS: Structure, Function
and Genetics. 35 (1999) 403–419
51. R. Nussinov and H.J. Wolfson: Efficient detection of three-dimensional motifs
in biological macromolecules by computer vision techniques. Proc. Natl. Acad.
Sci. USA. 88 (1991) 10495–10499
52. M. Rarey, B. Kramer, and T. Lengauer: Time-efficient docking of flexible ligands
into active sites of proteins. In 3’rd Int. Conf. on Intelligent Systems for Mol.
Bio. (1995) 300–308
53. M. Rarey, B. Kramer, T. Lengauer, and G. Klebe: A fast flexible docking method
using incremental construction algorithm. J. Mol. Biol. 261 (1996) 470–489
54. M. Rarey, S. Wefing, and T. Lengauer: Placement of medium-sized molecular
fragments into active sites of proteins. J. Computer-Aided Molecular Design. 10
(1996) 41–54
55. B. Sandak, R. Nussinov, and H.J. Wolfson: An automated robotics-based tech-
nique for biomolecular docking and matching allowing hinge-bending motion.
Computer Applications in the Biosciences (CABIOS). 11 (1995) 87–99
56 Haim J. Wolfson

56. B. Sandak, R. Nussinov, and H.J. Wolfson: Docking of conformationally flexible

molecules. In Seventh Symposium on Combinatorial Pattern Matching. Lecture
Notes in Computer Science 1075. (1996) 271–287
57. B. Sandak, R. Nussinov, and H.J. Wolfson: A method for biomolecular structural
recognition and docking allowing conformational flexibility. J. Comp. Biol. 5(4)
(1998) 631–654
58. D. Sankoff and J. B. Kruskal: Time Warps, String Edits and Macromolecules:
The Theory and Practice of Sequence Comparison. Addison-Wesley, 1983
59. J. T. Schwartz and M. Sharir: Identification of Partially Obscured Objects in
Two Dimensions by Matching of Noisy ‘Characteristic Curves’. The Int. J. of
Robotics Research. 6(2) (1987) 29–44
60. J.C. Setubal and J. Meidanis: Introduction to Computational Molecular Biology.
PWS, Boston, 1997
61. M. Shatsky, Z. Y. Fligelman, R. Nussinov, and H. J. Wolfson: Alignment of
flexible protein structures. In Int. Conf. on Intell. Systems for Mol. Bio., (2000)
329–343
62. G. Stockman: Object recognition and localization via pose clustering. J. of Com-
puter Vision, Graphics, and Image Processing. 40(3) (1987) 361–387
63. W. R. Taylor and C. A. Orengo: Protein structure alignment. J. Mol. Biol. 208
(1989) 1–22
64. C. J. Tsai, B. Ma, Y.Y. Sham, S. Kumar, H. J. Wolfson, and R. Nussinov:
A hierarchical building-block-based computational scheme for protein structure
prediction. IBM J. of Research and Development. 43(3-4) (2001) 513–523
65. C.J. Tsai, S.L. Lin, H.J. Wolfson, and R. Nussinov: A dataset of protein-protein
interfaces generated with a sequence order independent comparison technique.
J. Mol. Biol. 260 (4) (1996) 604–620
66. I.A. Vakser: Protein docking for low resolution structures. Protein Engineering.
8 (1995) 371–377
67. I.A. Vakser: Main chain complementarity in protein recognition. Protein Engi-
neering. 9 (1996) 741–744
68. I.A. Vakser, O.G. Matar, and C.F. Lam: A systematic study of low resolution
recognition in protein–protein complexes. Proc. Natl. Acad. Sci. USA. 96 (1999)
8477–8482
69. G. Verbitsky, R. Nussinov, and H. J. Wolfson: Flexible structural comparison
allowing hinge bending, swiveling motions. PROTEINS: Structure, Function
and Genetics. 34 (1998) 232–254
70. P.H. Walls and J.E. Sternberg: New algorithms to model protein-protein recog-
nition based on surface complementarity; applications to antibody-antigen dock-
ing. J. Mol. Biol. 228 (1992) 227–297
71. H. J. Wolfson: Generalizing the generalized Hough transform. Pattern Recogni-
tion Letters. 12(9) (1991) 565–573
72. H. J. Wolfson and R. Nussinov: Structural Bioinformatics group.
http://bioinfo3d.cs.tau.c.il
Geometric Methods for Protein Structure
Comparison

Carlo Ferrari and Concettina Guerra

Department of Information Engineering, University of Padova,

Via Gradenigo 6a, 35131 Padova, Italy
tel: 039-049-8277821 fax: 039-049-8277826
{carlo,guerra}@dei.unipd.it

1 Introduction
Protein structural comparison is an important operation in molecular biology
and bionformatics. It plays a central role in protein analysis and design. As
proteins fold in three dimensional space, assuming a variety of shapes, a careful
characterization of their geometry is needed to study their function which
is known to be related to the shape. Moreover, the comparison of protein
structures is essential to infer evolutionary information.
The problem of comparing three-dimensional structures has been widely
studied in other disciplines such as computer vision and image processing,
robotics, astronomy and some core methods have migrated from these disci-
plines to bionformatics.
There are many instances of the protein comparison problem that have
been addressed; they include: 1) protein pairwise comparison, 2) protein clas-
sification, to organize all known structures in a biologically relevant groups,
3) searching for common folding patterns and three-dimensional motifs, 4)
studying of protein interaction to identify binding sites for drug design.
From the application point of view, it is important to mention how the
growth of the Protein Data Bank (PDB) asks for effective automatic pro-
cedures for classification and search of the database elements. Currently the
PDB contains more than 17,000 structures and this number is rapidly growing.
The protein comparison may involve different levels of representations of
the three dimensional protein structures, from the atomic level to the level of
secondary structures. Most methods presented in the literature deal with a
protein representation in terms of atomic coordinates
and therefore with a matching problem that uses as basic elements sets of
points (atoms).

C. Guerra, S. Istrail (Eds.): Protein Structure Analysis and Design, LNBI 2666, pp. 57-82, 2003.
 Springer-Verlag Berlin Heidelberg 2003
58 Carlo Ferrari and Concettina Guerra

Other approaches are based on secondary structures, i.e. α helices and β

strands, that play an important role in the functional behavior of a protein.
The secondary structure elements are represented as vectors in 3D space.
An alignment of α helices and β strands may be used for fast retrieval of
folds or motifs from the PDB. On the other hand, the comparison of secondary
structures can be used as the first step in a two step comparison procedure
that first identifies possible candidate solutions in a fast way and then refines
the solutions taking into consideration the atomic descriptions of proteins.
Another advantage of a structural comparison of secondary structures is that
it allows to study the folding process by tracing the evolution of the fold from
the molden state.
Most of existing approaches allow to detect global similarity between entire
proteins as well as local similarity ([1], [20], [21], [23], [26], [27], [41], [46], [51]).
The integration of strategies operating at different levels of representations
appears very promising to achieve robustness and efficiency. Extensive surveys
on the subject of protein comparison exist enphasizing different aspects of the
general problem [6], [37], [50].
In this paper, we review some of the theoretical results on the computa-
tional complexity of the algorithms designed to obtain optimal solutions to
the problem of matching sets of points using specific metrics. From a theo-
retical point of view, the problem has been extensively studied in the area of
computational geometry, where it is often formulated as the problem of find-
ing correspondences between sets of geometric features (for instance, points or
segments). From these studies it appears that, in most practical cases, exact
algorithms are too time consuming to be useful. Thus, approximate algorithms
are considered that are computationally practical and at the same time are
guaranteed to produce solutions that are within a certain bound from optimal.
Furthermore, we discuss methods for the estimation of rigid transforma-
tions under different metrics such as the Root Mean Square Deviation (RMSD)
and the Hausdorff distance. Geometric indexing techniques prove their effec-
tiveness in searching large protein databases and they are presented in details.
Finally graph-theoretic protein modeling is reviewed as it is useful in designing
algorithms for substructure identification and comparison.
Throughout the paper, we will use pure geometric information, ignor-
ing other properties associated with atoms. Chemical properties, such as hy-
drophobicity, charge, etc. may be important in protein comparison and often
they can be easily incorporated in a matching procedure. The use of such
properties may help reduce the computation time by allowing pruning of the
possible associations at early stages of the processing. However, we will not be
consider these other properties in this chapter. Applications of protein com-
parison are an important subject; since they discussed in the two previous
chapters of this volume, they are not considered here.
 Geometric Methods for Protein Structure Comparison 59

2 Protein Description

A protein is a sequence of aminoacids linked by peptide bonds. An aminoacid

consists of a carbon atom Cα to which are attached a hydrogen atom, an amino
group, and a carboxyl group. The 20 aminoacids differ in the side chain or
residue attached to the Cα atom. The peptide bonds between the aminoacids
in the chain join the carboxyl group of one aminoacid with the amino group of
the next eliminating water in the linking process. The sequence of aminoacids
is generally referred to as the primary structure of a protein. Its length varies
from a few tens to few thousands aminoacids. A different level of protein
representation, known as secondary structure, describes a protein in terms of
recurrent regular substructures, such as the α helices and the β strands. The
tertiary structure is the packing of the structural elements into the 3D shape.
The protein may contain several chains forming its quaternary structure. For
a survey of the protein architecture see [5], [38].

Fig. 1. The volumetric representation of protein 1rpa

The volumetric representation of a protein is displayed in figure 1 where

all atoms are shown as balls; the secondary structure elements of the same
protein are displayed as ribbons in figure 2.
Arrangements of the secondary structures α helices and β strands are the
basis for the protein structural classification of SCOP [44]. In the SCOP clas-
sification hierarchy, the fold level corresponds to the last level of the hierarchy,
the other two being family and superfamily. Proteins sharing a fold have the
same major secondary structures but do not necessarily have a common evolu-
tionary relationship, unlike proteins clustered into families and superfamilies.
The similarity in the arrangements of secondary structures in a fold may be
due to the physical and chemical properties of the packing of the proteins.
60 Carlo Ferrari and Concettina Guerra

Fig. 2. The secondary structure representation of protein 1rpa

Fig. 3. The ribbon representation of protein 1bxa showing a β sandwich

Fig. 4. The ribbon representation of protein 3por forming a β barrel

 Geometric Methods for Protein Structure Comparison 61

Common structural arrangements of secondary structures or motifs have

been identified within the folds and they include, among others, the β-
sandwich and the β-barrel, as seen in figure 3 and 4, respectively.
Approaches to protein comparison use different protein structural de-
scriptions. A complete structural description is given by the 3D coordinates
(x, y, z) of the individual atoms of a protein. Often only the Cα atoms of the
aminoacids, that form the so-called backbone of a protein, are considered for
comparison.
A more compact description is in terms of the linear vectors associated to
the structural elements helices and β strands. While most of the comparison
approaches are based on the atomic description of a protein, the secondary
structure description may provide a fast method to retrieve substructures or
motifs from large protein databases. Furthermore, it is often used as a first
step when searching in a database for the most similar protein with respect
to a target protein. In fact, hypotheses of similarity for the target protein
are generated in a fast and efficient way based on secondary structures only;
such hypotheses are further verified by a more refined and costly process
that is only applied to those hypothesized proteins. This two-step procedure
may considerably speed up the protein comparison when large databases are
involved.
For most proteins in the PDB, secondary structures are annotated by
the original depositor who provides the starting and ending residue numbers
of all secondary structures. However programs have been designed for the
assignment of the secondary structures from the PDB files and for the analysis
of the overall and residue-by-residue geometry of a protein [10], [34], [35].

Fig. 5. The vectorial representation of protein kinase CK2

Several programs have also been developed to yield the vectorial represen-
tation of a protein [17], [41]. Singular-value decomposition (SVD) is a stan-
dard routine [4], [19] to find the axes of α helices and the best fit segments
62 Carlo Ferrari and Concettina Guerra

for the β strands. In this routine typically only Cα atoms are used. Other
simpler methods derive the vector associated to a β strand either by directly
connecting the starting and ending residues of the β strand assignments or
by connecting two points that are computed as the average points of few of
the extreme residues on both sides of the strand [51]. This second approach
is less sensitive to curved or kinked structures. Figure 5 shows the vectorial
representation of protein kinase CK2, where each segment is displayed as a
cylinder of fixed radius.

3 Structural Comparison: Problem Formulation

The general matching problem can be informally defined as follows: Given two
sets of geometric features, either points or line segments, determine the largest
common subsets, i.e. the subsets of maximum size, that are geometrically
similar. In the case of proteins, points correspond to atoms and segments are
the axes of the secondary structures.
There are many variants of the matching problem that have been consid-
ered in many different contexts. We now give more formal definitions of the
problem with varying degrees of computational complexity.

Problem 1. Consider two sets of geometric elements A = {a1 , a2 , · · · , an }

and B = {b1 , b2 , · · · , bm } in three-dimensional space and assume that they
have the same cardinality, i.e. n = m, and that the element ai corresponds to
the element bi . Find the transformation g between the two sets that minimizes
a given distance metric D over all rigid body transformations T , i.e.
minT D(T (A), B)
Problem 2. Consider two sets of geometric elements A = {a1 , a2 , · · · , an }
and B = {b1 , b2 , · · · , bm } in three-dimensional space. Find the transformation
g between the two sets that minimizes a given distance metric D over all rigid
body transformations T , i.e.
minT D(T (A), B)

This problem differs from the previous one because no correspondence is

known a priori between the elements of the two sets.
Problem 3. Given two sets A = {a1 , a2 , · · · , an } and B = {b1 , b2 , · · · , bm } in
three-dimensional space and a real number δ > 0, find a maximum-cardinality
set of pairs of elements, one element in A and the second one in B, such that
the distance d between each pair of elements is at most δ.
In this problem we are interested in the largest subset of corresponding ele-
ments of A and B (generally, in practical applications the sets are required to
be above a certain size).
 Geometric Methods for Protein Structure Comparison 63

Different metrics have been used in the literature to determine the struc-
ture similarity between geometric objects. The most common metric is the
RMSD (Root Mean Square Deviation) defined for point sets as follows:


D(A, B) = RM SD(A, B) = ( i=1,n d(ai , bi )2 )1/2
where d is the Euclidean distance between two points, and assuming that
the sets have the same cardinality n, ai corresponds to bi .
The RMSD distance is useful when comparing very similar structures,
as those produced during the christallographic analysis or NMR at different
stages of the process. Its disadvantage becomes apparent in the presence of
outliers when the proteins are not structurally close. The existence of even
few outliers may significantly alter the value of the distance and therefore the
determination of the optimal superposition of the two structures.
A second important definition between two point sets is based on the use
of contact maps [33]. The chapter by G. Lancia and Sorin Istrail in this book
deals extensively with contact maps and they are not further discussed here.
Another definition is the Hausdorff metric widely used in the area of com-
puter vision and image processing, in astronomy and extensively studied in
the field of computational geometry. The Hausdorff distance H(A, B) between
A and B is:
H(A, B) = max(h(A, B), h(B, A)))
where h(A, B) is the one-way Hausdorff distance from A to B given by:
 
h(A, B) = maxai ∈A minbj ∈B d(ai , bj )
In the following we discuss different approaches to solve the above three
problems with different metrics. Problem 1 and its solutions are presented in
section 5. Problem 2 with the Hausdorff distance as metric is considered in
section 6.Problems 3 is reviewed in sections 8.

4 Representation of Rigid Transformations

A large number of methods have been proposed in the literature to compute
the rigid body transformation between two sets of 3D points. They differ with
respect to the transformation representation, and the minimization procedure.
A survey by Sabata and Aggarwal [49] lists several representation of transfor-
mations and approaches to solve this problem using both closed form solutions
and iterative solutions. The book [16] gives a clear description of many issues
related to rigid transformations with enphasis on visualization aspects.
Here we describe several representations of rigid transformations while the
next section is devoted to review methods to compute them.
A rigid motion of an object is a motion that preserves the distances be-
tween object points. The net movement of a rigid body from one configuration
64 Carlo Ferrari and Concettina Guerra

to another configuration (via a rigid motion) is called a rigid displacement. A

rigid transformation applied to an object model represents a rigid displace-
ment of the object itself. Rigid transformations form the theoretical infras-
tructure both for studying actual objects motions and for predicting possible
(or hypothetical) motions. In the early 1800s Chasles and Poinsot proved
that every rigid body displacement can be realized by a rotation about an
axis combined with a translation parallel to that axis. This motion is what
it is usually referred as a screw motion. Different motion representations are
presently used (mainly in computer vision, computer graphics and robotics),
that can be roughly classified as local or global representations [43].
Let us start by formally introducing the definition of rigid body transfor-
mation. A rigid body transformation is a mapping g : R3 → R3 , that must
satisfy the properties:
 g(q) − g(p) = q − p  for all points p, q ∈ R3
g(v × w) = g(v) × g(w) for all vectors v, w ∈ R3
The former condition says that lengths are preserved and the latter con-
dition says that internal reflection is not allowed. As a consequence of the
above definitions, rigid body transformations also preserve the inner product,
in particular, orthogonal vectors are transformed to orthogonal vectors. In
general, a rigid body transformation takes right-handed orthonormal coordi-
nate frames to right-handed orthonormal coordinate frames. It is important
to point out that even if the distance between points and the cross product
between vectors are fixed, particles in a rigid body can move related to each
others, because they can rotate (but not translate) with respect to each other.
Then the motion of a body can be described by the motion of any one point
and the rotation of the body around this point. Hence a right-handed Carte-
sian coordinate frame can be attached to some point of the body and the
motion of individual points can be traced from the motion of the body frame
and the motion of the frame attachment point. Due to its importance we first
consider pure rotational motion.
Pure rotational motion in R3 can usually be described by a proper 3x3
matrix, that can be defined by stacking next to each other, the coordinates
of the principal axes of a coordinate frame B (the body frame) relative to a
coordinate frame A (the inertial frame). Such a matrix is called a rotational
matrix: its columns are mutually orthonormal and its determinant is +1. The
set of all 3x3 matrices that satisfy these two conditions is denoted by SO(3),
where SO stands for Special Orthogonal. SO(3) is a group under the matrix
multiplications, with the identity matrix I as the identity element. SO(3) is
the rotation group of R3 . A 3x3 rotation matrix can be seen as

 
r11 r12 r13
 r21 r22 r23 
r31 r32 r33
 Geometric Methods for Protein Structure Comparison 65

with nine different parameters, the orthonormality conditions (for the ma-
trix columns) add three more constraints, while the sign of the cross product
adds other three constraints. These six constraint equations, reduce the degree
of freedom to three, that is, only three parameters are needed to completely
represent a pure rotation in R3 . The matrix coefficients rij can be expressed
in term of these three parameters.
A rotation matrix R ∈ SO(3) represents a rigid body transformation. In
fact it can be proved that it preserves distances and orientations, that is:
•  Rq − Rp = q − p  for all points p, q ∈ R3
• R(v × w) = Rv × Rw for all v, w ∈ R3
Moreover a rotation matrix can be seen as an operator that takes the co-
ordinates of a point (or vector) from a frame to another. Let pb the coordinate
of a point P with respect to the frame B, and Rab the rotation matrix, the
coordinates of P with respect to the frame A are given as:
pa = Rab pb
The pure rotation operator is a linear operator (with the additional con-
straint that it is orthonormal). A sequence of two (or more) rotations will
result in a single combined rotation and conversely a given rotation can be
decomposed using two or more rotations. Rotation matrices can be combined
to form new rotation matrices using matrix multiplication. If a frame C has
orientation Rbc relative to frame B and B has orientation Rab from frame A,
then the orientation of C with respect to A is given by:
Rac = Rab Rbc
−1 T
In particular, as we could expect, Rab Rba = I and Rba = Rab = Rab .
An important result about rotations is the Euler Theorem that establishes
that any rotation R ∈ SO(3) is equivalent to a rotation about a given axis
ω ∈ R3 ( ω = 1), by an angle θ ∈ [0, 2π). In fact, it is possible to represent
the motion of a single point p rotating about ω at a constant unit velocity,
with the following differential equation:
ṗ(t) = ω × p(t) = ω̂p(t)
Solving this equation gives the expression for a single rotation about ω by
θ, that is R(ω, θ) = eω̂θ . The matrix ω̂ is defined as follows:
 
0 ω3 −ω2
 −ω3 0 ω1 
ω2 −ω1 0
where ω T = [ω1 , ω2 , ω3 ] and it has the property that ω̂ T = −ω̂. Such a matrix
is a skew-symmetric matrix. If  ω = 1, ω̂ is a unit skew-symmetric matrix.
It can be proved that the exponential eω̂θ can be rewritten in term of the
skew-symmetric matrix, resulting in the so-called Rodriguez’s formula:
66 Carlo Ferrari and Concettina Guerra

eω̂θ = I + ω̂ sin θ + ω̂ 2 (1 − cos θ)

This method represents rotations through the equivalent axis representa-
tion . This is perhaps the most intuitive way of representing rotations. How-
ever, it has some disadvantages. The representation is not unique: in fact



choosing ω = −ω and θ = 2π − θ gives the same rotation as ω and θ,
being the exponential map many-to-one. Moreover, singularities occur when
θ = 0 and the exponential equals I, in such a case ω cannot be determined.
These singularities create problems when computing the rotations. Finally the
transformations resulting from the composition of multiple rotations cannot
be easily computed.
Besides the canonical coordinates, there exist different coordinate systems
for representing the rotation group, mainly used in robotics systems. The fol-
lowing method of describing the orientation of the coordinate frame B relative
to A uses the Euler angles. At the beginning frames A and B are coincident.
First, the frame B is rotated about the z-axis by an angle α, then it is rotated
about the (new) y-axis by an angle β and finally B is rotated about the z-axis
by an angle γ. The triple of angles (α, β, γ) represents the overall rotation,
and the angles α, β, γ are called the ZYZ Euler angles.
These three rotations occurs at principal axes and the global rotation
matrix can be computed from the three rotation matrices related to the three
elementary rotations, that is, giving the specific values for α, β, γ it is easy
to compute both Rab and Rba . The converse question of whether the map
from SO(3) to α, β, γ is surjective is important. It can be proved that for any
R ∈ SO(3) it is possible to determine the Euler angles. This representation
suffers from the problem of singularity at R = I.
In order to solve this problem new methods should be studied. Rotations
in a 2D space can be represented by complex numbers on the unit circle.
When moving to a 3D space, it is possible to generalize this idea, by in-
troducing quaternions . Formally a quaternion Q is a 4-tuple of the form
< q0 , q1 , q2 , q3 >: where q0 is the scalar component of Q and −
→q = (q1 , q2 , q3 )
is the vector component of Q. Hence, Q = (q0 , q ) with q0 ∈ R and −
→
− →
q ∈ R3 .
The set of quaternions is a 4D vector space over the reals and it forms a
group with respect to quaternion multiplication (denoted “·”). Quaternions
multiplication is defined as follows:
Q · P = (q p − −
0 0
→
q ·−p, q −
→
0
→p +p −
→
q +−
0
→q ×−
→p)
The unit quaternions are the subset of all quaternions Q such that  Q =
1, where  Q 2 = q02 + q12 + q22 + q32
Each rotation matrix R = eω̂θ correspond to a unit quaternion defined
as Q = (cos(θ/2), ω sin(θ/2)). It can be proved that if Qab correspond to a
rotation of frame A to B and Qbc correspond to a rotation of frame B to C, then
the rotation between frame A to C is given by the quaternion Qac = Qab · Qbc .
An alternative representation of rotations, often used in computer vision, is
the unit quaternion, Given a unit quaternion Q = (q0 , −→q ) the corresponding
 Geometric Methods for Protein Structure Comparison 67

rotation is given by θ = 2 cos−1 (q0 ) and ω = k − →

q with k = 1/ sin(θ/2), if
θ= 0, ω = 0 otherwise.
The 3x3 rotation matrix R in terms of the unit quaternion is directly given
by:
 2 2 
q0 q1 − q2 2 − q3 2 0 0
R= 0 1 0
0 01
Quaternions give a global parametrization of SO(3), at the cost of using four
numbers instead of three to represent a rotation. Since the quaternions space
has a group structure that directly corresponds to that of rotations, they
provide an efficient representation without suffering from singularities.
Rigid body displacements usually are not limited to pure rotations (even
if pure rotations represent an important subset), but they generally consist of
rotations and translations. Pure translations have a very simple representa-
tion: given two co-oriented frames A and B, with pab the representation in A
of the origin of B, for any point q ∈ R3 , qa = pab + qb . Pure translations can
be represented by 3D vectors.
A rigid body motion (a rigid body displacement) then can be represented
by pab ∈ R3 and R ∈ SO(3). The Cartesian product of R3 with SO(3),
represents all the rigid body motions and it is denoted as SE(3) (that stands
for special Euclidean group):
SE(3) = {(p, R) : p ∈ R3 , R ∈ SO(3)}
Each element of SE(3) serves both as a specification of a rigid body place-
ment (with respect to a fixed environment frame) and as a transformation
taking the coordinates of a point from one frame to another. If a is a 3D point
and a is its corresponding point after a rigid transformation is applied, then
the following relation holds:
a = Rab a + pab
The above can be expanded into:
      
ax r11 r12 r13 px
 ay  =  r21 r22 r23  +  py 
az r31 r32 r33 pz
The transformation of points can be usefully represented using the homo-
geneous representation . The homogeneous representation maps points and
vectors in a 4D space, by adding a forth coordinate. The homogeneous coor-
dinates of a point q = (q1 , q2 , q3 ) are q̄ = (q1 , q2 , q3 , 1), while the homogeneous
coordinates of a vector v = (v1 , v2 , v3 ) are v̄ = (v1 , v2 , v3 , 0). Then a rigid
transformation becomes:
   
a Rab pab a
ā = = = g ab ā
1 01 1
The 4 × 4 matrix ‘g ab is called homogeneous representation of g =
(pab , Rab ) ∈ SE(3). Within homogeneous representation we obtain a linear
68 Carlo Ferrari and Concettina Guerra

representation of rigid body motions, and the standard matrix multiplication

is the composition rule for rigid body motion.
Homogeneous coordinates are useful for representing twists . A twist is a
couple (v, ω̂), with v ∈ R3 and ω̂ a skew-symmetric matrix. A twist can be
written as:

ω̂ v
ξˆ =
00

Of course ξˆ ∈ R4x4 , while the twist coordinates ξ = (v, ω) ∈ R6 . It can be

proved that the exponential of a twist multiplied by a scalar, is an element
of SE(3), that is, it represents a rigid transformation. Conversely it can be
proved that every rigid transformation can be written as the exponential of
some twist, that is, given a g ∈ SE(3), there exists a twist ξ̂ and θ ∈ R such
ˆ It is important to remind the reader that the exponential
that g = exp(ξθ).
map is many-to-one and hence the choice of ω and θ may not be unique for
solving the rotational component of the motion.
The concept of twist helps us to show that every motion is a screw motion.
A screw motion is a motion which consists of a rotation about an axis in space
by an angle of θ, followed by a translation along the same axis by an amount
d. This motion is called a screw motion since it remind us to the actual motion
of a screw that rotate and translates about the same axis. A screw consists
of an axis l, a pitch h and a magnitude M , while a screw motion represents
rotation by an amount θ = M , about the axis l followed by a translation
parallel to the axis l by an amount hθ. The overall rigid displacement can be
computed and it results as:
 ω̂θ
e (I − eω̂θ )q + hθω
g=
01
If we choose v = −ω × q + hω, the twist coordinates ξ = (v, ω) generate
the given screw motion. Going one step further it can be proved that it is
possible to define a screw associated with every twist. Finally it is possible to
conclude that every rigid body motion can be realized by a rotation about an
axis combined with a translation parallel to that axis.

5 Determination of 3D Rigid Transformations

In this section we review solutions for problem 1, as formulated in section 3.
Consider two sets of points A = {a1 , a2 , · · · , an } and B = {b1 , b2 , · · · , bm } in
3D space with the same cardinality. Assume that point correspondences are
known, i.e. that point ai corresponds to the point bi . The problem is to derive
the rigid body transformation that optimally maps A into B. This problem is
known as the absolute orientation problem. Points measurements are affected
by some noise and errors, due both to the estimation of the point coordinates
and to the determination of the point correspondences. Formally speaking
 Geometric Methods for Protein Structure Comparison 69

this problem can be stated as a minimization problem, according to the least

square error criterion.
In the literature there exist a large numbers of algorithms that compute
3-D rigid transformation between two sets of geometrical features. For our pur-
poses, it is interesting to review some of the most popular closed-form solution
using correspondent points [39]. The rotational and translational components
are computed as solutions to a least square formulation of the problem. The
proposed approaches differ mainly in the transformation representation. The
first was developed by Arun, Huang and Blostein [4] and it is based on com-
puting the singular value decomposition (SVD) of a matrix.
The problem has been formalized as a particular version of the well-studied
orthogonal Procrustes problem , that can be stated as it follows:
minimize  A − BR  with respect to R, subject to RT R = I
where A, B ∈ Rm×3 are given by the set of 3-D vectors ai , and the set of 3-D
vectors bi , respectively, and R ∈ R3×3 , is orthogonal.
This problem is equivalent to the following one:
maximize trace(RT B T A) with respect to R, subject to RT R = I
In fact

(A − BR)(A − BR)T = trace(AT A) + trace(B T B) − 2trace(RT B T A)

This problem can be approached using “The Singular Value Decomposition

(SVD) Theorem”, that can be stated as follows: Theorem. If A is a real m-by-

n matrix then there exist two orthogonal matrices U ∈ Rm×m and V ∈ Rn×n
such that U T AV = diag(σ1 , . . . , σp ), p = min{m, n} where σ1 ≥ σ2 ≥ . . . ≥
σp ≥ 0. The problem of maximizing trace(RT B T A) can be solved through

the computation of the SVD of B T A. In fact let Σ be the SVD of B T A, that

is
Σ = U T (B T A)V = diag(σ1 , σ2 , σ3 )
and we define a new orthogonal matrix Z = V T RT U . Then we obtain:
trace(RT B T A) = trace(RT U U T (B T A)V V T = trace(RT U ΣV T ) then

trace(RT U ΣV T ) = trace(V T RT U ΣV T V = trace(ZΣ)

The previous expression can be rewritten as:

3
trace(RT U ΣV T ) = trace(ZΣ) = zii σi
i=1

as Z is orthogonal the best choice for R is obtained when Z becomes the

identity matrix. Hence R = U V T
70 Carlo Ferrari and Concettina Guerra

This general method proves itself useful for computing 3-D rigid transfor-
mation estimation between two sets of corresponding points. The algorithm
can be sketched as a two step algorithm. The first step computes
Nthe op- T
timal rotation matrix R using the 3x3 correlation matrix H = i=1 ai bi
T
through its singular value decomposition (H = U diag(σ1 , σ2 , σ3 )V ) obtain-
ing R = V U T . The second step computes the optimal translation vector as
P = a − Rb.
A similar approach computes the eigenvalues of a proper derived matrix
(the orthonormal matrices ) instead. It was proposed by Horn, Hilden and
Negahdaripour [25]. With this method the correlation matrix H is firstly
computed. However, rather than computing its SVD, a polar decomposition
is used, such that H = RS, where S = (HH T )1/2 . The optimal rotation is
given by:
1 1 1
R = H T √ u1 uT1 + √ u2 uT2 + √ u3 uT3
λ1 λ2 λ3
where {λi } and {ui } are the eigenvalus and eigenvectors of the matrix HH T .
Representing rotations using unit and dual quaternions gives two more
techniques that have been proposed respectively by Horn [28] and by Walker,
Shao and Voltz [56]. The former method asks to rewrite the minimization
problem in the quaternion framework. A new 4x4 matrix can be constructed
from the correlation matrix H as:

H00 + H11 + H22 H12 − H21 H20 − H02 H01 − H10

− − − 
K=

H 12 H 21 H 00 H 11 H 22 H 01 + H 10 H20 + H02
 H20 − H02 H01 + H10 H11 − H00 − H22 H12 + H21 
H01 − H10 H20 + H02 H12 + H21 H22 − H11 − H00

The optimal rotation is the eigenvector related to the largest positive eigen-
value of K.
The latter method is the most significantly different of the four. It was
designed ot minimize the equation;

L
N
Σ 2 = i=1 αi  n1 i − Rn2 i 2 + i=1 βi  ai − Rbi − P 2
where {n1i } and {n2i } are two sets of corresponding unit normal vectors,
and {αi }, {β} are weighting factors reflecting data reliability. Dual quater-
nions for representing both rotation and translation are used and again the
minimization problem can be rewritten in this new framework, resulting in
new equations involving the parametrization of the dual quaternions. Again
optimal values for R and P can be computed.
These four algorithms can be compared with respect to their accuracy,
stability and efficiency [40]. Experimentations shows that no one algorithm
is superior in all case. In fact difference in accuracy (on nondegenerate 3-D
point sets) is almost insignificant. Stability is more discriminant instead. The
SVD and the unit quaternion method are very similar and usually the most
stable. In terms of efficiency, the orthonormal matrix looks quicker with small
data sets, while the dual quaternions method is superior with larger data
 Geometric Methods for Protein Structure Comparison 71

sets (according to proper computer memory configuration). In conclusion, the

SVD should provide the best stability and accuracy, even if is not as efficient
as dual quaternions with large data set. Otherwise the unit quaternions can
be chosen on smaller data set for slightly better speed performance.

6 Geometric Pattern Matching

In this section we focus on the Hausdorff distance, defined in section 3, and
review both theoretical and practical approaches to compute it.
Much work has been done on the computation of the Hausdorff distance.
In the area of computational geometry exact algorithms have been studied for
the problem of deciding whether there exist a transformation that maps one
set of points into another set within a given distance. Fundamental robust-
ness issues are discussed in [3]. Chew et al. [10] have considered the problem
of matching point-sets in a d-dimensional space using the Hausdorff distance
under translation only. For the case d = 3, they provide exact solutions in
O(n3 log2 n) time. Extensions [11] to the more general case of Euclidean mo-
tion and of sets of segments have obtained exact solutions in O(n6 log2 n)
time. However they are limited to the case of planar sets.
Exact algorithms cannot be used in most practical applications where mea-
surement errors and noise are present; furthermore, the high computational
complexity of the exact algorithms make them impractical for use in real prob-
lems. For these reasons, approximate solutions for the case of point sets, both
in 2-dimensional and in 3-dimensional space, have been considered [15].
In the field of computer vision, an efficient multi-resolution technique for
comparing images using the Hausdorff distance has been presented in [29]
where the space of possible transformations is limited to translations and scal-
ing; in [48] the above technique is extended to affine transformations. Affine
transformations are used in [24] for matching point sets. Other approaches
to matching sets of segments in 3D space based on various techniques and
metrics are given in [9], [24], [31]
The computation of the Hausdorff distance does not necessarily produce
a one-to-one correspondence between the elements of the two sets; it may
happen, in fact, that multiple elements in one set are associated with a sin-
gle element of the other set. This is unlike most existing object recognition
methods that give an explicit pairing.
Approaches have been proposed for the computation of the Hausdorff dis-
tance between sets of segments associated to secondary structures [10]. The
standard Hausdorff distance provides a good metric over point sets but does
not preserve the notion of relevant subsets like the segments. To keep infor-
mation relative to the line segments in the definition of the distance function
an alternative definition of the Hausdorff metric between sets of segments has
been introduced in [23], together with efficient approximate algorithms for its
computation Given two sets A = {a1 , a2 , · · · , am } and B = {b1 , b2 , · · · , bn } of
72 Carlo Ferrari and Concettina Guerra

line segments ai and bj , the Segment Hausdorff distance HS (A, B) between

A and B is:
HS (A, B) = max(hS (A, B), hS (B, A))),
where hS (A, B) is the one-way Segment Hausdorff distance given by:

hS (A, B) = max min H({ai }, {bj })

ai ∈A bj ∈B

The matching strategy is essentially an alignment that selects a few “rep-

resentative” segments of the set A and computes a rigid transformation based
on an hypothesized correspondence between the representative segments of A
and a group of segments of B. It then verifies the hypothesis by computing
the distance measure for such a transformation. The above steps are repeated
for all possible groups of segments of A. More precisely, the algorithm looks
for a rigid body transformation g (translation plus rotation) that minimizes
the distance between two sets of segments, A and B and it consists of the
following three main steps:
Step 1 determine a translation P ;
Step 2 determine a rotation R;
STep 3 evaluate the distance between g(A) and B, where g is the combined
transformation.
The rigid body transformation is obtained by selecting three representa-
tives for each of the two sets A and B that are affine independent elements.
First, a representative segment a for A is randomly chosen. This repre-
sentative is paired with each element b of B. For each such pair (a, b), the
translation P is defined by taking the mid-point am of a into the mid-point
bm of b. This choice of the translation minimizes the distance between the
transformed segment P (a) and b.
To define the rotation, two additional independent elements of A are
needed. The second representative a is chosen as the segment containing
the point af farthest from am . The third representative segment a is chosen
so that it contains the point ad at maximum distance from the line am af . It
is easy to see that the points af and ad must each be an endpoint of some
segment. The condition that is enforced is the affine independence of the three
points am , af and ad . These choices bind the error due to the approximation.
The next step of the algorithm is to choose the segments b and b of B in
all the m2 possible ways. For each b and each endpoint of b , consider the
rotation that has origin in am and that makes af and am to become collinear
with the endpoint of b . Define R as the one of the above rotations that min-
imizes the distance between af and the endpoints of b . Then define R to be
the rotation about the axis am af that brings ad closest to an endpoint of b .
Apply the transformations R R P (A). Finally choose over all the triplets b,
b and b the transformation g that resulted in the smallest distance.
 Geometric Methods for Protein Structure Comparison 73

The time complexity of the overall algorithm is O(mn3 log n). In fact,
using the Hausdorff metric, the nearest neighbor query in a set of segments
(to identify the segment of B “closest” to a segment of A), reduces to a nearest
neighbor query among points in R6 that can be performed in optimal O(log n)
time within a known error bound. It is shown that the error introduced with
the approximation is within a bounded factor from optimal. This bound is
the same as the bound obtained in [15] for the simpler case of point sets.
Experiments have been conducted on several proteins and the results were
consistent with previous studies. As an example presented in [22], figure 6
shows the superposition of two sets of segments associated to proteins 1rpa
and 1rpt, with very similar structures.

Fig. 6. The alignment of proteins 1rpa (red segments) and 1rpt (green segments)

7 Indexing Techniques
Indexing techniques, initially proposed in the field of computer vision by Wolf-
son et al., have found interesting applications in the area of bioinformatics.
Indexing or geometric hashing provides a way to efficiently search a large
database of proteins by storing redundant transformation invariant informa-
tion about the proteins in a hash table, from which this information can be
easily retrieved. The construction of the hash table, that constitutes the most
complex part of the entire process, is done off-line at a preprocessing stage.
Indexing techniques have been applied to compare proteins at different
levels of representations [2], [8], [13], [14], [32], [36], [55] (see also the chapter
by H. Wolfson of this volume).
One major distinction of the comparison approaches is whether they are
order dependent or order independent, in other words whether the use the
74 Carlo Ferrari and Concettina Guerra

order of the elements along the protein chain as a constraint in the correspon-
dence process. Indexing techniques do not take into consideration the order of
elements (either points or secondary structures) along the chain and therefore
fall into the category of order-idenpendent methods.
For matching 3D point sets, quadruples of points are used to define ref-
erence frames or bases in which the coordinates of all other points remain
invariant. Models are stored into the table by considering all possible combi-
nations of quadruples of points as bases and using the invariant coordinates
of the remaining points to index the table. At recognition time, if the correct
quadruple of points is chosen from the image points, the candidate matches
are efficiently retrieved from the corresponding entries of the hash table.
Here we concentrate on the application of hashing techniques at the sec-
ondary structure level involving transformation invariant properties of vectors
associated to the secondary structures. The programs 3dSEARCH [51] and
3d-Lookup [26], based on hashing, compute geometric properties of pairs of
secondary structures. They both construct the hash table by a procedure that
consists of the following steps for the insertion of a protein in the database:
Step 1. For each pair of vectors of the protein, compute a reference frame
or coordinate system identified by the two endpoints of one vector and by
the orientation of the other vector.
Step 2. For each remaining vector in the protein, compute its coordinates
in the reference frame defined in the previous step 1.
Step 3. The coordinates are quantized into fixed size interval and used to
access the entry of the table corresponding to those coordinates where the
following pair of information is stored: 1) name of the protein that hashed
into it; 2) identifiers of the two vectors used as reference frame.
Once the hash table is built, each secondary structure vector from the given
query structure is simultaneously compared to the entire library of target
structures by simply indexing into this table. Thus, to compare a query protein
to all target proteins the above step 1 and 2 are repeated for the query protein.
Step 3 is replaced by the following:
Step 3’ The coordinates are quantized into fixed size interval and used to
access the entry of a 3d table corresponding to those coordinates where a
vote is cast to every pair (protein name, two vector identifiers) present at
that entry.
At the end of the process the proteins in the table which obtained the most
votes are the candidates for matching.
A recently proposed approach [21] considers triplets of secondary struc-
tures rather than pairs to build the hash table. The three dihedral angles
associated to all triplets of secondary structures are used to index a hash ta-
ble. Let (si , sj , sk ) be a triplet of segments, where s corresponds either to an
α-helix or a β-strand. Let αsr be the dihedral angle formed by two segments
s and r. The dihedral angle between two segments is the angle formed by
 Geometric Methods for Protein Structure Comparison 75

the two planes perpendicular to the straight lines containing the segments
themselves and therefore is defined in the range [0, 180].
The triplet of segments (si , sj , sk ) is then described by the three angles
(αij , αjk , αki ). The three angles, quantized into uniform intervals, provide
three indeces for the table; a fourth index triplet type is used to access the
table; it depends on the types of the secondary structures in the triplets,
whether all α helices, one α helix and two β strands and so forth. Thus a
4-dimensional table is constructed during the set up phase of the method. No
explicit information is present in the tables about the order of the segments
along the polypeptide chain.
Each entry of the table keeps a record of each triplet hashed into it. The
record contains the following information: 1) the name of the protein, 2) the
identifiers for the three vectors, 3) the pairwise distances between the three
vectors. Such distances are used to filter incorrect hypotheses of associations
in the matching process, because false results could be obtained based on
angular information only. The distance is measured as the distance between
the middle points of two segments.
The construction of the table is computation intensive; it requires O(n3 )
time, for n secondary structures. Once it is built it allows fast retrieval of
candidate matches between the query protein and the proteins stored in the
database. The space requirements for this approach may be high. However,
the table is only partially occupied since the three angles are related by the
triangular inequality.
The table can be queried to find similarities in the arrangement of the sec-
ondary structures of a query protein with the proteins stored in the database.
A protein P is matched against the database of proteins by the following
procedure:
Step 1 For each triplet (si , sj , sk ) of secondary structures of P , compute the
three angles (α, β, γ) and the three distances of the associated segments.
Step 2 Access the cell of the hash table indexed by (α, β, γ, triplet type)
and tally a vote for each entry in the cell with similar distance values.
Step 3 Formulate and rank hypotheses of matching by determining the
proteins with the highest number of votes.
The verification of the hypothesized matches may be performed by a pair-
wise comparison between the proteins, either at the level of secondary struc-
tures [23] or by extending the matching to residue level.
Once compiled, the table can be used for different types of comparisons,
for instance for all-to-all structure comparison.
Experiments have been conducted that consisted in building the hash table
for all proteins in the PDB (approx. 14.000) have shown that the approach is
both robust and efficient. The construction of such a table for approximately
350 representative proteins from the PDB has led to interesting observations
about the distribution of the angles of the secondary structures which deviate
76 Carlo Ferrari and Concettina Guerra

from the distribution of randomly chosen vectors more significantly than one
would have expected [47].

8 Graph-Theoretic Approaches

Here we give a brief description of the use of graphs to represent protein

structures and of the techniques used to identify common substructures within
proteins or to search for specific patterns in databases of molecules. Early work
on protein matching based on graph theory was done by Brint and Willett
[7]. To compare two or more structures, the method generates a graph of
correspondences where each node represents a pair of atoms, one for each
protein, and an edge connects two nodes (a, b) and (a , b ) if the difference
of the distances d(a, a ) and d(b, b ) is below a certain tolerance. Problem 3
defined in section 3, that is finding a maximum-cardinality set of element pairs
such that the distance between each pair is at most a given value δ, can then
be restated as one of determining the maximal clique of a graph. Finding a
clique in a graph is a well known NP-complete problem.
In [57] graph isomorphism is used to search for pharmacophoric 3D pat-
terns in a database. The method of representation of a 3D structure is the
connection table, which contains the list of all atoms of the structure together
with the bond information that describes the way in which the atoms are
linked together. The connection table is basically a graph where the nodes
represent the individual atoms and the edges represent interatomic distances.
The graph is complete in the sense that there is an edge connecting every pair
of nodes. A variant of this graph is obtained by using angular information to
label the edges. The angle can be either the torsion angle or the valence angle
defined as the angle between three bonded atoms. The presence of a query
substructure in a database can be tested by means of a subgraph isomorphism.
Subgraph isormorphism is computationally expensive and therefore cannot be
used to test all entries in the database. Thus a screening strategy is suggested
to reduce the overall execution time by eliminating most of the entries of
the database from consideration by subgraph isomorphism. This screening is
analogous of an index technique in that it provides access to a small fraction
of the database by a preprocessing operation that groups all elements with
similar characteristics. An extension to flexible matching is provided as well.
The approach by Escalier at al. [12] finds the largest similar subsets of
atoms by recursively building subsets of increasing sizes, combining two sub-
sets of size k to build a subset of size k + 1. Two subsets can be combined
to form a larger one if they differ in one element only and their inter-atomic
distances are all below a given threshold. Thus, this problem is equivalent to
a clique finding problem. A suitable tree data structure allows an efficient im-
plementation of the merge operation. As stated by the authors, the approach
is suitable for small (less than 30 atoms) molecules. For larger structures such
as proteins, a brute force application of the algorithm may lead to unreason-
 Geometric Methods for Protein Structure Comparison 77

able execution times. Heuristics have to be introduced to reduce the amount

of computation. One such heuristic is to split the problem in two parts: the
first is to identify local fragments and the second is to assemble them together.
A graph-based approach was used in [20], [41] to compare secondary struc-
ture motifs in proteins. Proteins and motifs are represented as labelled graphs
with the nodes corresponding to the segments associated s to secondary struc-
tures and the arcs to the angular and spatial relations between segments. Sub-
graph isomorphism is used to identify common structural patterns in pairs of
proteins or to search for motifs in the PDB. The Ullman’s algorithm for sub-
graph isosmorphism [54] was found to be sufficiently fast for the search of
small motifs from he PDB. Chains of lines segments, corresponding to ei-
ther secondary structures or to linear representation of other fragments of the
backbone have also been considered [1].

9 Integration of Methods for Protein Comparison Using

Different Representations
Different representations offer richer source of information that can be used
in the comparison. This approach has been already investigated in the litera-
ture resulting in effective tools. Alignment of superfamily members has been
obtained through conservation of structural features such as solvent accessi-
bility, hydrogen bonding and the presence of secondary structures [42], [45],
[50], [53].
An hierarchical protein structure superposition using both secondary
structures and backbone atoms was recently proposed by Singh and Brutlag
[51]. The local alignment of secondary structures is obtained by a variation of
the Smith-Waterman dynamic programming algorithm [52]. A score function
is used in the dynamic programming to measure the degree of similarity be-
tween pairs of vectors (linear segments) and is an attribute that may be either
orientation independent (like the angle between two vectors within the same
protein) or orientation dependent (like the angle between two vectors corre-
sponding to two structures each belonging to one protein of the pair). Other
attributes may relate distances between segments within the same or differ-
ent proteins. The expression for the score function S used in this approach,
similar to that used by Gerstein and Levitt [18], is given by:
S= 2M
1+[d/d0 ]2
−M
where M is a weighting factor for the attribute being measured, d the attribute
value and d0 is the value at which the score should be 0. An important choice
of a dynamic programming approach is how to assign gap penalties. For sec-
ondary structure alignment it may be appropriate to decide to introduce no
penalty because often the deletion of a secondary structure is due to an incor-
rect assignment in the PDB or to a mutation that changed a single secondary
78 Carlo Ferrari and Concettina Guerra

structure element, say a strand, into two structures or converted a strand into
a turn.
Once an initial superposition of the secondary structures has been obtained
by the dynamic programming algorithm, it is refined by iteratively minimizing
the RMSD between pairs of nearest atoms from the two proteins. An inter-
esting feature of the approach is that it does not simply rely on RMSD for
judging the quality of the alignment but it takes into consideration also the
number of “well” aligned atoms. Well aligned atoms define the “core” of the
proteins and are selected as follows. Pairs of atoms, one atom from each pro-
tein, are selected so that each atom of the pair is the nearest atom of the other
atom of the pair in the other protein. Furthermore, to be included in the core,
such pairs of atoms have to satisfy the co-linearity property i.e. if (i, j) and
(h, k) are two such pairs and i < h then it must be j < k. Thus this method
is order-dependent, according to one major classification of protein compar-
ison approaches. The last step of the algorithm is to try to improve on the
superposition of the core atoms even at the cost of degrading the alignment
of the rest of the atoms.
The algorithm is efficient in terms of computational complexity and spends
most of the execution time on the secondary and atomic alignment and a small
fraction on the alignment of the core structures.

10 Conclusions
The problem of protein comparison can be successfully approached by first
considering the related geometric issues. In the paper, the power and limita-
tions of the different algorithms for protein structure comparison have been
reviewed and discussed. Most of them have already proved their utility in
computer vision and image processing, as well as in robotics, astronomy and
physics. Their use in molecular biology and bioinformatics opens new perspec-
tives for developing integrated methods for protein comparison, classification
and engineering. Even if the different methods have been introduced to be
used within different applications (characterized by different requirements),
they solve particular instances of a more general matching problem, as deeply
investigated in the area of computational geometry. The variety of protein
representations supports the reasoning both at the level of points (i.e. atomic
level) and at level of segments (or secondary structures). Estimation of rigid
transformations with different metrics is an important technique within the
protein structure comparison algorithms. Moreover geometric indexing tech-
niques prove their effectiveness in searching large protein databases. Finally,
graph-theoretic protein modeling helps in designing algorithms for substruc-
ture identification and comparison.
From the current research it has been recognized that the combination
of different methods and different protein representations may result in new
and effective algorithms with decreased computational complexity and better
 Geometric Methods for Protein Structure Comparison 79

speed. Solvent accessibility, hydrogen bonding and the presence of secondary

structures can be considered together in the alignment of superfamilies, while
a hierarchical protein structure superposition can be obtained using both sec-
ondary structures and backbones atoms. For a better characterization of the
proteins functions and their evolutionary information, geometric reasoning
should be coupled with some proper chemical consideration involving hy-
drophobicity, charge, etc. The goal is to use domain-specific information for
allowing a better pruning of the possible association choices at a very early
stage of the matching process.

11 Acknowledgements
Support for Guerra was provided in part by the Italian Ministry of Univer-
sity and Research under the National Project “Bioinformatics and Genomics
Research”, and by the Research Program of the University of Padova.
Support for Ferrari was provided in part by the Italian Ministry of Univer-
sity and Research and by the Research Program of the University of Padova.

References
1. Abagyan, R.A. , Maiorov, V.N. (1989). An Automatic Search for Similar Spa-
tial Arrangements of α helices and β-strands in globular proteins. Journal of
Molecular Structural Dynamics, 6, 5, 1045-1060.
2. Alesker, A., Nussinov, R., Wolfson H.J. (1996). Detection of non-topological
motifs in protein structures. Protein Engineering, 9, 5, 1103-1119.
3. Alter, T. D. (1992). Robust and efficient 3D recognition by alignment. Technical
Report AITR-1410, Massachusetts Institute of Technology, Artificial Intelligence
Laboratory.
4. Arun, K.S., Huang,T.S., Blostein, S.D. (1987). Least-square fitting of two 3-D
point sets. IEEE Trans. on Pattern Analysis and Machine Intelligence, 9, 5,
698-700.
5. Branden, C., Tooze, J. (1999) Introduction to Protein Structure, (second edi-
tion), Garland.
6. Brown, N.P., Orengo C.A., Taylor W.R. (1996). A protein structure comparison
methodology. Comp. Chem. , 27, 359-380.
7. Brint, A.T., Willett, P. (1987). Algorithms for the identification of three-
dimensional maximal common substructures. J. Chem. Inform. Comput. Sci.,
27, 152-156.
8. Califano, A., Mohan, R. (1992). Multidimensional indexing for recognizing visual
shapes. IEEE Trans. on Pattern Analysis and machine Intelligence, 16, 4, 373-
392.
9. Chen, H. H., and Huang, T. S. (1990) Matching 3d line segments with appli-
cation to multiple-objects motion estimation. IEEE Transactions on Pattern
Analysis and Machine Intelligence, 12, 1002-1008.
80 Carlo Ferrari and Concettina Guerra

10. Chew, P.L., Dor D.,pEfrat A., Kedem, K. (1995). Geometric pattern matching
in d-dimensional space, Algorithms-ESA ’95, 264-279.
11. Chew, L.P., Goodrich M.T., Huttenlocher,D.P., Kedem,K., Kleinberg, J.M.,
Kravets, D. (1997). Geometric pattern matching under Euclidean motion, Com-
putational Geometry. Theory and Applications, 7, 113-124.
12. Escalier V., Pothier, J., Soldano, H., Viari, A. (1998). Pairwise and Multi-
ple Identification of three-dimensional common substructures in proteins. J.
of Computational Biology, 5, 41-56,
13. Fischer, D., Bachar, O., Nussinov, R., and Wolfson, H. (1992). An efficient
automated computer vision based technique for detection of three dimensional
structural motifs in proteins. J. Biomol. Struct. Dyn., 9, 769-789.
14. Fischer, D., Tsai, C.J., Nussinov, R., and Wolfson, H. (1995). A 3D sequence-
independent representation of the protein data bank. Protein Engineering, 8,
981-997.
15. Goodrich, M.T., Mitchell, J.S.B., Orletsky, M.W. (1999) Practical Methods for
Approximate Geometric Pattern Matching Under Rigid Motion. IEEE Trans.
Pattern Analysis and Machine Intelligence, 21, 4, 371-379.
16. A. Glassner, ( editor), Graphics Gems, Academic Press, 1990.
17. Gerstein, M. (1992). A Resolution-Sensitive Procedure for Comparing Protein
Surfaces and its Application to the Comparison of Antigen-Combining Sites.
Acta Cryst., A8, 271-276.
18. Gerstein, M., Levitt, M. (1998). Comprehensive Assessment of automatic struc-
tural alignment against a manual standard, the scop classification of proteins.
Protein Science, 7, 445-456.
19. Golub, G.H., Van Loan C.F. , (1996) Matrix Computation, Johns Hopkins Uni-
versity Press.
20. Grindley, H.M., Artymiuk, P.J., Rice, D.W., and Willett, P. (1993). Identifi-
cation of tertiary structure resemblance in proteins using a maximal common
subgraph isomorphism algorithm. J. Mol. Biol., 229, 707-721.
21. Guerra, C., Lonardi, S., Zanotti, G., (2002). Analysis of proteins secondary
structures using indexing techniques. IEEE Proc First Int. Symposium on 3D
Data Processing Visualization and Transmission, 812-821.
22. C. Guerra, V. Pascucci. (1999) On matching sets of 3D segments”, Proceedings
of SPIE Vision Geometry VIII, 157-167.
23. Guerra, C., Pascucci, V. (1999) 3D segment matching using the Hausdorff dis-
tance. Proceedings of the IEEE Conference on Image Processing and its Appli-
cations, IPA99, 18-22.
24. Hagedoorn, M., Veltkamp, R. C. (1999). Reliable and efficient pattern matching
using affine invariant metric. Int. J. of Computer Vision, 31, (2/3), 203-225.
25. B. Horn, H. Hilden, S. Negahdaripour (1988). Closed-form solution of absolute
orientation using orthonormal matrices. J. Opt. Soc. Am., 5, 1127-1135.
26. Holm, L., Sander, C. (1995). 3D-Lookup: Fast protein structure database
searches at 90% reliability. Proc. Third Int. Conf. on Intell. Sys. for Mol. Biol.,
Menlo Park, 179-187.
27. Holm, L., Sander, C. (1996). Mapping the protein universe. Science, Science,
273, 595-602.
28. B. Horn, (1987). Closed-form solution of absolute orientation using unit quater-
nions. J. Opt. Soc. Am., 4, 629-642.
 Geometric Methods for Protein Structure Comparison 81

29. Huttenlocher, D. P., Klanderman, G. A., and Rucklidge, W. J. (1993) Compar-

ing images using the Hausdorff distance. IEEE Transactions on Pattern Analysis
and Machine Intelligence, 15, 9, 850-863.
30. Kabsch, W., Sander, C., (1983). Dictionary of protein secondary structure: pat-
tern recognition of hydrogen-bonded and geometrical features. Biopolymers, 22,
2577-2637.
31. Kanmgar-Parsi, B., and Kamgamr-Parsi, B. (1997) Matching sets of 3d line
segments with application to polygonal arc matching. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 19, 10, 1090-1099.
32. Y. Lamdan, J. T. Schwartz, H. J. Wolfson. (1990). Affine invariant model-based
object recognition, IEEE Trans. on Robotics and Automation, 578-589.
33. Lancia, G., Carr, R., Walenz, B., Istrail, S. (2001). Optimal PDB structure
alignments: A Branch-and-Cut algorithm for the maximum contact map overlap
problem. Proc. 5th ACM REsearch in COMputational Biology, 193-202.
34. Laskowski R.A., MacArthur M.W., Moss D.S., Thornton J.M. (1992). Stereo-
chemical duality of protein structure coordinates. Proteins, 12, 345-364.
35. Laskowski R.A., MacArthur M.W., Moss D.S., Thornton J.M. (1993).
PROCHECK: a program to check the stereochemical quality of protein struc-
tures. J. Appl. Cryst. , 26, 283-291.
36. Leibowitz, N., Fligelman, Z.Y., Nussinov, R., Wolfson, H.J. (1999). Multiple
structural alignment and core detection for geometric hashing. Proc. ISMB99,
Heidelberg, Germany, 169-177.
37. Lemmen, C., Lengauer, T. (2000). Computational methods for the structural
alignment of molecules. J. of Computer-Aided Molecular Design, 14, 215-232.
38. Lesk, A. M. (1991).Protein architecture: a practical approach. Oxford Univ.
Press, Oxford.
39. Lesk, A. (1994). Computational Molecular Biology. Encyclopedia of Computer
Science and Technology, 31, Marcel Dekker, Inc..
40. Eggert,D.W., Lorusso, A., Fisher, R.B. (1997). Estimating 3-D rigid body trans-
formations: a comparison of four major algorithms, Mach. Vis. and Applic., 9,
272-290.
41. Mitchell, E.M., Artymiuk, P.J., Rice, D.W. Willett, P. (1989) Use of Techniques
derived from graph theory to compare secondary structures motifs in proteins.
J. Molecular Biology, 212, 151-166.
42. Mizuguchi, K., Deane, C.M., Blundell, T.L., Johnson, M.S. and Overington,J.P.
(1998). JOY: protein sequence-structure representation and analysis. Bioinfor-
matics, 14, 617-623.
43. Murray, R.M, Li, X., Sastry, S.S., (1994). A Mathematical Introduction to
Robotic Manipulation, CRC Press.
44. Murzin, A.G., Brenner, S.E., Hubbard, T., Chothia, C. (1995) SCOP: a struc-
tural classification of proteins database for the investigation of sequences and
structures. J. Molecular Biology, 247, 536-540.
45. J. P. Overington, Z. Y. Zhu, A. Sali, M. S. Johnson, R. Sowdhamini, G. V.
Louie, and T. L. Blundell. (1993). Molecular recognition in protein families: a
database of aligned three-dimensional structures of related proteins. Biochem.
Soc. Trans., 21 , 3, 597-604.
46. Pennec, X., Ayache, N. (1998) A geometric algorithm to find small but highly
similar 3D substructures in proteins. Bionformatics, 14, 6, 516-522.
47. Platt, D.E., Guerra, C., Rigoutos, I., Zanotti, G. (2002). Global secondary struc-
ture packing angle bias in proteins. Manuscript.
82 Carlo Ferrari and Concettina Guerra

48. Rucklidge, W. J. (1997) Efficiently locating objects using the Hausdorff distance.
International Journal of Computer Vision, 24, 3, 251-270.
49. Sabata, B., Aggarwal, J.K. (1991). Estimatiom of motion from a pair of range
images: a review. Computer Vision, Graphics and Image Processing: Image Un-
derstanding, 54, 3, 309-324.
50. Sali A., Blundell, T.L. (1990). Definition of a General Topological Equivalence
in protein structures: a procedure involving comparisons of properties and re-
lationships through simulated annealing and dynamic programming, Journal of
Molecular Biology, 212, 403-428.
51. Singh,A.P., Brutlag, D.L., (1997). Hierarchical protein structure superposition
using both secondary structures and atomic representations. Proc. Fifth Int.
Conf. on Intell. Sys. for Mol. Biol., Menlo Park, 284-293.
52. Smith, T.F., Waterman, M.S., (1981). Identification of common molecular sub-
sequences. J. Mol. Biol., 147, 195-197.
53. Sowdhamini R., Burke D.F, Huang J-F, Mizuguchi, K. Nagarajaram H.A., Srini-
vasan N., Steward RE. and Blundell T.L. (1998). CAMPASS: A database of
structurally aligned protein superfamilies. Structure, 6, 9, 1087-1094.
54. Ullman, J.R. (1995) J. Assoc. Comp., 23, 31-42.
55. Verbitsky, G. Nussinov, R., Wolfson H.J. (1998). Structural comparisons allow-
ing hinge bendings, swiveling motions. Proteins, 34, 232-254.
56. Walker M.W, Shao L., Voltz R.A., Estimating 3-D Location Parameters Using
Dual Number Quaternions, CVGIP:Image Understanding, 54, 3, 1991, 358-367
57. Willett, P. (1995) Searching fore pharmacophoric patterns in databases of three-
dimensional chemical structures. J. of Molecular Recognition, 8, 290-303.
Identifying Flat Regions and Slabs
in Protein Structures

Mary Ellen Bock1 and Concettina Guerra2

1
Department of Statistics, Purdue University, West Lafayette, 47907 IN, USA
mbock@stat.purdue.edu
2
Dipartimento Elettronica e Informatica, Via Gradenigo 6/A, Padova, Italy or
Department of Computer Science, Purdue University,
West Lafayette, 47907 IN, USA
guerra@dei.unipd.it

Summary. Motivated by the problem of identifying flat regions in three dimen-

sional protein structures, we provide a new geometric approach for the extraction
of planar surfaces that compares favorably with existing approaches developed for
computer vision. Preliminary results obtained on proteins from different structural
classes are given.

Index Terms - Plane detection - Hough transform - Width - Protein Structure

analysis

1 Introduction
When we examine the 3D tangle of a protein string we sometimes observe
flat or planar regions. Certain large segments or groups of amino acids lie
roughtly in a plane or very narrow slab. The presence of these segments or
groups of amminoacids in a very narrow slab can be very useful for confirming
or suggesting potential secondary structures such as a β-strand. The episodic
occurence of this phenomenon motivated an interest in general algorithms
that quickly find flat regions in 3D point sets.
We formulate the problem of identifying flat regions as follows. Given a
set Γ of n points in 3D space and a non-negative constant , determine the
plane that is at a distance at most  from the maximal number of points of
Γ.
In other words, we want to determine a pair of parallel planes (slab) a distance
of 2 apart enclosing a maximal number of points of Γ . The pair of planes is
C. Guerra, S. Istrail (Eds.): Protein Structure Analysis and Design, LNBI 2666, pp. 83-97, 2003.
 Springer-Verlag Berlin Heidelberg 2003
84 Mary Ellen Bock and Concettina Guerra

not necessarily unique. For the extraction of multiple planar regions from the
input data, the same problem is solved repeatedly after the removal of the
points that are found to belong to the best slab.
We present a new method for the detection of planar regions in point
sets; we also give some heuristics that can help speed-up the process while
maintaining a good quality for the solution. We show that our geometric
method has many advantages over the Hough transform, a popular technique
to extract parametric shapes, such as lines, planes, and circles from images.
Our detection algorithm solves the problem as a maximum coverage prob-
lem in the parameter space. A subset of three points of Γ determines a region
in the parameter space containing the parameters of all planes within  dis-
tance from the three points. Consider the arrangement in the parameter space
formed by the regions determined by all subsets of three points of Γ . Define
the depth of a point in the parameter space as the number of regions that
contain the point. This is in fact the number of points of Γ within  distance
from the corresponding plane. The point of maximum depth is sought among
the vertices of the arrangement. Each such vertex gives the parameters of a
plane at distance either + or − from each of the three points.
The search for a vertex of maximum depth can be done by simply counting
for a plane corresponding to a vertex of the arrangement the number of points
of Γ within  distance from it and then choosing the plane that has the largest
count. This computation requires O(n) time for each plane. A more efficient
solution to this problem is also presented that avoids the above computation
for all planes, by exploiting the structure of the arrangement of the regions in
the parameter space. The algorithm has a O(n3 log n) time complexity.
The problem of detecting geometric primitives is a fundamental one in com-
puter vision; there it is often solved by the Hough transform, a method that
transforms the extraction problem into an intersection problem in the param-
eter space, that is is quantized in an accumulator array.
The most common application of the Hough transform in image processing
is for line detection. The method has been also applied to circle and ellipse
detection, and to the extraction of planes from range images. If storage space
is a concern this approach can be used only for parametric shapes with few pa-
rameters, since the accumulator array dimensionality grows with the number
of the parameters.
The plane extraction problem is related to a well-known problem in com-
putational geometry, the width problem, that is the determination of a pair of
parallel planes of minimum distance enclosing all points of a given set. Houle
and Toussain [9] give algorithms that produce the exact width of a point set
and run in O(n log n)time in
2 and in O(n2 ) in
3 . An efficient approxima-
tion to the width determination is given in [4], that can be used to perform the
metrology primitive of “flatness”. The requirement that all points are within
a single slab formed by the two parallel planes makes these approaches not
 Identifying Flat Regions and Slabs in Protein Structures 85

suitable for applications that require the extraction of multiple slabs with a
predefined width.

2 A Geometric Algorithm

Given a set Γ of n points in 3D-space and a non-negative real value , we

want to determine a plane P in
3 that is at a distance at most  from the
largest number of points of Γ .
Let the equation of plane P in
3 be
ax + by + cz + d = 0
where (a,b,c,d) is the 4-dimensional vector of parameters of P satisfying
a2 + b2 + c2 = 1.
We solve the 3D plane detection problem by solving an intersection prob-
lem in the
4 parameter space. To each three-dimensional point pi ∈ Γ we
associate a region Ri ∈
4 (as shown below). Each of the four-dimensional
points in Ri corresponds to the parameters of a plane in
3 at a distance at
most  from pi . If (a, b, c, d) is a four-dimensional point in the intersection of
the two regions Ri , Rj associated to the three-dimensional points pi , pj , then
the corresponding plane in
3 given by the equation ax + by + cz + d = 0 is
at a distance at most  from both points pi and pj in
3 . We define the depth
of a point in
4 to be the number of regions Ri that contain it and look for
a point in
4 with maximal depth. (The point is not unique.) Another way
of looking at this problem is to view it as a coloring problem. Each region is
colored by a different color; we say that a point is colored by a given color if
it belongs to a region with that color. The objective is to find a point that is
covered by the maximum number of different colors.
The region Ri in
4 associated to point pi = (xi , yi , zi ) in
3 can be
obtained as follows. First if a plane P
in
3 has parameters (a
, b
, c
, d
) and
if the plane contains pi then it satisfies:
a
xi + b
yi + c
zi + d
= 0.
The above equation also describes in
4 a hyperplane Qi 0 through the
origin with coefficient (xi , yi , zi , 1, 0). Let dpi ,P be the distance in
3 of point
pi from the plane P with parameters (a,b,c,d). That is,

dpi ,P = |axi + byi + czi + d| (1)

since a2 + b2 + c2 = 1.
The region Ri in
4 consists of points (a, b, c, d) satisfying:


axi + byi + czi + d +  ≥ 0
axi + byi + czi + d −  ≤ 0 (2)

 2
a + b 2 + c2 = 1
86 Mary Ellen Bock and Concettina Guerra

Ri has boundaries given by the two hyperplanes Qi +
and Qi −
with

respective coefficients (xi , yi , zi , 1, +) and (xi , yi , zi , 1, −) and the hyper-
cylinder
a 2 + b 2 + c2 = 1 (3)
The points (a, b, c, d) in Qi +
satisfy:
axi + byi + czi + d +  = 0
The points (a, b, c, d) in Qi −
satisfy:
axi + byi + czi + d −  = 0
Note that the two above hyperplanes Qi +
and Qi −
are parallel to the hy-
perplane Qi 0 and enclose it. Furthermore, both Qi +
and Qi −
have distance
from Qi 0 given by: √ 1
2 2 .
xi +yi +zi +1
We consider the arrangement of all regions Ri in
4 . It is easy to see that
region Ri cannot be contained in Rj , i  = j. This is because there always
exists a plane within  distance of pi that is farther than  from pj . Further-
more, the intersection of any three regions Ri ∩ Rj ∩ Rk , i, j, k = 1, ..n, is
non empty. In fact the intersection contains at least the point corresponding
to the parameters of the plane through the three points. The boundaries of
three regions intersect in at most 8 vertices, each obtained by intersecting
three hyperplanes, one for each region, and the hypercylinder. As far as the
depth information is concerned only the boundaries of the regions are rele-
vant. Thus we look for a point with maximum depth among the vertices of
the arrangement.
Consider the 8 vertices in
4 determined by Ri , Rj , and Rk . They corre-
spond to 8 candidate solution planes in
3 . However the triple of points pi ,
pj , and pk in
3 corresponding to the three regions does not lie in any of the
candidates planes. Each candidate plane can be described as the middle plane
of a pair of parallel planes a distance of 2  apart that encloses the maximum
number of points of Γ . This middle plane is parallel to the two members of
the pair and halfway between them. In the first two of the 8 cases, the triple of
points lies on one of the members of the pair, i.e. the first candidate optimal
plane is at distance  and the second at distance − from each of the three
points pi , pj , and pk . In the last 6 cases, two points of the original triple lie
on one member of the pair of enclosing planes while the third point lies on
the other member of the pair. In other words, by choosing an unordered pair
of points from the triple (and there are three possible choices), two planes are
determined that are between the pair of points and the remaining member of
the triple and that are at distance  (in absolute value) from each of the three
points.
The problem of finding a vertex of maximum depth has a simple solution
in the original space. For each vertex count the number of points in
3 within
 distance from the corresponding plane. Then select the vertex with maxi-
mum count. Since there are O(n3 ) vertices in the arrangement this procedure
 Identifying Flat Regions and Slabs in Protein Structures 87

takes O(n4 ) time. In the next section we present a more efficient algorithm
that exploits the structure of the arrangement of the regions and achieves
O(n3 log n) time complexity. We have implemented the algorithm also using
random sampling, that is the selection of a small number of triplets of points
from the input data set. The results are satisfactory in most cases; we do not
however discuss them here for lack of space.
For the template matching cost function, methods which generate only
the first two cases above for the candidate planes, i.e. the triple of points
always lies in one of the two planes which form the enclosing pair, can miss
25% of the points as seen in example 2.1. Again for the template matching
cost function, methods which always place each triplet of points in the middle
plane (and form the potentially optimal pair of planes by creating parallel
planes on either side of the middle plane a distance of  away) can miss (25%
) of the points as seen in example 2.2.
Example 2.1. Consider the following four points: P1 (0,.75,0), P2 (0, -
.75,0),P3 (1,0, .75), and P4 (1, 0, -.75). Let  = 0.5. The pair of parallel planes,
one with coefficients (1, 0, 0, −1) containing P3 and P4 and the second parallel
to it and containing P1 , encloses all four points (in fact P2 belongs to the
second plane). The solution generated by our method is the middle plane of
this pair with coefficient (1, 0, 0, −.5) and it is within distance 0.5 from all
four points.
On the other hand, restricting one of the two planes which form the pair
to always contain three of the four points misses one point. The middle plane
of such a pair has always distance greater or equal to 0.7 from the remaining
point. See for example, the plane (0.6, 0, −0.8, −0.5) which is middle plane of
a pair with one member containing P1 , P2 and P3 . Thus the maximal number
of points within distance 0.5 from such planes is 3.
Example 2.2. Consider the same four points as in the previous example:
P1 (0,.75,0), P2 (0, -.75,0),P3 (1,0, .75), and P4 (1, 0, -.75). Let  = 0.5. The
planes containing three of the four points always have the distance 1.2 from
the remaining point. Again the maximal number of points within distance 0.5
from such planes is 3.
An interesting feature of the algorithm is that its time complexity is in-
dependent of . Therefore it is the same even when high precision is required
(very small values of ). This is unlike the methods based on the Hough trans-
form for which the complexity in time and memory increases with the degree
of required precision.

3 An Improved Geometric Algorithm

In this section we present an efficient algorithm for finding a vertex of maxi-

mum depth in the arrangement of the Ri regions. We recall that a vertex is
determined by the intersection of three hyperplanes at the boundary of three
distinct regions Ri and the hypercylinder.
88 Mary Ellen Bock and Concettina Guerra

We first observe that the vertices of the arrangement lie on the O(n2 )
ellipses eij formed by the intersection of two hyperplanes Qi +
and Qj +
,
i, j = 1, ..., n, and the hypercylinder. More precisely, of the 8 vertices generated
by Ri , Rj and Rk 4 belong to the ellipse eij and 4 to the ellipse eik . (or to
the ellipse ejk ) On eij (eik ) one pair of vertices is obtained by intersecting
eij with Qk +
and the second pair with Qk −
(Qj +
and Qj −
). Each pair of
vertices on eij defines an arc consisting of points in Ri ∩ Rj ∩ Rk . Consider
all arcs on eij obtained for all different values of k. The depth of a vertex is
clearly given by 2 plus the number of arcs it belongs to.
Thus the problem of finding the vertex of maximum depth becomes that of
computing the maximum overlap in a set of arcs. This problem can be solved
using standard geometric techniques in O(k log k) time for k arcs [5]. Since
this step is repeated for each ellipse corresponding to a pair of regions Ri Rj ,
the algorithm has an overall O(n3 log n) time complexity.
The equation of the ellipse eij can be obtained as follows. Let
axi + byi + czi + d +  = 0
axj + byj + czj + d +  = 0
a 2 + b 2 + c2 = 1
be the equations of the hyperplanes Qi +
, Qj +
and the hypercylinder, re-
spectively.
The intersection of the two hyperplanes in
4 can be represented by:
a = sua + tva
b = sub + tvb
c = sua + tvc
d = − + sud + tvd
where s and t are parameters and u = (ua , ub , uc , ud ) and v = (va , vb , vc , vd )
are orthogonal unit vectors. We can choose the vector v so that vd = 0. The
intersection of the above plane with the hypercylinder a2 + b2 + c2 = 1 is given
by:
(u2a + u2b + u2c )s2 + (ua va + ub vb + uc vc )s · t + (va2 + vb2 + vc2 )t2 = 1
Since vd = 0 and u ⊥ v the equation above is rewritten as:
αs2 + t2 = 1
where α = 1 − u2d ≤ 1. The above is the equation of the ellipse on the plane.

4 Hough Transform

The Hough transform is a powerful technique developed in the area of image

processing for the detection of parametric curves in an image. It was originally
proposed for line detection from gray-level images [8] and later extended to
other parametric curves (circles, ellipses, etc.) [3] and to the recognition of
 Identifying Flat Regions and Slabs in Protein Structures 89

arbitrary shapes. There exist several variants of the basic Hough transform;
for recent surveys on the subject see [13], [16].
The standard formulation of the Hough transform for line detection is as
follows. Suppose we are given a set of image points - for instance edge points
or some other local feature points - and we want to determine subsets of
them lying on straight lines. The Hough transform associates an image point
with a line in the parameter space. The method is based on the property that
collinear points are transformed into lines intersecting at the same point in the
parameter space. The problem is then solved as an intersection problem in the
parameter space, that is quantized in an array of cells, called the accumulator
array. Specifically, the Hough transform consists of the following steps.
step 1. For each image point, determine the parameters of all lines through
it and accumulate the corresponding cells in the parameter space.
steo 2. Find the maximum (or the local maxima) in the accumulator array.
The corresponding parameters describe the detected line (lines).
An important property of the Hough transform is its insensitivity to noise
and to missing parts of lines. The disadvantages are its high computational
and memory requirements which increase with the resolution of the accumula-
tor array. Furthermore, since the array dimensionality grows with the number
of the curve parameters, this approach is applicable only to curves with few
parameters. To overcome these limitations, many approaches have been pro-
posed for image and vision vision applications, including probabilistic [11],
randomized [12] and fast [15] Hough approaches.
Some of these approaches formulate the Hough transform as a many-to-
one mapping from the original set of points into the parameter space, instead
of a one-to-many mapping as in the standard formulation. For example, in
the case of line detection that means determining for any two points of the
set the line passing through them and incrementing the corresponding entry
of the array in the parameter space. In general, for the detection of a curve
represented by m parameter equations, the original approach maps one point
into a m−1 hypersurface into the parameter space, while the second approach
maps m points which define a curve into one point of the parameter space.
The second approach is generally used in conjunction with random sampling
[6], that is the random selection of few subsets of n points from the original
input to reduce storage and computation time.
The Hough transform is easily extended to the detection of planes in a set
Γ of points in 3D space. A plane is parametrized by the following expression,
containing three parameters α, β and d:

cos α cos βx + cos α sin βy + sin αz + d = 0 (4)

90 Mary Ellen Bock and Concettina Guerra

Note that if we let

a = cos α cos β (5)

b = cos α sin β (6)
c = sin α

we have a2 + b2 + c2 = 1, that is the same constraint used in the previous

section.
The parameter space is quantized into h × k × t cells that form the ac-
cumulator array A; each entry of the array corresponds to a triple (α, β, d).
In other words the parameters α and β are discretized into h and k values,
respectively, taken at regular intervals in the range [0, π]. The parameter d,
which represents the distance of the plane from the origin, is discretized in t
values; the range of t values is chosen according to the point set.
The plane detection algorithm proceeds in two steps. Step 1 constructs
the accumulator array. It does so by computing for each point of the set Γ
and for each pair (α, β) the d value from equation 4 and incrementing the
corresponding entry in the accumulator array.
Step 2 of the algorithm determines the maximum in the accumulator array
A, that corresponds to the detected plane.
For a set with n points, the Hough transform requires O(n × h × k) time
for step 1 and O(h × k × t) for step 2.
Alternatively, the Hough transform for plane detection can be formulated
as a many-to-one mapping from the set of points into the parameter space.
For each triple of input points the plane through them is determined and only
the corresponding entry in the accumulator array is incremented. For this
alternative approach, the time for the construction of the accumulator array
is O(n3 ), while the time for searching the array to determine its maximum
value is still dependent on the quantization of the array, that is O(h × k × t).
As pointed out in the section 2, this method may fail to detect the best plane,
defined as the plane within a given distance from the maximal number of
points.

5 Performances of the Two Algorithms

The time and space complexity of the Hough transform depends on the num-
ber of points as well as on the quantization of the parameter space. For a
coarse quantization of the accumulator array the algorithm is very efficient.
There are however many applications requiring high resolution and therefore
a fine discretization of the parameter space. The time complexity of the geo-
metric algorithm is instead dependent only on the number n of input points
and independent of .
We have compared the two approaches on several input sets of varying sizes
and for different values of . To compare the two approaches in terms of time,
 Identifying Flat Regions and Slabs in Protein Structures 91

we have first experimentally determined the quantization of the parameter

space in the Hough approach that allows us to obtain results of the same
quality as the geometric algorithm. The quality of the result is expressed
by the number of points detected within  distance from the best plane. In
all experiments we observed that, with a proper choice of the quantization
of the accumulator array, the standard Hough transform obtained the same
maximum value as the geometric algorithm.
In Tables 1 and 2 the execution times of the two algorithms for small
sets of points are reported. The sets of points consist of fragments of varying
sizes of the backbone of the proteins 1KDU and 3PGM. For each of the two
 values, the required quantization step in the accumulator array for each of
the three parameters αstep , βstep and dstep is shown along with the overall
execution time THough . The execution time Tgeom of the geometric algorithm,
which is the same for all  values, is shown in column 2.
In all our tests the geometrical algorithm outperformed the Hough trans-
form. The time for the Hough transform is dominated by the complexity of
the second phase, i.e. finding the maximum in the accumulator array. As the
size of the set increases, the memory requirements of the Hough transform
increase significantly, and so does its time complexity. This is also because in
our examples the space occupancy of input points (in terms of the diameter)
increases with n requiring larger values for the parameter d.
There can be cases (depending on  ) where the Hough transform runs
faster. However, given a set of points it seems that is always possible to find a
value of  for which the geometrical algorithm outperforms the Hough trans-
form.

1KDU
=1
=2

Num. of points Tgeom THough αstep βstep dstep THough αstep βstep dstep
20 0.1 17.7 2 2 0.2 3.9 3.5 3.5 0.7
25 0.2 17.7 2 2 0.2 3.5 3.5 3.5 0.7
30 0.5 17.7 2 2 0.2 5.7 3.5 3.5 0.7
35 0.9 21.0 2 2 0.2 4.3 3.5 3.5 0.7
40 1.5 36.0 1.8 1.8 0.2 10.3 3 3 0.7
45 3.1 >3000 1.2 1.2 0.2 21.5 2.5 2.5 0.7

Table 1. Execution times of the two algorithms for fragments of the protein 1KDU.

3PGM
=1
=2

Num. of points Tgeom THough αstep βstep dstep THough αstep βstep dstep
20 0.1 0.3 9 9 0.5 0.4 9.9 9.9 0.6
30 0.6 1.4 7.2 7.2 0.3 0.5 9 9 0.5
40 1.8 10.6 3 3 0.2 2.8 5.6 5.6 0.3
50 4.4 36.2 1.8 1.8 0.2 27.9 2 2 0.5

Table 2. Execution times of the two algorithms for fragments of the protein 3PGM
92 Mary Ellen Bock and Concettina Guerra

6 Plane Detection in Proteins

One application of the above method is in improvement of protein secondary

structures assignment from atomic coordinates. This is an important task be-
cause it is often a preliminary step in the analysis of the protein structures; for
instance secondary structures can be used in protein structural comparison
to identify largest common substructures or to search for structural motifs
in the Protein data Bank (PDB) [1]. The task of assigning secondary struc-
tures may be based on different criteria which may lead to slightly different
solutions; they include hydrogen bond pattern, backbone torsion angles, and
inter-Ca atoms distances [7], [10], [14]. A recent comparison of three different
procedures for the assignment of secondary structures shows that they agree
in only about 60 % of the sequence sites in several proteins [2].
The algorithm we have presented may be used in identifying wrong as-
signments or, more importantly, in deriving properties that can be useful in
secondary structure prediction.
We have determined the widths of the β-strands and α-helices of a small set
of representative protein structures from the PDB database. The set includes
the following protein structures:
1aei 1eur 1emd 1gga 10vi 1pal 1tie 1acp 1azu 1bgt 1cbn 1cdg 1cll 1dat 1dfn 1fha
1gca 1gdl 1gpv 1hcv 1hpi 1ifb 1igf 1igm 1paz 1pml 1pyk 1rbp 1rpa 1rpr 1rpt 1sbp
1stp 1svc 1tim 1ttf 1utg 2mhr 2aza 2bop 2bpa 2cpl 2ctc 2cyh 2dri 2hla 2hnq 2liv
2paz 2pf2 2plh 2plt 2spt 2taa 3por 8paz
where the four-symbol strings represent the identifiers in the PDB.
Table 4 shows the different width values obtained for any given length of
the β-strands. Only the Ca atoms are included in the analysis. The lengths of
the β-strands considered range from 6 to 23 residues. Table 3 has the same
meaning except that the α-helices are considered. The lengths of the helices
range from 6 to 23 residues. Short helices and strands have not been considered
in this analysis. Repeated secondary structures in different chains of a protein
have been counted only once.
In table 3 there are only two values that are far from the average width
value. One such value (6.9) corresponds to the α-helix in figure 4. The pic-
ture clearly indicates a bad secondary structure assignment. The other value
corresponds to the α-helix in figure 3. On both α-helices the geometric al-
gorithm that finds the maximum number of points within  distance from a
plane was applied. For the α-helix of figure 3 the algorithm with a value of 
equal to the average width of all helices correctly identified the last fragment
of the secondary structure as being outside of the slab of average width. For
the α-helix of figure 4 the algorithm found one “outlier”, i.e. the residue with
a darker color in the figure. Figure 5 shows a β-strand with a width value
above average, and again the geometric algorithm found the extreme residues
outside the best slab of average width.
We believe that this analysis can be of help in the study of the relations
between residue-related parameters (such as the volume, etc.) and the 3D
 Identifying Flat Regions and Slabs in Protein Structures 93

geometry of the secondary structures that can help predict the secondary
structures from the sequence of amino-acids.
We have also applied the algorithm to the detection of slabs of given widths
enclosing the maximum numbe of points to several proteins from different
structural classes. In figures 6 and 7 we show the results for 6 proteins from
the PDB; in all cases we have considered the Ca atoms only.

  


 
       
           
   
      
 
     
 

 

 




   

Fig. 1. Widths of helices. (helix length on horizontal axis)




  
  
 
   
 
   
   
    
       
      
  
     



    
    

   
   
 




   

Fig. 2. Widths of β-strands. (strand length on horizontal axis)

94 Mary Ellen Bock and Concettina Guerra

Fig. 3. Protein 1emd has an helix with width larger than average

Fig. 4. Protein 1adb has an helix with width larger than average
 Identifying Flat Regions and Slabs in Protein Structures 95

Fig. 5. A β-strand of protein 2bpa with a width value larger than average

Fig. 6. (left)protein 2pf2:  = 1 (24 % of points,width= 18); (middle) protein 2bop:

 = 2 (41 % of points, width=16); (right) protein 1cbn:  = 1 ( 36 % of points,
width= 12)

7 Acknowledgements
Support for Guerra was provided in part by the Italian Ministry of Univer-
sity and Research under the National Project “Bioinformatics and Genomics
Research”, and by the Research Program of the University of Padova.
We would like to thank Ivano Tagliapietra for his help with this paper.
96 Mary Ellen Bock and Concettina Guerra

Fig. 7. (left) protein 1acp:  = 1 (23 % of points,width= 18); (middle) protein

4mdh:  = 1 (10 % of points,width= 46); (right) protein 2bpa:  = 2 (17 % of
points,width= 57)

References
1. F.C. Bernstein, T.F. Koetzle, G.J. Williams, E.F. Meyer, M.D. Brice, J.R.
Rodgers, O. Kennard, T. Shimanouchi, M. Tasumi. The Protein data bank: a
computer-based archival file for macromolecular structures. J. Mol. Biol.,vol.
112, pp. 535-542, 1977.
2. N. Colloc’h, C. Etchebest, E. Thoreau, B. Henrissat, J.P. Mornon. Comparison
of three algorithms for the assignment of secondary structures in proteins: The
advantage of a consensus assignment. Protein Eng. , vol. 6, pp. 377-382, 1993.
3. R.O. Duda, P.E. Hart. Use of the Hough transform to detect lines and curves in
pictures. Commun. ACM, , vol. 15, pp.11-15, 1972.
4. C.A. Duncan, M.T. Goodrich, and E.A. Ramos, Efficient Approximation and
Optimization Algorithms for Computational Metrology, 8th ACM-SIAM Symp.
on Discrete Algorithms (SODA), 1997.
5. H. Edelsbrunner. Algorithms in Combinatorial Geometry . Springer-Verlag, Hei-
delberg, Germany, 1987.
6. M. A. Fischler, R.C. Bolles. Random sample consensus: A paradigm for model
fitting with applications to image analysis and automated cartography. Computer
Vision, Graphics and Image Processing, vol. 24, pp. 381-395, n.6, 1981.
7. D. Frishman, P. Argos, Knowledge-based secondary structure assignment, Pro-
teins: structure, function and genetics, vol. 23, pp. 566-579, 1995.
8. P.V. Hough. 1962. Methods and means to recognize complex patterns. U. S.
patent 3.069.654.
9. M.E. Houle, G.T. Toussaint, Computing the width of a set, IEEE Trans. on
Pattern Analysis and Machine Intelligence, PAMI-10, 1988, pp. 761-765.
10. W. Kabsch, C. Sander, Dictionary of protein secondary structure: pattern recog-
nition of hydrogen-bonded and geometrical features, Biopolymers, vol. 22, pp.
2577-2637, 1983.
11. N. Kiryati, Y., Eldar, A.M., Bruckstein, A Probabilistic Hough Transform, Pat-
tern Recognition, vol. 24, pp. 303-316, 1991.
12. P. Kultanen, L. Xu,E. Oja, Randomized Hough Transform (Rht), Int. C. Pattern
Recognition, R-B(90), pp. 631-635, 1990.
13. J. Illingworth, J. Kittler, A survey of the Hough transform. Computer Vision,
Graphics and Image Processing, vol. 44, pp. 87-116, 1988
 Identifying Flat Regions and Slabs in Protein Structures 97

14. M. Levitt, J. Greer. Automatic identification of secondary structure in globular

protein. J. Mol. Biol. , vol. 114, pp. 181-239, 1977.
15. H. Li, M.A. Lavin, R.J. Le Master, Fast Hough transform. Computer Vision,
Graphics and Image Processing, vol 36, pp. 139-161, 1986.
16. V. F. Leavers, Which Hough Transform?, Image Understanding, vol. 58, pp.
250-264, 1993.
17. J. Princen, J. Illingworth, J. Kittler, A Hierarchical Approach To Line Extrac-
tion Based On The Hough Transform, Computer Vision, Graphics and Image
Processing:Image Understanding, vol. 52, pp. 57-77, 1990.
18. G. Roth, M. D. Levine, Extracting geometric primitives, Computer Vision,
Graphics and Image Processing:Image Understanding, vol. 58, n. 1, pp. 1-22,
1993
19. L. Xu, E. Oja, Randomized Hough Transform (Rht): Basic Mechanisms, Al-
gorithms, And Computational Complexities, Image Understanding vol. 57, pp.
131-154, 1993.
OPTIMA: A New Score Function for the
Detection of Remote Homologs

Maricel Kann and Richard A. Goldstein

Department of Chemistry, University of Michigan

Ann Arbor, MI 48109-1055 kann@ncbi.nlm.nih.gov

1 Abstract

A new method to derive a score function to detect remote relationships be-

tween protein sequences has been developed. The new score function, OP-
TIMA, was obtained after maximization of a function of merit representing
a measure of success in recognizing homologs of the newly sequenced pro-
tein among thousands of non-homolog sequences in the databases. We find
that the new score function obtained in such a manner performs better than
standard score functions for the identification of distant homologies.

2 Introduction
The amount of sequence data from the rapidly progressing genome projects
keeps growing exponentially. The similar fast growing speed of computers
has make computer aided tools to infer structure and function of the newly
sequenced proteins an integral part of modern molecular biology. In order
to understand the evolutionary history of the new sequences, aligning the
primary structure of the probe sequence with others in the database using so-
phisticated software such as FASTA [1], BLAST [2, 3], and SSEARCH [1, 4]
is one the most significant and widely used techniques. Sequences with a
high similarity score usually share a common structure and might have simi-
lar functions or mechanisms. For any pair of proteins, the optimal alignment
that maximizes the total score can be done quickly, using standard dynamic
programming techniques [5]. The maximum possible score for a given pair


RAG also with the Biophysics Research Division. phone: (734) 763-8013, fax:
(734)764-3323, email:richardgumich.edu.

C. Guerra, S. Istrail (Eds.): Protein Structure Analysis and Design, LNBI 2666, pp. 99-108, 2003.
 Springer-Verlag Berlin Heidelberg 2003
100 Maricel Kann and Richard A. Goldstein

of proteins is then used to determine whether the pair of proteins are homol-
ogous. This is often done by computing such quantities as p(Sr > x), the
probability that a random pair of proteins of the same length would have that
score or higher, E, the expected number of random proteins in the database
that would have at least that score, and P , the probability that there is at
least one random pair with a higher score. Smaller values of p(Sr > x), E,
and P indicate a higher likelihood that the given pair is in fact homologous.
The first commonly-used score matrices were the PAM (percent accepted
mutations) series developed by Dayhoff and co-workers [6]. Others such as
those developed by Gonnet et al (GCB) [7] and Jones et al. (JTT) [8] have
applied Dayhoff’s method to larger sequence datasets. Henikoff and Henikoff
used a dataset of aligned sequence blocks to construct his popular BLOSUM62
matrix [9]. Overington and coworkers used Henikoffs’ cluster method to create
a score matrix (STR) where the protein sequences were aligned based on their
observed structures [10].
The matrices described above, represent considerable improvements in the
search procedure, allowing the immediately detection of similarities among
approximately half of all the newly discovered genes. However, there is still a
challenging problem to overcome, when sequence similarity is low with those
score functions (below 25% sequence identity) distant relationships among
protein sequences that are in fact homologs, are impossible to detect.
In this paper, we describe a new procedure to generate a score function
optimized to detect distantly related pairs of protein sequences. Results with
the independent test sets demonstrate the superiority of the resulting score
function compared with other commonly-used score functions for the detection
of distant homologies.

3 Methods
Theory

We sequentially align a target protein At with each of the proteins in a dataset

of size D, achieving a distribution of scores {Sr } of the form


S= ni,j γi,j + ngap−I γgap−I + ngap−E γgap−E (1)
i,j

where ni,j refers to the number of times that amino acid type i is aligned with
amino acid type j, ngap−I is the total number of gaps in the alignment, ngap−E
is the total number of residues in each gap beyond one, γi,j , also known as
the score function, substitution matrix, or exchange residue matrix, represents
the contribution to the score for any amino acid match or mismatch, while
the gap penalties are given by γgap−I and γgap−E initialization and extention
of a gap, respectively . The score for the alignment of the target protein
and a putative homolog is x. To estimate the significance of this score, we
 OPTIMA: A New Score Function for the Detection of Remote Homologs 101

calculated the probability that this score or higher would be obtained by a

random match. We first compute the Z-score, defined as
x −
Sr 
Z=  (2)

Sr2  −
Sr 2

where the averages are over the alignments of the target protein with the
ensemble of random non-homologous proteins in the dataset. By using the
Z score, we automatically account for variations in the expected score with
the length of the proteins. We can represent the distribution of scores for
ungapped [3, 14, 15] and gapped alignments [16] by an extreme value distri-
bution (EVD) [17]. In this case, the probability that a given random score Sr
would be equal or greater than x is
 Sr
p(Sr > x) = 1 − ρ(x) dx
−∞
= 1 − exp (− exp (−αZ − β)) (3)

α = 1.282 and β = 0.5772... for a perfect EVD [18], although these parameters
are generally adjusted based on the observed distribution. For a search of a
database of size D, the expected number of scores between the target protein
and random pairs is equal to E = D p (Sr > x). In this paper, we use a
value of D = 100, 000. Assuming a Poisson distribution, the probability P of
observing at least one alignment with score equal to or greater than x is given
by
P = 1 − exp (−E) (4)
Both the E-value and the P -value depend upon the size of the database being
searched. E-values range from 0 to D, while P -values range from 0 to 1.
We are interested in optimizing the ability of a score function to discrimi-
nate between homologous and non-homologous pairs of proteins. That is, we
are interested in identifying a true homolog, and in having confidence in this
identification. Our confidence in a putative match is equal to the number of
correct matches divided by the number of matches, both correct and incor-
rect, with the same score or higher. Assuming that we have one true homolog
in the dataset, the average confidence C can be quantified as
1   −1
C= = 1 + D 1 − e− exp(−αZ−β) (5)
1+E
A C value close to 1 indicates a confident alignment, with C decreasing to
1/(1 + D) as the confidence of the alignment decreases. This represents our
average relative chance the match is to a true homolog. In this paper we
optimize the score function by maximizing
C averaged over the training set.
By optimizing
C we automatically focus on homologous pairs at the limit of
detection, reducing the dependence of the score function on homologies that
are either easily detectable (E  1) or overly distant (E ∼ D).
102 Maricel Kann and Richard A. Goldstein

Database Preparation

We are interested in optimizing our score functions for the detection of distant
homologs, beyond the capability of current score functions. We therefore need
a set of known homologs whose homology cannot be reliably determined with
standard pairwise sequence comparisons. For this purpose, we took advantage
of the Cluster of Orthologs Genomes (COG) database of Koonin and co-
workers [13]. A 900-pair training set was constructed of pairs of proteins
in the same COG that share less than 25% sequence identity. A 177-pair
disjoint test set was constructed in a similar manner from the COG database,
excluding all COGs that contributed to the training set, with each pair of
proteins again having less than 25% sequence identity. We also constructed
a test sets independently of the method used to construct the training setby
taking pairs of proteins identified as homologs in the PFAM database release
5.2 [19] In order to avoid overlap with the training and test sets derived
from the COG database, we ran a BLAST search [2] (using BLOSUM62 [9]
with -12,-2 for the gap penalties) of all the sequences in the PFAM database
against the 1077 pairs from the COGs that we were using either as the training
set or first test set, and excluded all PFAM families with any member with
similarity to these proteins (i.e. E < 10). From this set of protein sequences,
we selected 103 pairs that share less than 25% sequence identity as a second
test set of distantly related sequences.

Optimization of the Score Function

We are interested in maximizing the confidence value
C averaged over pro-
teins in the training set, where the calculation of C involves the distribution
of scores for the optimal alignment of the target proteins with homologous
and non-homologous proteins in the dataset. These optimal alignments are
themselves dependent upon the value of the score function. Thus an iterative
scheme is required.
We started with the BLOSUM62 matrix [9] and used the local dynamic
programming algorithm [5] to align each of the target proteins in the training
dataset against a homolog and a set of non-homologous proteins with a large
number of different gap penalties. We then calculated Z and C for each
pair of homologs, and averaged over the pairs in the training set to yield
C.
The highest C values were obtained with gap penalties of γgap−I = −12.0 and
γgap−E = −2.0. This scoring scheme (BL62(12,2)) was then used to gener-
ate an initial set of alignments of the target proteins with homologous and
non-homologous proteins. The observed distributions of the non-homologous
proteins were used to adjust the values of α = 1.31 and β = 0.74, similar to
the values expected (α = 1.282 and β = 0.5772) for a perfect EVD [18].
As multiplication of the score function by any constant does not change
Z or any of the other statistics, we fixed one score function (γgap−I ) equal to
 OPTIMA: A New Score Function for the Detection of Remote Homologs 103

−2.0, resulting in 211 adjustable parameters corresponding to the 210 possi-

ble pairs of amino acid types and the remaining gap penalty. Starting with
the BL62(12,2) score scheme and the corresponding set of aligned sequences,
we analytically calculated the approximate direction of a steepest descent for
the adjustable parameters assuming the alignments remained unchanged. We
then adjusted the parameters along that direction, realigning the sets of se-
quences at every point, until Armijo’s rule was satisfied [20]. The next appro-
priate direction of steepest descent was then recalculated. Approximately 10
cycles of optimization and re-alignments were performed until the score func-
tion converged. Performance was monitored by simultaneous calculations of

C averaged over the proteins in the test set of distant homologs derived from
the COG database. The statistics with the optimized score function indicated
that the appropriate values of α and β did not appreciably shift.

Results
The values of
C as averaged over the training set and distant COG homolog
test set during the optimization process are shown in Figure 1. The resulting
score function (OPTIMA) obtained after 10 iterations is shown in Table I.
Figure 2 shows the cumulative distribution of C values for COG distant-
homolog test set with the different score functions. As shown, the greater
discriminatory power of the OPTIMA score function is represented by the
larger fraction of the protein sequence pairs having larger values of C. That
implies a substantial improvement in our ability of making confident predic-
tions compared with other standard score functions. The optimal value for

A 36
R -9 56
N -19 4 59
D -20 -18 18 65
C 6 -29 -30 -30 99
Q -3 12 2 2 -30 46
E -10 3 3 20 -39 19 40
G 4 -18 7 -10 -29 -20 -23 67
H -19 3 12 -7 -29 3 2 -18 86
I -5 -28 -32 -34 -6 -30 -33 -41 -28 35
L -7 -20 -32 -43 -6 -23 -31 -42 -27 28 32
K -10 31 1 -4 -29 15 14 -18 -7 -31 -21 37
M -9 -10 -19 -30 -8 1 -21 -30 -19 12 24 -12 51
F -19 -30 -29 -33 -18 -29 -32 -32 -8 8 17 -29 2 57
P -5 -18 -17 -7 -30 -11 -7 -18 -18 -30 -33 -10 -21 -39 74
S 12 -11 10 4 -10 0 -1 2 -9 -20 -22 3 -10 -19 -8 36
T 0 -8 0 -10 -7 -7 -6 -17 -20 -8 -13 -8 -7 -18 -11 18 48
W -29 -29 -39 -40 -18 -19 -29 -19 -18 -28 -15 -30 -8 14 -38 -29 -19 110
Y -19 -15 -19 -20 -18 -9 -21 -29 20 -8 -2 -17 -9 37 -28 -19 -17 22 69
V 6 -32 -31 -31 -6 -19 -28 -30 -29 35 18 -23 10 0 -18 -22 6 -28 -9 38
A R N D C Q E G H I L K M F P S T W Y V

Table 1. OPTIMA score matrix achieved at the tenth iteration. The elements are
multiplied by ten to increase precision; corresponding gap penalties are -120 and
-20.

γgap−I was −11.97 with γgap−E fixed at −2.0. The small change in the gap
penalties indicates that most of the improvement comes from refinements of
104 Maricel Kann and Richard A. Goldstein

Score Matrix Gap penalties COGs Distant Homologs PFAM Distant Homologs
(Initiate/Extend) C p(Sr > x) P  C p(Sr > x) P 
OPTIMA -11.97/-2.0 0.736 0.004 0.283 0.710 0.007 0.311
1 BLOSUM62 [9] -12/-2 0.652 0.009 0.372 0.645 0.016 0.376
2 BLOSUM62 [9] -8/-0.5 [22] 0.248 0.024 0.800 0.312 0.033 0.737
3 PAM250 [6] -12/-2 0.480 0.017 0.549 0.669 0.012 0.359
4 PAM250 [6] -6/-1.3 [22] 0.013 0.072 0.999 0.035 0.061 0.988
5 GCB [7] -12/-2 0.647 0.007 0.377 0.703 0.022 0.324
6 GCB [7] -7.5/-0.4 [22] 0.023 0.061 0.997 0.030 0.042 0.987
7 STR [10] -12/-2 0.515 0.035 0.509 0.575 0.033 0.450
8 STR [10] -8/-0.5 [22] 0.172 0.041 0.866 0.281 0.034 0.774
9 JTT [8] -12/-2 0.517 0.009 0.516 0.642 0.022 0.392
10 JTT [8] -10.5/-0.4 [22] 0.076 0.035 0.958 0.127 0.036 0.916

Table 2. Comparison of the various score matrices and gap penalties on PFAM and
COG test sets of distant homologs (percentage identity less than 25%) as evaluated
with average confidence value (
C), average probability that a random score would
be higher than the known homolog (
p(Sr > x)), and average probability that at
least one of the random scores in a dataset of 100,000 proteins would be higher than
the known homolog (
P ). For the purpose of these comparisons, we use both the
standard default gap penalties as well as the gap penalties derived by Argos [22].

0.75

test

train
0.7
<C>

0.65

0.6

0 5 10 15 20
iteration

Fig. 1. Value of
C averaged over the training dataset and test dataset of distant
homologs during the optimization procedure.

the values of γi,j . As shown in Table II, OPTIMA has a significantly im-
proved average confidence (
C) value compared with other commonly-used
score matrices . This improvement is not confined only to values of C; both

p(Sr > x), the average probability that any random score would be higher
than the homolog, as well as
P , the average probability that at least one
 OPTIMA: A New Score Function for the Detection of Remote Homologs 105

random score is higher than the homolog, are both substantially decreased
compared with other matrices.

0.8
Fraction of protein sequences

0.6

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1
C

Fig. 2. Cumulative distribution of the C values for the various score matrices,
showing the fraction of all protein pairs in the COGs test set of distant homologs
with less than a given value of confidence. The various lines refer to the OPTIMA
score matrix ( – ); BLOSUM62 (12/2)[9] (· · · · · · ); GCB (12/2)[7] (− − −); STR
(12/2)[10] (− · − · − · −); JTT (12/2)[8] (− · · − · · −); and PAM250 (12/2)[6] (–).

Figures 3 A and B show the coverage or fraction of true positives vs. the
estimated number of false positives per query for the distant homolog COGs
and PFAM test sets, respectively, where the estimated number of false posi-
tives per query represents the expected number of random sequences with a
score greater than the pair of homologous sequences. The better performance
of OPTIMA can be seen from the large number of homologous pairs with
lower estimated number of false positives.
As a further test, we constructed a parametric plot where we calculated
the fraction of true positive homologs identified (coverage) and the fraction
of non-homologous pairs identified incorrectly as homologies (fraction of false
positives) as a function of the cut-off value of E. The results are shown for
the distant homolog COGs and PFAM test sets in Figures 3 C and D, respec-
tively. While this parametric plot may be compromised by the presence of
true homologs incorrectly labeled as non-homologs (false false-positives) in the
test sets, the qualitative agreement with the previous plots further supports
the performance of OPTIMA compared with the other score functions.
106 Maricel Kann and Richard A. Goldstein

Estimated false positives per query

0 0
10 10
A B
-2 -2
10 10

-4 -4
10 10

-6 -6
10 0 0.2 0.4 0.8 1 10 0 0.2 0.4 0.8 1
0.6 0.6
0
Coverage 0
Coverage
10 10
Fraction of false positives

-1 C -1 D
10 10

-2 -2
10 10

-3 -3
10 10

-4 -4
10 0.7 0.8 0.9 1 10 0.7 0.8 0.9 1
Coverage Coverage

Fig. 3. A) Expected number of false positives of different score matrices as a

function of the number for protein sequences pairs (Coverage) for the COG test
dataset of distant homologs. The plots show the fraction of the homologous pairs
with fewer than a given number of false positives of higher score (E) expected for
D = 100, 000. B) Similar plot for the PFAM test dataset of distant homologs. C)
Coverge vs. number of false positives for the COG test dataset of distant homologs.
For this plot, we calculated the E values for each homologous and non-homologous
pair of proteins in the test set, and then ranked these values from lower to higher. We
then considered all matches with a value of E lower than a given cut-off value, and
calculated the fraction of the true positives included in this set (Coverage) as well
as the fraction of non-homologous pairs also included (Fraction of false positives).
These two values were then plotted as a parametric plot. This approach, applied
to homology detection by Brenner et al. [11], is related to the Receiver Operating
Characteristic (ROC) measure [23, 24]. D) Similar plot for the PFAM test dataset
of distant homologs. All of the curves are designated as in Figure 2

Discussion
Most methods for constructing a score function rely on creating a dataset
of reliably-aligned sequences or sequence fragments and gathering statistics
on the relative number of times that each possible pair of amino acids are
aligned. In practice, however, we are interested in distinguishing optimally
(but possibly incorrectly) aligned homologs from optimal alignments of non-
homologs. Our approach towards generating a score matrix is to optimize
the ability of this matrix to do what we want to do: discriminate between
 OPTIMA: A New Score Function for the Detection of Remote Homologs 107

homologs and non-homologs. In order to do this we derived a measure of merit

of the score function, the average confidence of the homolog identification, and
maximized this measure over a set of homologous and non-homologous pairs
of proteins. Different measures of merit can be handled in a similar way. Our
score function still represents statistics derived from real homologous protein
sequences, and can be analyzed in terms of evolutionary substitutions and
the physical-chemical properties of the amino acids. In contrast to standard
derivations, the gap penalties can be treated as additional parameters to be
optimized. In tests with two disjoint set of test proteins, we are able to
demonstrate that this score function achieves greater success at discriminating
between homologs and non-homologs compared with standard score matrices.

Acknowledgements. The authors would like to thank Bin Qian for his collaboration
to this project and Stephen Altschul, Ting-Lan Chiu, William Pearson, Romesh
Saigal, Stephen Bryant and John Spouge for helpful discussions, also Todd Raeker
and Michael Kitson for computational assistance. Financial support was provided
by NIH Grant LM0577 and NSF equipment grant BIR9512955.

References
1. Pearson, W. R., Lipman, D. J. Improved tools for biological sequence analysis.
Proc. Nat. Acad. Sci. USA 85:2444–2448, 1988.
2. Altschul, S. F., Gish, W., W.Miller, Myers, E., Lipman, D. Basic local alignment
tool. J. Mol. Biol. 215:403–410, 1990.
3. Altschul, S. F., Gish, W. Local alignment statistics. Methods Enzymol. 215:460–
480, 1996.
4. Lipman, D. J., Pearson, W. R. Rapid and sensitive protein similarity searches.
Science 227:1435–1441, 1985.
5. Smith, T. F., Waterman, M. S. Identification of common molecular subsequences.
J. Mol. Biol. 147:195–197, 1981.
6. Dayhoff, M. O., Schwartz, R. M., Orcutt, B. C. A model of evolutionary change
in proteins. In: Atlas of Protein Sequence and Structure. Vol. 5. Dayhoff, M. O.
ed. Vol. 5. . National Biomedical Research Foundation 1978.
7. Gonnet, G. H., Cohen, M. A., Benner, S. A. Exhaustive matching of the entire
protein database. Science 256:1443–1445, 1992.
8. Jones, D. T., Taylor, W. R., Thornton, J. M. The rapid generation of mutation
data matrices from protein sequences. CABIOS 8:275–282, 1992.
9. Henikoff, S., Henikoff, J. G. Amino acid substitution matrices from protein
blocks. Proc. Nat. Acad. Sci. USA 89:10915–10919, 1992.
10. Overington, J., Donnelly, D., Johnson, M. S., Šali, A., Blundell, T. L.
Environment-specific amino-acid substitution tables: Tertiary templates and pre-
diction of protein folds. Protein Sci. 1:216–226, 1992.
11. Brenner, S. E., Chothia, C., Hubbard, T. J. P. Assessing sequence comparison
methods with reliable structurally identified distant evolutionary relationships.
Proc. Nat. Acad. Sci. USA 95:6073–6078, 1998.
108 Maricel Kann and Richard A. Goldstein

12. Altschul, S. F., Madden, T. L., Schaffer, A. A., Zhang, J., Zhang, Z., Miller,
W., Lipman, D. L. Gapped Blast and Psi-Blast: A new generation of protein
database search programs. Nucl. Acid Res. 25:3389–3402, 1997.
13. Tatusov, R. L., Galperin, M. Y., Koonin, E. V. The COG database: a tool for
genome-scale analysis of proteins functions and evolution. Nucleic Acids Res
28:33–36, 2000.
14. Karlin, S., Altschul, S. F. Methods for assessing the statistical significance of
molecular sequence features by using general scoring schemes. Proc. Nat. Acad.
Sci. USA 87:2264–2268, 1990.
15. Dembo, A., Karlin, S., Zeitouni, O. Limit distribution of maximal non-aligned
two-sequence segmental score. Ann. Prob. 22:2022, 1994.
16. Pearson, W. R. Empirical statistical estimates for sequence similarity searches.
J. Mol. Biol 276:71–84, 1998.
17. Gumbel, E. J. Statistics of Extremes. New York: Columbia University Press.
1958.
18. Gumbel, E. J. Statistics theory of extreme values and some practical applica-
tions. Washington: U.S. Government Printing Office: National Bureau of Stan-
dards Applied Mathematics Series 33. 1954.
19. Bateman, A., Birney, E., Durbin, R., Eddy, S. R., Finn, R. D., Sonnhammer,
E. L. L. Pfam 3.1: 1313 multiple alignments match the majority of proteins.
Nucleic Acids Research 27:260–262, 1999.
20. Dennis, J. E., Jr., and Schnabel, R. B. Numerical Methods for Unconstrained
Optimization and Nonlinear Equations. New York: Prentice-Hall. 1983.
21. Press, W. H., Teukolsky, S. A., Vetterling, W. T., Flannery, B. P. Numerical
Recipes in C: Cambridge University Press. 1992.
22. Vogt, G., Etzold, T., Argos, P. An assessment of amino acid exchange matrices
in aligning protein sequences: The twilight zone revisited. J. Mol. Biol. 249:816–
831, 1995.
23. Metz, C. E. ROC methodology in radiologic imaging. Invest. Radiol. 21:720–
733, 1986.
24. Swets, J. A. Measuring the accuracy of diagnostic systems. Science 240:1285–
1293, 1988.
A Comparison of Methods for Assessing the
Structural Similarity of Proteins

Dean C. Adams and Gavin J.P. Naylor

Dept. Zoology and Genetics, Iowa State University, Ames, IA 50011, U.S.A.
{dcadams, gnaylor}@iastate.edu

1 Introduction
The link between biological form and function is well known, and is assumed
to hold true at the molecular level. Since identifying similar protein struc-
tures is the first step in identifying similar functions, much effort has been
placed in developing methods to detect structural similarity. Several methods
exist, including: SCOP [8], the DALI algorithm (from the FSSP Database
[6]), the VAST algorithm (from the MMDB database [5]), and Root Mean
Square (RMS) superimposition [9]. The latter three provide quantitative met-
rics describing protein similarity on an objective, continuous scale. Statistical
analyses can then be performed on similarity scores for a set of proteins, to
obtain a plot of ’protein structure space’ [7]. Before such analyses are done
however, one must be sure that the metric used accurately represents similar-
ity.
In this paper, we describe the DALI Z-score and RMS-distance (DRMS )
metrics, and discuss their shortcomings. We then present a novel means of
comparing protein structures using Geometric Morphometric (GM) methods:
statistical shape methods borrowed from anatomy. Finally, we compare results
from these three methods for a data set of globin structures, and show that
the more intuitive GM method markedly outperforms existing techniques.

2 The DALI Algorithm

The DALI algorithm [6],[7] compares protein structures using two-dimensional
matrices, where each element in the matrix (dij ) is the Euclidean distance


Work supported in part by a National Science Foundation Postdoctoral Fellowship
in Biological Informatics: DBI-9974207 to DCA.

C. Guerra, S. Istrail (Eds.): Protein Structure Analysis and Design, LNBI 2666, pp. 109-115, 2003.
 Springer-Verlag Berlin Heidelberg 2003
110 Dean C. Adams and Gavin J.P. Naylor

between the ith and j th residues for that protein. Distance matrices are aligned
in pairwise fashion, and n homologous residues are identified. The structural
similarity for the two proteins (A and B) is then calculated as:
 ∗ 2


 |dA B 
ij − dij |
d
− 20ijÅ
S= 0.2 − e (1)
i j
d∗ij

where d∗ij is the mean distance for those residues (a standardized version of S,
the Z-score, is also calculated). Z-scores are calculated for all protein pairs, and
the best three-dimensional ordination of the structure space is found through
an eigen-decomposition (correspondence analysis) of the Z-score matrix, where
similar proteins are close together, and dissimilar proteins are far apart.
Though Z-scores quantify some aspects of structural similarity, details of
this metric warrant careful scrutiny. First, Z-scores are generated from pair-
wise alignments, so different residues can be used for each pair. Thus, values
in the Z-score matrix represent different aspects of structural similarity, and
are not directly comparable. Second, the metric contains a dissimilarity cut-off
(0.2) to eliminate protein comparisons > 20%. However, most protein com-
parisons in a large database are > 20%, yielding negative scores, which are
arbitrarily truncated to zero. An eigen-analysis of such data will explain little
of the variation in few dimensions, and a low dimensional ordination from
this analysis will fail to capture the essence of ’structure space.’ Finally, the
exponential term in the metric downweights contributions from residues far
from one another. This results in Z-scores for self-comparisons that are not the
same for each protein, implying that some proteins are more ’perfectly’ similar
to themselves than others, which is nonsensical. Thus, DALI Z-scores are not
a true similarity metric, and statistical analyses of them are unpredictable.

3 The Root Mean Square Algorithm

Root Mean Square (RMS) methods assess structural similarity using a least
squares (LS) criterion. First, two proteins (X & Y) are structurally aligned
to identify the set of n homologous residues [4]. Next, they are translated to
a common location, and are rotated so that homologous residues line up as
closely as possible [9]. Finally, the Euclidean distance (DRMS ) between them
is calculated: 
 n 3



DRMS =  (Xij − Yij )2 (2)
i=1 j=1

where Xij and Yij are the coordinate sets for the ith residue. DRMS is cal-
culated for all protein pairs, and the best three-dimensional ordination of the
structure space is found through an eigen-decomposition (principal coordi-
nates analysis) of the DRMS matrix.
 Comparison of Methods for Assessing Protein Structures 111

RMS methods are appealing because DRMS makes intuitive sense: unlike
proteins have a large DRMS , while similar proteins align quite well and have
a small DRMS . It is also a true distance measure, because all self-comparisons
of proteins yield an identical value of zero (no structural differences). Like Z-
scores however, DRMS is calculated in pairwise fashion, so different residues
can be used for each protein pair, rendering DRMS scores incomparable.

4 Geometric Morphometrics
Both DALI Z-scores and DRMS can be used to generate a map of protein struc-
ture space. However, both have methodological difficulties which limit their
utility. Interestingly, these same difficulties have already been addressed in a
completely different discipline: Geometric Morphometrics (GM). GM meth-
ods were originally developed to analyze anatomical structures (e.g., skulls),
but may easily be adapted to compare macromolecular structures. First, a set
of homologous points recorded on all specimens are superimposed using gen-
eralized Procrustes analysis (GPA), which translates specimens to a common
location, scales them to unit size, and optimally rotates them (in a LS sense)
[3],[10]. Shape variables are then generated for each specimen, which may be
used in statistical analyses [2]. Additionally, Procrustes distance (DP ROC )
between two specimens (X & Y) can be calculated as:



−1 
 n
3
DP ROC = 2 sin (Xij − Yij ) /2
2 (3)
i=1 j=1

where Xij and Yij are the aligned coordinates for the ith residue. DP ROC is
calculated for all protein pairs, and the best three-dimensional plot of protein
shape space is found through an eigen-decomposition (principal coordinates
analysis) of this data. Although GM and RMS are quite similar, they differ
in two important respects. First, size is mathematically held constant in GPA
(not in RMS), and second, GPA superimposes all specimens simultaneously.

5 Comparison of Methods
To compare the three methods described above we used a representative set of
protein structures. We extracted all globin sequences (as of 12/10/1999) from
the Protein Data Bank, and separated them into their individual chains, so
that monomeric and non-monomeric globins could be used. Structural simi-
larity among the 560 chains was then assessed using each of the three methods
(Z-scores, DRMS , DP ROC ). Pairwise structural alignments were calculated in
the DALI domain dictionary [7] (http://www2.ebi.ac.uk/dali) and both DALI
Z-scores and DRMS scores were obtained for each protein pair. For GM, we
112 Dean C. Adams and Gavin J.P. Naylor

aligned the amino acid sequences with Clustal W [11] and deleted all gaps,
yielding 96 homologous residues [1]. We then superimposed the structural data
for these residues with GPA, and generated DP ROC for each protein pair.
The ability of each metric to capture structural variation was assessed
using multivariate ordination methods. The DALI Z-score matrix was sum-
marized using correspondence analysis (as per [7]), and the DRMS and DP ROC
matrices were summarized using principal coordinates analysis. The percent-
age of variation explained by the first three dimensions from the ordination
analysis was compared for each method, and their ability to identify bio-
logically meaningful clusters was assessed through a visual inspection of the
ordination plots.
Using DRMS , the 1st three dimensions of structure space explained 76.1%
of globin chain variability. Inspection of this ordination plot revealed separa-
tion of a few individual chains (mostly hemoglobin chains), but no obvious
groups were identifiable (Fig. 1). Thus, although DRMS explained much of the
variation, it was unable to identify any biologically meaningful globin clusters.

Fig. 1. Three-dimensional view of globin structure space from DRM S .

Using DALI Z-scores, the 1st three dimensions of structure space explained
33.5% of globin chain variation, and to describe an equivalent amount of vari-
ation to DRMS (76%), 56 dimensions of the ordination were needed. Further,
the ordination plot revealed no obvious clusters of globin chains (Fig. 2). Thus,
DALI Z-scores were much less effective at summarizing structural variability,
and were unable to reveal biologically interpretable clusters of proteins.
Using DP ROC , the 1st three dimensions of the GM shape space explained
76.6% of the variation, which was similar to that found with DRMS . How-
ever, the ordination plot revealed remarkable separation of globin chains into
 Comparison of Methods for Assessing Protein Structures 113

Fig. 2. Three-dimensional view of globin structure space from DALI Z-scores.

identifiable groups. These groups corresponded to meaningful biological par-

titions of the data set, including: bacterial hemoglobins, clam hemoglobins,
ferric hemoglobins, hemoglobin α (& β)-chains, lamprey hemoglobins, leghe-
moglobins, and myoglobins (Fig. 3). Thus, much more biological information
is obtained using GM, as compared to either DALI Z-scores or DRMS .

2 4

5
6 8
3

Fig. 3. Three-dimensional view of globin shape space from DP ROC . Labels corre-
spond to the following groups: 1: bacterial hemoglobins, 2: clam hemoglobins, 3: fer-
ric globins, 4: hemoglobin α-chains, 5: hemoglobin β-chains, 6: lamprey hemoglobins,
7: leghemoglobins, and 8: myoglobins.

6 Discussion
In this paper we described two metrics used for assessing structural similar-
ity (DALI Z-scores and DRMS ), and described how geometric morphometric
(GM) methods, commonly used in anatomical studies, may also be employed
to compare protein structures. We then compared the ability of three metrics
114 Dean C. Adams and Gavin J.P. Naylor

to summarize structural variation in a set of globin structures. DALI Z-scores

explained very little (33.5%) of the total variation in three dimensions, and
were unable to identify any globin clusters. DRMS explained significantly more
variation (76.1%) in three dimensions, but it too was unable to identify clus-
ters of globins. On the other hand, DP ROC explained a large proportion of
the variation (76.6%) in three dimensions, and was able to identify biological
clusters of globins (e.g., bacterial hemoglobins, leghemoglobins, myoglobins,
etc.). Further, all but 1 globin chain was correctly classified to its biological
group (1 hemoglobin α-chain was classified as hemoglobin β-chain).
These results suggest that GM methods may be more useful for extracting
meaningful biological information from protein structures than are either the
DALI or RMS methods. Why might this be the case? It seems that DALI
Z-scores are predisposed not to identify meaningful structural variation: they
are calculated in pairwise fashion, and are not a true similarity measure. Fur-
ther, the arbitrary similarity cut-off predisposes them to identify many protein
comparisons as ’maximally’ different. Explaining the performance of DRMS
however, is more difficult. DRMS is a true distance measure, so it does not
suffer the same problem as Z-scores. Further, DRMS and DP ROC differ algo-
rithmically in only two respects: DRMS is a linear distance (where DP ROC is
curve-linear), and the DP ROC protocol standardizes the size of each specimen
(but for globins, size accounts for a tiny portion of variation). The only other
difference between the RMS and GM methods is that DRMS is calculated from
homologous residues from pairwise structural alignments, whereas DP ROC is
calculated from homologous residues common to all structures (found from a
multiple alignment). This assures that the information contained in DP ROC is
consistent among comparisons, which is an explicit requirement of any statis-
tical analysis. It appears then, that the GM protocol, using homology defined
for all proteins simulataneously, provides the best chance for identifying nat-
ural clusters of structurally-similar proteins.

References
1. D. C. Adams and G. J. P. Naylor: A new method for evaluating the structural
similarity of proteins using geometric morphometrics. in S. Miyano, R. Shamir,
and T. Takagi (eds.) ( Currents in computational molecular biology). Universal
Academy Press, Tokyo 2000.
2. D. C. Adams and F. J. Rohlf: Ecological character displacement in Plethodon:
biomechanical differences found from a geometric morphometric study. Proc.
Natl. Acad. Sci. U.S.A. 97 (2000) 4106–4111.
3. F. L. Bookstein: Morphometric Tools for Landmark Data: Geometry and Biol-
ogy. Cambridge University Press, Cambridge 1991.
4. M. Gerstain and M. Levitt: Comprehensive assessment of automatic structural
alignment against a manual standard, the scop classification of proteins. Protein
Sci. 7 (1998) 445–456.
5. J.-F. Gibrat, T. Madej, and S. H. Bryant: Surprising similarities in structure
comparison. Curr. Opin. Struct. Biol. 6 (1996) 377–385.
 Comparison of Methods for Assessing Protein Structures 115

6. L. Holm and C. Sander: Protein structure comparison by alignment of distance

matrices. J. Mol. Biol. 233 (1993) 123–138.
7. L. Holm and C. Sander: Mapping the protein universe. Science 273 (1996) 595–
602.
8. A. G. Murzin, S. E. Brenner, T. Hubbard, and C. Chothia: SCOP: A struc-
tural classification of proteins database for the investigation of sequences and
structures. J. Mol. Biol. 247 (1995) 536–540.
9. S. T. Rao and M. G. Rossman: Comparison of super-secondary structures in
proteins. J. Mol. Biol. 76 (1973) 241–246.
10. F. J. Rohlf and D. E. Slice: Extensions of the Procrustes method for the optimal
superimposition of landmarks. Syst. Zool. 39 (1990) 40–59.
11. J. D. Thompson, D. G. Higgins, and T. J. Gibson: Clustal W: Improving the
sensitivity of progressive multiple sequence alignment through sequence weight-
ing, position specific gap penalties and weight matrix choice. Nucl. Acids Res.
22 (1994) 4673–4680.
Prediction of Protein Secondary Structure at
High Accuracy Using a Combination of Many
Neural Networks

Claus Lundegaard1 , Thomas Nordahl Petersen1 , Morten Nielsen1 ,

Henrik Bohr2 , Jacob Bohr2 , Søren Brunak2 , Garry Gippert1 , and Ole Lund1
1
Structural Bioinformatics Advanced Technologies A/S, Agern Allé 3, DK-2970
Hørsholm, Denmark
2
Structural Bioinformatics Inc., SAB, San Diego, California

1 Summary
A protein secondary structure prediction protocol involving up to 800 neural
network predictions has been developed by SBI-AT. An overall performance
of 80% is obtained for three-state (helix, strand, coil) DSSP categories. Input
to primary-layer neural networks includes sequence profiles, relative residue
position, relative chain length, and amino-acid composition. Secondary struc-
ture predictions are made for three consecutive residues simultaneously – a
technique which we describe as ‘output expansion’ – which boosts the perfor-
mance of second-layer structure-to-structure networks. Independent network
predictions arise from 10-fold cross validated training and testing of 1032 pro-
tein sequences at both primary and secondary network layers. Network output
activities are converted to probabilities. Finally, 800 different predictions are
combined using a novel balloting procedure.

2 Introduction
Prediction of three-dimensional protein structure from primary amino-acid
sequence is one of the biggest challenges in structural biology today. One
step toward solving this problem is by increasing the accuracy of secondary
structure predictions for subsequent use as input to ab initio calculations or
threading algorithms. The predictions are of significant importance to pro-
tein fold recognition and homology modeling, and the degree to which the
secondary structure can be determined has become a benchmark for protein

C. Guerra, S. Istrail (Eds.): Protein Structure Analysis and Design, LNBI 2666, pp. 117-122, 2003.
 Springer-Verlag Berlin Heidelberg 2003
118 Claus Lundegaard et al.

Fig. 1. Architecture of first-layer neural networks used for secondary structure

predictions (A) without output expansion, (B) with output expansion.

structure prediction. The most successful method of secondary structure pre-

diction so far has been the use of artificial neural networks. Early methods
for these predictions relied on the use of a single protein sequence [5, 3].
However, a significant increase in performance can be obtained by using se-
quence profiles as demonstrated with the PHD method developed by Rost and
Sander [12, 13]. This method performed best in the CASP2 experiment with
a mean Q3 (three state prediction: helix, strand, coil) of 74%. In CASP3 the
PSI-PRED method [7] performed best with a Q3 performance approximately
seven percentage points better than a version of the PHD method similar to
the one used in CASP2 [10]. The major difference between theese two meth-
ods is that Rost and Sander used probability scores obtained by a single Blast
search [1], whereas Jones used the probability scores obtained from several
iterative search rounds using PSI-Blast [2]. Studies have shown that an in-
creased performance in secondary structure prediction can also be obtained
by combining several estimators, such as neural networks [12, 6]. A combina-
tion of up to eight neural networks has been shown to increase the accuracy,
but a saturation point was reached in the sense that adding more networks
would not increase the performance substantially [4]. The present paper is a
description of the performance of an improved procedure for combining up to
800 predictions generated using differently trained neural networks [11].

3 Methods
3.1 Neural Networks
A feed-forward neural network architecture was used, with input, hidden and
output layers (Figure 1). Input consists of sequence profiles in window sizes
 Prediction of Protein Secondary Structure at High Accuracy 119

of 15, 17, 19 and 21 residues surrounding position i, relative residue position,

relative chain length and per-chain amino-acid composition. Hidden layer sizes
of 50 and 75 units were used.
In conventional networks the output layer (Figure 1A) consists of sigmoidal
activities for H, E and C only for position i. Using output expansion the output
layer (Figure 1B) consists of HEC activities also for positions neighboring i.
First-layer network output is fed into a second-layer structure-to-structure
network using an input window size of 17, hidden size 40, and output HEC
activities for residue i.

3.2 Data Set Used for Network Training

Network training data was prepared from the PDB version of August 1999.
A set of 1168 protein chains remained after an extensive quality check [11],
application of a strict pairwise sequence similarity threshold [9], and removal
of transmembrane spanning chains. Eight-category DSSP secondary structure
assignments [8] were reduced to three-category H (helix), E (extended strand)
and C (all other) for use in network training. Sequence profiles were obtained
using iterative BLAST searches [2]. The set of 1168 protein chains was fur-
ther homology reduced against the evaluation set prior to network training,
yielding a set of 1032 chains representing ∼200,000 amino-acid residues.

3.3 Evaluation Data Set

Prediction performance was measured using a benchmark set of 126 protein

chains [12]. Several sequences in the 126 set were found to contain structural
gaps (missing residues) in the corresponding PDB entry. Missing segments
obtained from SWISS-PROT sequences were included as context information
for residues with DSSP structural assignments.

3.4 Balloting Procedure

Balloting consists of taking the weighted average of the subset of predictions

with ‘confidence’ greater than a specified threshold, expressed as a Z-score
with respect to the mean. The subset of networks is in general different for
each query chain. Per-chain, per-prediction confidence is the mean difference
between highest and next-highest probability among three-category predic-
tions [11].

4 Results
A performance of 80% mean per-residue secondary structure prediction accu-
racy was obtained using a benchmark set of 126 protein sequences [11, 12].
120 Claus Lundegaard et al.

SBI−AT
Psi−Pred
PHD server

population 0.20

0.10

0.00
40.0 50.0 60.0 70.0 80.0 90.0 100.0
HEC performance (% correct)

Fig. 2. Distribution of per-chain secondary structure prediction accuracy for SBI-

AT (red), PSI-PRED (black) and PHD (green) methods.

Per-chain accuracies range from 55-100% (Figure 2) with an overall mean of

80.5%. Competitive prediction servers yield a mean of 78.8% (PSI-PRED, [7])
and 76.3% (PHD, [12, 13]).

4.1 Probability Transformation

HEC probabilities are obtained from network output activities via matrix
transformation, which provides an additional 1.2% in performance over un-
transformed activities. A conversion matrix is obtained for each of 800 network
combinations by processing the training data set.

4.2 Output Expansion

Output expansion is a procedure developed at SBI-AT which provides training

hints to neural networks by predicting the secondary structure for more than
one residue at a time. Predicting for three contiguous residues provides an
additional 0.5% in performance over single-residue predictions.
 Prediction of Protein Secondary Structure at High Accuracy 121

80.2

80.0
Q3 performance (%)

79.8
A
WA
SD1
SD2
79.6 SD3

0 40 80 120 160 200

Number of predictions

Fig. 3. Performance vs. Number of predictions and balloting threshold. Mean and
standard deviations were obtained using random selections with ‘replacement’, which
tends to underestimate performance.

4.3 Combinations of Many Predictions

Prediction accuracy increases asymptotically (Figure 3) as more predictions

are combined to obtain HEC probabilities. Using more than 200 predictions
provides an additional 0.2% in performance over using only 10 predictions [4].
An efficient balloting procedure was developed in which ‘confident’ predic-
tions dominate in a weighted average of secondary structure probabilities.

References
1. Altschul, S. F., Gish, W., Miller, W., Myers, E. W. & Lipman, D. J. (1990) J.
Mol. Biol. 215, 403–410.
2. Altschul, S. F., Madden, T. L., Schffer, A. A., Zhang, J., Zhang, Z., Miller, W.
& Lipman, D. J. (1997) Nucleic Acids Res. 25, 3389–3402.
3. Bohr, H., Bohr, J. & Brunak, S., et al. (1988) FEBS Lett 241, 223–228.
4. Chandonia, J. M. & Karplus, M. (1999) Proteins 35, 293–306.
5. Chou, P. Y. & Fasman, G. D. (1974) Biochemistry 13, 211–222.
6. Cuff, J. A. & Barton, G. J. (1999) Proteins 34, 508–519.
7. Jones, D. T. (1999) J. Mol. Biol. 292, 195–202.
8. Kabsch, W. & Sander, C. (1983) Biopolymers 22, 2577–2637.
122 Claus Lundegaard et al.

9. Lund, O., Frimand, K., Gorodkin, J., Bohr, H., Bohr, J., Hansen, J. & Brunak,
S. (1997) Protein Engineering 10, 1241–1248.
10. Orengo, C.A., Bray, J.E., Hubbard T., LoConte L. & Sillitoe I. (1999) Proteins
Suppl 3, 149–170.
11. Petersen, T. N., Lundegaard, C., Nielsen, M., Bohr, H., Bohr, J., Brunak, S.,
Gippert, G. P. & Lund, O. (2000) Proteins 41, 17–20.
12. Rost, B. & Sander, C. (1993) J. Mol. Biol. 232, 584–599.
13. Rost, B. & Sander, C. (1994) Proteins 19, 55–72.
Self-consistent Knowledge-Based Approach to
Protein Design

Andrea Rossi1 , Cristian Micheletti1 , Flavio Seno2 , and Amos Maritan1

1
SISSA, INFM and the Abdus Salam International Center, Trieste - Italy
2
INFM - Dipartimento di Fisica, Università di Padova, Padova - Italy

1 Introduction
Two of the most investigated problems in molecular biology are protein fold-
ing and design. These problems stem from Anfinsen’s discovery [1] that the
sequence of amino acids of a naturally-occurring protein uniquely specifies
its thermodynamically stable native structure. The protein folding challenge
consists of predicting the native state of a protein from its sequence of amino
acids, while in protein design one is concerned to identify the amino acid se-
quences folding into a pre-assigned native conformation. This last issue, hav-
ing obvious practical and evolutionary significance, has attracted considerable
attention and effort of experimentalists and theorists [2, 3, 4, 5, 6, 7, 8].
The difficulty of the protein design problem is enormous because, in prin-
ciple, a rigorous approach [3, 7] would entail a simultaneous exploration of
both the family of viable sequences and the family of physical conformations.
By doing so, it would be possible to find the sequences having lower energy in
the target structure than in any other conformation. Stated mathematically,
to design a target structure Γ , one needs to identify the sequence of amino
acids, s, that maximizes the “occupation probability” according to Boltzmann
statistics:
exp (−βHs (Γ )) exp (−βHs (Γ ))
Ps (Γ ) =

= (1)

exp (−βHs (Γ )) Zs
{Γ }

evaluated at a suitable physiological temperature, 1/β = kB T . {Γ
} denotes

the family of conformations that can house the sequence s and Hs (Γ ) is the


energy of the sequence in the conformation Γ .
A first obstacle in using Eq. (1) is the difficulty of determining Hs (Γ ).
However, even assuming the correct knowledge of H, it would be impossible
to carry out an exhaustive search of the sequence maximizing Ps (Γ ), due to
C. Guerra, S. Istrail (Eds.): Protein Structure Analysis and Design, LNBI 2666, pp. 123-129, 2003.
 Springer-Verlag Berlin Heidelberg 2003
124 Andrea Rossi et al.

the computational difficulty of determining of Zs accurately. By writing Zs =

exp(log Zs ) and taking the first order term in its cumulant (high-temperature)
expansion, the condition of maximizing Ps yields

Ws (Γ ) ≡ Hs (Γ ) − Hs   0 , (2)

where the average · · ·  is carried out over all the conformations Γ that can
house s. Two main problems remain unsolved before the minimization of
Ws (Γ ) in sequence space can be exploited in automated contexts to design
protein structures. In fact neither the functional form of the effective energy
Hs is known, nor its average, · · · , is easily computable being dependent on
a large set of unknown conformations. In the next section we will introduce
a plausible simple form for Hs and will determine the corresponding Ws by
using data from a set of real proteins.

2 The Design Strategy

To represent a protein we shall use the common coarse-grained modelling
where each amino acid is identified by a centroid placed on the β carbon (α
carbon for Glycine).
Furthermore, we shall also partition the 20 types of amino acids into a
restricted number of classes. More precisely we will adopt the partition where
Ala, Ile, Leu, Met, Phe, Pro, Trp, and Val are in the same class ( hydrophbic),
Asn, Cys, Gln, Gly, Ser, Thr, and Tyr are in the class of the neutral polar
and, finally, Arg, His, Lys, Asp, and Glu are in the charged polar class.
This simplification stems mainly from the observation that most amino
acids in natural proteins can be substituted by “chemically equivalent” ones
without disrupting native folds [9]. Hence, within the present design scheme
we aim at predicting the classes of amino acids designing a given structure.
As in ref. [8], the putative solution could then be fine-grained into 20 amino
acids alphabet by using steric packing and solvation constraints.
Finally, we introduce a suitable (free) energy scoring function. The most
popular choice adopted in simplified models is the pairwise-interaction form

Hs (Γ ) = ∆ij (Γ )B(si , sj ) , (3)
i<j

where i, j are the positions along the sequence of the amino acids and the sum
is taken over all possible pairs. B(si , sj ) represents the interaction strength
of the amino acid pair si and sj . However, only amino acids that are close
enough will interact in a non-negligible way. This is enforced with a suitable
weight function, or contact map.
Due to the linear dependence of the energy H on the contact map (the
only factors that contain geometric information about structures), the r.h.s.
of Eq. (2) can be re-casted into the following forms:
 Self-consistent Knowledge-Based Approach to Protein Design 125

Hs  = ∆ij B(si , sj ) , (4)
i<j

The average contact map, ∆(i, j), was obtained by collecting data on the
contact map of a variety of naturally occurring proteins. It turns out that
∆(i, j) is well reproduced as a (decreasing) function of the sequence sep-
aration, |i − j| when |i − j| < 16. Contacts between residues with sequence
separation larger than 16 are rather rare hence were modeled by assuming a
constant frequency of occurrence, ∆(0) . The value of ∆(0) is a free parameter
that is to be tuned separately
for each protein length so that the average
number of overall contacts, i,j ∆i<j , matches the one observed in nature.
By using Eqs. (2) and (3), the scoring funcion Ws defined in Eq. (1) can
be re-written as:

Ws (Γ ) = (∆ij (Γ ) − ∆ij )B(si , sj ) , (5)
i<j

where the B(si , sj ) are parameters that have been set by requiring that the
scores Ws (Γ ) associated with the native sequences of 30 proteins on their own
native state be as small as possible, consistently with Eq. (2).

3 Results and Discussion

The design of a target structure, Γ , is carried out by minimizing the score,
Ws (Γ ) over the possible sequences of amino acids classes. The exploration of
sequence space is carried out within a stochastic scheme (simulated annealing,
the elementary move being the random mutation of a fraction of residues from
one class to another).
The minimization of Ws (Γ ) is achieved by varying a control parameter,
T , that influences the rate at which sequences with lower and lower values
of Ws (Γ ) are accepted. It is instructive to carry out our design scheme on
naturally occurring target structures. By doing so, it would be possible to
compare the putative design solution with the one adopted by nature. A
possible complication comes from the fact that proteins with sequence identity
as low as 30 % can share the same fold [10, 11]. After the coarse graining into
the three classes mentioned before, this threshold value for homology becomes
55 %. Thus, at the simplest level, a match of about 55 % between our design
solution and the one adopted by nature can be considered succesful. This value
is remarkably close to the best design scores achieved with our procedure (data
not shown). This does not imply automatically that our solutions are viable.
Site-directed mutagenesis experiments [13] have shown that a few protein sites
do not tolerate any substitutive mutation at all (otherwise the native state
would be destabilized). It should then be checked whether such key residues,
which are conserved in homologous proteins, are conserved also by our design
strategy.
126 Andrea Rossi et al.

80
1erv
1fow
70 2imm
2mcm
3wrp
60

Success (%)
50

40

30

20
-2.5 -2 -1.5 -1 -0.5 0 0.5
W

Fig. 1. The success as a function of the cost function W (s, Γt ) = Hs (Γt ) −
Hs  per
site. Success is defined here by majority rule on a sampling of hundred (decorrelated)
sequences. The value of the cost function for the respective wild-type sequences is
between −0.48 and −0.78.

A further difficulty may stem from the fact that naturally occurring se-
quences may not have evolved to provide maximum thermodynamic stability,
Ws (Γ ) . Under these circumstances, one would observe that design solutions
which extremize Ws (Γ ) give a worse match with the natural protein than se-
quences with a higher Ws (Γ ) score. For this reason we chose to test the success
rate not only for the minimum value of W (s), but also for other sequences. In
particular it is interesting to compare all the sequences s with W (s) < W (s∗ ),
where s∗ is the wild-type sequence. For each annealing temperature we extract
100 decorrelated sequences and make statistical analysis on this sequence set.
We evaluate the average of W (s) for this set and a “super-sequence” by ap-
plying a pointwise majority rule to this set: for each site we assign the most
frequent amino acid class observed in this sequence set at the given location.
It appears that, indeed, the highest matching with the native sequence, is not
obtained for the lowest value of Ws (Γ ) , but for higher ones as shown in Fig. 1
. Furthermore one can see (figure not shown, see [12] for details) that the sites
that are assigned unambiguosly already at high values of Ws (Γ ) (i.e. early in
our stochastic minimization), have a high probability to match the natural
solution (for protein 1erv for the top 40 sites, there are 32 correct matches!).
It is tempting to conjecture that the residues that are assigned with very
little uncertainty by our design procedure (conserved design residues) could
also correspond to conserved residues in nature, i.e. amino acids that play a
fundamental role in the folding process.
This hypothesis is confirmed by the results of design attempts on two
heavily investigated proteins: barnase and chymotrypsin inhibitor [13, 14, 15].
 Self-consistent Knowledge-Based Approach to Protein Design 127

Fig. 2. Scatter plot of the threshold value of the control parameter, T , at which
the designed class of each residue in barnase (sequentially labelled on the x axis) is
assigned with uncertainty less than 50 %. Lower values on the y axis correspond to
early locked sites in our design procedure. Circled dots represent sites belonging to
core1, core2 or core3 .

For each of them several experimental results are available which pinpoint the
key folding sites. For both proteins we found a striking agreement between
the conserved residues in our approach and the ones identified in mutagenesis
experiments. For example (see Fig. 2), for barnase, the residues for which the
class-locking occurs at high values of the control parameter, T , (i.e. large val-
ues of Ws (Γ ) ) correlate very well with the hydrophobic core1 which Fersht[14]
identified as the initiator of the folding transition. Analogous results hold for
CI2, where the top six conserved residues in our scheme contain the three
residues (ALA35, ILE76, LEU68) indicated by Itzhaki et al.[15] as the most
important in the folding process.
These striking results serve a two-fold purpose. On one hand they confirm
the validity of the present design approach; on the other they also show some of
its possible applications, in connection with the prediction of folding nucleus.

4 Summary
To summarize, we carried out automated protein design attempts over some
PDB conformations by introducing several novel strategies to identify optimal
energy-cost functions and select putative design solutions. A mere comparison
of designed sequences with the PDB ones gives a success rate between 40% and
55% when working with three classes of amino acids: a value well above the
random-guessing threshold. This success rate is not improving by introduc-
128 Andrea Rossi et al.

ing more sophisticated energy functions, suggesting that important features

of real proteins are neglected by short range Hamiltonians. Nevertheless, a
statistical analysis of a wider set of possible solutions, shows how the design
procedure could be used to predict, with a high confidence, at least a sub-set
of protein sites. These residues can be related to the conserved sites obtained
by a statistical analysis of naturally occurring homologous sequences. More-
over, for two specific proteins (barnase and chymotrypsin inhibitor), these
highly predictable sites correspond with a very good precision, to the folding
nucleus, which is crucial for the folding process.

Acknowledgments This work was supported by INFM (PAIS project) and

MURST (COFIN99).

References
1. C. Anfinsen. Principles that govern the folding of protein chains. Science,
181:223–239, 1973.
2. S. Sun, R. Brem, R. Chan, and K. Dill. Designing amino acid sequences to fold
with good hydrophobic cores. Protein Eng., 8:1205–1213, 1995.
3. F. Seno, M. Vendruscolo, J. Banavar, and A. Maritan. Optimal protein design
procedure. Phys. Rev. Lett., 77:1901–1904, 1996.
4. J. Deutsch and T. Kurosky. New algorithm for protein design. Phys. Rev. Lett.,
76:323–326, 1996.
5. B. Dahiyat and S. Mayo. De novo protein design: fully automated sequence
selection. Science, 278(5335):82–87, 1997.
6. C. Micheletti, F. Seno, A. Maritan, and J. Banavar. Design of proteins with
hydrophobic and polar amino acids. Proteins: Structure Function and Genetics,
32:80, 1998.
7. C. Micheletti, J. Banavar, A. Maritan, and F. Seno. Protein structures and opti-
mal folding from a geometrical variational principle. Phys. Rev. Lett., 82:3372–
3375, 1999.
8. A. G. Street and L. S. Mayo. Computational protein design. Structure with
folding and design, 7:R105–109, 1999.
9. S. Kamtekar, J. Schiffer, H. Xiong, J. Babik, and M. Hecht. Protein design by
binary patterning of polar and nonpolar amino acids. Science, 262:1680–1685,
1993.
10. C. Chothia and A. M. Lesk. The relation between the divergence of sequences
and structures in proteins. EMBO J., 5:823–826, 1986.
11. C. Sander and R. Schneider. Database of homology-derived protein structures
and the structural meaning of sequence alignment. Proteins, 9:56–68, 1991.
12. A. Rossi, C. Micheletti, F. Seno and A. Maritan Self-consistent knowledge-based
approach to protein design. Bioph. J., 80, 2001 (in press).
13. A. Fersht. Structure and mechanism in proteinscience: a guide to enzyme catal-
ysis and protein folding. W.H. Freeman, New York, 1999.
14. A. R. Fersht. Optimization of rates of protein folding - the nucleation conden-
sation mechanism and its implications. Proc. Natl. Acad. Sci. USA, 92:10869–
10873, 1995.
 Self-consistent Knowledge-Based Approach to Protein Design 129

15. L. S. Itzhaki, D. E. Otzen, and A. R. Fersht. The structure of the transition

state for folding of chymotrypsin inhibitor 2 analysed by protein engeeniring
methods: evidence for a nucleation-condensation mechanism for protein folding.
J. Mol. Biol., 254:260–288, 1995.
Protein Structure from Solid-State NMR

John R. Quine and Timothy A. Cross

Florida State University and National High Magnetic Field Laboratory

Summary. This article deals with mathematical questions arising from the deter-
mination of protein structure from data obtained by solid-state nuclear magnetic
resonance (NMR). Solid-state NMR holds the promise of revealing the structure of
membrane proteins in a lipid bilayer. The derivation of protein structure from NMR
data has most often been done using proteins in liquid state, and the mathematical
analysis has been done using distance geometry and distance matrices. The math-
ematical analysis for solid state NMR uses orientational constraints rather than
distance constraints, and matrices of inner products rather than distance matri-
ces. Solving the structure from the data requires supplying a sequence of signs, a
situation somewhat analogous to the necessity to supply the phases to solve a struc-
ture from x-ray crystallographic data. Other problems in solving for the structure
arise from the condition that the gram determinants be non-negative, and this is
analogous problem in distance geometry that the distance matrix must satisfy the
conditions of the Cayley-Menger theorem.

1 Discrete Curves
It is convenient to think of a protein as a collection of discrete curves. A
discrete curve is a sequence of points p0 , . . . , pn in three dimensional space,
which can be thought of as atomic coordinates. The backbone is naturally
a discrete curve consisting of points representing the atoms –C1 –N–Cα –C1 –
proceeding from N-terminus to C-terminus. Although the atoms in side chains
have no natural sequential order, side chains are also formed from collection
of discrete curves.
A version of the Frenet frame for differentiable space curves can be given
for discrete space curves. Let
sj = |pj+1 − pj |
and define a unit tangent vector at pj , j = 0, . . . , n − 1, by
pj+1 − pj
tj = . (1)
sj
C. Guerra, S. Istrail (Eds.): Protein Structure Analysis and Design, LNBI 2666, pp. 131-137, 2003.
 Springer-Verlag Berlin Heidelberg 2003
132 John R. Quine and Timothy A. Cross

The discrete curve can be reconstructed from the sequences {tj } and {sj } by

k−1
pk − p1 = sj tj . (2)
j=1

If tj−1 and tj are not parallel, binormal and normal vectors can be given by

tj−1 × tj
bj = nj = bj × tj (3)
|tj−1 × tj |

and a Frenet frame by

Fj = (tj , nj , bj ) . (4)
Generally, vectors will be thought of as column vectors, and frames as a se-
quence of three linearly independent column vectors considered as columns
of a non-singular 3 × 3 matrix. Orthogonal frames correspond to orthogonal
matrices and right-handed orthogonal frames to rotation matrices.
¿From crystallographic studies of small molecules [2], geometric parame-
ters related to the discrete Frenet frame, such as the sequences {tj−1 · tj } and
{sj }, are known to be independent of the particular protein, and dependent
only on the types of atoms involved in the bond. Thus, bond angles C1 – N–
Cα and C1 –C–N are approximately trigonal, 120◦ , and bond angles N–Cα –C1
approximately tetrahedral, 109◦, and bond lengths have standard values in
the range 1 to 1.5 Angstroms, depending on the types of atoms in the bond.
These standard values of bond angles and lengths were of key importance
in the initial investigations of protein secondary structure by Pauling and his
co-workers. It is also of key importance in structure determination by solid
state nuclear magnetic resonance (ss-NMR). The information available from
ss-NMR experiments concerns the values of tj · B, i.e., the cosines of the an-
gles of unit bond directions with the unit vector B giving the direction of the
magnetic field. We refer to these as bond direction cosines. An equation in one
or more bond direction cosines is referred to as an orientational constraint.
Each observation gives an orientational constraint which is a quadratic equa-
tion in one or two of these bond direction cosines. Orientational constraints
can be combined with information about the standard values for bond angles
and bond lengths to get structural information about the protein.
Since the observation of hydrogens is a key NMR tool, bond angles involv-
ing hydrogen atoms are also important in ss-NMR structure determination,
and standard values are used also for these bond angles involving hydrogens
[6]. This is a major difference from x-ray crystallography where hydrogen
atoms are not seen.
The discrete Frenet frames are also related to the usual torsion angles
used in molecular structure. The relationship of one Frenet frame to the next
is given by
Fj+1 = Fj R3 (θj+1 ) R1 (τj ) (5)
 Protein Structure from Solid-State NMR 133

where θj = − arccos(tj−1 · tj ), and τj is the torsion angle about the bond

direction tj , and where
   
1 0 0 cos θ − sin θ 0
R1 (θ) = 0 cos θ − sin θ R3 (θ) =  sin θ cos θ 0 (6)
0 sin θ cos θ 0 0 1

are rotations about the x and z axes respectively. Note that θj is the exte-
rior bond angle at pj . Thus the discrete curve can be reconstructed up to a
Euclidean motion from the sequences {sj }, {tj−1 · tj } and the torsion angles
sequence {τj }, the bond lengths, bond angles and torsion angles.
The discrete Frenet frame at pj is related to one or more molecular frames
that can be defined there. A molecular frame can be defined for an atom
bonded to two others. Suppose atom B is bonded to atoms A and C, then a
right-handed orthogonal molecular frame M can be constructed by setting
pA − pB pC − pB u1 × u2
u1 = u2 = b= (7)
|pA − pB | |pC − pB | |u1 × u2 |
and forming the frame
M = (u1 , b × u1 , b) . (8)
This molecular frame is especially useful if the electron cloud is considered to
be symmetric the the plane of the three atoms, as is the case in the peptide
plane. In this case the local chemical and electric properties, as they relate to
NMR, are often assumed to be symmetric in the plane of u1 and u2 .
The frame M is related to the Frenet frame for the sequence pA , pB and
pC or the sequence pC , pB by a rotation leaving the plane of u1 and u2 fixed.

2 Tensors and NMR

The physics of NMR is quantum mechanics. An NMR experiment is the ob-
servation of the precession of nuclear spins in the presence of a magnetic field.
The equation for this precession is derived from the Zeeman Hamiltonian. A
typical solid-state NMR observable σ is described by a symmetric tensor T
so that σ = B TB, where B is the unit direction of the magnetic field. The
analytical expression for this tensor comes from a perturbation of the Zeeman
Hamiltonian for the magnetic field, and only the second order perturbation is
significant. This is why the observable is given by a quadratic function of the
coordinates of B.
In NMR the signal from an ensemble of molecules is observed. In solid-
state NMR, the molecular frames are fixed with respect to the the direction
B, as opposed to liquid state NMR where the molecular frames are randomly
oriented with respect to B. In our solid-state NMR experiments, the molecules
are held fixed in membrane bilayers that are pressed between glass plates [1].
Thus in ss-NMR, for each piece of data there is an ensemble of tensors fixed
134 John R. Quine and Timothy A. Cross

with respect to the magnetic field direction. This is in contrast to liquid state
NMR, where for each piece of data there is an ensemble of tensors T = R DR
where D is a fixed diagonal tensor and R is a random rotation with respect
to the laboratory frame. This rotation is random function both in the time
variable and in the statistical measure of the ensemble.
Because the NMR observable is a change in spin precession frequency
principally due to the chemical environment of the observed nucleus, each of
these tensors can be thought of as fixed in a molecular frame. Thus M TM is
constant and can be determined by powder experiments where the ensemble
of tensors is kept fixed with respect to time, but random with respect to the
statistical measure of the ensemble.
Generally, one of the principal axes of T is aligned with u1 × u2 and

M TM = R3 (β) diag(σ1 , σ2 , σ3 ) R3 (β) (9)

where diag indicates a diagonal matrix and σ1 , σ2 , and σ3 are the principal
values of the tensor. The principal values and the parameter β are determined
from the powder experiment. A consequence of (9) is that for fixed bond angle,
and hence fixed u1 · u2 , the NMR observable B TB is a quadratic in u1 · B
and u2 · B, and setting it equal to its observed value gives and orientatioal
constraint.
The most common observables in our ss - NMR experiments are the
quadrupolar splitting, the dipolar splitting, and the chemical shift. The
quadrupolar and dipolar splittings give rise to zonal harmonic tensors with
2σ1 = −σ2 = −σ3 , and for these there is a unique principal major axis along a
bond direction. For the chemical shift tensors, the principal values are usually
distinct with σ1 the largest value, so that the maximum chemical shift occurs
in the u1 , u2 plane of the molecular frame.

3 Structure from Orientational Constraints

We turn to the question of whether it is possible to solve for a structure from

orientational constraints. In fact, the analytical solution includes some unde-
termined signs, ±1. The situation is somewhat analogous to crystallography,
where the analytic solution for the electron density map as an inverse Fourier
transform includes a collection of undetermined absolute value 1 complex
numbers, or phases.
As shown above, orientational constraints come from the expression of
NMR observables as quadratic tensors, giving quadratic equations in the
bond-orientation cosines {B · tj } where {tj } are unit tangent vectors along a
discrete curve and B is the unit direction of the magnetic field. With enough
of this kind of data, the quadratic equations can be solved up to undetermined
±1 signs coming from quadratic formula. We refer to these signs as quadratic
indices
 Protein Structure from Solid-State NMR 135

Let
κj = B · tj βj = tj−1 · tj (10)
Proceeding under the assumption that the quadratic indices are known and
the sequence κj has been found, we solve for the unit vectors {tj }. Once the
unit tangent vectors are found, the atom coordinates can be found from the
sequence {sj } of bond lengths using (2).
We can solve recursively for {tj } under the condition that κ2j
= 1 for all
j. Recall that we also assume that the bond angles are known, so that the
sequence βj is known.
To solve for the sequence {tj }, let
 
1 κj−1 κj
Gj = κj−1 1 βj  (11)
κ j βj 1

be the matrix of inner products (gram matrix) of the vectors (B, tj−1 , tj ). Let

gj = det Gj = 1 − κ2j−1 − κ2j − βj2 + 2κj−1 κj βj , (12)

be the sequence of determinants. Now choose a lab frame (i, j, k) with B = k,

then check that

t0 = 1 − κ20 i + κ0 k
1  √ 
tj = (βj − κj κj−1 ) tj−1 + (κj − βj κj−1 ) k + εj gj tj−1 × k .
(1 − κj−1 )
2

(13)
where εj = ±1 solves (10) recursively for the sequence {tj } of unit vectors
[7]. Note that
√
B · (tj × tj−1 ) = εj gj , (14)
i.e., εj is the sign of the above triple product. The εj are referred to as chi-
ralities.
The requirements for solving for the sequence {tj } using (13) are
1. κ2j
= 1 for j = 1, . . . , n
2. gj ≥ 0 for j = 1, . . . , n − 1,
3. the choice of a sequence εj of chiralities.
Each of these requirements can lead to some problems in practice. The first
requirement can be a problem if one of the unit bond vectors to be determined
is parallel to B. In this case the sequence {tj } cannot be determined uniquely.
If, for example, one unit tangent vector is parallel to B then all of the sub-
sequent unit vectors can be rotated about B by the same rotation without
changing the bond angles or orientation constraints.
The second requirement is due to the fact that the gramian determinant
of three linearly independent vectors must be positive definite, since it gives
136 John R. Quine and Timothy A. Cross

the metric tensor. If some of the gj are computed to be negative, then some of
the data is bad, or the assumption about the bond angle is bad. The situation
is analogous to distance geometry where the distance matrix must be positive
semi-definite by the Cayley-Menger Theorem [3].
The third requirement indicates again there is another choice of signs re-
quired to solve for the structure from the orientational constraints.
The method of dealing with these problems varies. Occasionally the data
is good enough so that a sequence of orientations and bond angles can be sup-
plied so that 1. and 2. above are satisfied. The chiralities and the quadratic in-
dices must then be supplied and an initial structure obtained. There may be a
set of reasonable choices for these indices and several intitial structures. These
structures should then be refined using an energy penalty and a stereochem-
ical energy function in a way analogous to x-ray crystallographic refinement
[4] [5].

4 Acknowledgment
This work was supported by the National Science Foundation, DMB 9986036
(JRQ and TAC), DMS 9972858 (JRQ). This work was largely performed at
the National High Magnetic Field Laboratory supported by NSF Cooperative
Agreement (DMR 9527035) and the State of Florida.

References
1. T. A. Cross and J. R. Quine. Protein structure in anisotropic environments:
development of orientational constraints. Concepts in Magnetic Resonance, 12,
2000.
2. R. A. Engh and R. Huber. Accurate bond and angle parameters for x–ray protein
structure refinement. Acta Crystallogr. A, 47:392–400, 1991.
3. T. F. Havel and A. W. M. Dress. Distance geometry and geometric algebra.
Foundations of Physics, 23:1357–1374, 1993.
4. R. R. Ketchem, K.-C. Lee, S. Huo, and T. A. Cross. Macromolecular structural
elucidation with solid-state NMR–derived orientational constraints. J. Biomol.
NMR, 8:1–14, 1996.
5. R. R. Ketchum, B. Roux, and T. A. Cross. High-resolution polypeptide structure
in a lamellar phase lipid environment from solid state NMR derived orientational
constraints. Structure, 5:1655–1669, 1997.
6. A. Kvick, A.R. Al-Karaaghouli, and T.F. Koetzle. Deformation electron density
of α- glycylglycine at 82 K. I. the neutron diffraction study. Acta Crystallogr.,
B33:3796–3801, 1977.
7. J. R. Quine and T. A. Cross. Protein structure in anisotropic environments:
unique structural fold from orientational constraints. Concepts in Magnetic Res-
onance, 12, 2000.
 Protein Structure from Solid-State NMR 137

t1
(a)

B t2
t1
(b)

(1)
B
t2

(2) t1
t2
(c)

Fig. 1. The recursion formula (13) for j = 2 is illustrated in the figure. The sphere
represents the set of unit vectors. The equations κ2 = B · t2 and β2 = t1 · t2 indicate
that given t1 , the the vector t2 lies on both of two circles about B and t1 .

a) If g2 < 0 the circles do not intersect and there is no solution.

b) If g2 = 0 then the circles intersect at one point and there is only one possibility
for t2 .
c) If g2 > 0 then there are two possibilities for t2 distinguished by the chirality
(1) (2)
ε2 . The vector t2 corresponds to ε2 = −1 and the vector t2 corresponds to
ε2 = 1.
Learning Effective Amino-Acid Interactions

Flavio Seno1 , Cristian Micheletti2 , Amos Maritan2 , and

Jayanth R. Banavar3
1
INFM - Dipartimento di Fisica, Università di Padova, Padova - Italy
2
SISSA, INFM and Abdus Salam Centre for Theoretical Physics, Trieste - Italy
3
Department of Physics, Penn-State University, University Park -USA

1 Introduction

1.1 Background

Proteins are heteropolymers of 20 amino acids with genetically determined

sequences. Most proteins assume specific globular conformations under physi-
ological conditions. Several experiments indicate that this unique native state
is thermodynamically stable and encoded unambiguously by the sequence of
amino acids [1].
Two challenging problems arise from these facts: protein folding and pro-
tein design. Protein folding consists of predicting the native state from the
known sequence of amino acids; protein design addresses the possibility of pre-
dicting which sequence admits a pre-assigned biologically viable conformation
as its native state[1]. Tackling these problems with a “first principles” entail-
ing the correct treatment of the quantum mechanical behaviour of the large
number of atoms constituting a protein and its surrounding solvent is beyond
the present computational capabilities. A natural way to avoid dealing with
too many microscopic degrees of freedom is to introduce a reduced represen-
tations where each amino acid is represented by one or a few interaction sites.
The main difficulty with such simplified representations is the introduction
of an effective energy function that captures the essential qualitative physical
and chemical features of the folding process. This involves both choosing the
right form for the energy function as well as determining the precise value of
the coarse-grained interaction potentials between the amino acids.
In this study we describe a method that allows to determine of these
potentials and to verify the reliability of the chosen parametrisation of the
energy function.

C. Guerra, S. Istrail (Eds.): Protein Structure Analysis and Design, LNBI 2666, pp. 139-145, 2003.
 Springer-Verlag Berlin Heidelberg 2003
140 Flavio Seno et al.

1.2 Determination of Interaction Potentials

Let us consider a coarse-grained description of a protein whose energy is de-
scribed by an unspecified effective energy function depending on a set of p
parameters (e.g. interaction potentials, chemical potentials etc.) (α1 , . . . , αp ).
The particular parametrisation is crucial: one would like to capture the key
ingredients such as size and steric constraints, the polar or hydrophobic na-
ture of the amino acids, electric charge and hydrogen bonding with as few
parameters as possible. Assuming that a physical parametrisation is avail-
able, one could extract the parameters by observing[2, 3] that, given a protein
sequence S, the correct set of {ᾱ} ought to recognize its native state Γ as
having lower energy than any arbitrary, though physically viable and distinct
from Γ , conformation Γ d of the same length (decoy conformations), namely:
H(S, Γ, ᾱ) < H(S, Γ d , ᾱ) Γ
= Γ d , (1)
where H(S, Γ, α) is the energy of the sequence S mounted on Γ obtained
with the energy function H and the parameters {α}. In other words, the
correct potentials must be chosen from the region of parameter space satisfying
inequalities (1) for any choice of viable decoys, {Γ d}.
The Protein Data Bank (PDB) contains the native state structures of a
large number of amino acid sequences. For each of them we could, in principle,
collect a large set of inequalities such as (1) as long as we are able to find
suitable decoy conformations. The goal is then to infer the energy function
and the set of parameters that satisfy these inequalities for all native states
and associated decoys. These scheme has the appealing feature of showing
whether the chosen parametrisation of H can be appropriate or not. In fact,
the impossibility of finding parameters recognizing the native states as the
lowest energy ones with respect to the decoys would signal that the form
for the energy function is unsuitable. The key difficulty to implement this
approach is the generation of relevant decoys, i.e. structures that are really
competitive with the native one to be ground-states of S. In the next section
we introduce a new procedure to generate significant competitive structures.

1.3 An Iterative Strategy

The key idea [4] of our method is the observation that the native states of
proteins must at least satisfy the requirement of being located at the bottom
of a smooth free energy minimum with a wide basin of attraction [5, 6] . This
suggests a straightforward approach:
a) begin with an initial guess of the potential parameters;
b) start from the native states of several proteins and carry out an unbiased
Monte Carlo (or molecular dynamics) simulation (say at zero temperature)
and determine several accessible local minima for each of the proteins.
c) Modify the potential parameters in such a way as to destabilize these
conformations in favour of the known native state, according to eq. (1).
d) Iterate this procedure by returning to (b).
 Learning Effective Amino-Acid Interactions 141

After several iterations, one would expect to converge to a set of potential

parameters which best capture the shape of the free energy landscape in the
vicinity of all the native state structures. Our method adopts the thermody-
namic stability scheme described before but with the proviso that the decoys
would be generated by an explicit and simple dynamical process. The struc-
tures so generated are guaranteed to be stringent competitors of the native
structure. Since the scheme is both flexible and optimal, it can be used to
compare the performance of many different scoring functions or parametriza-
tions and hence select the most promising one for ab initio folding simulations.
There are obviously many ways of implementing this idea and we will present a
few schemes here that we have tried and which yield remarkably good results.

2 Models and Techniques

2.1 The Model and the Dynamics

We have used a highly simplified model of a protein in which the amino

acid residue backbone is represented by a self-avoiding chain of connected C α
atoms. The C α positions of a real protein are mapped on suitably chosen
sites of an FCC lattice[7, 8] with unit spacing equal to the typical separation
of subsequent C α atoms (3.8 Å). This framework (XFCC model) allows for
a faithful representation of protein backbones, since the coarse-grained C α
positions are typically within 1 Å of the crystallographic positions [7]. The
choice of the proper FCC sites is done at the expenses of slight variations in
the peptide bond lengths but it allows to preserve the typical torsion angles
found in real protein molecules. For each amino acid we then define also the
position of the side centroid (C β atom) using standard geometrical rules[9].
We used a pairwise energy function, involving only interactions between
pairs of amino acids:
 2 

4.65
H= ∆(rij ) · (Si , Sj ) + 10 · r · − 1 · Ω(rij ) (2)
rij

1 1 6.5 − r
where : Ω(r) = Θ(4.65 − r) ∆(r) = + tanh (3)
2 2 2
Θ is the step function and rij is the distance (in Angstroms) between the
Cα atoms of the i-th and j-th amino acid. ∆ denotes the distance-dependent
strength of interactions between the i-th and the j-th amino acids along the
sequence mounted on the structure Γ ,  is the interaction matrix and Si
denotes the type of the ith amino acid in the sequence. Finally, r is a repulsive
term that penalizes cases where two non-consecutive pairs of C α ’s are closer
than 4.65 Å.
142 Flavio Seno et al.

The dynamics in conformation space is carried out using a Monte Carlo

technique. A new conformation is accepted according to the Metropolis rule.
At each attempted MC step, we move up to 2 of consecutive protein residues
to unoccupied positions. The new positions have to satisfy suitable physical
constraints that we deduced by the analysis of the CA and CB positions of an
ensemble of over hundred single-chain proteins. More precisely: the separation
d between two consecutive C α atoms (measured in Å) must remain in the
range 2.6 < d < 4.7; two non-consecutive C β atoms must not be closer than
2 Å; two non-consecutive C α and C β atoms must not be closer than 2 Å; the
chain length is allowed to fluctuate by up to a maximum value of 4 Å.

2.2 Perceptron Learning

A convenient way to find the optimal potentials (given the decoys confor-
mation) is the use of the perceptron algorithm [10] for the optimization of
a set of linear inequalities. In our case, the inequalities are of the form:
H(S, Γ ) − H(S, Γd ) < 0 . This expression can be rewritten as

20

20
(ndij − nnij )(i, j) + r (ndr − nnr ) ≡ aij (Γd ) · (i, j) + ar · r
i>j=1 i>j=1
≡ Q(Γd , ) > 0 . (4)
where nn,d
ij denotes the number of native/decoy contacts involving amino acids
of types i and j and nn,d
r denotes the strength of the native/decoy repulsive
term. Given the native state Γ and the sequence S, the 211 entries of ai,j (Γd )
plus ar depend on the geometrical properties of the decoys. For a given set of
M inequalities to be satisfied simultaneously, it is convenient to identify the
one (denoted with l) that, with the trial potentials is the worst satisfied one:

Q(Γl , ) < Q(Γk , ) k = 1, . . . , M, k
= m (5)

The selection of the conformation l can be done both when Q(Γl , ) is
negative (not all the inequalities are satisfied) and when it is positive (all
the inequalities are satisfied). Once l has been determined, one updates the
trial potentials, (i, j) (or r ) by adding a quantity proportional to (aij (Γl ))
(or ar (Γl ). With these new potentials, each inequality is re-valuated and the
cycle repeated. This method can be shown to converge to the optimal solu-
tion: Q(Γl , ) reaches a constant value [10], which can be of either sign. If it
is negative, no set of potentials can be found that consistently satisfies all
inequalities in the set (unlearnable problem).

3 Results
We began by considering a single protein, PDB code: 1vcc, which has 77
residues. Starting from a set of random potentials, we generated 30 decoys for
 Learning Effective Amino-Acid Interactions 143

which we computed the average root mean square deviation (RMSD),ḡ native
structure and its variance ∆ḡ. With these decoys, we found the potentials by
applying the perceptron algorithm. With the new interaction parameters, we
generated 30 more decoys and kept repeating this procedure.
In Fig. 1 we show the ḡ, as a function of the number of iterations. It is
remarkable that, with such a simplified model, ḡ can be decreased dramatically
from the initial value of about 6 Å to around 1.5-2.0 Å, which is just over the
order of the experimental uncertainty! This provides a nice demonstration of
the fact that it is possible to stabilize the native state in the native basin
within a low uncertainty.

Fig. 1. Results of the iterations on the single protein 1vcc: we show the asymptotic
RMSD of the decoy structures as a function of iteration.

Next, we attempted a task considerably more difficult: to stabilize the

native states of twenty proteins simultaneously. The proteins were chosen
among a list of non-redundant representatives of the main protein folds.
At each iteration step, we generated 5 decoys for each of them. We saw
an improvement as the iterations went on, although not as pronounced as for
the single 1vcc protein with ḡ decreasing to a value of 3.8 ± 0.5 Å.
As recommended by Lazaridis and Karplus [11] as an independent test of
the quality of our potentials we assessed their performance on a set of seven
proteins (PDB codes: 1ctf, 1r69, 1sn3, 2cro, 3icb, 4pti, 4rxn) unrelated to
those used for extracting the potentials and for which more than 600 stringent
decoys structures (for each protein) have been derived[9]. This study ought
to reflect the portability of our potentials, i.e. their applicability in contexts
different from which they have been derived. For each protein we compute
the energy of the the native state Eg and the energy of all the decoys Ei
(i = 1, . . . , M ) (where M is the number of decoys for each single protein) by
using our potentials. With the correct potentials, Eg should be always lower
than any other Ei . The native state is always among the best 5 to 10 %. This is
a highly non-trivial result since it is generally difficult to get such correlations
even employing specially designed energy scoring functions [9].
144 Flavio Seno et al.

As a final test we introduced a slightly more sophisticated model where

we consider interactions between all possible pairs of CA and CB. Using this
second model, we get a clear improvement[4] of the results. This example is
helpful in illustrating the possibility of using our novel optimization technique
in selecting the “most physical” energy parameterization.

4 Conclusions

We have demonstrated how one may extract effective interaction potentials

between amino acids in a coarse-grained description of a protein. The method
relies on the possibility of finding a set of competitive decoys of the native
state. We outlined an iterative procedure to generate these decoys which at-
tempts to stabilize, at least locally, the native state. The results obtained
with simple forms of the energy function are very promising – we were able
to stabilize a set of 20 proteins to an average distance of less than 4 Å and
moreover, the potentials, when tested on completely unrelated decoys, yield
results which are remarkably good especially in view of the simplicity of the
approach.

Acknowledgements.
This article was supported by grants from INFM and MURST-COFIN99.

References
1. T.E. Creighton: Proteins: structure and molecular properties. W.H. Freeman
ed., New York 1992
2. G.M. Crippen: Prediction of protein folding from amino acid sequence over dis-
crete conformation space. Biochemistry 30 (1991) 4232-4237
3. F. Seno, A. Maritan and J.R. Banavar: Interaction potentials for protein folding.
Proteins: Structures, Function and Genetics 30 (1998) 244-248
4. C. Micheletti, F. Seno, A,. Maritan & J.R. Banavar: Learning effective amino acid
interaction through iterative stochastic technique. In press Proteins: Function,
Structure and Genetics (2000).
5. Bryngelson J. D. and Wolynes, P. G: Spin glasses and the statistical mechanics
of protein folding, Proc. Natl. Acad. Sci. USA 84 (1987) 7524-7528
6. C. Micheletti, J. R. Banavar, A. Maritan and F. Seno: Protein structures and
optimal folding from a geometrical variational principle. Phys. Rev. Lett. 82
(1999) 3372-3375
7. D.G. Covell and R. Jernigan: Conformations of folded proteins in restricted
spaces. Biochemistry 19 (1990) 3287
8. B.H. Park and M. Levitt: The complexity and accuracy of discrete state models
of protein structure. J. Mol. Biol. 249 (1995) 493-507
9. B.H. Park and M. Levitt: Energy functions that discriminate X-ray and near-
native folds from well-constructed decoys. J. Mol. Biol. 258 (1996) 367-392
 Learning Effective Amino-Acid Interactions 145

10. W. Krauth & M. Mezard: Learning algorithms with optimal stability in neural
networks. J. Phys. A20 (1987) L745-L752
11. T. Lazaridis and M. Karplus: Effective energy functions for protein structure
prediction. Curr. Op. in Struct. Biol. 10 (2000) 139-145
Proteinlike Properties of Simple Models

Yves-Henri Sanejouand and Georges Trinquier

CRPP, Avenue Albert Schweitzer, 33600 Pessac, and IRSAMC, Université Paul
Sabatier, 118 route de Narbonne, 31062 Toulouse Cédex, France.

During the last decade, it has been shown that several properties of natural
proteins can be captured by simple models, such as 2-D or 3-D lattice models
[1, 2, 5, 6, 9, 11, 13]. In these models, the protein is figured as a chain of beads
occupying the sites of a lattice in a self avoiding way.

1 The 3x3x3 Cubic Lattice Model

Fig. 1. One of the 103346 possible compact conformations of sequence

PHP4 HPHPHP15 H. The hydrophobic (H) residues are the darkest ones.

In this section, we will focuss on some of the proteinlike properties exhibited

by the compact conformations of chains of 27 monomers located at all sites
of a 3 × 3 × 3 cubic lattice, when only two kinds of monomers are considered,

C. Guerra, S. Istrail (Eds.): Protein Structure Analysis and Design, LNBI 2666, pp. 147-153, 2003.
 Springer-Verlag Berlin Heidelberg 2003
148 Yves-Henri Sanejouand and Georges Trinquier

either hydrophobic (H) or polar (P) ones. Within the frame of this model,
there are 227
1.3.108 sequences and 103346 conformations[2], that is, 103346
different ways for a chain to go through all 27 lattice sites, going from a
site to a neighboring one at each step. One of these so-called hamiltonian
paths is shown in Fig. 1, for the following sequence: PHP4 HPHPHP15 H. Of
course, Fig. 1 can also be viewed as showing another possible conformation, the
reverse-labeled one, for the corresponding reverse-labeled sequence, namely,
HP15 HPHPHP4 HP. If such pairs of conformations are assumed to be identical,
the number of different compact conformations drops to 51704[6].
The fact that a sequence can be ”threaded” in a given structure in two different
ways, the position of the first residue in the first way being the position
of the last residue in the other way, is certainly not a property of natural
proteins, which are made with asymetric building blocks, namely, amino acids
of the L-series. As a consequence, the α−helix, one of their major structural
elements, is right-handed, and left-handed helices have not yet been observed
in the tridimensional structures of natural proteins. Note that the reverse-
labeled sequence of a given protein could quite well be synthesized. If buildt
with amino acids of the D-series, its tridimensional structure should be the
same as that of the natural protein, as far as the positioning of the amino
acid sidechains is concerned, except for proline residues, whose sidechain is
involved in a five-member ring with the backbone. However, while complete
syntheses of all-D proteins have been performed, to our knowledge, up to
now, following the seminal work of Shemyakin and its collaborators [12], only
small reverse-labeled all-D peptides, now called ”retro-inverso” peptides[3],
have been synthesized.
In order to study the sequence-structure relationship in the case of a given
model, it is necessary to choose an energy function. For lattice models, the
most usual one has the following form:


H= Eij ∆(ri − rj )
i<j

where ∆(ri − rj ) = 1 if monomers i and j are close neigbors in the lattice,

and ∆(ri − rj ) = 0 otherwise, and where Eij depends on the nature of the
interacting monomers. A popular choice for the Eij values has been Eij =
EHH = −, when monomers i and j are both hydrophobic, and Eij = 0
otherwise[2]. However, it has been noted that the effective interaction between
amino acids in natural proteins is nearly additive[7], that is:
Eij
Ei + Ej
where Ei is a characteristics of the amino acids, strongly correlated with
their hydrophobicity. When the strictly additive case is considered, with for
instance Ei = EH = −1 and Ei = EP = 0, the sequence-structure relationship
in the model exhibits some remarkable features. Noteworthy, if the energy is
determined for all sequences and for all 103346 conformations, it is found that
only 122750 of the sequences (0.09%) have non-degenerate ground states, that
 Proteinlike Properties of Simple Models 149

Table I. 60 remarkable sequences of the 3 × 3 × 3 cubic lattice model. When the

additive potential is used, the ground state of each of them is one of the remark-
able structures known for this model. The 60 remaining remarkable sequences are
the reverse-labelled of these ones. These 120 sequences are the only ones with five
hydrophobic residues and a non-degenerate ground state. For the sake of clarity, P
residues are indicated by dots.

...............H.H.H.H..H. ...H..H.H...........H.H....
.............H.H...H.H..H. ...H..H.H.........H.....H..
...........H.....H.H.H..H. ...H..H.H.........H.H......
.........H.......H.H.H..H. ...H..H.H.......H.H........
.........H.....H...H.H..H. ...H..H.H.....H...H........
.........H...H.H.H..H..... ...H..H.H...H.............H
.........H.H.......H.H..H. ...H..H.H.H.........H......
.........H.H..H......H...H ...H..H.H.H.....H..........
.........H.H..H....H.....H ..H.............H.H.H....H.
.........H.H.H.....H....H. ..H...........H...H.H..H...
.........H.H.H...H......H. ..H...........H.H..H....H..
.........H.H.H...H..H..... ..H......H..H...H.H........
.........H.H.H.H........H. ..H.....H...........H.H..H.
.......H.....H.H.H..H..... ..H.....H......H..H.H......
.......H.H..H........H...H ..H.....H....H..H.H........
.......H.H..H......H.H.... ..H.....H..H..H.H..........
.......H.H..H....H.......H ..H.....H.H.H............H.
.......H.H.H..H......H.... ..H....H..............H.H.H
.......H.H.H.H......H..... ..H....H..H...H.H..........
......H..H.H.H.H.......... ..H....H..H.H.............H
.....H.......H.H.H..H..... ..H...H.............H.H..H.
..H..............H...H.H.H ..H...H....H..H.H..........
..H......H.....H.H.H...... ..H..H..H.H.H..............
..H....H.............H.H.H ..H.H...............H.H..H.
..H....H.....H.H.H........ ..H.H.............H.H..H...
..H....H.H.H.......H...... ..H.H.H.H................H.
..H..H...............H.H.H .H................H.H.H...H
..H..H.......H.H.H........ .H............H.H.H.......H
..H..H.H...............H.H .H........H.H.H...........H
..H..H.H.............H...H .H....H.H.H...............H

is, for each of them, a given conformation is the lowest-energy one, while, for all
of them, ∆, the energy gap between this conformation and the second lowest-
energy one, is: ∆ = 2. Strikingly, out of the 103346 possible conformations,
only 120 (0.11%) are found to be possible ground states. All of these so-
called remarkable structures are characterized by a large ”designability”[6],
as measured by Ns , the number of sequences they are the ground state of, Ns
ranging between 513 and 2306.
150 Yves-Henri Sanejouand and Georges Trinquier

In Fig. 1 the top structure[6] is shown, that is, the conformation which is
the ground state of the largest number of sequences. Among these 2306 se-
quences, PHP4 HPHPHP15 H has the smallest count of hydrophobic residues.
As a matter of fact, each remarkable structure is the ground state confor-
mation of a single five-hydrophobic residues sequence (all given in Table I),
and the corresponding five residues are always located as shown in Fig. 1,
that is, one of them being at the cube center, the four other ones being at
the center of facets that are non bonded to the residue at the cube center.
Furthermore, in all sequences sharing a given ground state, these five residues
are always hydrophobic[13]. The latter point helps to clarify why there are
120 remarkable conformations. Indeed, if a sequence like PHP4 HPHPHP15 H
has a non-degenerate ground state, this means that, out of the 103346 pos-
sible ones, there is only one way to bring its five hydrophobic residues close
together, so that each of them can interact with at least another hydrophobic
one[13]. In other words, it is from a topological point of view that the 120
remarkable conformations of the 3 × 3 × 3 cubic lattice model are atypical[8].
This is a quite satisfactory property of the model, since it means that the
same 120 conformations are also expected to be remarkable when different
Eij values are chosen. For instance, with Eij = EHH = −2 − γ, Eij = EHP =
−1, Eij = EP P = 0, and γ = 0.3, the conformation shown in Fig. 1 is still the
top structure, but with Ns = 3794[6], and the picture is now more complicated
than in the case of the additive potential. The difference between the additive
and nearly additive cases comes from the fact that the departure from addi-
tivity lifts the degeneracy of many sequences. Now energy gaps of ∆ = 0 + nγ
and ∆ = 2 ± nγ are observed, with n = 1, 2, etc[4]. For instance, with γ = 0.3,
nearly half of the sequences whose ground state is one of the remarkable con-
formations have energy gaps lower than 1.0[13]. Moreover, while 4.75% of the
sequences have a unique ground state, for the vast majority of them it is not
one of the remarkable conformations, and the corresponding energy gap is, on
average, close to 0.3, i.e., the γ value[6]. When only sequences with energy
gaps larger than 1.0 are considered, the picture obtained with the additive
potential is restored, despite a few minor differences arising from the overlap
of the energy gaps splitting around the ∆ = 0 and ∆ = 2 cases[13].
An interesting proteinlike property of the model allows for the determination
of all large gap sequences (i.e., with ∆ = 2, in the case of the additive po-
tential) whose ground state is a given remarkable conformation, without the
need of any huge enumeration[13]. To do so, starting for instance from the
corresponding remarkable five hydrophobic residues sequence given in Table
I, all 27 singly-mutated sequences are generated. Then, among them, those
with the same single ground state are retained. Next, for each sequence of this
subset, all 27 singly-mutated sequences are generated, and so on, until no new
sequence can be retained. What is shown through the success of such a proto-
col is that the Ns sequences of each of the 120 remarkable structures belong
to a ”neutral island” of the sequence space[13]. Note that such a property has
also been found in the case of square lattice models[1].
 Proteinlike Properties of Simple Models 151

2 N-Soft-Spheres Models

Fig. 2. One of the possible compact conformations of sequence P2 H2 P4 H3 P4 H2 P2 .

Note that such a geometry is close to the cubo-octahedral one found to be the
average architecture of residue packing in protein structures[10].

Despite its many successes, the 3 × 3 × 3 cubic lattice model has several
drawbacks. In particular, the cubic geometry is a very unlikely one for a
polymer. Moreover, with such a geometry, one of the most spectacular prop-
erty of proteins, namely, their ability to fold into a given structure within a
limited amount of time, has to be studied during the course of Monte Carlo-
Metropolis processes[9, 11], that is, with a set of elementary displacements
chosen a priori. Noteworthy, with standard sets of displacements, it is diffi-
cult for a polymer to jump from a compact conformation to another, since
such a collective motion, during which several monomers are displaced, has
to be performed with local moves only, one or two monomers at a time. To
overcome this latter difficulty, and to allow for the study of the folding pro-
cess during the course of Molecular Dynamics simulations, it is necessary to
consider off-lattice models. However, with such models, one of the most ap-
pealing advantage of the 3×3×3 cubic lattice model is usually lost: the ability
to know, after some simple enumeration, which is the ground state and the
energy gap of any given sequence. The off-lattice model shortly introduced in
this section has been designed so as to retain most of the advantages of the
3 × 3 × 3 cubic lattice model. In this model, the polymer is represented as
a chain of N Lennard-Jones spheres of radius R2 , linked by bonds of length
R, and the sequence of the polymer is specified by ij , the well depth of the
interaction energy between two monomers. For instance, one can consider the
following case: ij = Eij , where Eij is the same as in lattice models. Thus,
the energy function is as follows:
152 Yves-Henri Sanejouand and Georges Trinquier

R 12 R6
H= −(Eij + Ec )( −2 )
i<j
rij rij

where rij is the distance between monomers i and j, and where Ec is a ”com-
paction term”, large enough to make sure that the lowest energy conformations
can be only one among the most compact. Such a term was also used in order
to study the folding process of 27-mers on a cubic lattice[11]. While in this
latter case the most compact geometry is the 3 × 3 × 3 cubic one, in the case
of N Lennard-Jones spheres, though usually more complicated, it is often a
very well defined one[14], like in the N = 19 case, shown in Fig. 2.
For such a family of models also, all the hamiltonian paths can be determined,
as well as the energy gap of any given sequence, and folding simulations of
large gap sequences can be performed. Such simulations are currently under
way.

References
1. E. Bornberg-Bauer. How are model protein structures distributed in sequence
space? Biophys. J., 73(5):2393–2403, 1997.
2. H. Chan and K. Dill. Compact polymers. Macromolecules, 22(12):4559–4573,
1989.
3. M. Chorev and M. Goodman. Recent developments in retro peptides and pro-
teins -an ongoing topochemical exploration. Trends Biotechnol., 13:438–445,
1995.
4. M. Ejtehadi, N. Hamedani, H. Seyed-Allaei, V. Shahrezaei, and M. Yahyane-
jad. Stability of preferable structures for a hydrophobic-polar model of protein
folding. Phys. Rev. E, 57(3):3298–3301, 1998.
5. N. Go. The consistency principle in protein-structure and pathways of folding.
Adv. Biophys., 18:149–164, 1984.
6. H. Li, R. Helling, C. Tang, and N. Wingreen. Emergence of preferred structures
in a simple model of protein folding. Science, 273:666–669, 1996.
7. H. Li, C. Tang, and N. Wingreen. Nature of driving force for protein folding: A
result from analyzing the statistical potential. Phys. Rev. letters, 79(4):765–768,
1997.
8. H. Li, C. Tang, and N. Wingreen. Are protein folds atypical? Proc. Natl. Acad.
Sci. USA, 95(9):4987–4990, 1998.
9. V. Pande and D. Rokhsar. Folding pathway of a lattice model for proteins.
Proc. Natl. Acad. Sci. USA, 96(4):1273–1278, 1999.
10. G. Raghunathan and R. Jernigan. Ideal architecture of residue packing and its
observation in protein structures. Protein Sci, 6(10):2072–2083, 1997.
11. A. Sali, E. Shakhnovich, and M. Karplus. How does a protein fold ? Nature,
369:248–251, 1994.
12. M. Shemyakin, Y. Ovchinnikov, and V. Ivanov. Topochemical investigations of
peptide systems. Angew. Chem. Int. Ed. Engl., 8:492–499, 1968.
13. G. Trinquier and Y.-H. Sanejouand. New protein-like properties of cubic lattice
models. Phys. Rev. E, 59(1):942–946, 1999.
 Proteinlike Properties of Simple Models 153

14. D. Wales and J. Doye. Global optimization by basin-hopping and the lowest
energy structures of lennard-jones clusters containing up to 110 atoms. J. Phys.
Chem. A, 101:5111–5116, 1997.
List of Participants

1. Adams Dean, Iowa State University (USA),

dcadams@iastate.edu
2. Adebiyi Ezekiel, Universitaet Tuebingen (DE),
adebiyi@informatik.uni-tuebingen.de
3. Bernasconi Anna, Ist. Matematica Comput. CNR (IT),
bernasconi@imc.pi.cnr.it
4. Bertonati Claudia, ”La Sapienza ” University of Rome (IT),
claudia@medea.ccr.rm.cnr.it
5. Bertrand Hugues-Olivier, University of Rennes (FR),
hbertrand@msi-eu.com
6. Bock Mary Ellen, Purdue University (USA),
mbock@stat.purdue.edu
7. Bockmayr Alexander, University Henri Poincare (FR),
Alexander.Bockmayr@loria.fr
8. Bortoluzzi Stefania, University of Padova (IT),
stefibo@bio.unipd.it
9. Bosotti Roberta, Pharmacia Upjohn (IT),
Roberta.Bosotti@eu.pnu.com
10. Califano Andrea, IBM Yorktown (USA),
acal@us.ibm.com
11. Chiusano Maria Luisa, Second University of Naples (IT),
marilu@crisceb.area.na.cnr.it
12. Dill Ken, University of California, San Francisco (USA),
dill@maxwell.ucsf.edu
13. Fransson Linda, Royal Institute of Technology (Sweden),
lindaf@biochem.kth.se
14. Fuchs Bernhard, University of Bonn (DE),
fuchsb@cs.uni-bonn.de
15. Houstis Nicholas, Harvard/MIT (USA),
houstis@genome.wi.mit.edu
156 List of Participants

16. Jensen Lars, Juhl Technical Univ. of Denmark (DK),

ljj@cbs.dtu.dk
17. Kaderali Lars, Cologne University (DE),
kaderali@zpr.uni-koeln.de
18. Kann Maricel, University of Michigan (USA),
kann@ncbi.nlm.nih.gov
19. Kannan Natarajan, Indian Institute of Science (INDIA),
kannan@mbu.iisc.ernet.in
20. Kein Holger, Universität zu Köln (DE),
holger.klein@uni-koeln.de
21. Ketterle Christophe , Genset S.A. (FR),
Christophe.Ketterle@genset.fr
22. Khyvon Karine, City University, London (UK),
kyvon@soi.city.ac.uk
23. Kutluay Esin, Cornell Med School (USA),
esk2001@mail.med.cornell.edu
24. Galbi David, Galbi Research (USA),
david galbi@yahoo.com
25. Gaither Larry, Univ. of Missouri-Columbia (USA),
lagc4b@missouri.edu
26. Giordano Suzanne, SUNY Stony Brook (USA),
scgiorda@cedarcrest.edu
27. Goldstein Steve, Genetics Computer Group (USA),
steveg@gcg.com
28. Grimm Vera, University of Cologne (DE),
vera.grimm@uni-koeln.de
29. Guerra Concettina, University of Padova (IT),
guerra@dei.unipd.it
30. Istrail Sorin, Celera Genomics, Rockville, MD (USA),
sorin.istrail@celera.com
31. Lancia Giuseppe, University of Padova (IT),
lancia@dei.unipd.it
32. Lesk Arthur, University of Cambridge Clinical School (GB),
aml2@mrc-lmb.cam.ac.uk
33. Levitt Michael, Stanford University School of Medicine (USA),
michael.levitt@stanford.edu
34. Lonardi Stefano, University of Padova (IT),
stelo@artemide.dei.unipd.it
35. Lundegaard Claus, University of Copenhagen (DK),
clundegaard@strubix.dk
36. Macchiarulo Antonio, University of di Perugia (IT),
antonio@asia.chimfarm.unipg.it
37. Martin Oliver, University of Cologne (DE),
oliver.martin@uni-koeln.de
 List of Participants 157

38. Miller David, William Paterson University (USA),

damiller@pilot.njin.net
39. Moult John, University of Maryland Biotechnology Institute (USA),
moult@umbi.umd.edu
40. Nicolotti Nicola, University of Bari (IT),
nicolotti@farmchim.uniba.it
41. Pavesi Giulio, University of Milano (IT),
pavesi@disco.unimib.it
42. Pellegrini Marialuisa, University of Bari (IT),
pellegrini@farmchim.uniba.it
43. Quine John R., Florida State University (USA),
quine@zeno.math.fsu.edu
44. Rantanen Ville-Veikko, University of Turku (Finland),
vrantane@btk.utu.fi
45. Rossi Andrea, SISSA/ISAS (IT),
rossi@cm.sissa.it
46. Ruth Erich, University of Miami (USA),
eruth@cs.miami.ed
47. Schliep Alexander, University of Cologne (DE),
schliep@zpr.uni-koeln.de
48. Schoenhuth Alexander, Cologne University (USA),
schoenhuth@zpr.uni-koeln.de
49. Seno Flavio, University of Padova (IT),
seno@pd.infn.it
50. Shatsky Maxim, Tel-Aviv University (Israel),
maxshats@math.tau.ac.il
51. Shin Whanchul, Seoul National University (KR),
nswcshin@plaza.snu.ac.kr
52. Shlyakhova Marianna, Cornell University (USA),
mis2024@mail.med.cornell.edu
53. Sjunnesson Fredrik, Lunds University (Sweden),
fredrik.sjunnesson@thep.lu.se
54. Tiuryn Jerzy, Warsaw University (Poland),
tiuryn@mimuw.edu.pl
55. Zimmermann. Olav, University of Cologne (DE),
o.zimmermann@uni-koeln.de
56. Wolfson Haim J. , Tel Aviv University (Israel),
wolfson@post.tau.ac.il