Advanced

Data Mining
Technologies in
Bioinformatics
Hui-Huang Hsu
Tamkang University, Taipei, Taiwan
Hershey London Melbourne Singapore
IDEA GROUP PUBLISHING
Acquisitions Editor: Michelle Potter
Development Editor: Kristin Roth
Senior Managing Editor: Amanda Appicello
Managing Editor: Jennifer Neidig
Copy Editor: Nicole Dean
Typesetter: Marko Primorac
Cover Design: Lisa Tosheff
Printed at: Integrated Book Technology
Published in the United States of America by
Idea Group Publishing (an imprint of Idea Group Inc.)
701 E. Chocolate Avenue
Hershey PA 17033
Tel: 717-533-8845
Fax: 717-533-8661
E-mail: [email protected]
Web site: http://www.idea-group.com
and in the United Kingdom by
Idea Group Publishing (an imprint of Idea Group Inc.)
3 Henrietta Street
Covent Garden
London WC2E 8LU
Tel: 44 20 7240 0856
Fax: 44 20 7379 0609
Web site: http://www.eurospanonline.com
Copyright 2006 by Idea Group Inc. All rights reserved. No part of this book may be repro-
duced, stored or distributed in any form or by any means, electronic or mechanical, including
photocopying, without written permission from the publisher.
Product or company names used in this book are for identification purposes only. Inclusion of the
names of the products or companies does not indicate a claim of ownership by IGI of the
trademark or registered trademark.
Library of Congress Cataloging-in-Publication Data
Advanced data mining technologies in bioinformatics / Hui-Hwang Hsu, editor.
p. cm.
Summary: "This book covers research topics of data mining on bioinformatics presenting the
basics and problems of bioinformatics and applications of data mining technologies pertaining to the
field"--Provided by publisher.
Includes bibliographical references and index.
ISBN 1-59140-863-6 (hardcover) -- ISBN 1-59140-864-4 (softcover) -- ISBN 1-59140-865-2
(ebook)
1. Bioinformatics. 2. Data mining. I. Hsu, Hui-Huang, 1965-
QH324.2.A38 2006
572.8'0285--dc22
2006003556
British Cataloguing in Publication Data
A Cataloguing in Publication record for this book is available from the British Library.
All work contributed to this book is new, previously-unpublished material. The views expressed in
this book are those of the authors, but not necessarily of the publisher.
Advanced Data Mining
Technologies in Bioinformatics
Table of Contents
Preface .......................................................................................................................... vi
Chapter I
Introduction to Data Mining in Bioinformatics ............................................................. 1
Hui-Huang Hsu, Tamkang University, Taipei, Taiwan
Chapter II
Hierarchical Profiling, Scoring and Applications in Bioinformatics ........................ 13
Li Liao, University of Delaware, USA
Chapter III
Combinatorial Fusion Analysis: Methods and Practices of Combining Multiple
Scoring Systems ......................................................................................................... 32
D. Frank Hsu, Fordham University, USA
Yun-Sheng Chung, National Tsing Hua University, Taiwan
Bruce S. Kristal, Burke Medical Research Institute and weill Medical College of
Cornell University, USA
Chapter IV
DNA Sequence Visualization...................................................................................... 63
Hsuan T. Chang, National Yunlin University of Science and Technology,
Taiwan
Chapter V
Proteomics with Mass Spectrometry .......................................................................... 85
Simon Lin, Northwestern University, USA
Salvatore Mungal, Duke University Medical Center , USA
Richard Haney, Duke University Medical Center , USA
Edward F. Patz, Jr., Duke University Medical Center , USA
Patrick McConnell, Duke University Medical Center, USA
Chapter VI
Efficient and Robust Analysis of Large Phylogenetic Datasets ............................... 104
Sven Rahmann, Bielefeld University, Germany
Tobias Mller, University of Wrzburg, Germany
Thomas Dandekar, University of Wrzburg, Germany
Matthias Wolf, University of Wrzburg, Germany
Chapter VII
Algorithmic Aspects of Protein Threading .............................................................. 118
Tatsuya Akutsu, Kyoto University, Japan
Chapter VIII
Pattern Differentiations and Formulations for Heterogeneous Genomic Data
through Hybrid Approaches ...................................................................................... 136
Arpad Kelemen, The State University of New York at Buffalo, USA &
Niagra University, USA
Yulan Liang, The State University of New York at Buffalo, USA
Chapter IX
Parameterless Clustering Techniques for Gene Expression Analysis ................... 155
Vincent S. Tseng, National Cheng Kung University, Taiwan
Ching-Pin Kao, National Cheng Kung University, Taiwan
Chapter X
Joint Discriminatory Gene Selection for Molecular Classification of Cancer ....... 174
Junying Zhang, Xidian University, China
Chapter XI
A Haplotype Analysis System for Genes Discovery of Common Diseases .............. 214
Takashi Kido, HuBit Genomix, Inc., Japan
Chapter XII
A Bayesian Framework for Improving Clustering Accuracy of Protein
Sequences Based on Association Rules ................................................................... 231
Peng-Yeng Yin, National Chi Nan University, Taiwan
Shyong-Jian Shyu, Ming Chuan University, Taiwan
Guan-Shieng Huang, National Chi Nan University, Taiwan
Shuang-Te Liao, Ming Chuan University, Taiwan
vii
Chapter XIII
In Silico Recognition of Protein-Protein Interactions: Theory and Applications .... 248
Byung-Hoon Park, Oak Ridge National Laboratory, USA
Phuongan Dam, University of Georgia, USA
Chongle Pan, University of Tennessee, USA
Ying Xu, University of Georgia, USA
Al Geist, Oak Ridge National Laboratory, USA
Grant Heffelfinger, Sandia National Laboratories, USA
Nagiza F. Samatova, Oak Ridge National Laboratory, USA
Chapter XIV
Differntial Association Rules: Understanding Annotations in Protein Interaction
Networks ................................................................................................................... 269
Christopher Besemann, North Dakota State University, USA
Anne Denton, North Dakota State University, USA
Ajay Yekkirala, North Dakota State University, USA
Ron Hutchison, The Richard Stockton College of New Jersey, USA
Marc Anderson, North Dakota State University, USA
Chapter XV
Mining BioLiterature: Toward Automatic Annotation of Genes and Proteins ........ 283
Francisco M. Couto, Universidade de Lisboa, Portugal
Mrio J. Silva, Universidade de Lisboa, Portugal
Chapter XVI
Comparative Genome Annotation Systems ............................................................... 296
Kwangmin Choi, Indiana University, USA
Sun Kim, Indiana University, USA
About the Authors ..................................................................................................... 314
Index ........................................................................................................................ 324
vi
Bioinformatics is the science of managing, analyzing, extracting, and interpreting infor-
mation from biological sequences and molecules. It has been an active research area
since late 1980s. After the human genome project was completed in April 2003, this area
has drawn even more attention. With more genome sequencing projects undertaken,
data in the field such as DNA sequences, protein sequences, and protein structures are
exponentially growing. Facing this huge amount of data, the biologist cannot simply
use the traditional techniques in biology to analyze the data. In order to understand the
mystery of life, instead, information technologies are needed.
There are a lot of online databases and analysis tools for bioinformatics on the World
Wide Web. Information technologies have made a tremendous contribution to the field.
The database technology helps collect and annotate the data. And the retrieval of the
data is made easy. The networking and Web technology facilitates data and information
sharing and distribution. The visualization technology allows us to investigate the
RNA and protein structures more easily. As for the analysis of the data, online se-
quence alignment tools are quite mature and ready for use all the time. But to conquer
more complicated tasks such as microarray data analysis, protein-protein interaction,
gene mapping, biochemical pathways, and systems biology, sophisticated techniques
are needed.
Data mining is defined as uncovering meaningful, previously unknown information
from a mass of data. It is an emerging field since mid 1990s boosted by the flood of data
on the Internet. It combines traditional databases, statistics, and machine learning
technologies under the same goal. Computer algorithms are also important in speeding
up the process while dealing with a large amount of data. State-of-the-art techniques in
data mining are, for example, information retrieval, data warehousing, Bayesian learn-
ing, hidden Markov model, neural networks, fuzzy logic, genetic algorithms, and sup-
port vector machines. Generally, data mining techniques deal with three major prob-
lems, i.e., classification, clustering, and association. In analyzing biological data, these
three kinds of problems can be seen quite often. The technologies in data mining have
been applied to bioinformatics research in the past few years with quite a success. But
Preface
vii
more research in this field is necessary since a lot of tasks are still undergoing. Further-
more, while tremendous progress has been made over the years, many of the fundamen-
tal problems in bioinformatics are still open. Data mining will play a fundamental role in
understanding the emerging problems in genomics and proteomics. This book wishes
to cover advanced data mining technologies in solving such problems.
The audiences of this book are senior or graduate students majoring in computer sci-
ence, computer engineering, or management information system (MIS) with interests in
data mining and applications to bioinformatics. Professional instructors and research-
ers will also find that the book is very helpful. Readers can benefit from this book in
understanding basics and problems of bioinformatics, as well as the applications of
data mining technologies in tackling the problems and the essential research topics in
the field.
The uniqueness of this book is that it covers important bioinformatics research topics
with applications of data mining technologies on them. It includes a few advanced data
mining technologies. Actually, in order to solve bioinformatics problems, there is plenty
of room for improvement in data mining technologies. This book covers basic concepts
of data mining and technologies from data preprocessing like hierarchical profiling,
information fusion, sequence visualization, and data management, to data mining algo-
rithms for a variety of bioinformatics problems like phylogenetics, protein threading,
gene discovery, protein sequence clustering, protein-protein interaction, protein inter-
action networks, and gene annotations. The summaries of all chapters of the book are
as follows.
Chapter I introduces the concept and the process of data mining, plus its relationship
with bioinformatics. Tasks and techniques of data mining are also presented. At the
end, selected bioinformatics problems related to data mining are discussed. It provides
an overview on data mining in bioinformatics.
Chapter II reviews the recent developments related to hierarchical profiling where the
attributes are not independent, but rather are correlated in a hierarchy. It discusses in
detail several clustering and classification methods where hierarchical correlations are
tackled with effective and efficient ways, by incorporation of domain specific knowl-
edge. Relations to other statistical learning methods and more potential applications
are also discussed.
Chapter III presents a method, called Combinatorial Fusion Analysis (CFA), for analyz-
ing combination and fusion of multiple scoring systems. Both rank combination and
score combination are explored as to their combinatorial complexity and computational
efficiency. Information derived from the scoring characteristics of each scoring system
is used to perform system selection and to decide method combination. Various appli-
cations of the framework are illustrated using examples in information retrieval and
biomedical informatics.
Chapter IV introduces various visualization (i.e., graphical representation) schemes of
symbolic DNA sequences, which are basically represented by character strings in con-
ventional sequence databases. Further potential applications based on the visualized
sequences are also discussed. By understanding the visualization process, the re-
searchers will be able to analyze DNA sequences by designing signal processing algo-
rithms for specific purposes such as sequence alignment, feature extraction, and se-
quence clustering.
viii
Chapter V provides a rudimentary review of the field of proteomics as it applies to mass
spectrometry, data handling and analysis. It points out the potential significance of the
field suggesting that the study of nuclei acids has its limitations and that the progres-
sive field of proteomics with spectrometry in tandem with transcription studies could
potentially elucidate the link between RNA transcription and concomitant protein ex-
pression. Furthermore, the chapter describes the fundamentals of proteomics with mass
spectrometry and expounds the methodology necessary to manage the vast amounts
of data generated in order to facilitate statistical analysis.
Chapter VI considers the prominent problem of reconstructing the basal phylogenetic
tree topology when several subclades have already been identified or are well-known
by other means, such as morphological characteristics. Whereas most available tools
attempt to estimate a fully resolved tree from scratch, the profile neighbor-joining (PNJ)
method focuses directly on the mentioned problem and has proven a robust and effi-
cient method for large-scale data sets, especially when used in an iterative way. The
chapter also describes an implementation of this idea, the ProfDist software package,
and applies the method to estimate the phylogeny of the eukaryotes.
Chapter VII provides an overview of computational problems and techniques for pro-
tein threading. Protein threading can be modeled as an optimization problem. This
chapter explains the ideas employed in various algorithms developed for finding opti-
mal or near optimal solutions. It also gives brief explanations of related problems:
protein threading with constraints, comparison of RNA secondary structures, and pro-
tein structure alignment.
Chapter VIII introduces hybrid methods to tackle the major challenges of power and
reproducibility of the dynamic differential gene temporal patterns. Hybrid clustering
methods are developed based on resulting profiles from several clustering methods.
The developed hybrid analysis is demonstrated through an application to a time course
gene expression data from interferon--1a treated multiple sclerosis patients. The re-
sulting integrated-condensed clusters and overrepresented gene lists demonstrate that
the hybrid methods can successfully be applied.
Chapter IX discusses the issue of parameterless clustering technique for gene expres-
sion analysis. Two novel, parameterless and efficient clustering methods that fit for
analysis of gene expression data are introduced. The unique feature of the methods is
that they incorporate the validation techniques into the clustering process so that high
quality results can be obtained. Through experimental evaluation, these methods are
shown to outperform other clustering methods greatly in terms of clustering quality,
efficiency, and automation on both of synthetic and real data sets.
Chapter X introduces gene selection approaches in microarray data analysis for two
purposes: cancer classification and tissue heterogeneity correction. In the first part,
jointly discriminatory genes which are most responsible to classification of tissue
samples for diagnosis are searched for. In the second part, tissue heterogeneity correc-
tion techniques are studied. Also, non-negative matrix factorization (NMF) is employed
to computationally decompose molecular signatures based on the fact that the expres-
sion values in microarray profiling are non-negative. Throughout the chapter, a real
world gene expression profile data was used for experiments.
Chapter XI introduces computational methods for detecting complex disease loci with
haplotype analysis. It argues that the haplotype analysis, which plays a major role in
ix
the study of population genetics, can be computationally modeled and systematically
implemented as a means for detecting causative genes of complex diseases. The expla-
nation of the system and some real examples of the haplotype analysis not only provide
researchers with better understanding of current theory and practice of genetic asso-
ciation studies, but also present a computational perspective on the gene discovery
research for the common diseases.
Chapter XII presents a Bayesian framework for improving clustering accuracy of pro-
tein sequences based on association rules. Most of the existing protein-clustering
algorithms compute the similarity between proteins based on one-to-one pairwise se-
quence alignment instead of multiple sequences alignment. Furthermore, the traditional
clustering methods are ad-hoc and the resulting clustering often converges to local
optima. The experimental results manifest that the introduced framework can signifi-
cantly improve the performance of traditional clustering methods.
Chapter XIII reviews high-throughput experimental methods for identification of pro-
tein-protein interactions, existing databases of protein-protein interactions, computa-
tional approaches to predicting protein-protein interactions at both residue and protein
levels, various statistical and machine learning techniques to model protein-protein
interactions, and applications of protein-protein interactions in predicting protein func-
tions. Intrinsic drawbacks of the existing approaches and future research directions are
also discussed.
Chapter XIV discusses the use of differential association rules to study the annota-
tions of proteins in one or more interaction networks. Using this technique, the differ-
ences in the annotations of interacting proteins in a network can be found. The concept
is extended to compare annotations of interacting proteins across different definitions
of interaction networks. Both cases reveal instances of rules that explain known and
unknown characteristics of the network(s). By taking advantage of such data mining
techniques, a large number of interesting patterns can be effectively explored that
otherwise would not be.
Chapter XV introduces the use of Text Mining in scientific literature for biological
research, with a special focus on automatic gene and protein annotation. The chapter
describes the main approaches adopted and analyzes systems that have been devel-
oped for automatically annotating genes or proteins. To illustrate how text-mining tools
fit in biological databases curation processes, the chapter also presents a tool that
assists protein annotation. At last, it presents the main open problems in using text-
mining tools for automatic annotation of genes and proteins.
Chapter XVI surveys systems that can be used for annotating genomes by comparing
multiple genomes and discusses important issues in designing genome comparison
systems such as extensibility, scalability, reconfigurability, flexibility, usability, and data
mining functionality. Further issues in developing genome comparison systems where
users can perform genome comparison flexibly on the sequence analysis level are also
discussed.
x
I would like to thank all the authors for contributing their thoughts and research results
to this book. Without their contributions, this book would not be possible. They are all
experts of data mining in bioinformatics. I did learn a lot from them. Further thanks go to
the authors who also served as a reviewer for other chapters. Their comments and
suggestions are really valuable in improving the manuscripts.
Next, I would like to acknowledge the efforts of the staffs at Idea Group Inc. Special
thanks go to Renee Davies for inviting me to initiate this project, Dr. Mehdi Khosrow-
Pour for reviewing and approving the project, Jan Travers for supervising the produc-
tion of the book, and Kristin Roth for coordinating the production of the book. Kristin
helped me a lot in answering authors questions and providing me needed documents
and solutions. She made the whole process much easier for me.
I am also grateful to two of my colleagues at Tamkang University. Prof. Timothy K. Shih
is a senior professor of computer science and information engineering and a long time
friend of mine. Thanks to his encouragement on the initiation of this project. I appreci-
ate his help and friendship throughout the years. Prof. Chien-Chung Cheng of chemis-
try and life science has been helping me understand biological problems. Cooperation
with him in the past two years has been a great pleasure.
Finally, I would like to express my appreciations to my wife Maggie. Without her sup-
port and patience, this work would not have been done smoothly.
Hui-Huang Hsu, PhD
Tamsui, Taipei, Taiwan
December 2005
Acknowledgments
Introduction to Data Mining in Bioinformatics 1
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Chapter I
Introduction to
Data Mining in
Bioinformatics
Hui-Huang Hsu, Tamkang University, Taipei, Taiwan
Abstract
Bioinformatics uses information technologies to facilitate the discovery of new
knowledge in molecular biology. Among the information technologies, data mining is
the core. This chapter first introduces the concept and the process of data mining, plus
its relationship with bioinformatics. Tasks and techniques of data mining are then
presented. At the end, selected bioinformatics problems related to data mining are
discussed. Data mining aims at uncovering knowledge from a large amount of data. In
molecular biology, advanced biotechnologies enable the generation of new data in a
much faster pace. Data mining can assist the biologist in finding new knowledge from
piles of biological data at the molecular level. This chapter provides an overview on
the topic.
2 Hsu
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Introduction
Progress of information technologies has made the storage and distribution of data much
easier in the past two decades. Huge amounts of data have been accumulated at a very
fast pace. However, pure data are sometimes not-that-useful and meaningful because
what people want is the knowledge/information hidden in the data. Knowledge/informa-
tion can be seen as the patterns or characteristics of the data. It is much more valuable
than data. Thus, a new technology field has emerged in the mid 1990s to deal with the
discovery of knowledge/information from data. It is called knowledge discovery in
databases (KDD) or simply data mining (DM) (Chen et al., 1996; Fayyad et al., 1996).
Although knowledge and information can sometimes be distinguished, we will treat them
as the same term in this chapter.
Data pile up like a mountain. But most of them are not that useful, just like earths and rocks
in a mountain. The valuables are metals like gold, iron, or diamond. Just like the miner
wants to dig out the valuables from the earth and rock, the data miner uncovers useful
knowledge/information by processing a large amount of data. Formal definitions of KDD
and DM have been given in different ways. Here are three examples: Knowledge
discovery in databases is the nontrivial process of identifying valid, novel, potentially
useful, and ultimately understandable patterns in data (Fayyad, 1996, p. 20). And, Data
mining is the process of extracting valid, previously unknown, comprehensible, and
actionable information from large databases and using it to make crucial business
decisions (Simoudis, 1996, p. 26). Or, to be simpler, data mining is finding hidden
information in a database (Dunhum, 2003, p. 3). Uncovering hidden information is the
goal of data mining. But, the uncovered information must be:
1. New: Common sense or known facts are not what is searched for.
2. Correct: Inappropriate selection or representation of data will lead to incorrect
results. The mined information needs to be carefully verified by domain experts.
3. Meaningful: The mined information should mean something and can be easily
understood.
4. Applicable: The mined information should be able to be utilized in a certain problem
domain.
Also, in Simoudis (1996), crucial business decision making is emphasized. That is because
the cost of data mining is high and was first applied in business problems, e.g., customer
relationship management, personalized advertising, and credit card fraud detection.
There is actually a slight difference between KDD and DM. Data mining is the algorithm/
method used to find information from the data. Meanwhile, KDD is the whole process
including data collection, data preprocessing, data mining, and information interpreta-
tion. So DM is the core of KDD. However, data preprocessing and information interpre-
tation are also very important. Without proper preprocessing, the quality of data might
be too bad to find meaningful information. Also, without correct interpretation, the mined
information might be mistakenly used. Thus, the cooperation between data mining
Introduction to Data Mining in Bioinformatics 3
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
technicians and problem domain experts is very much needed when a data mining task
is performed. In real life, the terms KDD and DM are used interchangeably. Data mining
is more frequently referred to because it is shorter and easier to comprehend.
With the goal of uncovering information from data, data mining uses technologies from
different computer and information science fields. The three major ones are databases,
machine learning, and statistics. Database technology manages data for convenient
selection of the data. Machine learning technology learns information or patterns from
data in an automatic way, and statistics finds characteristics or statistical parameters of
the data. Besides, information retrieval deals with selecting related documents by
keywords. It is more related to text processing and mining. Also, algorithms and parallel
processing are essential in data mining for a large amount of data. They speed up the
search for useful information. But they will not be discussed in this chapter. Interested
readers are referred to books of those two topics in computer science.
With the advent of high-throughput biotechnologies, biological data like DNA, RNA,
and protein data are generated faster than ever. Huge amounts of data are being produced
and collected. The biologist needs information technology to help manage and analyze
such large and complex data sets. Database and Web technologies are used to build
plenty of online data banks for data storage and sharing. Most of the data collected have
been put on the World Wide Web and can be shared and accessed online. For example,
GenBank (http://www.ncbi.nlm.nih.gov/Genbank/), EMBL Nucleotide Sequence Data-
base (http://www.ebi.ac.uk/embl/), and Protein Data Bank (PDB) (http://www.rcsb.org/
pdb/). To know the updated numbers of complete genomes, nucleotides, and protein
coding sequences, the reader can check the Genome Reviews of EMBL-EBI (http://
www.ebi.ac.uk/GenomeReviews/stats/). The reader is also referred to Protein Data Bank
for the number of known protein structures. As for the analysis of the data, data mining
technologies can be utilized. The mystery of life hidden in the biological data might be
decoded much faster and more accurately with the data mining technologies. In the
following, what kind of problems data mining can deal with, both in general and
specifically in bioinformatics, are introduced. Some classical techniques of data mining
are also presented.
Data Mining Process
This section gives an overview of the whole process of data mining. As emphasized in
the introduction section, data mining methods is the core of the process, but other phases
of the process are also very important. Each phase of data mining needs interaction with
human experts because the objective is usually inexact. This might be time-consuming,
but will ensure that the mined information is accurate and useful. The process consists
of four phases: data collection, data preprocessing, data mining, and information
interpretation. Throughout the process, visualization also plays an important role. Next,
further explanations of each phase and visualization issues are presented.
4 Hsu
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Data collection: Data is the raw material for mining information. Only with proper
data can useful information be mined. Domain experts are needed for the selection
of data for certain problems. So after raw data of a certain problem are collected,
those fields/attributes directly related to the target information are selected.
Data preprocessing: After data are carefully selected, preprocessing is needed.
That would include removal of erroneous data and transformation of data. Errone-
ous data, or so called outliers in statistics, should be detected and deleted.
Otherwise, they would result in incorrect results. Also, transformation of data to
a suitable representation is needed. For example, data can be encoded into a vector
representation. Also, data can be transformed from a high dimensional space to a
lower dimension to find more important features and to reduce the data mining effort
at the next phase.
Data mining: This phase usually involves building a model for the data. Different
algorithms and techniques can be used here. However, for a certain task, suitable
techniques should be chosen. The characteristics of different tasks and the state-
of-the-art techniques for those tasks will be presented in the next two sections.
Information interpretation: First, the mined information needs to be interpreted by
human experts. The interpreted results are then evaluated by their novelty,
correctness, comprehensibility, and usefulness. Only the information passing
through this filtering process can be used in real applications.
Visualization: Visualization technology demonstrates the data or the mined
information in graphics to let the user (miner) more easily capture the patterns
residing in the data or information. It can be used throughout the data mining
process incorporating with data preprocessing, data mining, and information
interpretation. This is very important for the presentation of data and information
to the user because it enhances the comprehensibility and thus the usefulness.
Readers interested in information visualization are referred to the book edited by
Fayyad et al. (2002).
It should be noticed that data mining is costly. Therefore, besides going through the
above mentioned steps with care, an estimate of the data mining project in advance is
necessary. Although certain degree of uncertainty is always involved in a data mining
project, with the cooperation of the data mining expert and the domain expert, better
Figure 1. KDD process

Data Collection

Attributes
Selection
Data Preprocessing

Outlier Removal &
Transformation
Data Mining

Algorithms &
Model
Information
Interpretation
Expert Evaluation
& Applications
Visualization/Graphical Representation
Introduction to Data Mining in Bioinformatics 5
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
understanding of the problem with a clearer objective can be achieved. This increases
the possibility of success of the data mining project.
Data Mining Tasks
Typical tasks of data mining are discussed in this section. The list might not be complete.
We can say that any form of interesting pattern or information that can be discovered
from the data can formulate a specific data mining task. But, of course, some of them might
overlap with others. In the following, four major data mining tasks, namely, classification,
clustering, association, and summarization, are presented with examples. Text mining
handles textual data. It is not a unique task comparable with the four tasks. However, text
mining is included and discussed in this section because it is essential in mining
information from the huge biological/medical literature.
Classification
Classification decides the class/group for each data sample. There should be at least two
classes and the classes are predefined. The input of a classification model is the attributes
of a data sample and the output is the class that data sample belongs to. In machine
learning, it takes supervised learning to build such a model. That is, a set of data with
known classes (training data) is needed to estimate the parameters of the classification
model. After the parameters are set, the model can be used to automatically classify any
new data samples. For example, iris species can be classified based on their measurement.
The input attributes are petal length, petal width, sepal length, and sepal width. The
output classes are setosa, versicolor, and verginica.
Prediction is also a type of classification. To predict if it will rain tomorrow, it is like
classifying the weather into two classes: Raining and Not Raining. Furthermore, to
predict the value of a stock would be a classification of many classes within a certain
range (possible values). In statistics, regression is a way to build a predictive model with
a polynomial. Prediction is often applied to time series analysis. From the previously
known patterns of change, future trends can be decided.
Another kind of classification is deviation analysis. It is for the detection of significant
change. So there are two classes for such analysis: Significant change and Insignificant
change. Credit card fraud detection is a case of deviation detection. When the buying
behavior of a customer is suddenly changed, it might be that the credit card is stolen and
needs to be detected right away to avoid possible losses.
Clustering
Clustering is to group similar data into a finite set of separate clusters/categories. This
task is also referred to as segmentation. In machine learning, it requires unsupervised
6 Hsu
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
learning. That is, the number of clusters and the categories are not known in advance.
A clustering tool simply measures the similarity of the data based on their attribute values
and put similar data into the same cluster. For example, Amazon.com groups its customers
into different clusters according to the books they buy. Customers in the same cluster
should have the same taste in book reading. So the most bought books by the customers
in one cluster can be recommended to the customers in the same cluster, who have not
bought the books.
It is sometimes confusing between clustering and classification. Both of them put data/
examples into different groups. The difference is that in classification ,the groups are
predefined and the task is to decide which group a new data sample should belong to.
In clustering the types of groups and even the number of groups are not known and the
task is to find the best way to segment all the data.
The major problem of clustering is the decision of the number of clusters. For some
methods, the number of clusters needs to be specified first. For example, to use the K-
Means algorithm, the user should input the number of clusters first. The centers of the
clusters are chosen arbitrarily. Then, the data iteratively move between clusters until
they converge. If the user is not satisfied with the results, another clusters number is then
tried. So this kind of method is a trial-and-error process.
Another solution to the decision of the number of clusters is to use hierarchical
clustering. There are two approaches for this: the top-down approach and the bottom-
up approach. For the top-down approach, the data are separated into different clusters
based on different criteria of the similarity measurement. At first, the whole data belong
to a big cluster. At last, each data sample is a single cluster. There will be different levels
of clustering in between the two. For the bottom-up approach, the data are group into
different clusters from clusters of single data sample until all the data are of one cluster.
Association
Another task of data mining is to search for a set of data within which a subset is
dependent on the rest of the set. This task is also referred to as link analysis or affinity
analysis. An association rule can be written as A B where both A and B are a data set.
If such an association rule is identified, it means that when A is seen, B will also be seen
with a high probability. For example, in a supermarket transaction database, an associa-
tion rule Diapers Beer was discovered from the transactions on Friday night. That
means people who bought diapers on Friday nights also bought beer. This kind of
relationship between products is very helpful to marketing decision making. And the
process is called market basket analysis.
An association rule A B can be identified when both the support and the confidence
of the rule are larger than respective thresholds. The support of the association rule is
the ratio of the number of transactions containing both A and B over the total number
of transactions in the database. The confidence of the association rule is the proportion
of the number of transactions containing both A and B over the total number of
Introduction to Data Mining in Bioinformatics 7
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
transactions containing A. The two thresholds are actually difficult to decide and need
domain expertise. The size of the transaction database and the number of product items
will affect the setting of proper thresholds. Another important issue of association rule
mining is how to speed up the identification procedure of association rules. Different
algorithms have been developed (Chen et al., 1996).
Summarization
A simple description of a large data set is called its summarization or characterization.
It is desirable that representative information of a data set can be obtained, so that we
can easily have a general view of the data set. The representative information can be a
subset or general properties/characteristics of the data. For example, in a class of
students, the average height 176.5 cm or a student with the height can be used to
characterize the students in the class. With a diversity of data, it is somewhat difficult
to find good representative information. Classical statistics can be applied to find some
useful parameters to represent the data. However, summarization is not so popular or
frequently used as classification, clustering, and association.
Text Mining
When the data to be mined are text instead of numerical data, the task is called text mining.
It is originated from information retrieval (IR) of the library science. Keywords are used
to find related documents from a document database. It is based on the similarity
measures between the keywords and the documents. These similarity measures can also
be applied to the comparison of a document with all other documents. In the Internet era,
IR technique is also used for search engine to find related Web pages. It is also referred
to as Web content mining.
For more advanced applications of text mining; classification, clustering, and association
techniques are utilized. The relationship between documents or terms within the docu-
ments can be investigated via these techniques. Natural language processing (NLP)
is also involved for semantic understanding of neighboring words (phrases or sen-
tences). Online question-answering systems have been developed based on natural
language processing and text mining. If a user asks a question in natural language, the
system can search a lot of Web pages to find possible answers to it. Text mining is also
important for bioinformatics. Plenty of publications in biology have been accumulated
for more than a century. It is desired that new research findings can be combined or cross-
referenced with known facts in the literature. However, in the biology literature,
different terms in different documents might refer to the same thing. Also, the amount
of biology literature is really huge. To overcome these problems, new technologies need
to be developed and utilized. In all, text mining can facilitate the re-examination of the
biology literature to link the facts that were already known.
8 Hsu
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Data Mining Techniques
Many concepts and techniques are useful for the same goal of mining hidden information
from data. Among them algorithms, databases, statistics, machine learning, and informa-
tion retrieval are essential. Algorithms and parallel processing techniques are mainly to
accelerate the mining process. Information retrieval techniques can be applied to text
mining and is introduced in the subsection of text mining. Here, we will concentrate on
databases, statistics, and machine learning.
Databases and Data Warehousing
Database techniques arrange data in a structured form. This facilitates the retrieval of a
subset of the data according to the user query. To organize data in a structured form
definitely helps the data mining process. But in data mining tasks, the data are usually
in semi-structured or unstructured forms. Also, data mining is aimed at uncovering
information like customers of a certain background usually have a good credit, not like
customers with last name Bush. The goals of databases and data mining are quite
different. But the well-developed database technologies are very useful and efficient in
storage and preprocessing of the collected data.
Data warehousing technologies were developed to accommodate the operational data
of a company. For a large amount of data from heterogeneous databases, a data
warehouse is a repository of operational databases and is used to extract needed
information. It is said to contain informational data comparing to the operational data in
the database. The information is a combination of historical data and can be used to
support business decision making.
Statistics
Statistics has sound theoretical ground developed since the 16
th
century. It estimates
various parameters to provide a general view of the data. It is essential in characterizing
the data. Also, the parameters can be helpful to decide if the discovered information is
significant. In the following, we introduce some basic concepts of statistics that are
related to data mining (Kantardzic, 2003).
To characterize or summarize a set of data, there are some parameters that the reader
should be familiar with. They are mean, median, mode, variance, and standard deviation.
Also, a box plot is a graphical representation of the distribution of the data. The range
of the data is divided into four parts (quartiles). Each part has the same number of data.
Data outside a certain range from the median can then be viewed as outliers.
Next, we would like to discuss the Bayesian classifier. It is based on the well-known
Bayes theorem P(Y|X) = [P(X|Y) P(Y)]/ P(X) where P(Y) is the prior probability and
P(Y|X) is the posterior probability. The classification of a new data sample can be decided
Introduction to Data Mining in Bioinformatics 9
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
by estimating the probabilities of the classes (Y
i
) given the new data sample (X),
respectively. The class with the highest probability is the predicted class of the new data
sample.
Regression is to build a mathematical model for some known temporal data and use the
model to predict the upcoming values. This task would be very difficult because the
trend of the data is usually nonlinear and very complicated. Many parameter estimates
are involved if the model is nonlinear. Therefore, to simplify the problem, a linear model
is usually used. A linear model is to use a straight line to estimate the trend of the data.
This is called linear regression. For a set of data with a nonlinear nature, it can be
assumed that the data trend is piecewise linear. That is, in a small period, the data trend
is about a straight line.
Although statistical theories are widely used in data mining, the philosophies of data
mining and statistics are different. In statistics, hypotheses are usually made first and
they are proved to be valid or invalid by applying the theories on the data. On the
contrary, data mining does not exactly know what it is looking for.
Machine Learning
Machine learning is a long-developed field in artificial intelligence (AI). It focuses on
automatic learning from a data set. A suitable model with many parameters is built first
for a certain domain problem and an error measure is defined. A learning (training)
procedure is then used to adjust the parameters according to the predefined error
measure. The purpose is to fit the data into the model. There are different theories for the
learning procedure, including gradient decent, expectation maximization (EM) algo-
rithms, simulated annealing, and evolutionary algorithms. The learning procedure is
repeated until the error measure reaches zero or is minimized. After the learning procedure
is completed with the training data, the parameters are set and kept unchanged and the
model can be used to predict or classify new data samples.
Different learning schemes have been developed and discussed in the machine learning
literature. Important issues include the learning speed, the guarantee of convergence,
and how the data can be learned incrementally. There are two categories of learning
schemes: (1) supervised learning and (2) unsupervised learning. Supervised learning
learns the data with an answer. Meaning, the parameters are modified according to the
difference of the real output and the desired output (the expected answer). The classi-
fication problem falls into this category. On the other hand, unsupervised learning learns
without any knowledge of the outcome. Clustering belongs to this category. It finds data
with similar attributes and put them in the same cluster.
Various models like neural networks (NN), decision trees (DT), genetic algorithms
(GA), fuzzy systems, and support vector machines (SVM) have proved very useful in
classification and clustering problems. But machine learning techniques usually handles
relatively small data sets because the learning procedure is normally very time-consum-
ing. To apply the techniques to data mining tasks, the problem with handling large data
sets must be overcome.
10 Hsu
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Bioinformatics Problems
Related to Data Mining
In this section, selected bioinformatics problems that are suitable for using data mining
technologies are introduced. Various bioinformatics problems and their trends and
possible solutions via data mining technologies are presented and discussed in the rest
of the book. Here, only a few typical problems are chosen to illustrate the relationship
between bioinformatics and data mining technologies. For an introduction to all kinds
of bioinformatics problems, the reader is referred to Lesk (2002).
Protein Structure Prediction
With so many known proteins, the functions of most proteins are still waiting for further
investigation. A proteins function depends upon its structure. If the structure of a
protein is known, it would be easier for the biologist to infer the function of the protein.
However, it is still costly to decide the structure of a protein via biotechnologies. On the
contrary, protein sequences are relatively easy to obtain. Therefore, it is desirable that
a proteins structure can be decided from its sequence through computational methods.
A protein sequence is called the primary structure of the protein. Hydrogen bonding of
the molecules results in certain substructures called the secondary structure. Interac-
tions between secondary structures assemble them into the tertiary structure. If a
protein is composed by more than one unit, it is a quaternary structure. Researchers had
produced pretty good results in predicting secondary structures from a protein se-
quence. However, it is still difficult and imprecise to determine the tertiary and quaternary
structures.
Two major approaches have been taken for such structure prediction. The first one is to
use machine learning algorithms to learn the mapping function from the sequence
segments to the secondary structures. With the learned mapping function, the second-
ary structures can be induced from the sequence segments of a new protein. With the
knowledge of secondary structures, the tertiary structure of the protein can be further
inferred. This kind of problem is exactly a prediction/classification problem in data
mining. The second approach applies the physical laws of the forces between different
molecules of the protein. Angle constraints of the chemical bonds are formulated and an
optimization process can be used to find the optimal solution of the angles and thus the
tertiary structure can be decided.
Gene Finding
A gene is a possible protein-coding region. Genes only compose a small part of the whole
genome. With a complete genome, we would like to know where the genes are located.
It is to identify the beginnings and endings of genes. It is also called gene annotation.
There are also two approaches for this problem. The first one is to use machine learning
Introduction to Data Mining in Bioinformatics 11
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
techniques to learn from the patterns of subsequences of known genes and then predict/
classify the genome into genes and non-genes. The second approach, called ab initio
methods, concentrates more on the properties of the DNA sequences. Machine learning
can also be applied in this approach. The two approaches can be combined to produce
better prediction results.
Current computational methods on gene finding are more accurate on bacterial/prokary-
otic genomes. The reason is that bacterial/prokaryotic genes are continuous. In contrast,
eukaryotic genes are much more complicated because many introns, which will not
express, reside in a gene. A splicing procedure is needed to result in a continuous gene.
But, the splicing mechanism is not easy to understand. So it is still hard to annotate genes
of a eukaryotic genome.
Protein-Protein Interaction
There are many different kinds of proteins within a cell of an organism. To understand
the physiological functions of the cell, we need to know first how the proteins interact
with each other and what the products of the interactions are. The interactions can be
physical or functional. High-throughput biotechnologies were developed for such a
task. Also, computational methods were investigated. With the information of protein
interactions, a protein interaction network can be built. This kind of network is very
essential in understanding the regulation of metabolic pathways and thus can be applied
to the design of new drugs. To identify the interactions among proteins, various
statistical or machine learning techniques can be used. Furthermore, to gain more
information from a lot of known protein interaction networks is another important task.
These are association problems in data mining.
Phylogenetics
People have been interested in finding out the relationships between different species.
A tree structure can be used to illustrate such a relationship. Besides the appearances
of the species, studies in anatomy and physiology have contributed to this taxonomy.
However, with the knowledge of the species at the molecular level, in other words, the
DNA and protein data, new findings are expected from this knowledge. And the field has
advanced from the comparison of phenotypes to the comparison of genotypes. By
comparing certain DNA or protein sequence data, the distance of two species can be
determined. Species with shorter distances should be put in the same group because they
are closer to each other. This is a classical problem of clustering. Hierarchical clustering
can be used here to build a phylogenetic tree for some species with known corresponding
DNA or protein sequences.
Viruses from animals cause a lot of human diseases. Such diseases have been a serious
problem for human beings. Some of them are deadly and even incurable. For example,
AIDS was from apes, new variant Creutzfeldt-Jacob disease (CJD) was from cows (Mad
Cow Disease or Bovine Spongiform Encephalopathy, BSE), Foot-and-Mouth disease
12 Hsu
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
was from pigs, and Bird Flu/Avian Flu was from chickens and birds. Comparison of
corresponding genes of different species might be able to help invent new methods or
drugs for the control of such diseases. Thus, finding the relationship of different species
at the genomic level has become an important issue.
Summary
Data can be collected much more easily nowadays and are abundant everywhere. But,
what is lacking is knowledge. How to dig useful knowledge from large amount of data is
really essential. Technologies of different disciplines have emerged under the same goal
of mining knowledge from data in the past 10 years. They have been applied to problems
of different fields with success. On the other hand, accomplishment of the human genome
project in 2003 began a new era for biological research. With plenty of DNA, RNA, and
protein data, it would be difficult for the biologist to uncover the knowledge residing in
the data with traditional methods. Information technologies can be very helpful in the
tasks. Thus, the term bioinformatics became popular. Among the technologies, data
mining is the core. In this chapter, basic concepts of data mining and its importance and
relationship with bioinformatics were introduced. These provide a brief overview to the
topic so that the reader can go on to more advanced technologies and problems in the
subsequent chapters.
References
Chen, M.-S., Han, J., & Yu, P. S. (1996). Data mining: An overview from a database
perspective. IEEE Trans. on Knowledge and Data Eng., 8(6), 866-883.
Dunham, M. H. (2003). Data mining: Introductory and advanced topics. Upper Saddle
River, NJ: Prentice Hall.
Fayyad, U. M., Piatetsky-Shapiro, G., Smyth, P., & Uthurusamy, R. (Ed.). (1996). Advances
in knowledge discovery and data mining. Menlo Park; Cambridge, MA: AAAI/
MIT Press.
Fayyad, U. M. (1996). Data mining and knowledge discovery: Making sense out of data.
IEEE Expert, 11(5), 20-25.
Fayyad, U., Grinstein, G. G., & Wierse, A. (Eds.). (2002). Information visualization in
data mining and knowledge discovery. San Francisco: Morgan Kaufmann.
Kantardzic, M. (2003). Data mining: Concepts, models, methods, and algorithms.
Piscataway, NJ: IEEE Press.
Lesk, A. M. (2002). Introduction to bioinformatics. New York: Oxford University Press.
Simoudis, E. (1996). Reality check for data mining. IEEE Expert, 11(5), 26-33.
Hierarchical Profiling, Scoring and Applications in Bioinformatics 13
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Chapter II
Hierarchical
Profiling, Scoring
and Applications in
Bioinformatics
Li Liao, University of Delaware, USA
Abstract
Recently, clustering and classification methods have seen many applications in
bioinformatics. Some are simply straightforward applications of existing techniques,
but most have been adapted to cope with peculiar features of the biological data. Many
biological data take a form of vectors, whose components correspond to attributes
characterizing the biological entities being studied. Comparing these vectors, aka
profiles, are a crucial step for most clustering and classification methods. We review
the recent developments related to hierarchical profiling where the attributes are not
independent, but rather are correlated in a hierarchy. Hierarchical profiling arises in
a wide range of bioinformatics problems, including protein homology detection,
protein family classification, and metabolic pathway clustering. We discuss in detail
several clustering and classification methods where hierarchical correlations are
tackled in effective and efficient ways, by incorporation of domain-specific knowledge.
Relations to other statistical learning methods and more potential applications are
also discussed.
14 Liao
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Introduction
Profiling entities based on a set of attributes and then comparing these entities by their
profiles is a common, and often effective, paradigm in machine learning. Given profiles,
frequently represented as vectors of binary or real numbers, the comparison amounts to
measuring distance between a pair of profiles. Effective learning hinges on proper and
accurate measure of distances.
In general, given a set A of N attributes, A = {a
i
|i = 1, , N}, profiling an entity x on A
gives a mapping p(x)
N
, namely, p(x) is an N vector of real values. Conveniently, we
also use x to denote its profile p(x), and x
i
the i-th component of p(x). If all attributes in
A can only have two discrete values 0 and 1, then p(x) {0,1}
N
yields a binary profile.
The distance between a pair of profiles x and y is a function: D(x, y) . Hamming
distance is a straightforward, and also one of the most commonly used, distance
measures for binary profiles; it is a simple summation of difference at each individual
component:
D (x, y) =
i
n
d(i) (1)
where d(i) = | x
i
- y
i
|. For example, given x = (0, 1, 1, 1, 1) and y = (1, 1, 1, 1, 1), then D(x,
y) =
i=1
5
d(i) = 1+0+0+0+0 = 1. A variant definition of d(i), which is also very commonly
used, is that d(i) = 1 if x
i
= y
i
and d(i) = -1 if otherwise. In this variant definition, D(x, y)
=
i=1
5
d(i) = -1+1+1+1+1 = 3.
The Euclidean distance, defined as
n 2
i i i
D = (x y ) , has a geometric representation: a
profile is mapped to a point in a vector space where each coordinate corresponds to an
attribute. Besides using Euclidean metric, in vector space the distance between two
profiles is also often measured as dot product of the two corresponding vectors: x y =

i
n
x
i
y
i
. Dot product is a key quantity used in Support Vector Machines (Vapnik 1997,
Cristianini & Shawe-Taylor 2000, Scholkopf & Smola 2002). Many clustering methods
applicable to vectors in Euclidean space can be applied here, such as K-means.
While Hamming distance and Euclidean distance are the commonly adopted measures
of profile similarity, both of them imply an underlying assumption that the attributes are
independent and contribute equally in describing the profile. Therefore, the distance
between two profiles is simply a sum of distance (i.e., difference) between them at each
attribute. These measures become inappropriate when the attributes are not equally
contributing, or not independent, but rather correlated to one another. As we will see,
this is often the case in the real-world biological problems.
Intuitively, nontrivial relations among attributes complicate the comparisons of profiles.
An easy and pragmatic remedy is to introduce scores or weighting factors for individual
attributes to adjust their apparently different contribution to the Hamming or Euclidean
distance between profiles. That is, the value of d(i) in equation (1) now depends not
only on the values of x
i
and y
i
, but also on the index i. Often, scoring schemes of this type
are also used for situations where attributes are correlated, sometimes in a highly
nonlinear way. Different scoring schemes thereby are invented in order to capture the
Hierarchical Profiling, Scoring and Applications in Bioinformatics 15
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
relationships among attributes. Weighting factors in these scoring schemes are either
preset a priori based on domain knowledge about the attributes, or fixed from the training
examples, or determined by a combination of both. To put into a mathematical framework,
those scoring based approaches can be viewed as approximating the correlations among
attributes, which, without loss of generality, can be represented as a polynomial function.
In general, a formula that can capture correlations among attributes as pairs, triples,
quadruples, and so forth, may look like the following:
D =
i
n
d(i) +
ij
n
d(i)c(i,j)d(j) +
ijk
n
d(i)d(j)d(k)c(i,j,k) + (2)
where the coefficients c(i,j), c(i,j,k), , are used to represent the correlations. This is much
like introducing more neurons and more hidden layers in an artificial neural network
approach, or introducing a nonlinear kernel functions in kernel-based methods. Because
the exact relations among attributes are not known a priori, an open formula like equation
(2) is practically useless: as the number of these coefficients grows exponentially with
the profile size, solving it would be computationally intractable, and there would not be
enough training examples to fit these coefficients.
However, it was found that the situation would become tractable when these correlations
could be structured as a hierarchy a quite loose requirement and readily met in many
cases as we shall see later. In general, a hierarchical profile of entity x can be defined as
p(x) {0,1}
L
, where L stands for the set of leaves in a rooted tree T. A hierarchical profile
is no longer a plain string of zeros and ones. Instead, it may be best represented as a tree
with the leaves labeled by zeros and ones for binary profiles, or by real value numbers
for real value profiles.
As the main part of this chapter, we will discuss in detail several clustering and
classification methods where hierarchical profiles are coped with effectively and effi-
ciently, noticeably due to incorporation of domain specific knowledge. Relations to other
statistical learning methods, e.g., as kernel engineering, and more possible applications
to other bioinformatics problems are also discussed, towards the end of this chapter.
Hierarchical Profilings in Bioinformatics
Functional Annotations
Hierarchical profiling arises naturally in many bioinformatics problems. The first example
is probably from the phylogenetic profiles of proteins and using them to assign functions
to proteins (Pellegrini et al., 1999). To help understand the key concepts and motivations,
and also to introduce some terminologies for later discussions, we first briefly review the
bioinformatics methodologies for functional annotation.
Determining protein functions, also called as functional annotation, has been and
remains a central task in bioinformatics. Over the past 25 years many computational
16 Liao
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
methodologies have been developed toward solving this task. The development of these
computational approaches can be generally broken into four stages by both the
chronological order and algorithmic sophistication (Liao & Noble, 2002). For our purpose
here, we can categorize these methods into three levels according to the amount and type
of information used.
The methods in level one compare a pair of proteins for sequence similarity. Among them
are the landmark dynamic programming algorithm by Smith and Waterman (1980) and its
heuristic variations BLAST (Altschul et al., 1990) and FASTA (Pearson, 1990). The
biological reason behind these methods is protein homology; two proteins are homolo-
gous if they share a common ancestor. Therefore, homologous proteins are similar to each
other in the primary sequence and keep the same function as their ancestors, until the
evolutionary divergence mutations, deletions, or insertions from the ancestral protein
(or gene, to be more precise) is significant enough to cause any change. A typical way
to annotate a gene with unknown function is to search against a database of genes whose
functions are already known, such as GenBank (http://www.ncbi.nlm.nih.gov/Genbank),
and assign to the query gene the function of a homologous gene found in the database.
The next levels methods use multiple sequences from a family of proteins with same or
similar functions, in order to collect aggregate statistic for more accurate/reliable
annotation. Profiles (Gribskov et al., 1987) and hidden Markov models (Krogh et al., 1994;
Durbin et al., 1998) are two popular methods to capture and represent these aggregate
statistics from whole sequences for protein families. More refined and sophisticated
methods using aggregate statistics are developed, such as PSI-BLAST (Altschul et al.,
1997), SVM-Fisher (Jaakkola et al., 1999, 2000), Profile-Profile (Sadreyev & Grishi, 2003;
Mittelman et al., 2003), SVM-Pairwise (Liao & Noble, 2002, 2003). The aggregate statistic
may also be represented as patterns or motifs. Methods based on patterns and motifs
include BLOCKs (Henikoff & Henikoff, 1994), MEME (Bailey & Elkan, 1995), Meta-
MEME (Grundy et al., 1997), and eMotif (Nevill-Manning et al., 1998).
The third levels methods go beyond sequence similarity to utilize information such as
DNA Microarray gene expression data, phylogenetic profiles, and genetic networks. Not
only can these methods detect distant homologues homologous proteins with
sequence identity below 30%, but they also can identify proteins with related functions,
such as those found in a metabolic pathway or a structural complex.
Given a protein, its phylogenetic profile is represented as a vector, where each compo-
nent corresponds to a genome and takes a value of either one or zero, indicating
respectively the presence or absence of a homologous protein in the corresponding
genome. Protein phylogenetic profiles were used in Pellegrini, Marcotte, Thompson,
Eisenberg, and Yeates (1999) to assign protein functions based on the hypothesis that
functionally linked proteins, such as those participating in a metabolic pathway or a
structural complex, tend to be preserved or eliminated altogether in a new species. In
other words, these functionally linked proteins tend to co-evolve during evolution. In
Pellegrini et al. (1999), 16 then-fully-sequenced genomes were used in building the
phylogenetic profiles for 4290 proteins in E. coli genome. The phylogenetic profiles,
expressed as 16-vector, were clustered as following; proteins with identical profiles are
grouped and considered to be functionally linked, and two groups are called neighbors
when their phylogenetic profiles differ by one bit. The results based on these simple
clustering rules supported the functional linkage hypothesis. For instance, homologues
Hierarchical Profiling, Scoring and Applications in Bioinformatics 17
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
of ribosome protein RL7 were found in 10 out of 11 eubacterial genomes and in yeast but
not in archaeal genomes. They found that more than half of the E. coli proteins with the
RL7 profile or profiles different from RL7 by one bit have functions associated with the
ribosome, although none of these proteins share significant sequence similarity with the
RL7 protein. That is, these proteins are unlikely to be annotated as RL7 homologues by
using sequence similarity based methods.
There are some fundamental questions regarding the measure of profile similarity that can
affect the results from analysis of phylogenetic profiles. Can we generalize the definition
of similar profiles? In other words, can we devise a measure so we can calculate similarity
for any pair of profiles? What threshold should be adopted when profile similarity is used
to infer functional linkage?
A simple measure for profile similarity first brought up was Hamming distance. In
Marcotte, Xenarios, van Der Bliek, and Eisenberg (2000), phylogenetic profiles are used
to identify and predict subcellular locations for proteins; they found mitochondrial and
non-mitochondrial proteins.
The first work that recognizes phylogenetic profiles as a kind of hierarchical profiling is
Liberles, Thoren, von Heijne, and Elofsson (2002), a method that utilizes the historical
evolution of two proteins to account for their similarity (or dissimilarity). Evolutionary
relationships among organisms can be represented as a phylogenetic tree where leaves
correspond to the current organisms and internal nodes correspond to hypothetical
ancient organisms. So, rather than simply counting the presence and absence of the
proteins in the current genomes, a quantity called differential parsimony is calculated
that minimize the number of times when changes have to be made at tree branches to
reconcile the two profiles.
Comparative Genomics
Another example of hierarchical profiling arises from comparing genomes based on their
metabolic pathway profiles. A main goal of comparing genomes is to reveal the evolu-
tionary relationships among organisms, which can be represented as phylogenetic trees.
Originally phylogenetic trees were constructed based on phenotypic particularly
morphological features of organisms. Nowadays, molecular reconstruction of phylo-
genetic trees is most commonly based on comparisons of small sub-unit ribosomal RNA
(16S rRNA) sequences (Woese, 1987). The small sub-unit rRNAs is orthodoxly used as
gold standard for phylogeny study, mainly due to two factors: their ubiquitous presence
and relative stability during evolution. However, the significance of phylogenies based
on these sequences have been recently questioned with growing evidence for extensive
lateral transfer of genetic material, a process which results in blurring the boundaries
between species. Phylogenetic trees based on individual protein/gene sequence analy-
sis thus called gene tree are not all congruent with species trees based on 16S rRNA
(Eisen, 2000). Attempts of building phylogenetic trees based on different information
levels, such as pathways (Dandekar et al., 1999; Forst & Schulten, 2001) or some particular
molecular features such as folds (Lin & Gerstein 2000), have also led to mixed results of
congruence with the 16S rRNA-based trees.
18 Liao
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
From a comparative genomics perspective, it makes more sense to study evolutionary
relationships based on genome-wide information, instead of a piece of the genome, be
it an rRNA or a gene. As more and more genomes are fully sequenced, such genome-wide
information becomes available. One particularly interesting type of information is the
entire repertoire of metabolic pathways in an organism, as the cellular functions of an
organism are carried out via these metabolic pathways. A metabolic pathway is a chain
of biochemical reactions, together fulfilling certain cellular functions. For example,
Glycolysis is a pathway existed in most cells, which consists of 10 sequential reactions
converting glucose to pyruvate while generating the energy that the cell needs. Because
most of these reactions require enzymes as catalyst, therefore in an enzyme centric
scheme, pathways are represented as sequences of component enzymes. It is reasonable
to set the necessary condition for a pathway to exist in an organism as that all the
component enzymes of that pathway are available. Enzymes are denoted by enzyme
commission (EC) numbers which specifies the substrate specificity. Most enzymes are
proteins. Metabolic pathways in a completely sequenced genome are reconstructed by
identifying enzyme proteins that are required for a pathway (Gaasterland & Selkov, 1995;
Karp et al., 2002).
The information about presence and absence of metabolic pathways in genomes can be
represented as a binary matrix, as shown in Figure 1, where an entry (i,j) = 1/0 represents
whether pathway j is present/absent in genome i. Therefore, each row serves as a profile
of the corresponding genome (termed metabolic pathway profiles), and each column
serves as a profile of the corresponding pathway (termed phyletic profiles). It is
reasonable to believe that valuable information about evolution and co-evolution is
embedded in these profiles, and comparison of profiles would reveal, to some degree, the
evolutionary relations among entities (either genomes or pathways) represented by
these profiles.
Once again, the attributes used for building these profiles are not independent but
correlated to each other. Because different metabolic pathways may be related to one
another in terms of physiological functions, e.g., one pathways absence/presence may
be correlated with another pathways absence/presence in a genome, these relationships
Figure 1. Binary matrix encoding the presence/absence of pathways in genomes; O
1
to O
m
represent m genomes, and P
1
to P
n
represent n pathways
P
1
P
2
P
n
O
1
O
2
O
m
0 1
1 1
0 1
0
1
1
. . .
. . .
.

.

.
.

.

.
.

.

.
. . .
.

.

.
Hierarchical Profiling, Scoring and Applications in Bioinformatics 19
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
among pathways, as attributes of metabolic pathway profiles, should be taken into
account when comparing genomes based on their MPPs. The relationships among
various pathways are categorized as a hierarchy in the WIT database (Overbeek et al.,
2000), which can be found at the following URL (http://compbio.mcs.anl.gov/puma2/cgi-
bin/functional_overview.cgi).
Comparing Hierarchical Profiles
In the last section, we showed that the data and information in many bioinformatics
problems can be represented as hierarchical profiles. Consequently, the clustering and
classification of such data and information need to account for the hierarchical correla-
tions among attributes when measuring the profile similarity. While a generic formula like
equation (2) posits to account for arbitrary correlations theoretically, its demand of
exponentially growing amount of training examples and lacking of an effective learning
mechanism render the formula practically useless. In hierarchical profiles, however, the
structure of relations among attributes is known, and sometimes the biological interpre-
tation of these relations is also known. As a result, the learning problems will become
rather tractable.
P-Tree Approach
Realizing that the hierarchical profiles contain information not only in the vector but also
in the hierarchical correlations, it is natural to first attempt at treating them as trees. How
to compare trees is itself an interesting topic with wide applications, and has been the
subject of numerous studies. In Liao, Kim, and Tomb, (2002), a p-tree based approach
was proposed to measure the similarity of metabolic pathway profiles. The relationships
among pathways are adopted from the WIT database and represented as a Master tree.
About 3300 known metabolic pathways are collected in the WIT database and these
pathways are represented as leaves in the Master tree. Then, for each genome, a p-Tree
is derived from the Master tree by marking off leaves whose corresponding pathways are
absent from the organism. In this representation, a profile is no longer a simple string of
zeros and ones, where each bit is treated equally and independently. Instead, it is mapped
into a p-Tree so that the hierarchical relationship among bits is restored and encoded in
the tree.
The comparison of two p-Trees thus evaluates the difference between the two corre-
sponding profiles. To take into account the hierarchy, a scoring scheme ought to weight
(mis)matches at bits i and j according to their positions in the tree, such as i and j are
sibling, versus i and j are located distantly in the tree. For that, the (mis)matches scores
are transmitted bottom-up to the root of the Master tree in four steps: (1) overlay the two
p-Trees; (2) score mismatches and matches between two p-Trees and label scores at the
corresponding leaves on the master tree; (3) average scores from siblings (weight
breadth) and assign the score to the parent node; (4) iterate step 3 until the root is reached.
20 Liao
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Figure 2. A comparison of two organisms orgi and orgj, with respect to the three
pathways p1, p2, and p3 (weighted scoring scheme, Case 1); Panel A featuring the
master tree, where each pathway is represented by a circle and each square represents
an internal node; in Panel B, the presence(absence) of a pathway is indicated by 1(0);
Panel C features the p-trees, where a filled (empty) circle indicates the presence
(absence) of the pathway; Panel D, the bottom-up propagation of the scores is
illustrated; a cross (triangle) indicates a mismatch (match)
p1 p2
p3
A. Master Tree
p1 p2 p3
orgi
orgj
1 0 1
0 1 1
B. Profiles for orgi and orgj
C. p-Trees of orgi and orgj
-1 -1
1
(-1-1)/2=0
(0+1)/2=1
D. Scoring
Figure 3. Case 2: A comparison of two organisms org
i
and org
j
, with respect to the three
pathways p1, p2, and p3 (weighted scoring scheme, Case 2); the same legends as in
Figure 2
p1 p2
p3
A. Master Tree
p1 p2 p3
orgi
orgj
1 1 0
1 1 1
B. Profiles for orgi and orgj
C. p-Trees of orgi and orgj
1 1
-1
(1+1)/2=1
(1-1)/2=0
D. Scoring
Hierarchical Profiling, Scoring and Applications in Bioinformatics 21
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
An algorithm implementing these steps is quite straightforward and has a time complexity
linear with the size of the Master tree.
To demonstrate how the algorithm works, let us look at an example of two organisms, org
i
and org
j
, and three pathways p
1
, p
2
, and p
3
. Two hypothetical cases are considered and
are demonstrated in Figures 2 and 3 respectively. In case one, org
i
contains pathways
p
1
and p
3
, and org
j
contains p
2
and p
3
. The metabolic pathway profiles for org
i
and org
j
are shown in panel B and their corresponding p-Trees are displayed in panel C. In the
panel D, two p-Trees are superposed. Matches and mismatches are scored at the leaves
and the scores are propagated up to the root. In case two, org
i
contains pathways p
1
and
p
2
, whereas org
j
contains all three pathways, and similar calculation is shown in Figure
3. In this example, given the topology of the master tree, the two cases have the same final
score 0 in the p-Tree scoring scheme. This is in contrast to the variant Hamming scoring
scheme, where the score for case 1 equals to (-1-1+1) = -1 and the score for case 2 equals
to (1+1-1) = 1. Evidently, the p-Tree scoring scheme has taken into account the pathway
relationships present in the Master tree: p
1
and p
2
, being in the same branch, are likely
to have a similar biochemical or physiological role, which is distinct from the role of the
p
3
pathway.
Using pathway hierarchical categories in the WIT database and the scoring scheme
described above, we compared 31 genomes based on their metabolic pathway profiles
Figure 4. MPP-based tree and 16S rRNA-based tree for 31 genomes; neighbor-joining
program from Phylip package is used to generate these trees
22 Liao
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
(MPP) (see Figure 4). Relations among 31 genomes are represented as a MPP-based tree
and are compared with the phylogeny based on 16s rRNA. While the MPP-based tree
is congruent with the 16S rRNA-based tree at several levels, some interesting discrep-
ancies were found. For example, the extremely radiation resistant organism, D. radiodurans
is positioned in E. coli metabolic lineage. The noted deviation from the classical 16s
rRNA phylogeny suggests that pathways have undergone evolution transcending the
boundaries of species and genera. The MPP-based trees can be used to suggest
alternative production platform for metabolic engineering. A different approach to this
problem on the same dataset was later proposed in Heymans and Singh (2003).
Parameterized Tree Distance
Evolution of pathways and the interaction with evolution of the host genomes can be
further investigated by comparing the columns (phyletic profiles) in the binary matrix in
Figure 1. The intuition is that the co-occurrence pattern of pathways in a group of
organisms would coincide with the evolutionary path of these host organisms, for
instance, as a result of the emergence of enzymes common to those pathways. Therefore,
such patterns would give useful information, e.g., organism-specific adaptations. These
co-occurrences can be detected by clustering pathways based on their phyletic profiles.
Again, the phyletic profiles here are hierarchical profiles, and the distance measure
between them should be calculated using the scheme that we developed above. In Zhang,
Liao, Tomb, and Wang (2002), 2719 pathways selected from 31 genomes in the WIT
database were clustered into 69 groups of pathways, and pathways in each group co-
occur in the organisms.
Further insights were achieved by studying the evolution of enzymes that are compo-
nents in co-occurred pathways. For completely sequenced genomes, sequences of these
enzymes are available and can be used to build individual gene trees (in contrast to
species trees). Comparisons of the gene trees of component enzymes in co-occurred
pathways would serve as a basis for investigating how component enzymes evolve and
whether they evolve in accordance with the pathways. For example, two pathways p1 =
(e1, e2, e3) and p2 = (e4, e5) co-occur in organisms o1, o2, and o3. For component enzyme
e1 of pathway p1, a gene tree T
e1
is built with e1s homologues in organisms o1, o2 and
o3, by using some standard tree reconstruction method such as neighbor joining
algorithm (Felsenstein, 1989). This can be done for other component enzymes e2 to e5
as well. If, as an ideal case, gene trees T
ei
for i = 1 to 5 are identical, we would then say
that the pathway co-occurrence is congruent to speciation. When gene trees for
component enzymes in co-occurred pathways are not identical, comparisons of gene
trees may reveal how pathways evolve differently, possibly due to gene lateral transfer,
duplication or loss. Recent work also shows that analysis of metabolic pathways may
explain causes and evolution of enzymes dispensability.
To properly address tree comparisons, several algorithms were developed and tested
(Zhang et al., 2002). One similarity measure is based on leaf overlap. Let T1 and T2 be two
trees. Let S1 be the set of leaves of T1 and S2 be the set of leaves of T2. The leaf-overlap-
based distance is defined as
Hierarchical Profiling, Scoring and Applications in Bioinformatics 23
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
D
x
(T1, T2) = |S1 S2 | / |S1 S2 |, (3)
where |.| denotes the set cardinality. A more elaborated metric, called parameterized
distance as an extension from the editing distance between two unordered trees, was
proposed to account for the structural difference between ordered and rooted trees. A
parameter c is introduced to balance the cost incurred at different cases such as deleting,
inserting, and matching subtrees. A dynamic programming algorithm is invoked to
calculate the optimal distance. For example, in Figure 5, the distance between D1 and p
is 1.875 and the distance between D2 and p is 1.813 when the parameter c is set at value
of 0.5. On the other hand, the editing distance between D1 and p and between D2 and
p are both 6, representing the cost of deleting the six nodes not touched by the dotted
mapping lines in Figure 5. Noting that D1 differs from D2 topologically, this example
shows that the parameterized distance better reflects the structural difference between
trees than the editing distance. In Zhang (2002), the 523 component enzymes of these
2719 pathways were clustered, based on parameterized tree distance with c being set at
0.6, into exactly the same 69 clusters based on pathway co-occurrence. This suggests
that our hypothesis about using co-occurrence to infer evolution is valid, at least
approximately.
Figure 5. Illustration of parameterized distances between trees. Red dotted lines map
tree P to trees D1 and D2, the parameter C = 1 gives the editing distances; the editing
distance between D1 and P is 6, because six deletions of nodes other than a, b, and c
in D1 will make D1 identical to tree P, although D1 and D2 are topologically different,
their editing distances to P are both equal to 6
e
a
f b k
h g
k
a
b
c
c
e
a
f b k
h g
a
b
c c
k
1.875 6 D1, P
D2, P
d
y
()
1.8125 6
C=0.5 C=1
D1
D2
P
P
24 Liao
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Tree Kernel Approach
As shown before, use of phylogenetic profiles for proteins has led to methods more
sensitive to detect remote homologues. In the following we discuss classifiers using
support vector machines to explore similarity among phylogenetic profiles as hierarchi-
cal profiles.
As a powerful statistical learning method, support vector machines (SVMs) have
recently been applied with remarkable success in bioinformatics problems, including
remote protein homology detection, microarrray gene expression analysis, and protein
secondary structure prediction. SVMs have been applied to problems in other domains,
such as face detection and text categorization. SVMs possess many nice characteristics
a good learning method shall have: it is expressive; it requires fewer training examples;
it has an efficient learning algorithm; it has an elegant geometric interpretation; and
above all, it generalizes well. The power of SVMs comes partly from the data represen-
tation, where an entity, e.g., a protein, is represented by a set of attributes instead of a
single score. However, how those attributes contribute to distinguishing a true positive
(filled dot in Figure 6) from a true negative (empty circle) may be quite complex. In other
words, the boundary line between the two classes, if depicted in a vector space, can be
highly nonlinear (dashed line in the left panel of Figure 6), and nonetheless, it is the goal
of a classifier to find this boundary line. The SVMs method will find a nonlinear mapping
that transform the data from the original space, called input space, into a higher
dimensional space, called feature space, where the data can be linearly separable (right
panel of Figure 6). The learning power of SVMs comes mostly from the use of kernel
functions, which define how the dot product between two points in the feature space can
be calculated as a function of their corresponding vectors in the input space. As the dot
product between vectors is the only quantity needed to find a class boundary in the
Figure 6. Schematic illustration of nonlinear mapping of data from input space to
feature space for a SVM
F ( )
x
i
x
j
F (x
j
)
F (x
i
)
Hierarchical Profiling, Scoring and Applications in Bioinformatics 25
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
feature space, kernel functions therefore contain sufficient information, and more
importantly they avoid explicit mapping to high dimensional feature space; high dimen-
sionality often poses difficult problems for learning such as overfitting, thus termed the
curse of dimensionality. The other mechanism adopted by SVMs is to pick the boundary
line that has the maximum margin to both classes. A maximum margin boundary line has
low Vapnik-Chervonenkis dimension, which ensures good generalization. Because of the
central role played by kernel functions, how to engineer kernel functions to incorporate
domain specific information for better performance has been an active research activity.
It is worth noting that, as compared to other similar learning methods such as artificial
neural networks, the SVMs require fewer training examples, which is a great advantage
in many bioinformatics applications.
Vert (2002) proposed a tree kernel to compare not just the profiles themselves but also
their global patterns of inheritance reflected in the phylogenetic tree. In other words, the
tree kernel takes into account phylogenetic histories for two genes when in evolution
they transmitted together or not rather than by just comparing phylogenetic profiles
organism per organism, at the leaves of the phylogenetic tree. A kernel function is thus
defined as
K(x, y) =
i=1 to D

i
(x)
i
(y) (4)
where
i
(x) is an inheritance pattern i for profile x. An inheritance pattern of x gives an
explanation of x, that is, the presence (1) or absence (0) of gene x at each current genome
is the result of a series of evolutionary events happened at the ancient genomes. If we
assume a gene is possessed by the ancestral genome, at each branch of the phylogenetic
tree which corresponds to a speciation, the gene may either be retained or get lost. An
inheritance pattern corresponds to a series of assignments of retain or loss at all branches
such that the results at leaves match the profile of x. Because of the stochastic property
of these evolutionary events, we cannot be certain whether a gene is retained or lost.
Rather, the best we know may be the probability of either case. Let F
i
(x) = (x | i), which
is the probability that profile x can be interpreted by the inheritance pattern i, then K (x,
y) =
i=1 to D
P(x|i) P(y|i) is the joint probability that both profiles x and y are resulted
from all possible pattern i. Intuitively, not all possible inheritance patterns occur at the
same frequency. Let P(i) be the probability that pattern i actually occurs during the
evolution, then the so called tree kernel is refined as
K (x, y) =
i=1 to D
P(i) P(x|i) P(y|i) (5)
The formalism of using joint probability as kernels first appeared in other applications,
such as convolution kernels (Watkins, 1999; Haussler, 1999). Because the number of
patterns D grows exponentially with the size of the phylogenetic tree, an efficient
algorithm is needed to compute the kernel. Such an algorithm was developed in Vert
(2002), which uses post-order traversals of the tree and has a time complexity linear with
respect to the tree size.
26 Liao
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
To test the validity of the tree kernel method, phylogenetic profiles were generated for
2465 yeast genes, whose accurate functional classification are already known, by BLAST
search against 24 fully-sequenced genomes. For each gene, if a BLAST hit with E-value
less than 1.0 is found in a genome, the corresponding bit for that genome is then assigned
as 1, otherwise is assigned as 0. The resulting profile is a 24-bit string of zeros and ones.
Two assumptions were made in calculating the tree kernel for a pair of genes x and y.
Although their exact probabilities may never be known, it is reasonable to assume that
losing a gene or obtaining a new gene is relatively rare as compared to keeping the status
quo. In Vert, the probability that an existing gene is retained at a tree branch (i.e.,
speciation) is set at 0.9, and the probability that a new gene is created at a branch is set
at 0.1. It was further assumed that such a distribution remains the same at all branches
for all genes. Even with these crude assumptions, in the cross validation experiments on
those 2465 yeast genes, the tree kernels classification accuracy already significantly
exceed that of a kernel using just the dot product x y =
i=1 to 24
x
i
y
i
.
Extended Phylogenetic Profiles
The tree-kernel approachs improvement at classification accuracy is mainly due to
engineering the kernel functions. In Narra and Liao (2004, 2005) further improvement is
attained by both data representations and kernel engineering. As the reader should have
been convinced by now, these phylogenetic profiles contain more information than just
the string of zeros and ones; the phylogenetic tree provides relationships among the bits
in these profiles. In Narra and Liao, a two-step procedure is adopted to extend phyloge-
netic profiles with extra bits encoding the tree structure: (1) a score is assigned at each
internal tree node; (2) the score labeled tree is then flatten into an extended vector. For
an internal tree node in a phylogenetic tree, as it is interpreted as ancestor of the nodes
underneath it, one way to assign a score for it is to take the average of the scores from
its children nodes. This scoring scheme works top-down recursively until the leaves are
reached: the score at a leaf is just the value of the corresponding component in the
hierarchical profile. The same scoring scheme was also used in p-tree approach. Unlike
p-Tree approach that keeps just the score the root node and thus inevitably causes
information loss, the scores at all internal nodes are retained and then mapped into a
vector via a post-order tree traversal. This vector is then concatenated to the original
profile vector forming an extended vector, which is called tree-encoded profile. The
scheme works for both binary and real-valued profiles. In order to retain information, real
value profiles for yeast genes are used; the binary profiles for tree-kernel are derived real
value profiles by imposing a cutoff at E-values
Given a pair of tree-encoded profiles x and y, the polynomial kernel is used classification:
K(x, y) = [1+ s D(x, y)]
d
(6)
where s and d are two adjustable parameters. Unlike ordinary polynomial kernels, D(x, y)
is not the dot product of vectors x and y, but rather, a generalized Hamming distance for
real value vectors:
Hierarchical Profiling, Scoring and Applications in Bioinformatics 27
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
D(x, y) =
i=1 to n
(S(|x
i
-y
i
|) (7)
where the ad hoc function S has value 7 for a match, 5 for a mismatch by a difference less
then 0.1, 3 for a mismatch by a difference less than 0.3, and 1 for a mismatch by a difference
less than 0.5. The values in the ad hoc function S are assigned based on the E-value
distribution of the protein dataset. The methods are tested on the same data set by using
the same cross-validation protocols as in Vert. The classification accuracy of using the
extended phylogenetic profiles with E-values and polynomial kernel generally outper-
forms the tree-kernel approach at most of the 133 functional classes of 2465 yeast genes
in Vert.
More Applications and Future Trends
We have seen in the last two sections some problems in bioinformatics and computa-
tional biology where relationships can be categorized as hierarchy, and how such
hierarchical structure can be utilized to facilitate the learning. Because of the central role
played by evolution theory in biology, and the fact that phylogeny is natively expressed
as a hierarchy, it is no surprise that hierarchical profiling arises in many biological
problems.
In Siepel and Haussler (2004), methods are developed that combine phylogenetic and
hidden Markov models for biosequence analysis. Hidden Markov models, first studied
as a tool for speech recognition, were introduced to bioinformatics field around 1994
(Krogh et al., 1994). Since then, hidden Markov models have been applied to problems
in bioinformatics and computational biology including gene identification, protein
homology detection, secondary structure prediction, and many more.
In sequence modeling, hidden Markov models essentially simulate processes along the
length of the sequence, mostly ignoring the evolutionary process at each position. On
the other hand, in phylogenetic analysis, the focus is on the variation across sequences
at each position, mostly ignoring correlations from position to position along their
length. From a hierarchical profiling viewpoint, each column in a multiple sequence
alignment is a profile, not binary but 20ary, and provides a sampling for the current
sequences. It is a very attractive idea to combine these two apparently orthogonal
models. In Siepel and Haussler, a simple and efficient method was developed to build
higher-order states in the HMM, which allows for context-sensitive models of substitu-
tion, leading to significant improvements in the fit of a combined phylogenetic and
hidden Markov model. Their work promises to be a very useful tool for some important
biosequence analysis applications, such as gene finding and secondary structure
prediction.
Hierarchical relations also exist in applications where phylogeny is not the subject. For
example, in Holme, Huss, and Jeong (2003), biochemical networks are decomposed into
sub-networks. Because of the inherent non-local features possessed by these networks,
a hierarchical analysis was proposed to take into account the global structure while doing
28 Liao
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
decomposition. In another work (Gagner et al., 2004), the traditional way of representing
metabolic networks as a collection of connected pathways is questioned: such represen-
tation suffers the lack of rigorous definition, yielding pathways of disparate content and
size. Instead, they proposed a hierarchical representation that emphasizes the gross
organization of metabolic networks in largely independent pathways and sub-systems
at several levels of independence.
While hierarchical relations are widely existed, there are many applications where the
topology of the data relation can not be described as a hierarchy but rather as a network.
Large scale gene expression data are often analyzed by clustering genes based on gene
expression data alone, though a priori knowledge in the form of biological networks is
available. In Hanish, Zien, Zimmer, and Lengauer (2002), a co-clustering method is
developed that makes use of this additional information and has demonstrated consid-
erably improvements for exploratory analysis.
Summary
As we have seen, hierarchical profiling can be applied to many problems in bioinformatics
and computational biology, from remote protein homology detection, to genome com-
parisons, to metabolic pathway clustering, wherever the relations among attribute data
possess structure as a hierarchy. These hierarchical relations may be inherent or as an
approximation to more complex relationships. To properly deal with such relationships
is essential for clustering and classification of biological data. It is advisable to heed the
relations that bear artificially hierarchical structure; hierarchical profiling on these
relations may yield misleading results.
In many cases, the hierarchy, and the biological insight wherein embodied, can be
integrated into the framework of data mining. It consequently facilitates the learning and
renders meaningful interpretation of the learning results. We have reviewed the recent
developments in this respect. Some methods treat hierarchical profile scoring as a tree
comparison problem, some as an encoding problem, and some as a graphical model with
a Bayesian interpretation. The latter approach is of particular interest, since most
biological data are stochastic by nature.
A trend is seen in bioinformatics that combines different methods and models so the
hybrid method can achieve a better performance. In the tree-kernel method, the hierar-
chical profiling and scoring are incorporated as kernel engineering task of the support
vector machines. In sequence analysis, the phylogenetic techniques and hidden Markov
models are combined to account for the relationships exhibited in sequences that either
method alone can not handle properly. As more and more biological data with complex
relationships being generated, it is reasonable to believe the hierarchical profiling will
see more applications, either serve by itself as a useful tool for analyzing these data, or
serve as a prototype for developing more sophisticated and powerful data mining tools.
Hierarchical Profiling, Scoring and Applications in Bioinformatics 29
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
References
Altschul, S. F., Gish, W., Miller, W., Myers, E., & Lipman, D. J. (1990). Basic local
alignment search tool. Journal of Molecular Biology, 215, 403-410.
Altschul, S. F., Madden, T. L., Schffer, A. A., Zhang, J., Zhang, Z., Miller, W., et al. (1997).
Gapped BLAST and PSI-BLAST: A new generation of protein database search
programs. Nucleic Acids Res., 25, 3389-3402.
Bailey, T. L., & Elkan, C. P. (1995). Unsupervised learning of multiple motifs in biopoly-
mers using EM. Machine Learning, 21(1-2), 51-80.
Burges, C. J. C. (1998). A tutorial on support vector machines for pattern recognition.
Data Mining and Knowledge Discovery, 2, 121-167.
Cristianini, N., & Shawe-Taylor, J. (2000). An introduction to support vector machines
and other kernel-based learning methods. Cambridge, UK: Cambridge University
Press.
Dandekar, T., Schuster, S., Snel, B., Huynen, M., & Bork, P. (1999). Pathway alignment:
application to the comparative analysis of glycolytic enzymes. Biochemical
Journal, 343, 115-124.
Durbin, R., Eddy, S., Krogh, A., & Mitchison, G. (1998). Biological sequence analysis:
Probabilistic models of proteins and nucleic acids. Cambridge, UK: Cambridge
University Press.
Eisen, J. A. (2000). Horizontal gene transfer among microbial genomes: New insights from
complete genome analysis. Current Opinion in Genetics & Development, 10, 606-
611.
Felsenstein, J. (1989). PHYLIP Phylogeny Inference Package (Version 3.2). Cladistics,
5, 164-166.
Forst, C. V., & Schulten, K. (2001). Phylogenetic analysis of metabolic pathways. Journal
of Molecular Evolution, 52, 471-489.
Gaasterland, T., & Selkov, E. (1995). Reconstruction of metabolic networks using
incomplete information. In Proceedings of the Third International Conference on
Intelligent Systems for Molecular Biology (pp. 127-135). Menlo Park, CA: AAAI
Press.
Gagneur, J., Jackson, D., & Casar, G. (2003). Hierarchical analysis of dependence in
metabolic networks. Bioinformatics, 19, 1027-1034.
Gribskov, M., McLachlan, A. D., & Eisenberg, D. (1987). Profile analysis: Detection of
distantly related proteins. Proc. Natl. Acad. Sci. USA, 84 (pp. 4355-4358).
Grundy,W. N., Bailey, T. L., Elkan, C. P., & Baker, M. E. (1997). Meta-MEME: Motif-based
hidden Markov Models of biological sequences. Computer Applications in the
Biosciences, 13(4), 397-406.
Hanisch, D., Zien, A., Zimmer, R., & Lengauer, T. (2002). Co-clustering of biological
networks and gene expression data. Bioinformatics, 18, S145-S154.
30 Liao
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Haussler, D. (1999). Convolution kernels on discrete structures (Technical Report
UCSC-CRL-99-10). Santa Cruz: University of California.
Henikoff, S., & Henikoff, J.G. (1994). Protein family classification based on search a
database of blocks. Genomics, 19, 97-107.
Heymans, M., & Singh, A. J. (2003). Deriving phylogenetic trees from the similarity
analysis of metabolic pathways. Bioinformatics, 19, i138-i146.
Holme, P., Huss, M., & Jeong, H. (2003). Subnetwork hierarchies of biochemical path-
ways. Bioinformatics, 19, 532-538.
Jaakkola, T., Diekhans, M., & Haussler, D. (1999). Using the Fisher Kernel Method to
detect remote protein homologies. In Proceedings of the Seventh International
Conference on Intelligent Systems for Molecular Biology (pp. 149-158). Menlo
Park, CA: AAAI Press.
Jaakkola, T., Diekhans, M., & Haussler, D. (2000). A discriminative framework for
detecting remote protein homologies. Journal of Computational Biology, 7, 95-
114.
Karp, P.D., Riley, M., Saier, M., Paulsen, I. T., Paley, S. M., & Pellegrini-Toole, A. (2000).
The EcoCyc and MetaCyc databases. Nucleic Acids Research, 28, 56-59.
Krogh, A., Brown, M., Mian, I. S., Sjolander, K., & Haussler, D. (1994). Hidden Markov
models in computational biology: Applications to protein modeling. Journal of
Molecular Biology, 235, 1501-1531.
Liao, L., Kim, S., & Tomb, J-F. (2002). Genome comparisons based on profiles of metabolic
pathways, In The Proceedings of The Sixth International Conference on Knowl-
edge-Based Intelligent Information & Engineering Systems (pp. 469-476). Crema,
Italy: IOS Press.
Liao, L., & Noble, W. S. (2003). Combining pairwise sequence similarity and support
vector machines for detecting remote protein evolutionary and structural relation-
ships. Journal of Computational Biology, 10, 857-868.
Liberles, D. A., Thoren, A., von Heijne, G., & Elofsson, A. (2002). The use of phylogenetic
profiles for gene predictions. Current Genomics, 3, 131-137.
Lin, J., & Gerstein, M. (2002). Whole-genome trees based on the occurrence of folds and
orthologs: Implications for comparing genomes on different levels. Genome
Research, 10, 808-818.
Marcotte, E. M., Xenarios, I., van Der Bliek, A. M., & Eisenberg, D. (2000). Localizing
proteins in the cell from their phylogenetic profiles. Proc. Natl. Acad. Sci. USA, 97,
(pp. 12115-12120).
Mittelman, D., Sadreyev, R., & Grishin, N. (2003). Probabilistic scoring measures for
profile-profile comparison yield more accurate shore seed alignments.
Bioinformatics, 19, 1531-1539.
Narra, K., & Liao, L. (2004). Using extended phylogenetic profiles and support vector
machines for protein family classification. In The Proceedings of the Fifth Inter-
national Conference on Software Engineering, Artificial Intelligence, Network-
ing, and Parallel/Distributed Computing (pp. 152-157). Beijing, China: ACIS
Publication.
Hierarchical Profiling, Scoring and Applications in Bioinformatics 31
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Narra, K., & Liao, L. (2005). Use of extended phylogenetic profiles with E-values and
support vector machines for protein family classification. International Journal
of Computer and Information Science, 6(1).
Nevill-Manning, C. G., Wu, T. D., & Brutlag, D. L.(1998). Highly specific protein sequence
motifs for genome analysis. Proc. Natl. Acad. Sci. USA, 95(11), 5865-5871.
Noble, W. (2004). Support vector machine applications in computational biology. In B.
Scholkopf, K. Tsuda, & J-P. Vert. (Eds.), Kernel methods in computational biology
(pp. 71-92). Cambridge, MA: The MIT Press.
Overbeek, R., Larsen, N., Pusch, G. D., DSouza, M., Selkov Jr., E., Kyrpides, N., Fonstein,
M., Maltsev, N., & Selkov, E. (2000). WIT: Integrated system for high throughput
genome sequence analysis and metabolic reconstruction. Nucleic Acids Res., 28,
123-125.
Pearson, W. (1990). Rapid and sensitive sequence comparison with FASTP and FASTA.
Meth. Enzymol., 183, 63-98.
Pellegrini, M., Marcotte, E. M., Thompson, M. J., Eisenberg, D., & Yeates, T. O. (1999).
Assigning protein functions by comparative genome analysis: Protein phyloge-
netic profiles. Proc. Natl. Acad. Sci. USA, 96, (pp. 4285-4288).
Rabiner, L. R. (1989). A tutorial on hidden Markov models and selected applications in
speech recognition. Proc. IEEE, 77, 257-286.
Sadreyev, R., & Grishin, N. (2003). Compass: A tool for comparison of multiple protein
alignments with assessment of statistical significance. Journal of Molecular
Biology, 326, 317-336.
Scholkopf, B., & Smola, A. J. (2001). Learning with kernels: Support vector machines,
learning). Cambridge, MA: The MIT Press.
Siepel, A., & Haussler, D. (2004). Combining phylogenetic and hidden Markov Models
in biosequence analysis. J. Comput. Biol., 11(2-3), 413-428.
Smith, T. F., & Waterman, M. S.(1981). Identification of common molecular subse-
quences. Journal of Molecular Biology, 147, 195-197.
Vapnik, V. (1998). Statistical Learning Theory: Adaptive and learning systems for signal
processing, communications, and control. New York: Wiley.
Vert, J-P. (2002). A tree kernel to analyze phylogenetic profiles. Bioinformatics, 18, S276-
S284.
Watkins, C. (1999). Dynamic alignment kernels. In A. J. Smola, P. Bartlett, B. Scholkopf,
& C. Schuurmans (Ed.), Advances in large margin classifiers. Cambridge, MA: The
MIT Press.
Woese, C. (1987). Bacterial evolution. Microbial Rev., 51, 221-271.
Zhang, K., Wang, J. T. L., & Shasha, D. (1996). On the editing distance between
undirected acyclic graphs. International Journal of Foundations of Computer
Science, 7, 43-58.
Zhang, S., Liao, L., Tomb, J-F., Wang, J. T. L. (2002). Clustering and classifying enzymes
in metabolic pathways: Some preliminary results. In ACM SIGKDD Workshop on
Data Mining in Bioinformatics, Edmonton, Canada (pp. 19-24).
32 Hsu, Chung, & Kristal
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Chapter III
Combinatorial
Fusion Analysis:
Methods and Practices
of Combining
Multiple Scoring Systems
D. Frank Hsu, Fordham University, USA
Yun-Sheng Chung, National Tsing Hua University, Taiwan
Bruce S. Kristal, Burke Medical Research Institute and
Weill Medical College of Cornell University, USA
Abstract
Combination methods have been investigated as a possible means to improve
performance in multi-variable (multi-criterion or multi-objective) classification,
prediction, learning, and optimization problems. In addition, information collected
from multi-sensor or multi-source environment also often needs to be combined to
produce more accurate information, to derive better estimation, or to make more
knowledgeable decisions. In this chapter, we present a method, called Combinatorial
Fusion Analysis (CFA), for analyzing combination and fusion of multiple scoring. CFA
characterizes each Scoring system as having included a Score function, a Rank
function, and a Rank/score function. Both rank combination and score combination
are explored as to their combinatorial complexity and computational efficiency.
Combinatorial Fusion Analysis 33
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Information derived from the scoring characteristics of each scoring system is used to
perform system selection and to decide method combination. In particular, the rank/
score graph defined by Hsu, Shapiro and Taksa (Hsu et al., 2002; Hsu & Taksa, 2005)
is used to measure the diversity between scoring systems. We illustrate various
applications of the framework using examples in information retrieval and biomedical
informatics.
Introduction
Many problems in a variety of applications domains such as information retrieval, social
/ welfare / preference assignments, internet/intranet search, pattern recognition, multi-
sensor surveillance, drug design and discovery, and biomedical informatics can be
formulated as multi-variable (multi-criterion or multi-objective) classification, prediction,
learning, or optimization problems. To help obtain the maximum possible (or practical)
accuracy in calculated solution(s) for these problems, many groups have considered the
design and integrated use of multiple, (hopefully) complementary scoring schemes
(algorithms or methods) under various names such as multiple classifier systems (Ho,
2002; Ho, Hull, & Srihari, 1992, 1994; Melnik, Vardi, & Zhang 2004; Xu, Krzyzak, & Suen,
1992; Kittler & Alkoot, 2003), social choice functions (Arrow, 1963; Young, 1975; Young
& Levenglick, 1978), multiple evidences, Web page scoring systems or meta searches
(Aslam, Pavlu, & Savell, 2003; Fagin, Kumar, & Sivakumar, 2003; Diligenti, Gori, &
Maggini, 2004), multiple statistical analysis (Chuang, Liu, Brown, et al., 2004; Chuang,
Liu, Chen, Kao, & Hsu, 2004; Kuriakose et al., 2004), cooperative multi-sensor surveil-
lance systems (Collins, Lipton, Fujiyoshi, & Kanade, 2001; Hu, Tan, Wang, & Maybank,
2004), multi-criterion ranking (Patil & Taillie, 2004), hybrid systems (Duerr, Haettich,
Tropf, & Winkler, 1980; Perrone & Cooper, 1992), and multiple scoring functions and
molecular similarity measurements (Ginn, Willett, & Bradshaw, 2000; Shoichet, 2004;
Yang, Chen, Shen, Kristal, & Hsu, 2005). For convenience and without loss of generality,
we use the term multiple scoring systems (MSS) to denote all these aforementioned
schemes, algorithms, or methods. We further note the need for the word hopefully
above there are limited practical means of predicting which combinations will be
fruitful the problem we address in the remainder of this report.
The main purpose in constructing multiple scoring systems is to combine those MSSs
in order to improve the efficiency and effectiveness or increase the sensitivity and
specificity of the results. This purpose has been met; it has been demonstrated that
combining MSSs can improve the optimization results. Combination of multiple scoring
systems has been studied under different names such as classification ensemble (Ho,
2002; Ho et al., 1992, 1994; Kittler & Alkoot, 2003; Tumer & Ghosh, 1999; Xu et al., 1992),
evidence combination (Belkin, Kantor, Fox, & Shaw, 1995; Chuang, Liu, Brown, et al.,
2004; Chuang, Liu, Chen, et al., 2004), data / information fusion (Dasarathy, 2000; Hsu
& Palumbo, 2004; Hsu et al., 2002; Hsu & Taksa, 2005; Ibraev, Ng, & Kantor, 2001; Kantor,
1998; Kuriakose et al., 2004; Lee, 1997; Ng & Kantor, 1998, 2000), rank aggregation
(Dwork, Kumar, Naor, & Sivakumar, 2001; Fagin et al., 2003), consensus scoring (Ginn
et al., 2000; Shoichet, 2004; Yang et al., 2005), and cooperative surveillance (Collins et
34 Hsu, Chung, & Kristal
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
al., 2001; Hu et al., 2004). In addition, combination of MSSs has been also used in
conjunction with other machine learning or evolutional computation approaches such
as neural network and evolutional optimization (Garcia-Pedrajas, Hervas-Martinez, &
Ortiz-Boyer, 2005; Jin & Branke, 2005). We use the term combinatorial fusion to denote
all the aforementioned methods of combination.
Combination of multiple approaches (multiple query formulation, multiple retrieval
schemes or systems) to solving a problem has been shown to be effective in data fusion
in information retrieval (IR) and in internet meta search (Aslam et al., 2003; Belkin et al.,
1995; Dwork et al., 2001; Fagin et al., 2003; Hsu et al., 2002; Hsu & Taksa, 2005; Lee, 1997;
Ng & Kantor, 1998, 2000; Vogt & Cottrell, 1999). In performing classifier combination in
the pattern recognition (PR) domain, rules are used to combine the output of multiple
classifiers. The objective is to find methods (or rules) for building a hybrid classifier that
would outperform each of the individual classifiers (Ho, 2002; Ho et al., 1992, 1994; Melnik
et al., 2004). In protein structure prediction (PSP), results from different features are
combined to improve the accurate predictions of secondary classes or 3-D folding
patterns (C.-Y. Lin et al., 2005; K.-L. Lin et al., 2005a, 2005b). Biology may well represent
a major area of future needs with respect to combinatorial fusion and related concepts.
The last decade has seen an explosion in two data-driven concepts, so-called -omics level
studies and in silico approaches to modeling.
Omics approaches are approaches that attempt to take snapshots of an organism at a
specific level, for example, simultaneously measuring all the metabolites in a tissue and
reconstructing pathways. The -omics levels studies range from the four major areas (e.g.,
genomics the omics field of DNA analysis; transcriptomics the omics field of RNA
analysis; proteomicsthe omics field of protein analysis; metabolomics the omics
field of metabolite analysis) to very specific subareas, such as glycomics (omics
approaches to glycated proteins). Omics approaches represent a shift for two reasons:
(1) the amount of data inherent either prevents or forces modifications in traditional
data analysis approaches (e.g., t-tests being replaced by t-tests with false discovery rate
calculations); and (2) these omics level approaches lead to data-driven and/or hypoth-
esis-generating analyses, not (at least generally) to the testing of specific hypotheses.
Most importantly, by offering greatly improved ability to consider systems as a whole
and identify unexpected pieces of information, these approaches have opened new areas
of investigation and are, perhaps too optimistically, expected to offer fundamentally new
insights into biological mechanisms and fundamentally new approaches to issues such
as diagnostics and medical classification.
In silico simulations are also becoming a major focus of some biological studies. There
are at least two broad areas of simulation studies that are already playing major roles in
biological investigations:
in silico ligand-receptor (or drug-target) binding studies, and;
in silico simulations of physiological systems.
In each of these cases, the great advantage lies not in the qualitative change empowered
by technological advances, but in the quantitative savings of time and money. For
Combinatorial Fusion Analysis 35
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
example, obtaining and testing a small chemical library for binding or inhibition can
readily cost between $1 and $100 and up (per compound), considering assay costs and
obtaining (or synthesizing) the compounds. In contrast, once basic algorithms and
binding site models are in place, in silico screening is limited only by computational
costs, which are dropping exponentially. Similarly, in silico models of physiological
systems can be somewhat costly to develop, but they are capable of identifying potential
targets for intervention and to determine that other targets cannot work, saving tens or
hundreds of millions of dollars in failed trials.
Fully utilizing the potential power of these omics and in silico based approaches,
however, is heavily dependent on the quantity and quality of the computer resources
available. The increases in hardware capacity, readily available software tools, database
technology, and imaging and scanning techniques, have given the biomedical research
community large scale and diversified data sets and the ability to begin to utilize these
sets. The problem of how to manipulate, analyze, and interpret information from these
biomedical data is a challenging and daunting task. Gradually biologists are discovering
tools such as clustering, projection analyses, neural nets, genetic algorithms, genetic
programs, and other machine learning and data mining techniques built in other fields,
and these tools have found their way to the biomedical informatics domain.
A major issue here lies in the choice of appropriate tools; there is, for example, no clear
informatics pipeline for omics level studies. One can readily pick up many of software
tools, but their applicability for a given problem can be difficult to determine a priori. For
example, clustering is often powerful for microarray experiments, but we have shown it
is very limiting within some metabolomics experiments (Shi, et al., 2002a, 2002b). Similarly,
principal components analyses and its supervised cousin, Soft Independent Modeling
of Class Analogy (SIMCA) seem to work very well within defined cohorts, but they
breakdown in complex multi-cohort studies (Shi et al., 2002a, 2002b, 2004), a problem
apparently solvable using discriminant-based projection analyses (Paolucci, Vigneau-
Callahan, Shi, Matson, & Kristal, 2004). Thus, the choice of an analysis method must
often be determined empirically, in a slow, laborious step-wise manner. For the purpose
of this article, we will break these biological studies into two broad areas, one of
description (i.e., how do we best understand the data in front of us) and one of prediction
(i.e., how can we use this information to make predictions about, for example, which ligand
will bind or which person will become ill questions which in many ways are mathemati-
cally equivalent).
The report focuses on mathematical issues related to the latter of these two broad issues,
i.e., can we use CFA to improve prediction accuracy? The goal is a complex zero-sum
game we ideally want to further save time by reducing both false positives and false
negatives, while simultaneously increasing accuracy on continuous measurements (e.g.
binding strengths). In practice, it is almost certain that some trade-offs will have to be
made. To be useful we must, at a minimum, identify an approach which enables some
a priori decisions to be made about whether such fusion approaches are likely to
succeed. Otherwise we have done nothing but to add a layer of complexity between where
we are and where we need to be, without removing the time-consuming, laborious, and
inherently limited stages of empirical validation. The system we choose to focus on is
in virtual screening (VS), the use of in silico approaches to identify potentially optimal
binding ligands. VS is an area in which consensus scoring has been used in drug design
36 Hsu, Chung, & Kristal
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
and discovery and molecular similarity measurement for years, and in which data fusion
approaches have recently been a major focus of efforts. (see Shoichet, 2004; Yang et al.,
2005; and their references).
In this chapter, we present a method called combinatorial fusion analysis (CFA) which
uses the Cayley network Cay(S
n
, T
n
) on the symmetric group S
n
with generating set T
n
(Biggs & White, 1979; Grammatikakis, Hsu, & Kraetzl, 2001; Heydemann, 1997; Marden,
1995). We study the fusion and combination process in the set R
n
, called score space,
where R is the set of real numbers, and in the set S
n
, called rank space. In the next section,
we define rank and score functions and describe the concept of a rank/score graph
defined by Hsu, Shapiro and Taksa (Hsu et al., 2002; Hsu & Taksa, 2005). The combination
of two scoring systems (each with rank and score functions) is discussed and analyzed
in the context of a Cayley network. The next section also entails the property of combined
scoring system, performance evaluation and diversity issues, and various kinds of fusion
algorithms on rank functions. The section Data Mining Using Combinatorial Fusion
deals with mining and searching databases and the Web, and includes examples from
application domains in biomedical informatics, in particular in virtual screening of
chemical libraries and protein structure prediction. Then, in section Conclusion and
Discussion, we summarize our results with some discussion and future work.
Combinatorial Fusion
Multiple Scoring Systems
Successfully turning raw data into useful information and then into valuable knowledge
requires the application of scientific methods to the study of the storage, retrieval,
extraction, transmission, diffusion, fusion/combination, manipulation, and interpreta-
tion at each stage of the process. Most scientific problems are multi-faceted and can be
quantified in a variety of ways. Among many methodologies and approaches to solve
complex scientific problems and deal with large datasets, we only mention three: (a)
classification, (b) clustering, and (c) similarity measurement. Hybrid methods combining
(a), (b), and (c) have been used.
Large data sets collected from multi-sensor devices or multi-sources or generated by
experiments, surveys, recognition and judging systems are stored in a data grid G(n, m,
q) with n objects in D = {d
1
, d
2
, , d
n
}, m features/attributes/indicators/cues in G = {a
1
,
a
2
, , a
m
} and, possibly, q temporal traces in T = {t
1
, t
2
, , t
q
}. We call this three-
dimensional grid the data space (see Figure 1).
Since both m and q can be very big and the size of the datasets may limit the utility of
single informatics approaches, it is difficult to use/design a single method/system
because of the following reasons:
1. Different methods/systems are appropriate for different features / attributes /
indicators / cues and different temporal traces. There have been a variety of
Combinatorial Fusion Analysis 37
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
different methods/systems used/proposed in the past decades such as statistical
analysis and inference (e.g., t-test, non-parametric t-test, linear regression, analy-
sis of variance, Bayesian systems), machine learning methods and systems (e.g.,
neural networks, self-organizing map, support vector machines), and evolutional
computations (e.g., genetic algorithms, evolutionary programming).
2. Different features / attributes / indicators / cues may use different kinds of
measurements. Many different measurements have been used such as variates,
intervals, ratio-scales, and binary relations between objects. So in the data grid,
each column on the D-G plane a
j
can be an assignment of a score or a rank to each
object d
i
. For example, M(i, j) is the score or rank assigned to object d
i
by feature/
attribute/indicator/cue a
j
.
3. Different methods/systems may be good for the same problem with different data
sets generated from different information sources/experiments. When different
data sets generated from different information sources or experiments, different
methods/systems should be used according to the style of the source and the
nature of the experiment.
4. Different methods/systems may be good for the same problem with the same data
sets generated or collected from different devices/sources. Even when the data
sets are the same in the same experiments, different methods/systems should be
adopted according to a variety of multi-sensor/multi-sources.
Due to the complexity of the problem involved, items (3) and (4) indicate that each single
system/method, when applied to the problem, can be improved in performance to some
extent, but it is difficult to become perfect. Item (1) indicates that performance of a single
system / method may be optimal for some features/attributes/indicators/cues, but may
downgrade its performance for other features/attributes/indicators/cues.
Figure 1. Data space G(n, m, q)

t
q
t
2
t
1
d
1
d
2
d
n
a
1
a
2
a
m
DATA
SPACE
38 Hsu, Chung, & Kristal
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Recently, it has been demonstrated that combination of multiple systems/methods
improves the performance of accuracy, precision, and true positive rate in several
domains such as pattern recognition (Brown et al, 2005; Duerr et al., 1980; Freund, Iyer,
Schapire, & Singer, 2003; Garcia-Pedrajas et al., 2005; Ho, 2002; Ho et al., 1992, 1994; Jin
& Branke, 2005; Kittler & Alkoot, 2003; Perrone & Cooper, 1992; Triesch & von der
Malsburg, 2001; Tumer & Ghosh, 1999; Xu et al., 1992), microarray gene expression
analysis (Chuang, Liu, Brown, et al., 2004; Chuang, Liu, Chen, et al., 2004; Kuriakose et
al., 2004), information retrieval (Aslam et al., 2003; Belkin et al., 1995; Diligenti et al., 2004;
Dwork et al., 2001; Fagin et al., 2003; Hsu et al., 2002; Hsu & Taksa, 2005; Kantor, 1998;
Lee, 1997; Ng & Kantor, 1998, 2000), virtual screening and drug discovery (Shoichet,
2004; Yang et al., 2005; and references in both), and protein structure prediction (C.-Y.
Lin et al., 2005; K.-L. Lin et al., 2005a, 2005b).
There have been special meetings (such as the Workshop on Multiple Classifier
Systems), conferences (such as International Conference on Information Fusion),
societies (such as International Society of Information Fusion), and journals (e.g.:
Information Fusion [Dasarathy, 2000]) dedicated to the scientific study of fusion/
combination. The main title of the featured article The Good of the Many Outweighs the
Good of the One, by Corne, Deb, Fleming, & Knowles (2003) in IEEE Neural Networks
Society typifies the scientific and philosophical merits of and motivations for fusion/
combination.
Each system/method offers the ability to study different classes of outcomes, e.g., class
assignment in a classification problem or similarity score assignment in the similarity
measurement problem. In this chapter, we view the outcome of each system/method as
a scoring system A which assigns (a) an object as a class among all objects in D, (b) a
score to each object in D, and (c) a rank number to each object in D. These three outcomes
were described as the abstract, score, and rank level respectively by Xu et al. (1992). We
now construct the system grid H(n, p, q) with n objects in D = {d
1
, d
2
, , d
n
}, p systems
Figure 2. System space H(n, p, q)

t
q
t
2
t
1
d
1
d
2
d
n
A
1
A
2
A
p
SYSTEM
SPACE
Combinatorial Fusion Analysis 39
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
in H = {A
1
, A
2
, , A
p
}, and possibly, q temporal traces in T = {t
1
, t
2
, , t
q
}. We call this
three dimensional grid the system space for the multiple scoring systems (see Figure 2).
In the next section, we will define score function, rank function and the rank/score
function for each scoring system A. In the section following the next, rank and score
combination are defined and studied. Section Method of Combinatorial Fusion deals
with performance evaluation criteria and diversity between and among different scoring
systems.
Score Function, Rank Function and the Rank/Score
Graph
Let D = {d
1
, d
2
, ..., d
n
} be a set of n objects and N = {1, 2, 3, , n} be the set of all positive
integers less than or equal to n. Let R be the set of real numbers. We now state the
following three functions that were previously defined and studied by Hsu, Shapiro and
Taksa (2002) and Hsu and Taksa (2005).
Definition 1
(a) Score Function: A score function s is a function from D to R in which the function
s assigns a score (a real number) to each object in D. In a more formal way, we write
s: D R such that for every d
i
in D, there exists a real number s(d
i
) in R corresponding
to d
i
.
(b) Rank Function: A rank function r is a function from D to N such that the function
r : D N maps each element d
i
in D to a natural number r(d
i
) in N. The number r(d
i
)
stands for the rank number r(d
i
) assigned to the object d
i
.
(c) Rank/Score Function: Given r and s as rank and score function on the set D of
objects respectively, the rank/score function f is defined to be f : N R such that
f(i) = ) )( (
1
i r s

o = s(r
1
(i)). In other words, the score function s = r f o is the
composite function of the rank/score function and the rank function.
(d) Rank/Score Graph: The graph representation of a rank/score function.
We note that in several application domains, one has to normalize the score function
values before any combination can be performed. Hence it is quite natural to define the
two functions s and f in the way that each of them has [0, 1] = {x | x in R, 0 x 1} instead
of R as their function range. Other intervals of real numbers can be also used depending
on the situation and environment. We also note that since the rank function r' defined by
Hsu, Shapiro and Taksa (see Hsu et al., 2002; Hsu & Taksa, 2005) is the inverse of r defined
above, the rank/score function f would be such that f = s

r' instead of f = s

r'
1
.
At this point, we would like to comment on some perspectives regarding rank vs. score
function. Although these two functions both deal with the set of objects under study
(in the case of PR, IR, VS and PSP, these would be classes, documents, chemical
40 Hsu, Chung, & Kristal
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
compounds, and classes or folding patterns, respectively), their emphases are different.
The Score function deals more with the detailed data level while Rank function is more
relevant for or related to the decision level. In theory, the Score function depends more
on the variate data in the parametric domain while the Rank function depicts more on the
ordinal data in the non-parametric fashion. The comparison can go on for a long time as
score vs. rank, data level vs. decision level, variate data vs. ordinal data, and parametric
vs. non-parametric. Historically and from the discipline perspective, scores are used in
sciences, engineering, finance, and business, while ranks are used in social choices,
ordinal data analysis and decision science. However, in biomedical informatics, since
the data collected is large (and of multiple dimension) and the information we are seeking
from biological and physiological systems is complex (and multi-variable), the informa-
tion we find (or strive to find) from the relation between score and rank function would
become valuable in biological, physiological, and pharmaceutical study.
The concept of a rank/score graph which depicts the graph representation of a rank/score
function has at least three characteristics and advantages:
Remark 1
(a) Efficiency: When a score function s
A
is assigned resulting from scoring system A
by either lab work or field study (conducted in vivo or in vitro), treating s
A
as an
array and sorting the array of scores into descending order would give rise to the
rank function r
A
. The rank/score function can be obtained accordingly. If there are
n objects, this transformation takes O(n log n) steps.
(b) Neutrality: Since the rank/score function f is defined from N to R or from N to [0,
1], it does not depend on the set of objects D = {d
1
, d
2
, , d
n
}. Hence the rank/score
function f
A
of a scoring system A exhibits the behavior of the scoring system (scorer
or ranker) A and is independent of who (or which object) has what rank or what
Figure 3. Score function, Rank function and Rank/Score function of A and B
D d
1
d
2
d
3
d
4
d
5
d
6
d
7
d
8
d
9
d
10

s
A
(d
i
) 4 10 4.2 3 6.4 6.2 2 7 0 1
r
A
(d
i
) 6 1 5 7 3 4 8 2 10 9
(a) Score and Rank function for A

D d
1
d
2
d
3
d
4
d
5
d
6
d
7
d
8
d
9
d
10

s
B
(d
i
) 4 7 3 1 10 8 5 6 9 2
r
B
(d
i
) 7 4 8 10 1 3 6 5 2 9
(b) Score and Rank function for B

N 1 2 3 4 5 6 7 8 9 10
f
A
(i) 10 7 6.4 6.2 4.2 4 3 2 1 0
f
B
(i) 10 9 8 7 6 5 4 3 2 1
(c) Rank/Score function for A and B
Combinatorial Fusion Analysis 41
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
score. The rank/score function f also fills the gap of relationship between the three
sets D, N and R.
(c) Visualization: The graph of the rank/score function f
A
can be easily and clearly
visualized. From the graph of f
A
, it is readily concluded that the function f
A
is a non-
increasing monotonic function on N. The thrust of this easy-to-visualize property
is that comparison (or difference) on two functions f
A
and f
B
can be recognized by
drawing the graph of f
A
and f
B
on the same coordinate system.
Two examples of score functions, rank functions (derived from score function), rank/
score functions with respect to scoring systems A and B are illustrated in Figure 3 and
the rank / score graphs of f
A
and f
B
are included in Figure 4, where D = {d
i
| i = 1 to 10} and
s(d
i
) is in [0, 10].
Rank and Score Combination
As mentioned earlier, the combinations (or fusions) of multiple classifiers, multiple
evidences, or multiple scoring functions in the PR, IR, VS, and PSP domain has gained
tremendous momentum in the past decade. In this section, we deal with combinations of
two functions with respect to both score and rank combinations. The following defini-
tions were used by Hsu, Shapiro, and Taksa (2002) and Hsu and Taksa (2005).
Definition 2
(a) Score Combinations: Given two score functions s
A
and s
B
, the score function of the
score combined function s
F
is defined as s
F
(d) = )] ( ) ( [
2
1
d s d s
B A
+ for every object
d in D.
Figure 4. Rank/Score graphs of f
A
and f
B
0
2
4
6
8
10
12
1 2 3 4 5 6 7 8 9 10
rank
score
f








f
B
(i)

f
A
(i)
42 Hsu, Chung, & Kristal
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
(b) Rank Combinations: Given two rank functions r
A
and r
B
, the score function of the
rank combined function s
E
is defined as s
E
(d) = )] ( ) ( [
2
1
d r d r
B A
+ for every object d
in D.
Since each of the scoring systems A and B has their score and rank function s
A
, r
A
and
s
B
, r
B
respectively, each of the combined scoring systems E and F (by rank and by score
combination) has s
E
, r
E
and s
F
, r
F
, respectively. These functions can be obtained as
follows. The score function of the rank combination s
E
is obtained from r
A
and r
B
using
rank combination. Sorting s
E
into ascending order gives rise to r
E
. The score function of
the score combination s
F
is obtained from s
A
and s
B
using score combination. Sorting s
F
into descending order gives rise to r
F
. Hence for scoring systems E and F, we have the
score function and rank function s
E
, r
E
and s
F
, r
F
respectively.
We note the difference between converting from s
E
to r
E
and that from s
F
to r
F
. Since the
scoring systems E and F are rank and score combination of A and B respectively, the
transformation from s
E
to r
E
is by sorting into ascending order while that from s
F
to r
F
is
by sorting into descending order. The fusion architecture given by Hsu, Shapiro and
Taksa (Figure 5 in Hsu & Taksa, 2005) depicts the framework for the fusion method we
use. Figure 5(a) and (b) show the rank and score function of the rank and score
combination E and F (both related to the scoring systems A and B in Figure 3)
respectively. The reader might have noticed that in the combination (rank or score), a
simple average combination was used. Combination using weighted proportion on A and
B is studied in Hsu and Palumbo (2004).
Recall that the main purpose of doing fusion/combination is whether or not the fused
scoring system can outperform each individual scoring system in isolation. The concept
of performance has to be defined for a given study (or, arguably, a given class of
studies), so that it is meaningful to say outperform. Recall also that our fusion
framework is for application domains such as PR, IR, VS and PSP. In PR, each classifier
Figure 5. Score and Rank function of E and F

D d
1
d
2
d
3
d
4
d
5
d
6
d
7
d
8
d
9
d
10

s
E
(d
i
)

6.5 2.5 6.5 8.5 2 3.5 7 3.5 6 9
r
E
(d
i
)

6 2 4 9 1 3 8 7 5 10

(a) Score and rank function of E

D d
1
d
2
d
3
d
4
d
5
d
6
d
7
d
8
d
9
d
10
s
F
(d
i
) 4.0 8.5 3.6 2.0 8.2 7.1 3.5 6.5 4.5 1.5
r
F
(d
i
) 6 1 7 9 2 3 8 4 5 10

(b) Score and rank functions of F

Combinatorial Fusion Analysis 43
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
produces a ranking of a set of possible classes. When given an image, each of the
classifiers including the combined gives rise to a rank function and the class that was
ranked at the top is predicted by the classifier as the identity of the image. So the
performance (or accuracy) of the classifier A, written P(A), is the percentage of times
that this classification gives a correct prediction. From the perspective of biology and
medicine, P(A) may be defined as the reduction of false negatives (e.g., in cancer
diagnostic) or the reduction of false positives (e.g., in screens of related compounds
where follow-up is expensive).
In IR, a query Q is given. Then each of the ranking algorithms calculates similarity
between the query and the documents in the database of n documents D = {d
1
, d
2
. . ., d
n
}.
A score function s
A
for algorithm A is assigned and a rank function r
A
is obtained. The
performance of the algorithm for the query, written P
q
(A), is defined as the precision of
A at q with respect to the query Q. More specifically, the following is defined and widely
used in the information retrieval community.
Definition 3
(a) Let Rel(Q) be the set of all documents that are judged to be relevant with respect
to the query Q. Let |Rel(Q)| = q for some q, 0 q n. On the other hand, let A
(k)
= {d
| d in D and r
A
(d) k}.
(b) Precision of A at q. The performance of the scoring system A with respect to the
query Q is defined to be P
q
(A) = |Rel(Q) A
(q)
| / q, where q = |Rel(Q)|.
In VS, molecular compound libraries are searched for the discovery of novel lead
compounds for drug development and/or therapeutical treatments. Let L be the total
number of active ligands and T the total number of compounds in the database. Let L
h
be the number of active ligands among the T
h
highest ranking compounds (i.e., the hit
list). Then the goodness-of-hit (GH) score for a scoring system A is defined as (see Yang
et al., 2005):
( )

+
=
L T
L T
L T
T L L
A GH
h h
h
h h
1
4
3
) (
.
The GH score ranges from 0.0 to 1.0, where 1.0 represents a perfect hit list. The GH score
as defined in Yang et al. (2005) contains a coefficient to penalize excessive hit list size.
We will come back to this topic in section Virtual Screening and Drug Discovery.
In the protein structure prediction problem (PSP), one wishes to extract structural
information from the sequence databases as an alternative to determine the 3-D structure
of a protein using the X-ray diffraction or NMR (resource and labor intensive, expensive,
and, in practice, often difficult or impossible, particularly when one deals with variably
modified proteins). Given a protein sequence, the objective is to predict its secondary
structure (class) or its 3-D structures (folding patterns). The standard performance
44 Hsu, Chung, & Kristal
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
evaluation for the prediction of the n
i
protein sequences T
i
= {t
1
, t
2
, ..., t
n
i
} for the ith class
or ith folding pattern is the percentage accuracy rate Q
i
= Q(T
i
) = (p
i
/ n
i
) 100, where n
i
is the number of testing proteins in the ith class or ith folding pattern and p
i
is the number
of proteins within the n
i
protein sequences correctly predicted. The overall prediction
accuracy rate Q is defined as

=
=
k
i i i
Q q Q
1
, where q
i
= n
i
/ N, N is the total number of
proteins tested (i.e., N =

=
k
i i
n
1
) and k is the number of classes or folding patterns.
Now we come back to the fundamental issue of when the combined scoring system E or
F outperforms its individual scoring system A and B. The following two central questions
regarding combination and fusion were asked by Hsu, Shapiro, and Taksa (2002) and Hsu
and Taksa (2005):
Remark 2
(a) For what scoring systems A and B and with what combination or fusion algorithm,
P(E) (or P(F)) max{P(A), P(B)}, and
(b) For what A and B, P(E) P(F)?
Four important issues are central to CFA: (1) What is the best fusion algorithm / method
to use? (2) Does the performance of E or F depend very much (or how much) on the
relationship between A and B? (3) Given a limited number of primary scoring systems, can
we optimize the specifics of the fusion? and (4) Can we answer any or all of the previous
three issues without resorting to empirical validation? The general issue of combination
algorithm / method will be discussed in the next section. In this section, we simply use
the average combination regardless of rank or score combination.
Arguably, issues (1) and (2) may be considered primary issues, and issues (3) and (4)
secondary or derivative. We propose that issue (2) is as, if not more, important as issue
(1). It has been observed and reported extensively and intensively that the combined
scoring system E or F performs better than each individual scoring system A and B when
A and B are different, diverse, or orthogonal. In particular, Vogt and Cottrell (1999)
studied the problem of predicting the performance of a linearly combined system and
stated that the linear combination should only be used when the individual systems
involved have high performance, a large overlap of relevant documents, and a small
overlap of non-relevant documents. Ng and Kantor (2000) identified two predictive
variables for the effectiveness of the combination: (a) the output dissimilarity of A and
B, and (b) the ratio of the performance of A and B. Then Hsu, Shapiro and Taksa (2002)
and Hsu and Taksa (2005) suggested using the difference between the two rank/score
functions f
A
and f
B
as the diversity measurement to predict the effectiveness of the
combination. This diversity measurement has been used in microarray gene expression
analysis (Chuang, Liu, Brown, et al., 2004; Chuang, Liu, Chen, et al., 2004) and in virtual
screening (Yang et al., 2005). We will discuss in more details the use of graphs for rank/
score functions f
A
and f
B
in the diversity measurement between A and B in section Virtual
Screening and Drug Discovery.
Combinatorial Fusion Analysis 45
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Method of Combinatorial Fusion
As defined in section Score Function, Rank Function and the Rank/Score Graph, a rank
function r
A
of a scoring system A is an one-one function from the set D = {d
1
, d
2
, , d
n
}
of n objects to the set N of positive integers less than or equal to n. When considering
r
A
as a permutation of N, we have r
A
(D) = [r
A
(d
1
), r
A
(d
2
), , r
A
(d
n
)]. Without loss of clarity,
we write r instead of r
A
.
Since the set of all permutations of the set N is a group with the composition

as the
binary operation, called symmetric group S
n
, r(D) as a permutation of N is an element in
the group S
n
. Each element a in S
n
has two different ways of representation as [(1), (2),
..., (n)]=[
1
,
2
, ... ,
n
]=[
1

2

3
...
n
] called standard representation and as the product
of disjoint cycles each consisting of elements from N called cycle representation. The
general concept of a Cayley graph (or network) can be found in the book and article
(Grammatikakis et al., 2001; Heydemann, 1997).
Definition 4
(a) Example of Permutations: The permutation a(i) = 2i for i = 1, 2, 3 and 2i 7 for i
= 4, 5, 6 in S
6
can be written as [2 4 6 1 3 5] and (124) (365). The permutation b(i) =
i for i = 1, 4, 5, 6 and b(2) = 3, b(3) = 2 can be written as b = [1 3 2 4 5 6] = (23). Note
that in the cycle representation, we ignore the cycles that are singletons.
(b) In the group S
n
, the set of n 1 adjacent transpositions such as b= (23) in S
6
is
denoted as T
n
. In other words, T
n
consists of all cycles of length 2 which are adjacent
transpositions and T
n
= {(1 2), (2 3), , (n 1 n)} is a subset of S
n
. With this in mind,
we can define a Cayley network based on S
n
and T
n
:
Cayley network Cay (S
n
, T
n
): The Cayley network Cay(S
n
, T
n
) is a graph G(V, E) with
the node set V = S
n
and arc set E = {(, t o ) | in S
n
and t in T
n
}.
The concept of a Cayley network extends the group structure in S
n
to the graph structure
in Cay(S
n
, T
n
). By doing so, a distance measure between any two permutations (and hence
any two rank functions) is well defined in the context of applications that will prove to
be very useful in biomedical informatics. In fact, it has been mentioned by Hsu, Shapiro
and Taksa (2002) and Hsu and Taksa (2005) that the graph distance in Cay(S
n
, T
n
) is the
same as Kendalls tau distance in the rank correlation analysis (RCA) (see e.g., Kendall
&Gibbons, 1990; Marden, 1995). This striking coincidence supports the importance and
usefulness of using Cayley networks as a framework for fusion and combination.
Moreover, we point out that the combinatorial fusion we proposed and the rank
correlation studied by many researchers in the past bear similarity but have differences.
They are very similar because they all study ranks although one treats ranks as a function
and the other treats them as ordinal data or the order of the values of a random variable.
On the other hand, they are quite different. The CFA (combinatorial fusion analysis)
views the set S
n
as a rank space aiming to produce a dynamic process and reasonable
46 Hsu, Chung, & Kristal
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
algorithms to reach a better combined rank function (or in general, combined scoring
system). The RCA (rank correlation analysis) views the set S
n
as a population space
aiming to calculate the static correlation and significant P-value to reach a hypothesis
testing result.
Suppose we are given the following p rank functions A
j
obtained from the data set D =
{d
1
, d
2
, , d
n
}, A
j
, j = 1, 2, , p:
A
j
= (a
1j
, a
2j
, a
3j
, , a
nj
)
t
,
where V
t
is the transpose of the vector V. Let M
r
be the matrix, called rank matrix, with
dimension n p such that M
r
(i, j) = M(i, j) = the rank assigned to the object d
i
by scoring
system A
j
. Now, we describe the rank version of the combinatorial fusion problem.
Definition 5
Combinatorial Fusion Problem (rank version): Given p nodes A
j
, j = 1, 2, , p in the
Cayley network Cay(S
n
, T
n
) with respect to n objects D = {d
1
, d
2
, , d
n
}, find a node
A
*
in S
n

which performs as good as or better than the best of A
j
s in the sense of
performance as defined as accuracy, precision or goodness-of-hit in IR, VS and PSP
described previously.
There are several ways to find the candidates for the node A
*
when given the p nodes
A
j
, j = 1, 2, , p in S
n
. We briefly describe, in Definition 6, the following six types of methods
/ algorithms to fuse the given p nodes and generate the candidate node. All of these
approaches aim to construct a score function which, when sorted, would lead to a rank
function.
Definition 6
(a) Voting: Scoring function s
*
(d), d in D. The score of the object d
i
, s
*
(d
i
), is obtained
by a voting scheme among the p values M(i, j), j = 1, 2, , p. These include max,
min, and median.
(b) Linear Combination: These are the cases that s
*
(d
i
) is a weighted linear combina-
tion of the M(i, j)s, i.e.,

=
=
p
j j i
j i M w d s
1
*
) , ( ) ( for some weighted function so that

=
=
p
j j
w
1
1. When w
j
= 1/p, s
*
(d
i
) is the average of the ranks M(i, j)s, j = 1, 2, . . .,
p.
(c) Probability Method: Two examples are the Bayes rule that uses the information
from the given p nodes A
j
, j = 1, 2, , p to predict the node A
*
, and the Markov Chain
method that calculates a stochastic transition matrix.
(d) Rank Statistics: Suppose that the p rank functions are obtained by the p scoring
systems or observers who are ranking the n objects. We may ask: what is the true
ranking of each of the n objects? Since the real (or true) ranking is difficult to come
Combinatorial Fusion Analysis 47
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
by, we may ask another question: What is the best estimate of the true ranking
when we are given the p observations? Rank correlation among the p rank functions
can be calculated as W = 12S / [p
2
(n
3
- n)], where 4 / ) 1 (
2 2
1
2
+ =

=
n np R S
n
i i
and

=
=
p
j i
j i M R
1
) , ( . The significance of an observed value of W is then tested in the
(n!)
p
possible sets of rank functions.
(e) Combinatorial Algorithm: For each of the n objects and its set of p elements {M(i,
j) | j = 1, 2, , p} = C as the candidate set for s
*
(d
i
), one combinatorial algorithm
considers the power set 2
C
and explores all the possible combinatorial combina-
tions. Another algorithm treats the n objects as n vectors d
i
= (a
i1
, a
i2
, , a
ip
) where
a
ij
= M(i, j), i = 1, 2, , n and 1 a
ij
n. It then places these n vectors in the context
of a partially ordered set (Poset) L consisting of all the n
p
vectors (a
i1
, a
i2
, , a
ip
),
i = 1, 2, , n and a
ij
in N. The scores s
*
(d
i
), i = 1, 2, , n is then calculated based
on the relative position of the vector d
i
in the Poset L.
(f) Evolutional Approaches: Genetic algorithms and other machine learning tech-
niques such as neural networks and support vector machines can be used on the
p rank functions to process a (large) number of iterations in order to produce a rank
function that is closest to the node A
*
.
The voting schemes in Definition 6 (a) have been used in social choice functions (Arrow,
1963; Young & Levenglick, 1978). Linear combination and average linear combination in
6(b), due to their simplicity, have been widely used in many application domains (Kendall
& Gibbons, 1990; Kuriakose et al., 2004; Hsu & Palumbo, 2004; Hsu et al., 2002; Hsu &
Taksa, 2005; Vogt & Cottrell, 1999). In fact, the concept of Borda count, used by Jean-
Charles de Borda of the LAcademie Royale des Sciences in 1770, is equivalent to the
average linear combination. Dwork et al. (2001) used Markov chain method to aggregate
the rank functions for the Web. As described in Definition 6(d), the significance of S
depends on the distribution of S in the (n!)
p
possible set of rank functions. Due to the
manner that S is defined, it may be shown that the average linear combination gives a
best estimate in the sense of Spearmans rho distance (see Kendall & Gibbons, 1990,
Chapter 6). Combinatorial algorithms stated in Definition 6(e) have been used in Mixed
Group Ranks and in Rank and Combine method by researchers (Chuang, Liu, Brown, et
al., 2004; Chuang, Liu, Chen, et al., 2004; Melnik et al., 2004). Although genetic algorithms
such as GemDOCK and GOLD were used to study the docking of ligands into a protein,
the authors in Yang et al. (2005) use linear combination and the rank/score graph as a
diversity measurement. We will discuss the application in more details in next section.
Definition 6 lists six different groups of methods/algorithms/approaches for performing
combination. Here we return to the second issue raised by Remark 2. That is: What are
the predictive variables / parameters / criteria for effective combination? In accordance
with Remark 2, we focus on two functions A and B (i.e. p = 2) at this moment although the
methods / algorithms / approaches in Definition 6, are able to deal with the multiple
functions (p 2). We summarize, in Definition 7, the two variables for the prediction of
effective combination among two scoring systems A and B (Chuang, Liu, Brown, et al.,
2004; Chuang, Liu, Chen, et al., 2004; Hsu et al., 2002; Hsu & Taksa, 2005; Ng & Kantor,
2000; Vogt & Cottrell, 1999; Yang et al., 2005).
48 Hsu, Chung, & Kristal
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Definition 7
(a) The performance ratio, P
l
/ P
h
, measures the relative performance of A and B where
P
l
and P
h
are the lower performance and higher performance among {P(A), P(B)}
respectively.
(b) The bi-diversity between A and B, d
2
(A, B), measures the difference / dissimilarity
/ diversity between the two scoring systems A and B.
We note that in order to properly use diversity d
2
(A, B) as a predictive parameter for
effective combination of functions A and B, d
2
(A, B) might be defined to reflect different
combination algorithms and different domain applications. However, for the diversity
measurement to be effective, it has to be universal at least among a variety of data sets
in applications domain. Diversity measurement between two scoring systems A and B,
d
2
(A, B) have been defined and used (see Chuang, Liu, Brown, et al., 2004; Chuang, Liu,
Chen, et al., 2004; Ng & Kantor, 2000; Yang et al., 2005). We summarize in Definition 8.
Definition 8
Diversity Measure: The bi-diversity (or 2-diversity) measure d
2
(A, B) between two
scoring systems A and B can be defined as one of the following:
(a) d
2
(A, B) = d(s
A
, s
B
), the distance between score functions s
A
and s
B
. One example of
d(s
A,
s
B
) is the covariance of s
A
and s
B
,

Cov(s
A
, s
B
), when s
A
and s
B
are viewed as two
random variables,
(b) d
2
(A, B) = d(r
A
, r
B
), the distance between rank functions r
A
and r
B
. One example of
d(r
A
, r
B
) is the Kendalls tau distance as we defined in S
n
, and
(c) d
2
(A, B) = d(f
A
, f
B
), the distance between rank/score functions f
A
and f
B
.
We note that diversity measure for multiple classifier systems in pattern recognition and
classification has been studied extensively (Kuncheva, 2005).
Definition 9
In the data space G(n, m, q), m = 2, defined in Figure 1, given a temporal step t
i
in T = {t
1
,
t
2
, , t
q
} and the two scoring systems A and B, we define:
(a)
=
j B A t
j f j f B A d
i
| ) ( ) ( | ) , ( , where j is in N = {1, 2, , n}, as the function value of
the diversity score function d
x
(A, B) for t
i
;
(b) if we let i vary and fix the system pair A and B, then s
(A,B)
(x) is the diversity score
function, defined as s
(A,B)
(t
i
) = ) , ( B A d
i
t
, from T = {t
1
, t
2
, , t
q
} to R;
(c) sorting s
(A,B)
(x) into ascending order leads to the diversity rank function r
(A,B)
(x)
from T to {1, 2, , q}; and then
Combinatorial Fusion Analysis 49
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
(d) the diversity rank/score function f
(A,B)
(j) can be obtained as f
(A,B)
(j)= )( (
1
) , ( ) , (
r s
B A B A

o
) )( j = )) ( (
1
) , ( ) , (
j r s
B A B A

, where j is in {1, 2, , q};
(e) the diversity rank/score graph (or diversity graph) is the graph representation of
the diversity rank/score function f
(A,B)
(j) from {1, 2, , q} to R.
We note the difference between the rank/score function and the diversity rank/score
function. In the definition of rank/score function f
A
: N [0, 1] (see Definition 1(c)), the
set N is different from the set D which is in turn the set of objects (classes, documents,
ligands, and classes or folding patterns). The set N is used as the index set for the rank
function values. The rank/score function f
A
so defined describes the scoring (or ranking)
behavior of the scoring system A and is independent of the objects under consideration.
The diversity rank/score function (see Definition 9(d)) f
(A,B)
(j) is defined from Q = {1, 2,
. . ., q} to R (or [0, 1]). The set Q is different from the set T = {t
1
, t
2
, , t
q
} which is the set
of temporal steps under study. The set Q is used as the index set for the diversity rank
function values. The diversity rank/score function f
(A,B)
so defined describes the diversity
trend of the pair of scoring systems A and B and is independent of the specific temporal
step t
i
for some i under study.
Data Mining Using
Combinatorial Fusion
In this section, we present three examples of data mining using combinatorial fusion as
defined in the previous section. These three examples are from applications in informa-
tion retrieval (IR) (Hsu et al, 2002; Hsu & Taksa, 2005; Ng & Kantor, 1998, 2000),
consensus scoring in virtual screening (VS) (Yang et al., 2005), and protein structure
prediction (PSP) (C.-Y. Lin et al., 2005; K.-L. Lin et al., 2005a, 2005b). But before we
concentrate on special cases, we will further discuss the relation between rank and score
functions as defined in the previous section.
Rank/Score Transfer
Let M
r
(i, j) be the rank matrix defined before, where M
ij
in M
r
(i, j) is the rank assigned
to the object d
i
by scoring system A
j
. Let M
s
(i, j) be the n p score matrix defined similarly
with respect to the p score functions so that M
ij
(without ambiguity) in M
s
(i, j) is the score
value assigned to the object d
i
by scoring system A
j
. The algorithms and approaches
described in Definition 6 can be applied to the rank matrix M
r
. However, some of these
algorithms have been also applied to the score matrix M
s
. Up to this point in this chapter,
we have emphasized rank combination algorithms and considered ranks of objects as the
basic data of a given situation / experiment / work, regardless of the manner in which they
were obtained. However, in many situations, the ranking takes place according to the
score values of a variable or variate. It is, therefore, of considerable interest to study the
50 Hsu, Chung, & Kristal
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
relationship between M
r
(i, j) and M
s
(i, j). As we mentioned before, M
r
(i, j) can be derived
from M
s
(i, j) by sorting each column, A
j
, into descending order and assigning higher value
with higher rank (i.e., smaller number). One of the interesting questions is that: Is there
any difference between the information represented by M
s
(i, j) and that by M
r
(i, j)? In 1954,
A. Stuart showed the following (see Kendall & Gibbons, 1990, Chapters 9, 10):
Remark 3 (Correlation between Scores and Ranks)
With n = 25, the correlation between scores and ranks for a scoring system (ranker/scorer)
A is as high as 0.94 under the assumption of normal distribution and 0.96 for the uniform
distribution among the score values. These values increase when the sample size (i.e.,
the number n) increases and reach the limits of 0.98 and 1, for normal and uniform
distribution, respectively.
In light of this close relationship between ranks M
r
(i, j) and scores M
s
(i, j), we might expect
that operating on M
r
(i, j) and on M
s
(i, j) would draw the same conclusion. This appears to
be so in a number of special cases. But in general, it has to be approached with certain
amount of caution and care. It is clear, for example, that a few points with comparatively
high residuals (i.e. poor correlations) would not have major effects on the overall correlation
and correlation structure of the dataset, but these outliers may well be the key target of the
investigation. We list the special features of transforming from M
s
(i, j) to M
r
(i, j).
Remark 4 (Rank/Score Transfer)
When transforming from a score function s
A
to a rank function r
A
on n objects, we have:
(a) the dimension of sample space is reduced from R
n
for the score space to N
n
(and
then N! because of permutation) for the rank space, and
(b) the score function values have been standardized to the scale while the mean is
fixed for every rank function.
Remark 4(a) states that dimension reduction is obtained by a rank/score transform
process that gives rise to the concept of a rank/score function, and the rank/score graph
has at least three advantages: efficiency, neutrality, and visualization (see Remark 1).
Remark 4(b) states that in M
r
(i, j), the mean of each column (a rank function on the n
objects) is fixed to be (n + 1)/2. The same phenomenon is also true for some score data
under certain cases of specific functions and study objectives. However, we note that
when non-parametric and ordinal rank data is used, emphasis is not on the mean of the
data. Rather it is on the discrete order and position each of the rank data is placed.
We recall that fully ranked data on n objects are considered as rank functions on the
symmetric group of n elements S
n
. Since S
n
does not have a natural linear ordering,
graphical methods such as histograms and bar graphs may not be appropriate for
displaying ranked data in S
n
. However, a natural partial ordering on S
n
is induced in the
Cayley network Cay(S
n
, T
n
). Moreover, since a polytope is the convex hull of a finite set
of points in R
n-1
, the n! nodes in Cay(S
n
, T
n
) constitute a permutation polytope when
regarded as vectors in R
n
(see Marden, 1995; McCullagh, 1992; Thompson, 1992). In fact,
Combinatorial Fusion Analysis 51
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
the n! nodes of Cay(S
n
, T
n
) lie on the surface of a sphere in R
n-1
. The six nodes and twenty
four nodes of Cay(S
3
, T
3
) and Cay(S
4
, T
4
) are exhibited in Figure 6(a) and 6(b), respectively.
Information Retrieval
We now turn to the application of these data mining techniques to the information
retrieval domain. We use as an example the study by Ng and Kantor (1998, 2000) (We
call this the NK-study). Their exploratory analysis considered data from TREC compe-
tition with 26 systems and 50 queries for each system on a large but fixed database
consisting of about 1000 documents. The results from these 26 systems are then fused
in a paired manner. As such, there are [(26 25) / 2] 50 = 16, 250 cases of data fusion
in the training data set. In 3,623 of these cases, the performance measures, as P
100
, of the
combined system is better than the best of the two original systems. We refer to these
as positive cases. There are 9,171 negative cases where the performance of the combined
system is worse than the best of the two original systems. In order to understand these
two outcomes, two predictive variables are used. The first is the ratio of P
100
, r = P
l
/ P
h
(see Definition 7(a)). The second variable is the normalized dissimilarity z = d(r
A
, r
B
) (see
Figure 6. Nodes in Cay(S
3
, T
3
) and Cay(S
4
, T
4
), (a) six nodes in Cay(S
3
, T
3
), and
(b) 24 nodes in Cay(S
4
, T
4
)
(a)

123 213
231
321 312
132
(b)

1324
1234
2134
1342
3142
3124
1243
3214
1432 2314
2143
1423
3241
4132 2341
3412
4123
3421
4312
4213
2413
2431
4321
4231
52 Hsu, Chung, & Kristal
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Definition 8(b)). We summarize the results of the NK-study as follows. See Figure 7 for
an illustration.
Remark 5 (NK-study)
The results of the NK-study shows that (a) the positive cases tend to lie above the
diagonal line r + z = 1, and (b) the negative cases are more likely to scatter around the
line r + z = 1.
Remark 5 gives the general trend as to where the positive cases and negative cases
should fall. There are very few negative cases with small r and small z and comparatively
very few cases with high r and high z. Since the negative cases all spread around the line
r + z = 1, z approaches 0 as r approaches 1 and vice versa. This means that for the negative
cases, when the performances P
100
of the two IR systems are about the same, their rank
functions are similar to each other. For the positive cases, it was found that there are very
few cases with small r and z and comparatively few cases with large r and z. But as Remark
5(a) indicated, the positive cases are more likely to lie above the diagonal r + z = 1. This
indicates that systems with dissimilar (i.e., diverse) rank functions but comparable
performance are more likely to lead to effective fusion.
Virtual Screening and Drug Discovery
We now turn to the application of the data mining technique to biomedical informatics.
In particular, we discuss in more details the study by Yang et al. (2005) (we call this
paper the YCSKH-study). The study explores criteria for a recently developed virtual
screening technique called Consensus Scoring (CS). It also provides a CS procedure
for improving the enrichment factor in CS using combinatorial fusion analysis (CFA)
techniques and explores diversity measures on scoring characteristics between indi-
vidual scoring functions.
In structure-based virtual screening, a docking algorithm and a scoring function are
involved (see Shoichet, 2004, and its references). The primary purpose for a docking
program is to find out the most favorable combination of orientation and conformation
(Pose). It also requires a comparison of the best pose (or top few poses) of a given ligand
with those of the other ligands in chemical data base such that a final ranking or
ordering can be obtained. In essence, VS uses computer-based methods to discover new
ligands on the basis of biological structure. Although it was once popular in the 1970s
and 1980s, it has since struggled to meet its initial promise. Drug discovery remains
dominated by empirical screening in the past three decades. Recent successes in
predicting new ligands and their receptor-bound structure have re-invigorated interest
in VS, which is now widely used in drug discovery.
Although VS of molecular compound libraries has emerged as a powerful and inexpen-
sive method for the discovery of novel lead compounds, its major weakness the
inability to consistently identify true positive (leads) is likely due to a lack of
understanding of the chemistry involved in ligand binding and the subsequently
Combinatorial Fusion Analysis 53
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
imprecise scoring algorithms. It has been demonstrated that consensus scoring (CS),
which combines multiple scoring functions, improves enrichment of true positions.
Results of VS using CS have largely focused on empirical study. The YCSKH-study is
one attempt to provide theoretical analysis using combinatorial fusion (Yang et al., 2005).
The YCSKH-study developed a novel CS system that was tested for five scoring systems
(A, B, C, D, and E) with two evolutionary docking algorithms (GemDOCK and GOLD) on
four targets: thymidine kinase (TK), human dihydrofolate reductase (DHFR), and
estrogen receptors (ER) of antagonists and agonists (ERA). Their scoring systems
consist of both rank-based and score-based CS systems (RCS and SCS). They used the
GH (goodness-of-hit) score to evaluate the performance of each individual and
combined systems. That is, P(A) = GH score of the system A. Two predicative variables
are used: (a) the performance ratio PR(A, B) = P
l
/P
h
and (b) the diversity measure d
2
(f
A
,
f
B
) as defined in Definitions 7 and 8 (see Yang et al., 2005).
PR(A, B) = P
l
/ P
h
= min{P(A), P(B)} / max{P(A), P(B)},
and
( )
2 / 1
1
2
2
/ ) ( ) ( ) , (

=

=
n
j
B A B A
n j f j f f f d
,
where P(A) and P(B) are the performances of the two scoring systems A and B to be
combined.
Figure 7. The two predictive variables proposed in NK-study r and z, and the
regions that positive and negative cases are most likely to scatter around


Positive
Cases

Negative
Cases

1
z
0.5
0
0 0.5
r
1
54 Hsu, Chung, & Kristal
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Let g(x) denote the normalization function for the two predictive variables PR(A, B)
and d
2
(f
A
, f
B
) so that g(x) is in [0, 1]. Their results regarding bi-diversity and combination
of two scoring systems are summarized in the following remark where we use x and y
as the coordinates for g(P
l
/ P
h
)

and g(d
2
(f
A
, f
B
)) respectively (see Figure 8).
Remark 6 (YCSKH-study)
The YCSKH-study shows that numbers of positive and negative cases split into roughly
half and (a) most of the positive cases are located above the line x + y = 1 while none of
the few cases below the line have both x 0.30 and y 0.30., and (b) most of the negative
cases tend to be located below the line x + y = 1 while only one of the few cases above
the line have both x 0.50 and y 0.50. The exceptional case has g(d
2
(f
A
, f
B
)) 0.60 and
g(PR(A, B)) 0.95 but both P
l
and P
h
are very small.
Remark 6 reconfirms that combining two scoring systems (rank function or score
function) improves performance only if (a) each of the individual scoring systems has
relatively high performance and (b) the scoring characteristics of each of the individual
scoring systems are quite different. This suggests that the two systems to be combined
have to be fairly diverse so that they can complement each other, and the performance
of each system has to be good, although we cannot yet quantitate/constrain the quality
of good outside of our specific study.
Protein Structure Prediction
Following their previous work establishing a hierarchical learning architecture (HLA),
two indirect coding features, and a gate function to differentiate proteins according to
their classes and folding patterns, C.-Y. Lin et al. (2005) and K.-L. Lin et al. (2005a, 2005b)
have used combinatorial fusion to improve their prediction accuracy on the secondary
Figure 8. Positive and negative cases w.r.t. P
l
/ P
h
and d
2
(f
A
, f
B
)
1.00
0.90
0.80
0.70
0.60
0.50
0.40
0.30
0.20
0.10
0.00
0.00 0.20 0.40 0.60 0.80 1.00
g
(
d
2
(
f
A
,

f
B
)
)
g(P
l
/P
h
)
Combinatorial Fusion Analysis 55
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
class structure and/or 3-D folding patterns. Using 8 and 11 features respectively (i.e.,
scoring systems) and neural networks as a multi-class classifier to build HLA, they
adopted the radical basis function network (RBFN) model to predict folding pattern for
384 proteins. In other words, in the system space M(n, p, q) defined in Figure 2, the number
of scoring systems p = 8 or 11 and the number of temporal steps q = 384.
The work by C.-Y. Lin et al. (2005) using 8 features has an overall prediction rate of 69.6%
for the 27 folding categories, which improves previous results by Ding and Dubchak
(2001) of 56% and by Huang et al. (2003) of 65.5%. The work by K.-L. Lin et al. (2005a,
2005b), using CFA to facilitate feature selection using 11 features, achieves a prediction
accuracy of 70.9% for 27 folding patterns.
Both works utilize the concept of diversity rank/score graph (Definition 9(e)) to select
features to combine. In C.-Y. Lin et al. (2005), the scoring systems (features in this case)
E, F, G, H are selected from the eight features {A, B, C, D, E, F, G, H}. Then the diversity
rank/score functions for the six pairs of the four scoring systems E, F, G, H, are
calculated (see Figure 9), where the diversity rank/score graph for the pair (E, H) is
found to have the highest overall value across the q = 384 protein sequences tested.
Hence E and H are considered to have the highest diversity among the six pairs of scoring
systems. The best result was obtained by combining these two scoring systems E and
H.
The work by K.-L. Lin et al. (2005a, 2005b) uses 11 features and has selected H, I, K out
of 11 features {A, B, C, D, E, F, G, H, I, J, K} because of their higher individual performance.
The diversity rank/score function is then calculated for any pair among H, I, K (see Figure
10). From the graphs of these three functions, they conclude that the pair H, I has the
highest diversity across the q = 384 protein sequences tested. This pair of scoring
systems H and I are then used to perform combination to achieve the desired result.
Conclusion and Discussion
In this chapter, we have described a method, called Combinatorial Fusion Analysis
(CFA), for combining multiple scoring systems (MSS) each of which is obtained from a
set of homogeneous or heterogeneous features/attributes/indicators/cues. The method
(CFA) is based on information obtained from multi-variable (multi-criterion/multi-
objective) classification, prediction, learning or optimization problems, or collected
from multi-sensor / multi-source environments or experiments. We distinguished
between the data space (see Figure 1) which consists of using features/attributes/
indicators/cues to describe objects (e.g., pattern classes / documents / molecules /
folding patterns), and the system space (see Figure 2), which consists of using different
scoring systems (statistical methods, learning systems, combinatorial techniques, or
computational algorithms) to assign a score and a rank to each of the objects under
study. In the extreme case when each scoring system is a feature / attribute / indicator
/ cue, the system space coincides with the data space.
We use the concepts of score functions, rank functions and rank/score functions as
defined by Hsu, Shapiro and Taksa (2002) and Hsu and Taksa (2005) to represent a scoring
56 Hsu, Chung, & Kristal
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
system. Rank / score transfer and correlation between scores and ranks are described
and examined. We also described various performance measurements with respect to
different application domains. Various combination/fusion methods/algorithms of
combining multiple scoring systems have been explored. Theoretical analysis that gives
insights into system selection and methods of combination/fusion has been provided.
In particular, we observe that combining multiple scoring systems improves the perfor-
mance only if (a) each of the individual scoring systems has relatively high performance,
and (b) the individual scoring systems are distinctive (different, diverse or orthogonal).
We have initiated study on these two issues (a) and (b) for the special case of two
scoring systems A and B. Two predictive parameters are used. The first parameter is the
Figure 9. Diversity graphs for the six pairs of scoring systems E, F, G and H
Figure 10. Diversity graphs for the three pairs of scoring systems H, I and J.

0
0.5
1
1.5
2
2.5
3
3.5
1 17 33 49 65 81 97 113 129 145 161 177 193 209 225 241 257 273 289 305 321 337 353 369 385
Diversity Rank
D
i
v
e
r
s
i
t
y

S
c
o
r
e
HI
HK
IK
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
1 14 27 40 53 66 79 92 118 105 144 157 131 170 183 196 209 222 235 248 261 274 287 300 313 326 339 352 365 378
Diversity Rank
D
i
v
e
r
s
i
t
y

S
c
o
r
e
EF
EG
EH
FG
FH
GH
Combinatorial Fusion Analysis 57
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
performance ratio, P
l
/P
h
, which measures the relative performance of A and B where P
l
and P
h
are the lower performance and higher performance respectively. The second
parameter deals with the bi-diversity between A and B, d
2
(A, B), which measures the
degree of difference/dissimilarity/diversity between the two scoring systems A and B.
The bi-diversity (or 2-diversity) measure d
2
(A, B) between two scoring systems A and B
is defined as one of the following three possibilities: (1) d
2
(A, B) = d(s
A
, s
B
), the distance
between the score functions s
A
and s
B
of scoring system A and B; (2) d
2
(A, B) = d(r
A
, r
B
),
the distance between the rank functions r
A
and r
B
; and (3) d
2
(A, B) = d(f
A
, f
B
), the distance
between rank/score functions f
A
and f
B
. Diversity measures have been studied extensively
in pattern recognition and classification (Kuncheva, 2005; Brown et al., 2005). Diversity
measure defined in the form of rank function was used in information retrieval (Ng &
Kantor, 1998, 2000). The work of Hsu and Taksa (2005) and Yang et al. (2005) used rank/
score functions to measure diversity between two scoring systems in their study of
comparing rank vs. score combination and consensus scoring criteria for improving
enrichment in virtual screening and drug discovery. For the protein structure prediction
problem, the rank/score functions of A and B, f
A
and f
B
, are used for each protein sequence
p
i
, where

=
j B A p
j f j f B A d
i
| ) ( ) ( | ) , ( , j in N = {1, 2, , n}, is the diversity score function
for p
i
(C.-Y. Lin et al., 2005; K.-L. Lin et al., 2005a, 2005b).
Considering all protein sequences p
i
in P = {p
1
, p
2
, , p
q
}, the diversity score function,
s
(A,B)
(x) written as s
(A,B)
(p
i
) = ) , ( B A d
i
p
, is a function from P to R. Consequently, sorting
s
(A,B)
(x) into descending order leads to a diversity rank function r
(A,B)
(x) from P to Q = {1,
2, , q}. Hence the diversity rank/score function (or diversity function) f
(A,B)
(j) defined
as )) ( ( ) )( ( ) (
1
) , ( ) , (
1
) , ( ) , ( ) , (
j r s j r s j f
B A B A B A B A B A

= = o , where j is in Q, is a function from Q to R.
The diversity function and its graph play important roles in system selection and
method combination in the PSP problem. We note that the diversity function so defined
for PSP problem can be applied to other classification or prediction problems as well.
We illustrate the method of combinatorial fusion using examples from three different
domains IR, VS, and PSP. In all three applications, multiple scoring systems are used. The
issue of bi-diversity was discussed. The diversity rank/score function was calculated in
the protein structure prediction problem.
In summary, we have discussed the method of combinatorial fusion analysis developed
and used recently in pattern recognition (PR), information retrieval (IR), virtual screening
(VS), and protein structure prediction (PSP). Our current work has generated several
issues and topics worthy of further investigation. Among them, we list four:
1. We have so far emphasized more on bi-diversity (i.e., 2-diversity). How about tri-
diversity (i.e., 3-diversity)? How about higher level diversity measurement? Can
this be (or is this best) treated as a single optimization or a sequential series of bi-
diversity problems?
2. Our diversity score function ) , ( B A d
i
t
for the feature pair or scoring systems A
and B with respect to temporal step t
i
is defined using the variation of the rank/
score functions between A and B (i.e., d(f
A
, f
B
)). In general, the variation of the
score functions d(s
A
, s
B
) or the rank functions d(r
A
, r
B
) could be used to define the
diversity score function ) , ( B A d
i
t
.
58 Hsu, Chung, & Kristal
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
3. There have been several results concerning the combination of multiple scoring
systems in the past few years. Three important issues we would like to study are:
(a) selecting the most appropriate scoring systems to combine, (b) finding the best
way to combine (Definition 6), and (c) establishing the predictive variables for
effective combination. But due to page limitation and chapter constraint, we have
included only three examples in IR, VS and PSP with respect to all three issues (a),
(b,) and (c) described above.
4. We will explore more examples combining multiple heterogeneous scoring sys-
tems in the future. For example, we are working on multiple scoring systems which
are hybrid of classifier systems, prediction systems, and learning systems (such
as neural networks, support vector machines, or evolutionary algorithms).
Acknowledgments
D.F. Hsu thanks Fordham University, National Tsing Hua University, and Ministry of
Education of Taiwan, for a Faculty Fellowship and for support and hospitality during his
visit to NTHU in Spring 2004. B.S. Kristal acknowledges support from NIH and a NY State
SCORE grant.
References
Arrow, K.J. (1963). Social choices and individual values. New York: John Wiley.
Aslam, J.A, Pavlu, V., & Savell, R. (2003). A unified model for metasearch, pooling, and
system evaluation. In O. Frieder (Ed.), Proceedings of the Twelfth International
Conference on Information and Knowledge Management (pp. 484-491). New York:
ACM Press.
Belkin, N.J., Kantor, P.B., Fox, E.A., & Shaw, J.A. (1995). Combining evidence of multiple
query representation for information retrieval. Information Processing & Manage-
ment, 31(3), 431-448.
Biggs, N.L., & White, T. (1979). Permutation groups and combinatorial structures (LMS
Lecture Note Series, Vol. 33). Cambridge: Cambridge University Press.
Brown, G., Wyatt, J., Harris, R., & Yao, X. (2005). Diversity creation methods: a survey
and categorization. Information Fusion, 6, 5-20.
Chuang, H.Y., Liu, H.F., Brown, S., McMunn-Coffran, C., Kao, C.Y., & Hsu, D.F. (2004).
Identifying significant genes from microarray data. In Proceedings of IEEE
BIBE04 (pp. 358-365). IEEE Computer Society.
Chuang, H.Y., Liu, H.F., Chen, F.A., Kao, C.Y., & Hsu, D.F. (2004). Combination method
in microarray analysis. In D.F. Hsu et al. (Ed.), Proceedings of the 7
th
International
Combinatorial Fusion Analysis 59
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Symposium on Parallel Architectures, Algorithms and Networks (I-SPAN04)
(pp. 625-630). IEEE Computer Society.
Collins, R.T., Lipton, A.J., Fujiyoshi, H., & Kanade, T. (2001). Algorithms for cooperative
multisensor surveillance. Proceedings of the IEEE, 89 (10), 1456-1477.
Corne, D.W., Deb, K., Fleming, P.J., & Knowles, J.D. (2003). The good of the many
outweighs the good of the one: evolutional multi-objective optimization [Featured
article]. IEEE Neural Networks Society, 9-13.
Dasarathy, B.V. (2000). Elucidative fusion systemsan exposition. Information Fusion,
1, 5-15.
Diligenti, M., Gori, M., & Maggini, M. (2004). A unified probabilistic framework for web
page scoring systems, IEEE Trans. on Knowledge and Data Engineering, 16(1),
4-16.
Ding, C.H.Q., & Dubchak, I. (2001). Multi-class protein fold recognition using support
vector machines and neural networks. Bioinformatics, 17 (4), 349-358.
Duerr, B., Haettich, W., Tropf, H., & Winkler, G. (1980). A combination of statistical and
syntactical pattern recognition applied to classification of unconstrained hand-
written numerals. Pattern Recognition, 12(3), 189-199.
Dwork, C., Kumar, R., Naor, M., & Sivakumar, D. (2001). Rank aggregation methods for
the web. In Proceedings of the Tenth International World Wide Web Conference,
WWW 10 (pp. 613-622). New York: ACM Press.
Fagin, R., Kumar, R., & Sivakumar, D. (2003). Comparing top k-lists. SIAM Journal on
Discrete Mathematics, 17, 134-160.
Freund, Y., Iyer, R., Schapire, R.E., & Singer, Y. (2003). An efficient boosting algorithm
for combining preferences. Journal of Machine Learning Research, 4, 933-969.
Garcia-Pedrajas, N., Hervas-Martinez, C., & Ortiz-Boyer, D. (2005). Cooperative coevo-
lution of artificial neural network ensembles for pattern classification. IEEE Trans.
on Evolutional Computation, 9(3), 271-302.
Ginn, C. M. R., Willett, P., & Bradshaw, J. (2000). Combination of molecular similarity
measures using data fusion [Perspectives]. Drug Discovery and Design, 20, 1-15.
Grammatikakis, M.D., Hsu, D.F., & Kraetzl, M. (2001). Parallel system interconnections
and communications. Boca Raton, FL: CRC Press.
Heydemann, M.C. (1997). Cayley graphs and interconnection networks. In G. Hahn & G.
Sabidussi (Eds.), Graph symmetry (pp. 161-224). Norwell, MA: Kluwer Academic
Publishers.
Ho, T.K. (2002). Multiple classifier combination: Lessons and next steps. In H. Bunke &
A. Kandel (Ed.), Hybrid methods in pattern recognition (pp. 171-198). Singapore:
World Scientific.
Ho, T.K., Hull, J.J., & Srihari, S.N. (1992). Combination of decisions by multiple classifiers.
In H.S. Baird, H. Burke, & K. Yamulmoto (Eds.), Structured document image
analysis (pp. 188-202). Berlin: Springer-Verlag.
Ho, T.K., Hull, J.J., & Srihari, S.N. (1994). Decision combination in multiple classifier
system. IEEE Trans. on Pattern Analysis and Machine Intelligence, 16(1), 66-75.
60 Hsu, Chung, & Kristal
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Hsu, D.F., & Palumbo, A. (2004). A study of data fusion in Cayley graphs G(S
n
, P
n
). In:
D.F. Hsu et al. (Ed.), Proceedings of the 7
th
International Symposium on Parallel
Architectures, Algorithms and Networks (I-SPAN04) (pp. 557-562). IEEE Com-
puter Society.
Hsu, D.F., Shapiro, J., & Taksa, I. (2002). Methods of data fusion in information retrieval:
Rank vs. score combination (Tech. Rep. 2002-58). Piscataway, NJ: DIMACS
Center.
Hsu, D.F., & Taksa, I. (2005). Comparing rank and score combination methods for data
fusion in information retrieval. Information Retrieval, 8(3), 449-480.
Hu, W., Tan, T., Wang, L., & Maybank, S. (2004). A survey on visual surveillance of
object motion and behaviors. IEEE Trans. on Systems, Man, and Cybernetics
Part C: Applications and Review, 34(3), 334-352.
Huang, C.D., Lin, C.T., & Pal, N.R. (2003). Hierarchical learning architecture with
automatic feature selection for multi-class protein fold classification. IEEE Trans.
on NanoBioscience, 2(4), 503-517.
Ibraev, U., Ng, K.B., & Kantor, P.B. (2001). Counter intuitive cases of data fusion in
information retrieval (Tech. Rep.). Rutgers University.
Jin, X., & Branke, J. (2005). Evolutional optimization in uncertain environmentsa
survey. IEEE Trans. on Evolutional Computation, 9(3), 303-317.
Kantor, P.B. (1998, Jan). Semantic dimension: On the effectiveness of nave data fusion
methods in certain learning and detection problems. Paper presented at the
meeting of the Fifth International Symposium on Artificial Intelligence and Math-
ematics, Ft. Lauderdalee, FL.
Kendall, M., & Gibbons, J.D. (1990). Rank correlation methods. London: Edward Arnold.
Kittler, J., & Alkoot, F.M. (2003). Sum versus vote fusion in multiple classifier systems.
IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(1), 110-115.
Kuncheva, L.I. (2005). Diversity in multiple classifier systems [Guest editorial]. Informa-
tion Fusion, 6, 3-4.
Kuriakose, M.A., Chen, W.T., He, Z.M., Sikora, A.G., Zhang, P., Zhang, Z.Y., Qiu, W.L.,
Hsu, D.F., McMunn-Coffran, C., Brown, S.M., Elango, E.M., Delacure, M.D., &
Chen, F.A.. (2004). Selection and Validation of differentially expressed genes in
head and neck cancer. Cellular and Molecular Life Sciences, 61, 1372-1383.
Lee, J.H. (1997). Analyses of multiple evidence combination. In N.J. Belkin, A.D.
Narasimhalu, P. Willett, W. Hersh (Ed.), Proceedings of the 20
th
Annual Interna-
tional ACM SIGIR Conference on Research and Development in Information
Retrieval (pp. 267-276). New York: ACM Press.
Lin, C.-Y., Lin, K.-L., Huang, C.-D., Chang, H.-M., Yang, C.-Y., Lin, C.-T., & Hsu, D.F.
(2005). Feature selection and combination criteria for improving predictive accu-
racy in protein structure classification. In Proceedings of IEEE BIBE05 (pp. 311-
315). IEEE Computer Society.
Lin, K.-L., Lin, C.-Y., Huang, C.-D., Chang, H.-M., Lin, C.-T., Tang, C.Y., & Hsu, D.F.
(2005a). Improving prediction accuracy for protein structure classification by
Combinatorial Fusion Analysis 61
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
neural networks using feature combination. Proceedings of the 5
th
WSEAS Inter-
national Conference on Applied Informatics and Communications (AIC05)(pp.
313-318).
Lin, K.-L., Lin, C.Y., Huang, C.-D., Chang, H.-M., Lin, C.-T., Tang, C.Y., & Hsu, D.F.
(2005b). Methods of improving protein structure prediction based on HLA neural
networks and combinatorial fusion analysis. WSEAS Trans. on Information Sci-
ence and Application, 2, 2146-2153.
Marden, J.I. (1995). Analyzing and modeling rank data. (Monographs on Statistics and
Applied Probability, No. 64). London: Chapman & Hall.
McCullagh, P. (1992). Models on spheres and models for permutations. In Probabality
Models and Statistical Analysis for Ranking Data (M.A. Fligner, & J.S. Verducci,
Ed., pp. 278-283). (Lecture Notes in Statistics, No. 80). Berlin: Springer-Verlag.
Melnik, D., Vardi, Y., & Zhang, C.U. (2004). Mixed group ranks: preference and confi-
dence in classifier combination. IEEE Trans. on Pattern Analysis and Machine
Intelligence, 26(8), 973-981.
Ng, K.B., & Kantor, P.B. (1998). An investigation of the preconditions for effective data
fusion in information retrieval: A pilot study. In C.M. Preston (Ed.), Proceedings
of the 61
st
Annual Meeting of the American Society for Information Science (pp.
166-178). Medford, NJ: Information Today.
Ng, K.B., & Kantor, P.B. (2000). Predicting the effectiveness of nave data fusion on the
basis of system characteristics. Journal of the American Society for Information
Science. 51(13), 1177-1189.
Paolucci, U., Vigneau-Callahan, K. E., Shi, H., Matson, W. R., & Kristal, B. S. (2004).
Development of biomarkers based on diet-dependent metabolic serotypes: Char-
acteristics of component-based models of metabolic serotype. OMICS, 8, 221-238.
Patil, G.P., & Taillie, C. (2004). Multiple indicators, partially ordered sets, and linear
extensions: multi-criterion ranking and prioritization. Environmental and Ecologi-
cal Statistics, 11, 199-288.
Perrone, M.P., & Cooper, L.N. (1992). When networks disagree: Ensemble methods for
hybrid neural networks (Report AF-S260 045). U.S. Dept. of Commerce.
Shi, H., Paolucci, U., Vigneau-Callahan, K. E., Milbury, P. E., Matson, W. R., & Kristal,
B. S. (2004). Development of biomarkers based on diet-dependent metabolic
serotypes: Practical issues in development of expert system-based classification
models in metabolomic studies. OMICS, 8, 197-208.
Shi, H., Vigneau-Callahan, K., Shestopalov, I., Milbury, P.E., Matson, W.R., & Kristal B.S.
(2002a). Characterization of diet-dependent metabolic serotypes: Proof of principle
in female and male rats. The Journal of Nutrition, 132, 1031-1038.
Shi, H., Vigneau-Callahan, K., Shestopalov, I., Milbury, P.E., Matson, W.R., & Kristal,
B.S. (2002b). Characterization of diet-dependent metabolic serotypes: Primary
validation of male and female serotypes in independent cohorts of rats. The Journal
of Nutrition, 132, 1039-1046.
Shoichet, B.K (2004). Virtual screening of chemical libraries. Nature, 432, 862-865.
62 Hsu, Chung, & Kristal
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Thompson, G. L. (1992). Graphical techniques for ranked data. In Probabality Models
and Statistical Analysis for Ranking Data (M.A. Fligner, & J.S. Verducci, Ed., pp.
294-298). (Lecture Notes in Statistics, No. 80). Berlin: Springer-Verlag.
Triesch, J., & von der Malsburg, C. (2001). Democratic integration: self-organized
integration of adaptive cues. Neural Computation, 13, 2049-2074.
Tumer, K., & Ghosh, J. (1999). Linear and order statistics combinations for pattern
classification. In: Amanda Sharkey (Ed.), Combining artificial neural networks
(pp. 127-162). Berlin: Springer-Verlag.
Vogt, C. C., & Cottrell, G. W. (1999). Fusion via a linear combination of scores.
Information Retrieval, 1(3), 151-173.
Xu, L., Krzyzak, A., & Suen, C.Y. (1992). Methods of combining multiple classifiers and
their applications to handwriting recognition. IEEE Transactions on Systems,
Man, and Cybernetics, 22(3), 418-435.
Yang, J.M., Chen, Y.F., Shen, T.W., Kristal, B.S., & Hsu, D.F. (2005). Consensus scoring
for improving enrichment in virtual screening. Journal of Chemical Information
and Modeling, 45, 1134-1146.
Young, H.P. (1975). Social choice scoring functions. SIAM Journal on Applied Math-
ematics, 28(4), 824-838.
Young, H. P., & Levenglick, A. (1978). A consistent extension of Condorcets election
principle. SIAM Journal on Applied Mathematics, 35(2), 285-300.
DNA Sequence Visualization 63
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Chapter IV
DNA Sequence
Visualization
Hsuan T. Chang, National Yunlin University of Science and Technology, Taiwan
Abstract
This chapter introduces various visualization (i.e., graphical representation) schemes
of symbolic DNA sequences, which are basically represented by character strings in
conventional sequence databases. Several visualization schemes are reviewed and
their characterizations are summarized for comparison. Moreover, further potential
applications based on the visualized sequences are discussed. By understanding the
visualization process, the researchers will be able to analyze DNA sequences by
designing signal processing algorithms for specific purposes such as sequence
alignment, feature extraction, and sequence clustering, etc.
Introduction
Recently, the great progress of biotechnology makes the deoxyribonucleic acid (DNA)
sequencing more efficient. Huge amounts of DNA sequences of various organisms have
been successfully sequenced with higher accuracies. By analyzing DNA sequences, the
biological relationships such as homologous and phylogeny of different species can be
64 Chang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
investigated. However, the analysis of DNA sequences by the use of biological methods
is too slow for processing huge amount of DNA sequences. Therefore, the assistance
of computers is necessary and thus bioinformatics is extensively developed. Efficient
algorithms and implemented computer-based tools are desired to deal with the consid-
erable and tedious biomolecular data.
In general, DNA sequences are stored in the computer database system in the form of
character strings. In a human somatic cell, its haploid nuclear DNA contains 2.85 billion
base pairs (bps), in which a tremendous wealth of genetic information resides (Collins
et al., 2004). Distinguishing the differences and similarities among DNA sequences has
been a major task for biologists. Most of the sequences in their character strings are too
long to be displayed on the computer screen and, therefore, are very hard to be extracted
for any feature or characteristic.
Development of visualization techniques for presenting biological sequences has been
widely attempted (Roy, Raychaudhury, & Nandy, 1998; Loraine & Helt, 2002). Mehta &
Sahni (1994) proposed some efficient algorithms that make use of the compact symmetric
directed acyclic word graph (csdawg) data structure. Blumer, Blumer, Haussler,
McConnell, and Ehrenfeucht (1987) proposed the analysis and visualization of patterns
in long string. Some previous studies (Anastassiou, 2001; Berger, Mitra, Carli, & Neri,
2002; Wang & Johnson, 2002) have shown various methods (such as discrete Fourier
transform or wavelet transform) of transforming the symbolic DNA sequences to numeric
sequences for further processing. With the methods described above, the periodic
patterns existed in DNA sequences can be observed from the determined scalograms or
spectrograms. On the other hand, some methodologies (Cheever & Searls, 1989; Cork &
Wu, 1993; Wu, Roberge, Cork, Nguyen, & Grace, 1993) were proposed to depict symbolic
sequences by two-dimensional (2-D) images, three-dimensional (3-D) curves, or graphs.
The calculation in some methods was troublesome and required intensive computation.
Efficient and direct mapping methods are desired to convert the symbolic sequences into
the numeric sequences, and have them displayed in graphs.
Visualization (i.e., graphical representation) of DNA sequences provides corresponding
pseudo shapes in a global view, which makes the sorting, comparison, and feature
extraction based on the pseudo shape available. Visual recognition of differences among
related DNA sequences by inspection can be made through sequence visualization. The
graphical form can be viewed on a computer display or be printed on a piece of paper.
Therefore, global and local characterizations/features of sequences can be quickly
grasped in a perceivable form. Moreover, numerical characterizations (or features) of
sequences can be determined from the visualized data (Chang, Xiao, & Lo, 2005). The
extracted characterizations or features make the sequence data in a much more manage-
able fashion. Visualization is an alternative for DNA sequence representations. In
addition to complementing the limitations of conventional symbolic-based methods,
more powerful application tools would be expected in the near future.
There have been many researchers around the world working on this topic. The related
bioinformatics tools and databases of visualized DNA sequence also have been acces-
sible over the Internet. Therefore, the objective of this chapter is a literature review and
summarization of the visualization methods for DNA sequences. Further applications
based on the visualization methods will also be mentioned.
DNA Sequence Visualization 65
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Background
A DNA strand is a biomolecular polymer characteristic of four different bases, i.e.,
adenine (A), guanine (G), cytosine (C), and thymine (T). The length unit of a DNA
sequence is in base pair (bp) for a double-stranded DNA, or in nucleotide (nt) for a single
strand. With the rapid development and progress in bioinformatics and the completion
of more genome sequencing projects, many sequence databases have been constructed
and will be continuously constructed for sequence query.
It is not easy to directly access large amount of genomic DNA sequence data to perform
mathematical analysis. That is, character-based representation of DNA sequences
cannot provide immediately useful nor informative characterizations. To perform visu-
alization, symbolic-based sequences should be transformed into numeric-based ver-
sions. Different transformation will provide various characterizations of the visualized
sequences. Therefore, the transformation (or mapping, simply saying) is a critical issue
in sequence visualization.
There are many transformation methods for mapping the symbolic characters A, T, C, G,
to numerical values (Almeida & Vinga, 2002; Chen, Li, Sun, & Kim, 2003; Wang &
Johnson, 2002; Zhang et al., 2002). Those methods transformed DNA sequences into
one-dimensional sequence, 2-D, 3-D, higher dimensional, or some graphical/image-like
data. The data values can be real, imaginary, complex, or quaternion numbers. Therefore,
the resulting numerical characterizations of transformed sequences are model depen-
dent. The simplest way to transform the symbolic sequence to a numerical one is to
directly assign an integer (or real value) for each nucleotide. For example, one can select
the integers 1, 2, 3, and 4 for representing A, T, C, and G, respectively. On the other hand,
more complicated transformations such as a complex assignment or a statistical calcu-
lation were also proposed to derive different graphical representation of DNA se-
quences. In addition to providing a global view and to directly perform sequence
comparison, further analyses based on the visualized data, such as feature extraction for
concise data representation and data retrieval, and phylogenetic clustering, are all the
possible objectives. Therefore, the significance of the studies on DNA sequence
visualization is quite obvious.
Sequence Visualization Methods
The studies of DNA sequence visualization were initiated more than 20 years ago. From
the literature survey, Hamori & Ruskin (1983), Gates (1986), Jeffery (1990), Peng et al.
(1992), Zhang & Zhang (1994), and Roy & Nandy (1994) could be the pioneers in this
research area during the 1980s and 1990s. Roy, Raychaudhury, and Nandy (1998) gave
a comprehensive survey on the graphical representation methods and their applications
in viewing and analyzing DNA sequences. Randics group (Randic, 2000a; Randic &
Vracko, 2000; Randic et al., 2000; Guo, Randic, & Basak, 2001; Randic & Basak, 2001;
Randic & Balaban, 2003; Randic et al., 2003a; 2003b; 2003c; Randic, 2004; Randic, Zupan,
66 Chang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
& Balaban, 2004; Zupan & Randic, 2005) proposed various visualization methods after
2000. Many new perspectives and practical applications based on the visualized DNA
sequences are blossoming. In this section, we will briefly and historically review most
of the published visualization methods. Their methodologies, characterizations, and
their applications on bioinformatics will be discussed as well.
H-Curves
Most recent studies have referred to the first visualization method for DNA sequences
as the H-curves proposed by Hamori and Ruskin (1983). The authors claimed that the
important global features of long symbolic sequences, such as characteristic nucleotide
ratios of certain sections, repetition of longer fragments, sudden changes in prevailing
nucleotide compositions at certain locations, are not readily conveyed. The analysis,
recognition, comparison, mental recollection, and so forth, of long sequences without
long range features serving as visual clues become nearly impossible tasks for most
people.
In this method, a vector function g(z) is defined as:
Figure 1. 3-D H-curves: Homo sapiens B-cell CLL/lymphoma 2 (BCL2) mRNA sequence
(AC: NM_000657, 911 bp) and Human dihydrofolate reductase gene (AC: J00140, 979
bp) shown in blue/dark and green/light lines, respectively

DNA Sequence Visualization 67
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.

= +
=
=
= +
=
G if ,
C if ,
T if ,
A if ,
) (
z k j i
z k j i
z k j i
z k j i
z g
(1)
where i, j, and k are unit vectors pointing in the direction of the Cartesian x, y, and z axes,
respectively. The 3-D curve (H-curve) consisting of a series of n joined base vectors is
defined as

= =
n
n
z g z h H
1
, 1
). ( ) (
(2)
As shown in Eq. (1), four bases are represented by four directions in 2-D x-y plane. The
H-curve is constructed by moving one unit in the corresponding direction and another
for each unit in z direction. H-curve can uniquely represent a DNA sequence. In addition,
more advantages H-curves offer include: (1) H-curve preserves all the important charac-
teristics of the long DNA sequence; (2) The average nucleotide composition of a certain
fragment is an important sequence property, which is not directly evident from the
Figure 2(a). 2-D DNA walks: Homo sapiens B-cell CLL/lymphoma 2 (BCL2) mRNA
sequence (AC: NM_000657, 911 bp)

68 Chang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
symbolic sequence; (3) Visual comparisons and matching between different DNA
sequences are facilitated; (4) The information associated with the location of the end
point of an H-curve is provided. Figure 1 shows the H-curves of Homo sapiens B-cell CLL/
lymphoma 2 (BCL2) mRNA sequence (AC: NM_000657, 911 bp) and Human dihydrofolate
reductase gene (AC: J00140, 979 bp) shown in blue/dark and green/light lines, respec-
tively. The characterizations of two curves corresponding to the two sequences are
obviously quite different.
DNA Walks
DNA walks for representing the DNA sequence have been widely utilized since 1992
(Peng et al., 1992). The first work is the 1-D DNA walk, which uses binary alphabets. Let
a DNA sequence of length N
s
be denoted as {x[i], i=1,2,3...,N
s
.}. The presences of purines
(A, G) and pyrimidines (C, T) in DNA sequences correspond to the x[k] values +1 and
-1, respectively, where k denotes the position in the sequence. Then a DNA walk
sequence s[k] is defined as the cumulative sum of x[i], i.e.,

=
=
k
i
i x k s
1
]. [ ] [
(3)
Figure 2(b). 2-D DNA walks: Human dihydrofolate reductase gene (AC: J00140, 979 bp)

DNA Sequence Visualization 69
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Similarly, other two letter alphabets such as S-W and K-M can be employed to obtain the
1-D DNA walks. Note that the symbols S (strong), W (weak), K (keto), and M (amino) are
defined as follow: } , { }, , { }, , { }, , { T C Y G A R T A W G C S .
The dimensionality of the numerical DNA sequence can be more than one for a better
representation. For example, the 3-D vector notation for mapping A, T, C, and G, is similar
to Eq. (1). Furthermore, the complex representation for DNA walk can be defined as
follows:

=
G for ,
C for ,
T for , 1
A for , 1
) (
j
j
k x
. (4)
This complex DNA walk can be plotted in a 3-D Cartesian coordinate system by treating
the accumulated values of the real part and imaginary part, and the variable k as the values
for x, y, and z axes, respectively. Some repeat sequence can be found by projecting the
DNA walk down to the x-y plane (Berger et al., 2004). DNA walk representations can be
used to extract useful information, such as the long-range correlation information and
sequence periodicities, from DNA sequences. To elucidate further information, the
Figure 3. 3-D DNA walks: Homo sapiens B-cell CLL/lymphoma 2 (BCL2) mRNA
sequence (AC: NM_000657, 911 bp) and Human dihydrofolate reductase gene (AC:
J00140, 979 bp) shown in blue/dark and green/light lines, respectively

70 Chang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
wavelet transform analysis can be applied to 1-D or 2-D walks for observing dependen-
cies of nucleotide arrangement in long sequences (Arneodo, Bacry, Graves, & Muzy,
1995; Berger et al., 2004). Figures 2(a) and 2(b) show the 2-D DNA walks of Homo sapiens
B-cell CLL/lymphoma 2 (BCL2) mRNA sequence (AC: NM_000657, 911 bp) and Human
dihydrofolate reductase gene (AC: J00140, 979 bp), respectively. On the other hand,
Figure 3 shows the corresponding 3-D DNA walks of two sequences.
Z-Curves
Zhang and Zhang (1994) proposed the Z-curve representation to visualize DNA se-
quences in 1994. Consider a DNA sequence of N bases and let the accumulated numbers
of the bases A, T, C, and G be four positive integers A
n
, T
n
, C
n
, and G
n
, respectively. The
Z-curve consists of a series of nodes P
n
(n = 0, 1, 2,, N), whose coordinates are denoted
by x
n
, y
n
, and z
n
. Then these three coordinates are determined by the use of the four
integers A
n
, T
n
, C
n
, and G
n
. Their relationships are called the Z transform and are expressed
as below:
, 0,1,..., ], , [ , , ,
) ( ) (
) ( ) (
) ( ) (
N n N N z y x
G C T A z
T C C A y
T C G A x
n n n
n n n n n
n n n n n
n n n n n
=

+ + =
+ + =
+ + =
(5)
Figure 4. 3-D Z-curves: Homo sapiens B-cell CLL/lymphoma 2 (BCL2) mRNA sequence
(AC: NM_000657, 911 bp) and Human dihydrofolate reductase gene (AC: J00140, 979
bp) shown in blue/dark and green/light lines, respectively

DNA Sequence Visualization 71
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
where A
0
= T
0
= C
0
=G
0
=0 and thus x
n
=y
n
=z
n
=0. Given the coordinates of a Z-curve, the
corresponding DNA sequence can be reconstructed by the use of inverse Z-transform,
which is expressed as
, 0,1,..., ,
1 1 1
1 1 1
1 1 1
1 1 1
4
1

1
1
1
1
4
N n
z
y
x
n
T
G
C
A
n
n
n
n
n
n
n
=

(6)
where the relation A
n
+T
n
+C
n
+G
n
=n is used. Figure 4 shows the smoothed Z-curves of
Homo sapiens B-cell CLL/lymphoma 2 (BCL2) mRNA sequence (AC: NM_000657, 911 bp)
and Human dihydrofolate reductase gene (AC: J00140, 979 bp).
As shown in Eq. (5), some statistical characterizations of the DNA sequences can be
obtained directly from the Z-curves. For example, the distributions of purine/pyrimidine,
amino/keto, and strong-H bond/weak-H bond bases along the sequences correspond to
the three components of the Z-curve: x
n
, y
n
, and z
n
, respectively. Moreover, algorithms
for exons identification, gene finding, prediction, or recognition using the Fourier
transform performed on x
n
, y
n
, and z
n
components of the Z-curve are also possible (Guo,
Figure 5. Randics 3-D curves: Homo sapiens B-cell CLL/lymphoma 2 (BCL2) mRNA
sequence (AC: NM_000657, 911 bp) and Human dihydrofolate reductase gene (AC:
J00140, 979 bp) shown in blue/dark and green/light lines, respectively

72 Chang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Ou, & Zhang, 2003; Zhang & Zhang, 2004). Identification of replication origins in
genomes based on the Z-curve method is also reported by Zhang and Zhang (2005).
The Z-curve database now is also available on the Internet (Zhang, Zhang, & Ou, 2003).
In additional to draw and manipulate the Z-curve of a users input sequence online,
several software services and archives are accessible in the Web site: http://
tubic.tju.edu.cn/zcurve/.
Randics Visualization Methods
Since 2000, Randics research group had proposed several visualization schemes for
DNA sequences (Randic, 2000a; Randic & Vracko, 2000; Randic et al., 2000; Guo, Randic,
& Basak, 2001; Randic & Basak, 2001; Randic & Balaban, 2003; Randic et al., 2003a; 2003b;
2003c; Randic 2004; Randic, Zupan, & Balaban, 2004; Zupan & Randic, 2005). A 3-D
graphical representation method was first proposed (Randic et al., 2000). Then the
sequences can be analyzed based on the matrix characteristics (Randic, Kleiner, &
DeAlba, 1994; Randic 2000; Randic & Basak, 2001) extracted from the original DNA
sequences. In this method, the four vertices associated with a regular tetrahedron are
assigned to four nucleotides. The mapping between four nucleotides and corresponding
3-D coordinates is shown below:
T ) 1 1 1 (
C ) 1 1 1 (
G ) 1 1 1 (
A ) 1 1 1 (
+ + +
+
+
+
, ,
, ,
, ,
, ,
(7)
For a sequence, a 3-D curve is plotted by connecting all the points determined by
sequentially adding the (x, y, z) coordinates assigned for the nucleotides. Figure 5 shows
the 3-D curves of Homo sapiens B-cell CLL/lymphoma 2 (BCL2) mRNA sequence (AC:
NM_000657, 911 bp) and Human dihydrofolate reductase gene (AC: J00140, 979 bp). Note
that the 2-D representations in Nandys (1994) and Leongs (1995) methods can be
considered as the projections on (x, y) and (x, z) plane, respectively.
The problems of degeneracy and overlap in graphical representation of DNA sequence
may result in identical graphs for different sequences. Therefore, a 2-D graphical
representation of low degeneracy was proposed to solve the mentioned problems (Guo
& Basak, 2001). The basic idea is to introduce a positive integer d while assigning the four
special vectors in 2-D Cartesian (x, y) coordinates for nucleotides. One of the possible
axes systems for this 2-D graphical representation can be
DNA Sequence Visualization 73
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
C ) 1
1
(
G )
1
1 (
T ) 1
1
(
A )
1
1 (
+ +
+ +
+
+
,
d
d
,
,
d
d
,
(8)
The mathematical derivation for calculating the minimum length of the DNA sequence
that can form a loop in the graphical representation was performed. It shows that if d is
a larger even number, the corresponding graphical representation has a lower degen-
eracy. Figure 6 shows the 2-D curves of Homo sapiens B-cell CLL/lymphoma 2 (BCL2)
mRNA sequence (AC: NM_000657, 911 bp) and Human dihydrofolate reductase gene
(AC: J00140, 979 bp). The degeneracy problem is obviously reduced and much less than
that shown in Figure 2. Liu, Guo, Xu, Pan, and Wang (2002) gave some notes on 2-D
graphical representation with lower or no degeneracy.
A 2-D graphical representation (Randic et al., 2003a), which does not consider four
directions along the Cartesian coordinate axes, was then proposed in 2003. Four
Figure 6. 2-D curves with low degeneracy: Homo sapiens B-cell CLL/lymphoma 2
(BCL2) mRNA sequence (AC: NM_000657, 911 bp) and Human dihydrofolate reductase
gene (AC: J00140, 979 bp) shown in blue/dark and green/light lines, respectively

74 Chang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
horizontal lines separated by unit distance are used to represent four nucleotides and
the nucleotides constituting the considered sequence are denoted as dots and placed
on the corresponding lines. By connecting the adjacent dots, we can obtain a zigzag like
curve for sequence visualization. Figure 7 shows the zigzag-like curve of a given
sequence:
aggctggagagcctgctgcccgcccgcccgtaaaatggtcccctcggctggacagctcgccctgttcgct
ctgggtattgtgttggctgcg.
There is no degeneracy problem with this representation.
Another very compact 2-D visualization scheme (Randic et al., 2003c; Randic, 2004) based
on a template curve was proposed to inspect lengthy DNA sequences without requiring
large space. A zigzag spiral representing the worm curve is used as the template for
constructing compact graphical representation. Then four binary codes for the four
nucleotides are assigned as A=00, G=01, C=10, and T=11. Starting from the center of the
template, a binary code sequence is drawn by putting dark spots for ones and nothing
for zeros on the vertices of template. There is no degeneracy problem from this graphical
representation. On the other hand, only a square space with side approximately n is
required to visualize a DNA sequence of length n. Figures 8(a) and 8(b) show the 2-D
spiral worm curves of Homo sapiens B-cell CLL/lymphoma 2 (BCL2) mRNA sequence
Figure 7. 2-D zigzag like curve

DNA Sequence Visualization 75
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Figure 8(a). The 2-D spiral worm curves: Homo sapiens B-cell CLL/lymphoma 2 (BCL2)
mRNA sequence (AC: NM_000657, 911 bp)
Figure 8(b). The 2-D spiral worm curves: Human dihydrofolate reductase gene (AC:
J00140, 979 bp)


76 Chang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Figure 9(a). Chaos Game Representation (CGR): Homo sapiens B-cell CLL/lymphoma
2 (BCL2) mRNA sequence (AC: NM_000657, 911 bp)
(AC: NM_000657, 911 bp) and Human dihydrofolate reductase gene (AC: J00140, 979 bp),
respectively. They demonstrate a quite different visualization form from those demon-
strated in Figures 1-7.
Chaos Game Representation and Fractals
In addition to the curve form representation, more complicated patterns can be utilized
for DNA sequence visualization. Jeffrey (1990) developed the chaos game representation
(CGR) method, which uses an iterated function system (IFS) (Barnsley, 1988) to map the
character-based DNA sequence to a set of points in R
2
. The result points are not
connected by line segments but rather are displayed in a 2-D pattern image, which is a
picture of fractal structure. On the other hand, the fractal dimension of a DNA sequence
can be measured as one of sequence characterizations (Berthelsen, Glazier, & Skolnik,
1992). Sequence visualization through the iterated function systems (Tino, 1999) and
fractals (Hao, Lee, & Zhang, 2000; Ichinose, Yada, & Takagi, 2001) are also available. The
properties of the corresponding DNA sequence can be revealed because the sequence
basically is non-random and thus an attractor is visually observable. Figures 9 and 10
show the CGR and fractals (square), respectively, of Homo sapiens B-cell CLL/lymphoma
2 (BCL2) mRNA sequence (AC: NM_000657, 911 bp) and Human dihydrofolate reductase
gene (AC: J00140, 979 bp).
With the CGR, different fractal patterns can be plotted from groups of genes. The most
significant property of the CGR different from the curve form is that the adjacent bases

DNA Sequence Visualization 77
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Figure 9(b). Chaos Game Representation (CGR):Human dihydrofolate reductase gene
(AC: J00140, 979 bp)
Figure 10(a). Fractals: Homo sapiens B-cell CLL/lymphoma 2 (BCL2) mRNA sequence
(AC: NM_000657, 911 bp)


78 Chang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
in the sequence are not always plotted neighboring to each other. Therefore, a new metric
should be developed for measuring the similarity between two sequences.
One of the applications of the CRG for a DNA sequence is to discriminate the random and
natural DNA sequences (Jeffery, 1990). By measuring the histogram entropy of a CRG
image, the entropic profiles of DNA sequences can be constructed. On the other hand,
a more generalized visualization model based on IFS was proposed by Wu et al. (1993).
Three visual models based on the IFS were introduced: (1) H-curve; (2) Chaos game
representation; (3) W-curve. By controlling two parameters in the proposed model, one
of the three visual models can be selected.
Other Visualization Methods
A more detailed discussion on the 2-D graphical representation shown in Figure 8 is given
by Randic et al. (2003b). Yau et al. (2003) proposed a visualization method in which the
2-D graphical representation is shown without degeneracy. Li and Wang (2004) proposed
a 3-D representation of DNA sequences. The elements in the 3-D vectors assigned for
the four bases are all positive such that no overlap and intersection appear. Yuan, Liao,
and Wang (2003) proposed the 3-D representation of DNA sequences that can avoid
overlaps and intersect. The basic concept is that the z-axis coordinates are the accumu-
lated length of the bases. Chang et al. (2003) proposed a 3-D trajectory method based on
Figure 10(b). Fractals: Human dihydrofolate reductase gene (AC: J00140, 979 bp)

DNA Sequence Visualization 79
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
a regular tetrahedron, which is similar to Randics 3-D graphical representation method.
Liao, Tan, and Ding (2005) proposed a unique 4-D representation of DNA sequences,
which is different with Randics 4-D representation method (Randic & Balaban, 2003).
The studies on DNA sequence visualization are obviously flourishing.
Discussion and Future Trends
The most important issue in investigating various visualization methods is what the
direct applications are. In addition to some direct curve features, such as the geometrical
center and end points (Liao, Tan, & Ding, 2005), many methods employ the distance matrix
and its invariants (Randic, Kleiner, & DeAlba, 1994; Randic & Basak, 2001; Yuan, Liao,
& Wang, 2003) to extract the eigenvalues as the condensed sequences for comparison.
The protein coding measure, identification of replication origins and termination for some
bacterial and archaeal genomes based on the Z-curve method was reported (Yan, Lin, &
Zhang, 1998; Zhang & Zhang, 2005). As for the chaos game and fractal representation,
global distance and local similarity can be determined for sequence comparison (Almeida,
Carrico, Maretzek, Noble, & Fletcher, 2001) and for measuring phylogenetic proximity
(Deschavanne, Giron, Vilain, Fagot, & Fertil, 1999).
Finally, the possible further studies and future work based on the visualized DNA
sequences are summarized as follows:
1. Apply mathematical analysis or signal-processing algorithms such as the Fourier
and Wavelet transforms to the visualized sequence data for further sequence
analysis, clustering, and annotation (Bernaola-Galvan, Carpena, Roman-Roldan, &
Oliver, 2002; Dodin, Vandergheynst, Levoir, Cordier, & Marcourt, 2000; Issac et al.,
2002).
2. Develop better human machine interface for scientists in operating the bioinformatics
tools based on graphical representation of sequences (Ovcharenko, Boffelli, &
Loots, 2004).
3. Cooperate with computer graphics techniques for better graphical representation
of DNA sequences (Ghai, Hain, & Chakraborty, 2004; Montgomery et al., 2004).
4. Extract the features from the visualized/numerical sequences and then develop
feature databases for fast and efficient database retrieval. (Chang et al., 2005)
5. The investigation of the effects of insertion, deletion, substitution, reverse, and
complementary in sequences (Chang et al., 2003). Find more biological applications
and develop various visualization tools such as characterization and classifica-
tions of species, visualizing regular pattern in sequences (Loraine & Helt, 2002).
6. Adopt the existing visualization methods to protein sequences (Basu, Pan, Dutta,
& Das, 1997; Bayley, Gardiner, Willett, & Artymiuk, 2005; Randic & Krilov, 1997;
Randic, Zupan, & Balaban, 2004).
80 Chang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Summary
The aids of computer and network technologies provide better data access, acquisition,
update, and management on huge genomic sequences over the worldwide Internet
databases. Through the data visualization methods mentioned in this chapter, various
graphical representations of DNA sequences provide alternative observations, which
can derivate certain sequence characteristics that cannot be directly obtained from the
original character-based sequences. With the rapid increase of genomic sequences and
databases, the development of new techniques for sequence visualization, analysis,
characterization, comparison, and annotation is more and more important in a
bioinformatics scenario. Most of recent researches on DNA sequence visualization/
graphical representation have been briefly reviewed in this chapter. Based on the
visualized data formats, well known graphic/image processing and pattern recognition
techniques for exploring the intra- and inter-sequence relationships in genomic research
would be a significant trend.
References
Almeida, J. S., Carrico, J. A., Maretzek, A., Noble, P. A., & Fletcher, M. (2001). Analysis
of genomic sequences by Chaos Game Representation. Bioinformatics, 17(5), 429-
437.
Almeida, J. S., & Vinga, S. (2002). Universal sequence map (USM) of arbitrary discrete
sequences. BMC Bioinformatics, 3(6).
Anastassiou, D. (2001). Genomic signal processing. IEEE Signal Processing Magazine,
18, 8-20.
Arneodo, A., Bacry, E., Graves, P. V., & Muzy, J. F. (1995). Characterizing long-range
correlations in DNA sequences from wavelet analysis. Physics Review Letters,
74(16), 3293-3296.
Barnsley, M. F. (1988). Fractals everywhere. New York: Springer-Verlag.
Basu, S., Pan, A., Dutta, C., & Das, J. (1997). Chaos game representation of proteins.
Journal of Molecular Graphics and Modelling, 15, 279-289.
Bayley, M. J., Gardiner, E. J., Willett, P., & Artymiuk, P. J. (2005). A Fourier fingerprint-
based method for protein surface representation. Journal of Chemical Information
Model, 45(3), 696-707.
Bernaola-Galvan, P., Carpena, P., Roman-Roldan, R., & Oliver, J. L. (2002). Study of
statistical correlations in DNA sequences. Gene, 300, 105-115.
Berger, J., Mitra, S., Carli, M., & Neri, A. (2002). New approaches to genome sequence
analysis based on digital signal processing. Proceedings of IEEE Workshop on
Genomic Signal Processing and Statistics (GENSIPS), CP2-08.
DNA Sequence Visualization 81
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Berger, J., Mitra, S., Carli, M., & Neri, A. (2004). Visualization and analysis of DNA
sequences using DNA walks. Journal of the Franklin Institute, 341(1-2), 37-53.
Berthelsen, C. L., Glazier, J. A., & Skolnik, M. H. (1992). Global fractal dimension of human
DNA sequences treated as pseudorandom walks. Physics Review A, 45, 8902-8913.
Blumer, A., Blumer, J., Haussler, D., McConnell, R., & Ehrenfeucht, A. (1987). Complete
inverted files for efficient text retrieval and analysis. Journal of ACM, 34, 578-595
Chang, H. T., Lo, N. W., Lu, W. C., & Kuo, C. J. (2003). Visualization and comparison of
DNA sequences by use of three-dimensional trajectory. Proceedings of The First
Asia-Pacific Bioinformatics Conference (pp. 81-85).
Chang, H. T., Xiao, S. W., & Lo, N. W. (2005). Feature extraction and comparison of TDTs:
an efficient sequence retrieval algorithm for genomic databases. In Abstract Book
of The 3rd Asia Pacific Bioinformatics Conference (APBC2005) (pp. 86).
Cheever, E. A., & Searls, D. B. (1989). Using signal processing techniques for DNA
sequence comparison. Proceedings of the 1989 Fifteenth Annual Northwest
Bioengineering Conference (pp. 173-174).
Chen, J., Li, H., Sun, K., & Kim, B. (2003). Signal processing applications how will
bioinformatics impact signal processing research? IEEE Signal Processing Maga-
zine, 20(6), 16-26.
Collins, F. S., Lander, E. S., Rogers, J., Waterston, R. H., et al. (2004). Finishing the
euchromatic sequence of the human genome, Nature, 431, 931-945.
Cork, D. J., & Wu, D. (1993). Computer visualization of long genomic sequences.
Proceedings of the 4th Conference on Visualization, (pp. 308-315).
Deschavanne, P. J., Giron, A., Vilain, J., Fagot, G., & Fertil, B. (1999). Genomic signature:
Characterization and classification of species assessed by Chaos Game Represen-
tation of sequences. Molecular Biology Evolution, 16(10), 1391-1399.
Dodin, G., Vandergheynst, P., Levoir, P., Cordier, C., & Marcourt, L. (2000). Fourier and
wavelet transform analysis, a tool for visualizaing regular patterns in DNA se-
quences. Journal of theoretical Biology, 206, 323-326.
Gates, M. (1986). A simple way to look at DNA. Journal of Theoretical Biology, 119(3),
319-328.
Ghai, R., Hain, H., & Chakraborty, T. (2004). GenomeViz: visualization microvial genomes.
BMC Bioinformatics, 5, 198.
Guo, F.-B., Ou, H.-Y., & Zhang, C.-T. (2003). Z-CURVE: A new system for recognizing
protein-coding genes in bacterial and archaeal genomes. Nucleic Acids Research,
31(6), 1780-1789.
Guo, X., Randic, M., & Basak, S. C. (2001). A novel 2-D graphical representation of DNA
sequences of low degeneracy. Chemical Physics Letters, 350, 106-112.
Hamori, E., & Ruskin, J. (1983). H curves, A novel method of representation of nucleotide
series especially suited for long DNA sequences. The Journal of Biological
Chemistry, 258(2), 1318-1327.
Hao, B.-L., Lee, H. C., & Zhang, S.-Y. (2000). Fractals related to long DNA sequences and
complete genomes. Chaos, Solitons and Fractals, 11, 825-836.
82 Chang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Ichinose, N., Yada, T., & Takagi, T. (2001). Quadtree Representation of DNA Sequences.
Genome Informatics, 12, 510-511.
Issac, B., Singh, H., Kaur, H., & Raghave, G.P.S. (2002). Locating probable genes using
Fourier transform approach. Bioinformatics, 18(1), 196-197.
Jeffrey, H. J. (1990). Chaos game representation of gene structure. Nucleic Acids
Research, 18(8), 2163-2170.
Leong, P. M., & Morgenthaler, S. (1995). Random walk and gap plots of DNA sequence.
Cabios, 11, 503-507.
Li, C., & Wang, J. (2004). On a 3-D representation of DNA primary sequences. Combi-
natorial Chemistry and High Throughput Screening, 7, 23-27.
Liao, B., Tan, M., & Ding, K. (2005). A 4D representation of DNA sequences and its
application. Chemical Physics Letters, 402, 380-383.
Liu, Y., Guo, X., Xu, J., Pan, L., & Wang, S. (2002). Some notes on 2-D graphical
representation of DNA sequence. Journal of Chemical Information and Computer
Sciences, 42(3), 529-533.
Loraine, A. E., & Helt, G. A. (2002). Visualizing the genome: techniques for presenting
human genome data and annotations. BMC Bioinformatics, 3(19).
Mehta, D. P., & Sahni, S. (1994). Computing display conflicts in string visualization, IEEE
Transaction on Computers, 43, 350-361.
Montgomery, S. B., Astakhova, T., Bilenky, M, Birney, E., Fu, T., Hassel, M., et al. (2004).
Sockeye: a 3D environment for comparative genomics. Genome Research, 14, 956-
962.
Nandy, A. (1994). A new graphical representation and analysis of DNA sequence
structure: I. Methodology and application to globin genes. Current Science, 66(4),
309-314.
Ovcharenko, I., Boffelli, D., & Loots, G. G. (2004). eShadow: A tool for comparing closely
related sequences. Genome Research, 1191-1198.
Peng, C. K., Buldyrev, S. V., Goldberger, A. L., Havlin, S., Sciortino, F., Simons, M., et al.
(1992). Longrange correlations in nucleotide sequences. Nature, 356, 168-170.
Randic, M. (2000) On characterization of DNA primary sequences by a condensed matrix.
Chemical Physics Letters, 317, 29-34.
Randic, M. (2004) Graphical representations of DNA as 2-D map. Chemical Physics
Letters, 386, 468-471.
Randic, M., & Balaban, A. T. (2003). On a four-dimensional representation of DNA
primary sequences. Journal of Chemical Information and Computer Sciences,
43(2), 532-539.
Randic, M., & Basak, S.C. (2001). Characterization of DNA primary sequences based on
the average distances between bases. Journal of Chemical Information and
Computer Sciences, 41(3), 561-568.
Randic, M., & Krilov, G. (1997). Characterization of 3-D sequence of proteins. Chemical
Physics Letters, 272, 115-119.
DNA Sequence Visualization 83
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Randic, M., & Vracko, M. (2000). On the similarity of DNA primary sequences. Journal
of Chemical Information and Computer Sciences, 40(3), 599-606.
Randic, M., Kleiner, A. F., & DeAlba L. M. (1994). Distance/Distance matrices. Journal
of Chemical Information and Computer Science, 34, 277-286.
Randic, M., Vracko, M., Lers, N., & Plavsic, D. (2003a). Novel 2-D graphical representa-
tion of DNA sequences and their numerical characterization. Chemical Physics
Letters, 368, 1-6.
Randic, M., Vracko, M., Lers, N., & Plavsic, D. (2003b). Analysis of similarity/dissimilarity
of DNA sequences based on novel 2-D graphical representation. Chemical Physics
Letters, 371, 202-207.
Randic, M., Vracko, M., Nandy, A., & Basak, S. C. (2000). On 3-D graphical representation
of DNA primary sequences and their numerical characterization. Journal of
Chemical Information and Computer Science, 40(5), 1235-1244.
Randic, M., Vracko, M., Zupan, J., & Novic, M. (2003c) Compact 2-D graphical represen-
tation of DNA. Chemical Physics Letters, 373, 558-562.
Randic, M., Zupan, J., & Balaban, A. T. (2004) Unique graphical representation of protein
sequences based on nucleotide triplet codons. Chemical Physics Letters, 397,
247-252.
Roy, A., Raychaudhury, C., & Nandy, A. (1998). Novel techniques of graphical represen-
tation and analysis of DNA sequences a review. Journal of Bioscience, 23(1),
55-71.
Tino, P. (1999). Spatial representation of symbolic sequences through iterative function
systems. IEEE Transactions on Signal Processing, 29, 386-393.
Wang, W., & Johnson, D. H. (2002). Computing linear transforms of symbolic signals.
IEEE Transaction on Signal Processing, 50, 628-634.
Wu, D., Roberge, J., Cork, D. J., Nguyen, B. G., & Grace, T. (1993). Computer visualization
of long genomic sequence. Proceedings of IEEE Conference on Visualization (pp.
308-315).
Yan, M., Lin, Z. S., & Zhang, C. T. (1998). A new Fourier transform approach for protein
coding measure based on the format of the Z curve. Bioinformatics, 14(8), 685-690.
Yau, S. T., Wang, J., Niknejad, A., Lu, C., Jin, N., & Ho, Y. K. (2003). DNA sequence
representation without degeneracy. Nucleic Acids Research, 31(12), 3078-3080.
Yuan, C., Liao, B., & Wang, T.M. (2003). New 3D graphical representation of DNA
sequences and their numerical characterization. Chemical Physics Letters, 379,
412-417.
Zhang, C. T., Zhang, R., & Ou, H. Y., (2003). The Z curve database: a graphic represen-
tation of genome sequences. Bioinformatics, 19(5), 593599.
Zhang, R., & Zhang, C. T. (1994). Z curves, an intuitive tool for visualizing and analyzing
DNA sequences. Journal of Biomolecular Structure Dynamics, 11, 767-782.
Zhang, R., & Zhang, C. T. (2005). Identification of replication origins in archaeal genomes
based on the Z-curve method. Archaea, 1, 335-346.
84 Chang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Zhang, X. Y., Chen, F., Zhang, Y. T., Agner, S. C., Akay, M., Lu, Z. H., et al. (2002). Signal
processing techniques in genomic engineering. Proceedings of The IEEE, 90(12),
1822-1833.
Zupan, J., & Randic, M. (2005). Algorithm for coding DNA sequences into spectrum-
like and zigzag representations. Journal of Chemical Information Model, 45(2),
309-313.
Proteomics with Mass Spectrometry 85
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Chapter V
Proteomics with
Mass Spectrometry
Simon Lin, Northwestern University, USA
Salvatore Mungal, Duke University Medical Center, USA
Richard Haney, Duke University Medical Center, USA
Edward F. Patz, Jr., Duke University Medical Center, USA
Patrick McConnell, Duke University Medical Center, USA
Abstract
This chapter provides a rudimentary review of the field of proteomics as it applies to
mass spectrometry, data handling, and analysis. It points out the potential significance
of the field suggesting that the study of nuclei acids has its limitations and that the
progressive field of proteomics with spectrometry in tandem with transcription studies
could potentially elucidate the link between RNA transcription and concomitant
protein expression. Furthermore, we describe the fundamentals of proteomics with
mass spectrometry and expound the methodology necessary to manage the vast
amounts of data generated in order to facilitate statistical analysis. We explore the
underlying technologies with the intention to demystify the complexities of the nascent
field and to fuel interest by readers in the near future.
86 Lin, Mungal, Haney, Patz, & McConnell
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Introduction
The science of proteomics seeks to identify and characterize protein expression in
biological systems. It leverages technologies that have been around for decades, such
as mass spectrometry, liquid chromatography, and electrophoresis. Though the basic
science is mature, proteomics technology has made much progress over the last decade
(Aebersold & Mann, 2003). Proteomics has been gaining momentum as the limitations
of studying DNA and RNA alone have been documented (Greenbaum, Colangelo,
Williams, & Gerstein, 2003). Gene sequences themselves yield little information about
how much of its transcribed protein will be expressed and in what cellular states.
Furthermore, the nature and prevalence of alternative splicing has shown that studying
gene expression at the protein level can yield complementary knowledge at the nuclei acid
level.
Proteomics is a vast and complex field consisting of a variety of platforms. In this chapter,
we focus on mass spectrometry technology and data because of its wide use in the field
for protein identification and profiling experiments. The goal of identification is to find
out the amino acid sequence of extracted proteins; whereas the goal of profiling is to
quantify the expression level of proteins.
Proteomics data is measured in mass-to-charge ratios, termed m/z values. Peaks can be
identified when we plot measured intensities against m/z values. The location of a peak
corresponds to the chemical composition of the protein and thus can be used for protein
identification, and the amplitude of a peak carries information on the protein abundance,
which is the basis of profiling. Such data provides a unique challenge both to encode and
to analyze. Firstly, even modest-sized experiments provide gigabytes worth of data.
Encoding and exchanging data of this size is quite a technical challenge without mention
of the problem of analyzing such a large dataset. Secondly, the data itself has intrinsic
properties that make it particularly challenging to analyze and interpret. For example, one
common task in analyzing spectra data is to locate data peaks. However, what charac-
teristics of the peaks are most important: height, area under the peak, or some ratio
thereof? Furthermore, noise in a single spectrum and aggregating multiple spectra
provide unique statistical challenges.
There has been some work towards standardization of proteomics data. The Proteomics
Standards Initiative (PSI) has been created under the auspices of the Human Proteome
Organization (HUPO). PSI has working groups to develop data standards for protein-
protein interaction and mass spectrometry (Orchard et al., 2004). Though good inten-
tioned, the PSI mass spectrometry data standards are still in an abstract form. The
Minimum Information About Proteomics Experiments is a document describing the types
of data items needed to fully described proteomics experiments, but no data standard
exists. The PSI Object Model (PSI-OM) and PSI Markup Language (PSI-ML) have defined
a set of concepts and the relationships of those concepts, but not concrete data formats.
Finally, the PSI Ontology (PSI-Ont) is being developed as an extension to the Microarray
and Gene Expression Data Ontology (MGED Ontology) with the PSI namespace, but is
not yet complete. Proteomics data standardization is far from complete, and there is much
work to be done to define data formats and ontologies.
Proteomics with Mass Spectrometry 87
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Data
Production of Proteomics Data
Proteomics data is produced from mass spectrometry experiments, which measures the
atomic mass of peptides (see Figure 1). Classical proteomics study begins with 2D-gel
for protein separation prior to mass spectrometry identification of gel spots. First, a
biologic sample containing some proteins of interest is prepared. The proteins are then
digested, or split, into peptide fragments. These fragments are ionized just before they
enter the mass analyzer and detector, where their masses are measured. There are several
techniques for measuring the mass. One popular approach, termed Time-of-Flight or just
TOF, measures the amount of time an ion takes to reach a detector after it enters the mass
spectrometer. Currently, the most accurate method, Fourier-transform ion cyclotron
resonance (FTICR), involves measuring the frequency that ions circle the high magnetic
field of a super-conducting magnet. The resulting data from the mass spectrometer is a
list of masses and the relative abundance of the ions for that mass, termed the intensity.
A secondary, or tandem, mass spectrometry reading is sometimes taken in which the
peptide itself is then broken into its components and mass readings are taken. This allows
for a much more accurate reading (Steen and Mann, 2004). The fingerprint generated by
the mass spectrometry is then matched against a database of known proteins. The peak
height information is usually discarded during the database search.
Although mass spectrometry is often used to identify proteins first separated by 2D-gel,
several recent studies have suggests that direct analysis of intact protein (without gel
separation and digestion) with MALDI-TOF or SELDI-TOF can provide a measurement
of the protein expression level, as indicated by the peak heights (Rosenblatt et al., 2004).
The goal of a proteomics experiment is usually one of two objectives: accurately identify
proteins in a sample or create a predictive model from a set of proteomics spectra. These
are the protein identification and protein profiling problems, respectively. In protein
identification, the goal is to as accurately as possible measure the peptide fragments of
proteins of interest. The masses of the high intensity peptides are searched against a
Figure 1. Process flow for a typical proteomics experiment: Different options/
technologies are listed below each step (Steen and Mann, 2004, reprinted with
copyright permission from Nature Reviews Molecular Cell Biology)
Cell culture
SDS-Page
2D-gel
electrophoresis
Trypsin
Lys-C
Asp-N
Glu-C
Sample
preparation/
fractionation
Protein
digestion
Peptide
seperation
Peptide
HPLC
Ion exchange
Electrospray
ionization
MALDI
Sample
ionization
Spray
needle
Pepti se
i ons
Mass
spectrometry
Quadrupole
Time of flight
Quadrupole ion traps
FTICR
Mass spectrum
I
n
t
e
n
s
it
y
m/z
Data
analysis
PeptideSearch
Sequest
Mascot
Nature Reviews Molecular Cell Biology
88 Lin, Mungal, Haney, Patz, & McConnell
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
database to identify which proteins exist in the sample. In proteomics profiling studies,
the mass/intensity spectra of different samples are used to create a predictive model. The
data is fed into a machine learning algorithm which can later be used as a model to predict
what class a particular spectra falls into, for example diseased and healthy.
Overview of m/z Data
A spectrum collected from a proteomics experiment is a list of peaks that consist of a pair
of values. The first value is the mass-to-charge ratio, often referred to as the m/z value.
The charge is included because sometimes larger fragments can gain greater than a single
charge during the ionization step. The second value is the intensity, or abundance of the
peak. That intensity value is relative to a particular run of the mass spectrometer and
typically requires statistical normalization to be comparable to other runs. This full list
of m/z-intensity values makes up a proteomics spectra (see Figure 2).
As noted previously in Figure 1, there can be a third dimension to proteomics experi-
ments, such as when a mass spectrometer is used to analyze data that comes out of a liquid
chromatography (LC) data. In that case, each spectrum is associated with a retention time,
that is, the time that a sample takes to go through the chromatography column.
Successive scans come out at the intervals of several seconds. So, the data has three
dimensions overall, involving m/z values, LC retention times, and finally the MS intensity
values. In addition to making 2-dimensional (2-D) plots such as the one depicted in Figure
2, systems analyzing this data also make plots in 3-D.
Figure 2. A proteomics spectra has the mass values running along the x-axis and the
intensities along the y axis; the mass is actually the mass-to-charge ratio, as peptides
can gain more than a single charge when ionized. the intensity is typically measure by
counts of ions for each mass value (Steen and Mann, 2004, reprinted with copyright
permission from Nature Reviews Molecular Cell Biology)

800
600
400
200
0
400 600 800 500 700 900
m/z
I
n
t
e
n
s
i
t
y

(
c
o
u
n
t
s
)
617.28
617.78
618.28
300
200
100
0
615
m/z
620
716.41
615
534.75
435.77
Proteomics with Mass Spectrometry 89
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Important Metadata Considerations
In addition to the m/z-intensity values, it is important to keep track of other experiment-
related metadata. In the data modeling world, metadata is simply data about the data of
interest. For proteomics, there are some important categories of metadata.
Data Acquisition
It is important to keep track of how the data was acquired, from the biological sample
through any lab procedures used, and up to the ionization procedure. There is a class
of software called Laboratory Information Management Systems (LIMS) for tracking
these sorts of data. LIMS are usually quite large, complex, and expensive systems, and
when it comes to analysis, the information stored in a LIMS must be attached to the data
for processing. Data acquisition considerations can impact how data is processed,
analyzed, and interpreted.
Instrumentation
The equipment used to ionize the peptides and the mass-spectrometry equipment should
be tracked with the m/z-intensity data. This instrumentation can be an important
consideration when performing data processing and analysis, as the mass spectrometer
can have a large impact on the sensitivity and accuracy of the data.
Data Processing
As you will read in the Analysis section of this chapter, there is a full pipeline of processes
that can be run on the data. Many of these statistical analyses simply modify or reduce
the m/z-intensity values. This is important to keep track of because the data processing
can have a large impact on how the data should be interpreted.
The Proteomic Standard Initiative (PSI) is a group that is attempting to define the
Minimum Information About a Proteomics Experiment (MIAPE) that should be collected
(Orchard et al., 2004). At the time of writing, the MIAPE standard is only a draft document
outlining some metadata guidelines. However, when the document reaches maturity, it
will be an invaluable resource for proteomics metadata standards.
Available Data Standards
As of the writing of this chapter, there are two major efforts underway to provide a
standard encoding of proteomics data. The first effort called mzXML is by the Institute
for System Biology (ISB) (Pedrioli et al., 2004). The second effort called mzData is by the
90 Lin, Mungal, Haney, Patz, & McConnell
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Table 1. Example mzXML file (Notice the instrumentation and dataprocessing metadata
associated with the msRun and the metadata attributes associated with the scan; the
second row shows a standard SAX XML parser written in Java, and characters between
start and end elements are stored in a StringBuffer, attributes are parsed in the
startElement method, and the actual peak values should be parsed in the endElement
method.)
<?xml version="1.0" encoding="ISO-8859-1"?>
<msRun
xmlns="http://sashimi.sourceforge.net/schema/"
xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
xsi:schemaLocation="http://sashimi.sourceforge.net/schema/
http://sashimi.sourceforge.net/schema/MsXML.xsd"
scanCount="7198"
startTime="PT0S"
endTime="PT120S"
>
<parentFile fileName="myData.wiff" fileType="RAWData"
fileSha1="068adb0e388196bc03ad95cd7140c4cce233a181"
/>

<instrument
manufacturer="ABI" model="QSTAR"
ionisation="ESI" msType="Q-TOF"
cid="LE"
>
<software type="acquisition" name="Analyst" version="1.0"/>
</instrument>

<dataProcessing intensityCutoff="0">
<software type="conversion" name="mzStar" version="1.0.6"/>
<processingOperation name="min_peaks_per_spectra" value="1"/>
</dataProcessing>

<scan num="1" msLevel="1" peaksCount="4246" polarity="+"
retentionTime="PT1.00499999523163S" startMz="400" endMz="3500"
basePeakMz="447.352228504701" basePeakIntensity="8"
totIonCurrent="9016"
>
<peaks precision="32">Q8gbrEAAAABDyJx6QAAAAEPJF6 . . .</peaks>
</scan>

</msRun>

Proteomics Standards Initiative (PSI) (Orchard et al., 2004). Both these standards are very
similar and can be used interchangeably with some translation of formats, so we will focus
on mzXML because it was the first and is the most widely used. However, it there may
be convergence on mzData sometime in the future because it has the backing of a
standards group.
mzXML is a standard for encoding proteomics m/z-intensity data. All data is encoded in
the Extensible Markup Language (XML), which is a textual format for encoding arbitrary
data. The m/z-intensity values are encoded as in Base64, which is an efficient encoding of
binary data into text. Although Base64 is unreadable by humans, it provides the relatively
low overhead of encoding three bytes of data in four textual characters (Table 1).
Proteomics with Mass Spectrometry 91
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
mzXML also encodes some of the necessary metadata associated with instrumentation
and data processing. For each mass spectrometry run, there is an instrumentation element
that encodes the mass spectrometer manufacturer, model, ionization method, and mass
spectrometry type. Additionally, there is a dataProcessing element that provides a
description of the software employed, and a generic description of the operation utilized
(Table 1).
Data Parsing
At this time, most proteomics data is exchanged as comma delimited (CSV) text files (see
Table 2). A drawback of this approach is that it takes upwards of 12 bytes to encode four
pr i va t e c l a s s Mz Xml Ha ndl e r e xt e nds De f a ul t Ha ndl e r
{
pr i va t e St r i ngBuf f e r c ha r s = ne w St r i ngBuf f e r ( ) ; / / c ha r s be t we e n e l s
pr i va t e i nt pe a kCount = 0; / / numbe r of va l ue pa i r s t o de c ode
pr i va t e i nt pr e c i s i on = 0; / / 32 or 64
pr i va t e i nt s c a nNumbe r = 0; / / s c a n numbe r

publ i c voi d s t a r t El e me nt (
St r i ng na me s pa c e URI , St r i ng l oc a l Na me , St r i ng qNa me ,
At t r i but e s a t t s
)
t hr ows SAXExc e pt i on
{
c ha r s . de l e t e ( 0, c ha r s . l e ngt h( ) ) ;

/ / pa r s e out a t t r i but e va l ue s
i f ( qNa me . e qua l s ( " s c a n" ) ) {
pe a kCount = I nt e ge r . pa r s e I nt ( a t t s . ge t Va l ue ( " pe a ks Count " ) ) ;
s c a nNumbe r = I nt e ge r . pa r s e I nt ( a t t s . ge t Va l ue ( " num" ) ) ;
} e l s e i f ( qNa me . e qua l s ( " pe a ks " ) ) {
pr e c i s i on = I nt e ge r . pa r s e I nt ( a t t s . ge t Va l ue ( " pr e c i s i on" ) ) ;
}
}

publ i c voi d c ha r a c t e r s ( c ha r buf [ ] , i nt of f s e t , i nt l e n)
t hr ows SAXExc e pt i on
{
/ / s t or e va l ue s be t we e n s t a r t a nd e nd t a gs
c ha r s . a ppe nd( buf , of f s e t , l e n) ;
}

publ i c voi d e ndEl e me nt (
St r i ng na me s pa c e URI , St r i ng l oc a l Na me , St r i ng qNa me
)
t hr ows SAXExc e pt i on
{
i f ( qNa me . e qua l s ( " pe a ks " ) ) {
/ / de c ode ba s e 64 va l ue s ba s e d on pe a kCount a nd pr e c i s i on
/ / ha ndl e ba s e 64 va l ue s ba s e d on s c a nNumbe r
}

c ha r s . de l e t e ( 0, c ha r s . l e ngt h( ) ) ;
}
}

Table 1. continued
92 Lin, Mungal, Haney, Patz, & McConnell
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
bytes of data. The benefits of CSV files are that the files are convenient for people to read
and the ease of integration with existing software. For example, CSV files can be read into
statistical packages such as R or Mathematica with one line of code and can be imported
into Microsoft Excel with ease (see Table 2).
To be sure, CSV files of m/z and intensity values contain no metadata. Hence it is still
necessary to provide an additional file containing that data. It is possible to have this
file of attributes and attribute values also be in a CSV form. However, as of this writing,
when proteomics data is exchange using CSV files, most of the time, such files of metadata
are not included.
The other main approach involves encoding proteomics data using an XML-based
format. One major issue with encoding proteomics data in XML is that XML is more
technically difficult to parse than other text formats (see Table 1). This problem is
compounded by the fact that the peak values are stored in base64, which is not
immediately useful to the programmer. An extra step of decoding the base64 values is
needed. Luckily, there are numerous free base64 parsers available on the Web. And, to
ease the use of mzXML, ISB has created a suite of tools for converting and parsing
mzXML data. They have written converters from the binary output of many popular mass
spectrometers, libraries in C++ and Java for parsing mzXML files, and graphical viewers
Table 2. Example of comma-delimited proteomics data (the first column contains m/z
values and the second column contains intensity values; unlike mzXML, there is not
metadata associated with the data, the second row shows R code to read the m/z-
intensity data into a data frame for easy manipulation)
M/ Z, I nt e ns i t y
1999. 883455, 1207. 3
2000. 178622, 1279. 94
2000. 473811, 1324. 59
2000. 769021, 1360. 99
2001. 064254, 1412. 31
2001. 359508, 1486. 62
2001. 654784, 1457. 04
2001. 950082, 1442. 05
2002. 245402, 1376. 06
2002. 540743, 1331. 33
2002. 836107, 1344. 02
2003. 131492, 1410. 01
2003. 426899, 1427. 38
2003. 722328, 1461. 4
2004. 017778, 1514. 69
2004. 313251, 1557. 36
2004. 608745, 1567. 99
2004. 904261, 1597. 97
2005. 199799, 1593. 68
2005. 495359, 1585. 36
2005. 790940, 1527. 7
2006. 086544, 1474. 01
2006. 382169, 1448. 39
# R c ode t o r e a d a CSV f i l e i nt o a da t a f r a me
da t a = r e a d. c s v( f i l e , he a de r =TRUE)

Proteomics with Mass Spectrometry 93
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
for inspecting mzXML. However, one of ISBs most interesting contributions is the
ability to index an mzXML file and access scans in a non-sequential way.
It should be noted that the above two formats (CSV and XML-based formats ) most often
represent data messaging formats. In practice, pipelines used for high throughput
proteomics frequently take data directly from the raw vendor formats from manufacturers
and convert that to binary files or store it in databases, without going through the
intermediary step of converting to CSV or XML form.
Publicly Available Data Sources
As of the writing of this chapter, there are few large-scale proteomics efforts or data
repositories. A good starting point for finding public data is the Open Proteomics
Database, which is a repository containing four organisms and 1.2 million spectra (Prince,
Carl son, Wang, Lu, & Marcot t e, 2004). The mzXML Web si t e (ht t p: / /
sashimi.sourceforge.net/repository.html) has a small repository of useful designed and
other experiments. Finally, the PeptideAtlas site has proteomics data from four organisms
and hundreds of thousands of spectra (Desiere et al., 2004).
Open Source and Commercial Databases
Currently, there are no open-source or commercial databases for proteomics data
available. The Fox Chase Cancer Center is currently working on an open-source
implementation of a proteomics LIMS for managing proteomics data from the point-of-
view of the lab. This project is slated to be completed in year 2006. Many mass
spectrometer vendors provide software for a fee that will manage proteomics data as it
is produced by their instrumentation. The fee is usually nominal in comparison to the
price of the mass spectrometer itself, being 1%-5% of the cost of the machine, so many
labs choose this option.
Analysis
We will use the data set in the Duke Proteomics Data Mining Conference (September 23,
2002) to illustrate the data analysis challenges. A collection of 41 clinical specimens were
measured by MALDI-TOF, of which 17 were normal and 24 were lung cancer. The first
goal is to find out whether we can classify the patients based on proteomics measure-
ments. If so, the second goal is to find out protein biomarkers that can differentiate cancer
versus normal. Results from this conference were compiled into a special issue of
Proteomics, a journal by Wiley-VCH (volume 3, No. 9, 2003).
94 Lin, Mungal, Haney, Patz, & McConnell
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Raw Data and Preprocessing
Without appropriate preprocessing, some systematic bias (irrelevant information) can
drive the data mining process to get superficial results. This argument is supported by
observed differences in replicated measurements of the same biosample. The goal of data
preprocessing is to reduce the variance seen in replicates of the same biosample, but also
to preserve the difference among different biosamples. Variance caused by systematic
bias can be corrected with baseline removal and normalization.
Raw Data
Ideally, a protein corresponds to a vertical line on the spectrum, and its abundance is
represented by the height of the line. In reality, due to heterogeneous molecular weight
caused by isotopic elements, and stochastic ionic movements during the measurement
Figure 3. Part of spectra showing the difference between cancer and normal: Ten
replicate runs of each biosample are shown superimposed (left) and stacked (right)
1
5
0
0
1
0
0
0
5
0
0
0
0 500 1000 1500 2000
0 500 1000 1500 2000
2
0
0
0
4
0
0
0
6
0
0
0
8
0
0
0
1
0
0
0
0
Proteomics with Mass Spectrometry 95
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
process, the vertical line is broadened into a peak. A total of 20 raw spectra (10 replicates
each from a cancer and a normal specimen) are shown in Figure 3.
Baseline Correction
With both MALDI-TOF and SELDI-TOF data, we usually observe an elevated baseline
at the low m/z range. Figure 4 illustrates two scans of the same biological sample. Elevated
baseline of scan 3 can be removed by subtracting the local minima smoothed by a cubic
spline (Berndt, Hobohm, & Langen, 1999). From this example, we clearly see that without
appropriate baseline correction, it is pointless to compare the intensity difference
between normal and cancer samples.
Similar to background removal in microarray image analysis, baseline correction of
spectrum data have the following implications:
1. Some negative intensities values may be generated after baseline removal. These
values can be reset to zero if necessary.
2. Incorrect choice of algorithm and parameters can results in erroneous results: to
one extreme, excessive correction will remove all the peaks and result in a flat
spectrum with noise only; to the other extreme, inadequate correction will not
remove the systematic bias.
Figure 4. The effect of baseline removal on spectra

96 Lin, Mungal, Haney, Patz, & McConnell
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Normalization
As seen in Figure 5, the maximum amplitude of scan 3 and scan 8, which are from the same
biosample, are different even after baseline correction. This difference can be corrected
by a multiplication factor. This simple operation corresponds to scale-normalization of
microarray data processing. The utilities of other more complex normalization methods
are yet to be evaluated.
Smoothing
As with time series data, the spectrum data contains a lot of noise. A smoothing process
can help removing some superficial spikes (see more discussion below in the peak finding
section), and thus help the peak finding process (Figure 6). It can be achieved by moving
average, smoothing splines (Gans & Gill, 1984), Fourier transform filtering (Cameron &
Moffatt, 1984), or wavelet filtering (Cancino-De-Greiff, Ramos-Garcia, & Lorenzo-Ginori,
2002).
Notice that excessive smoothing can change the shape of the peak: the peak will be
shorter and broader with moving average operation. Extreme smoothing can destroy the
peak features by generating a flat line.
Figure 5. The effect of scale normalization on spectra

Proteomics with Mass Spectrometry 97
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Feature Extraction
Each patient is described by a collection of m/z values. These raw m/z data can be used
directly to drive the pattern analysis (Sorace & Zhan, 2003). However, since some spectra
contain from one hundred thousand to one million data points, it is common to reduce
the amount of data, and use more interesting spectrum features. Here we briefly discuss
two widely used feature extraction methods in analytical chemistry for spectrum data.
Binning
The simplest possible method for reducing the information used in subsequent analyses
involves binning the data. In this approach, all m/z values within a given window are
condensed to one point. The window size used can be either fixed or varied. Domain-
specific knowledge of the x-axis measurement accuracy can guide the selection the
window size (Petricoin, Rajapaske, Herman, Arekani, Ross, & Johann, et al., 2004). For
example, by a variable-length window, 350 000 raw data features were reduced to 7 084
features (Petricoin, Rajapaske, et al., 2004), before they were analyzed with genetic
algorithm-based feature selection and self-organizing map (SOM) classification.
Notice that the discussion of binning here is in the context of feature reduction of x-axis,
rather than the usual discussion of binning continuous input into ordinary variable of
the y-axis.
Figure 6. Smoothing (black line) the raw data (grey line) helps the identification of
the peaks (dots)

98 Lin, Mungal, Haney, Patz, & McConnell
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Peak Finding
Theoretically, a protein is observed as a peak on the spectrum. Usually, peak height is
used to drive the down-stream analysis. Other measurements of a peak include full width
at half maximum (FWHM), peak location, and area under the curve (Figure 7).
Peaks can be found by the local maximum or by modeling peaks using a parametric model.
In this context, however, the presence of noise can interfere with such algorithms for
finding local maxima. Thus, it is often used together with adequate smoothing (Morrey,
1968). Parametric models can handle noise and outliers more elegantly with a model fitting
process. However, their parametric assumptions can also in practice limit their utility
(Scheeren, Barna, & Smit, 1985).
Data Mining and Analysis
After appropriate data preprocessing and feature extraction, proteomics data mining
shares a lot of similarity with microarrays data mining. For preprocessing steps, general
signal processing techniques are usually relevant as we have already discussed.
Depending on whether the patient outcomes are available at the model-building process,
proteomics data mining can be classified as either a supervised learning (outcome is
known) or an unsupervised learning (outcome is unknown) process (Figure 8).
Figure 7. A peak can be characterized with the following parameters: Peak location
(m), peak height (h), full width at half maximum (FWHM), and area under the curve
(AUC)

Proteomics with Mass Spectrometry 99
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
These data mining methods are related to clinical applications in the following ways:
Clustering can provide new insights into disease classification and can discover
new subtypes of a disease.
Supervised machine learning methods can automatically classify new measure-
ments into predefined disease categories, such as cancer or normal. Survival
modeling is a special kind of supervised analysis used in cancer prognostics, where
the clinical outcome variable is time to death (Shoemaker & Lin, 2005).
A majority of data mining techniques have been applied to proteomics, we list some
representative papers from 2002-2004 in Table 3.
Open Source and Commercial Tools
A commercial tool of ProteomeQuest (Correlogic Inc., Bethesda, MD) was used in
(Petricoin, Rajapaske, et al., 2004c). It is a combination of genetic-algorithm-based feature
selection and self-organizing-map-based pattern recognition. Excellent classification
result was reported by using this tool (Petricoin, Ardekani, Hitt, Levine, Fusaro, &
Steinberg, et al., 2002), although there are later controversies of this study (Baggerly,
Morris, & Coombes, 2004; Petricoin, Fishman, Conrads, Veenstra, & Liotta, 2004).
Figure 8. Predictive data mining: Proteomics measurements (features) of patients are
linked to outcomes, such as cancer or norma; if the outcomes are not available at the
time of the modeling process, it becomes an unsupervised machine learning problem
Patient
n

Patient
2
Patient
1
Feature
p
Feature
2
Feature
1
Outcome
n

Outcome
2
Outcome
1
Outcomes
100 Lin, Mungal, Haney, Patz, & McConnell
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
There are several open source tools written in the R statistical language for analyzing
proteomics data, including PROcess in the Bioconductor package (http://
www.bioconductor.org), the mscalib at the CRAN distribution site (http://cran.r-
project.org/), and the Rproteomics under the caBIG project (http://cabig.nci.nih.gov).
Since these tools are all on the R computing platform, many modules can be mixed-and-
matched for practical uses.
Summary
The State-of-the-Art
The expression level of a single protein, such as PSA or Ca125, has been used for clinical
diagnostics of prostate and ovarian cancer. From a data mining point of view, these tests
are univariate classifiers. With the advance of proteomics, we are able to measure the
expression level of many proteins in one run. Thus, a more powerful multivariate pattern
can be used for clinical diagnostics. However, this approach faces typical data mining
challenges:
1. Accuracy and reproducibility of high-throughput measurements
2. Curse of high dimensionality: very large number of features vs. very small number
of training samples
3. Concerns of over-fitting the model
4. Interpretability of the potential black-box models
Even with these obstacles, current researches have shown the early promises of clinical
proteomics application (Carr, Rosenblatt, Petricoin, & Liotta, 2004; Petricoin, Wulfkuhle,
Espina, & Liotta, 2004).
Table 3. Representative data mining methods used in proteomics
Proteomics Study Data Mining Methodology Citation
Breast cancer multivariate logistic regression (Li et al., 2002)
Drug toxicity genetic algorithm and SOM (Petricoin, Rajapaske, et al., 2004)
Taxol resistance artificial neural network (Mian et al., 2003)
Renal carcinoma decision tree (Won et al., 2003)
Prostate cancer boosted decision tree (Qu et al., 2002)
Breast cancer unified maximum separation analysis (Li et al., 2002)
Colorectal cancer support vector machine (Yu et al., 2004)

Proteomics with Mass Spectrometry 101
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
The Future of Proteomics
Experimental studies have suggested there is not necessarily a good correlation between
the mRNA expression and the protein expression level (Greenbaum et al., 2003; Greenbaum,
Luscombe, Jansen, Qian, & Gerstein, 2001). Thus, proteomics measurements provide
additional information to microarray measurements. Combining proteomics with
transcriptomics seems logical but aggravates the high-dimensionality problem of data
mining and data analysis.
Integrating multiple omics data which include transcriptome, proteome, glycome,
interactome, and metabolome to study systems biology is a major challenge in
bioinformatics. Usually, information on different omics is available from separate data
sources. Data fusion provides a way to combine information for data mining. Proteome
adds a new dimension to this already complex problem (Yates, 2004). As the measurement
platforms of proteomics getting mature and becoming more accurate, we will expect to
see more discussion of data mining and data fusion challenges in system biology.
References
Aebersold, R., & Mann, M. (2003). Mass spectrometry-based proteomics. Nature, 422,
198-207.
Baggerly, K. A., Morris, J. S., & Coombes, K. R. (2004). Reproducibility of SELDI-TOF
protein patterns in serum: Comparing datasets from different experiments.
Bioinformatics, 20, 777-785.
Berndt, P., Hobohm, U., & Langen, H. (1999). Reliable automatic protein identification
from matrix-assisted laser desorption/ionization mass spectrometric peptide fin-
gerprints. Electrophoresis, 20, 3521-3526.
Cameron, D. G., & Moffatt, D. J. (1984). Deconvolution, derivation, and smoothing of
spectra using Fourier-transforms. Journal of Testing and Evaluation, 12, 78-85.
Cancino-De-Greiff, H. F., Ramos-Garcia, R., & Lorenzo-Ginori, J. V. (2002). Signal de-
noising in magnetic resonance spectroscopy using wavelet transforms. Concepts
in Magnetic Resonance, 14, 388-401.
Carr, K. M., Rosenblatt, K., Petricoin, E. F., & Liotta, L.A. (2004). Genomic and proteomic
approaches for studying human cancer: Prospects for true patient-tailored therapy.
Human Genomics, 1, 134-140.
Desiere, F., Deutsch, E. W., Nesvizhskii, A. I., Mallick, P., King, N. L., Eng, J. K., et al.
(2005). Integration with the human genome of peptide sequences obtained by high-
throughput mass spectrometry. Genome Biology, 6(1), R9.
Gans, P., & Gill, J.B. (1984). Smoothing and differentiation of spectroscopic curves using
spline functions. Applied Spectroscopy, 38, 370-376.
102 Lin, Mungal, Haney, Patz, & McConnell
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Greenbaum, D., Colangelo, C., Williams, K., & Gerstein, M. (2003). Comparing protein
abundance and mRNA expression levels on a genomic scale. Genome Biology, 4,
117.
Greenbaum, D., Luscombe, N. M., Jansen, R., Qian, J., & Gerstein, M. (2001). Interrelating
different types of genomic data, from proteome to secretome: oming in on function.
Genome Research, 11, 1463-1468.
Li, J., Zhang, Z., Rosenzweig, J., Wang, Y. Y., & Chan, D. W. (2002). Proteomics and
bioinformatics approaches for identification of serum biomarkers to detect breast
cancer. Clinical Chemistry, 48, 1296-1304.
Mian, S., Ball, G., Hornbuckle, J., Holding, F., Carmichael, J., Ellis, I., et al. (2003). A
prototype methodology combining surface-enhanced laser desorption/ionization
protein chip technology and artificial neural network algorithms to predict the
chemoresponsiveness of breast cancer cell lines exposed to Paclitaxel and Doxo-
rubicin under in vitro conditions. Proteomics, 3, 1725-1737.
Morrey, J. R. (1968). On determining spectral peak positions from composite spectra with
a digital computer. Analytical Chemistry, 40, 905-914.
Orchard, S., Hermjakob, H., Julian, R. K., Jr., Runte, K., Sherman, D., Wojcik, J., Zhu, W.,
& Apweiler, R. (2004). Common interchange standards for proteomics data: Public
availability of tools and schema. Proteomics, 4, 490-491.
Pedrioli, P. G., Eng, J. K., Hubley, R., Vogelzang, M., Deutsch, E. W., Raught, B., et al.
(2004). A common open representation of mass spectrometry data and its applica-
tion to proteomics research. Nature Biotechnology, 22(11), 1459-1466.
Petricoin, E., Wulfkuhle, J., Espina, V. & Liotta, L.A. (2004). Clinical proteomics: revolu-
tionizing disease detection and patient tailoring therapy. Journal of Proteome
Research, 3, 209-217.
Petricoin, E. F., Ardekani, A. M., Hitt, B. A., Levine, P. J., Fusaro, V. A., Steinberg, S. M.,
et al. (2002). Use of proteomic patterns in serum to identify ovarian cancer. Lancet,
359, 572-577.
Petricoin, E. F., Fishman, D. A., Conrads, T. P., Veenstra, T. D., & Liotta, L. A. (2004).
Proteomic pattern diagnostics: Producers and consumers in the era of correlative
science. BMC Bioinformatics. Retrieved February 1, 2006, from http://
www.biomedcentral.com/1471-2105/4/24/comments#14454
Petricoin, E. F., Rajapaske, V., Herman, E. H., Arekani, A. M., Ross, S., Johann, D., et al.
(2004). Toxicoproteomics: Serum proteomic pattern diagnostics for early detection
of drug induced cardiac toxicities and cardioprotection. Toxicologic Pathology,
32, 122-130.
Prince, J. T., Carlson, M. W., Wang, R., Lu, P., & Marcotte, E. M. (2004). The need for a
public proteomics repository. Nature Biotechnology, 22, 471-472.
Qu, Y., Adam, B. L., Yasui, Y., Ward, M. D., Cazares, L. H., Schellhammer, P. F., et al. (2002).
Boosted decision tree analysis of surface-enhanced laser desorption/ionization
mass spectral serum profiles discriminates prostate cancer from noncancer pa-
tients. Clinical Chemistry, 48, 1835-1843.
Proteomics with Mass Spectrometry 103
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Rosenblatt, K. P., Bryant-Greenwood, P., Killian, J. K., Mehta, A., Geho, D., Espina, V.,
et al. (2004). Serum proteomics in cancer diagnosis and management. Annual
Review of Medicine, 55, 97-112.
Scheeren, P. J. H., Barna, P., & Smit, H. C. (1985). A Software package for the evaluation
of peak parameters in an analytical signal based on a non-linear regression method.
Analytica Chimica Acta, 167, 65-80.
Shoemaker, J. S., & Lin, S. M. (2005). Methods of microarray data analysis IV. New York:
Springer.
Sorace, J. M., & Zhan, M. (2003). A data review and re-assessment of ovarian cancer
serum proteomic profiling. BMC Bioinformatics, 4, 24.
Won, Y., Song, H. J., Kang, T. W., Kim, J. J., Han, B. D., & Lee, S. W. (2003). Pattern
analysis of serum proteome distinguishes renal cell carcinoma from other urologic
diseases and healthy persons. Proteomics, 3, 2310-6.
Yates, J. R., 3
rd
. (2004). Mass spectrometry as an emerging tool for systems biology.
Biotechniques, 36, 917-9.
Yu, J. K., Chen, Y. D., & Zheng, S. (2004). An integrated approach to the detection of
colorectal cancer utilizing proteomics and bioinformatics. World Journal of Gas-
troenterology, 10, 3127-3131.
104 Rahmann, Mller, Dandekar, & Wolf
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Chapter VI
Efficient and Robust
Analysis of Large
Phylogenetic Datasets
Sven Rahmann, Bielefeld University, Germany
Tobias Mller, University of Wrzburg, Germany
Thomas Dandekar, University of Wrzburg, Germany
Matthias Wolf, University of Wrzburg, Germany
Abstract
The goal of phylogenetics is to reconstruct ancestral relationships between different
taxa, e.g., different species in the tree of life, by means of certain characters, such as
genomic sequences. We consider the prominent problem of reconstructing the basal
phylogenetic tree topology when several subclades have already been identified or are
well known by other means, such as morphological characteristics. Whereas most
available tools attempt to estimate a fully resolved tree from scratch, the profile
neighbor-joining (PNJ) method focuses directly on the mentioned problem and has
proven a robust and efficient method for large-scale datasets, especially when used in
an iterative way. We describe an implementation of this idea, the ProfDist software
package, which is freely available, and apply the method to estimate the phylogeny of
the eukaryotes. Overall, the PNJ approach provides a novel effective way to mine large
sequence datasets for relevant phylogenetic information.
Efficient and Robust Analysis of Large Phylogenetic Datasets 105
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Introduction
Basics
Phylogenetic analyses aim to reconstruct ancestral relationships (the phylogeny)
between different taxa, e.g., different species in the tree of life, by means of certain
morphological or molecular characters. The latter ones are available from the increasing
number of sequencing projects. Because of the wealth of data generated in these
projects, we need efficient methods to mine the sequence databases for relevant
phylogenetic information if we want to reconstruct the most likely phylogeny.
Over the last 50 years, researchers have proposed many methods for phylogenetic tree
reconstruction, based on different concepts and models. Each method has its strengths
and weaknesses. Distance-based methods, such as Neighbor-joining (Saitou and Nei,
1987) or improved variants thereof, e.g., WEIGHBOR (Bruno, Socci, & Halpern, 2000),
BIONJ (Gascuel, 1997), FASTME (Desper & Gascuel, 2002), are relatively fast (the
running times generally grows as a cubic polynomial of the number of taxa), but first
reduce the information contained in the characters to a matrix of distances. Character-
based methods, on the other hand, such as maximum parsimony (Camin & Sokal, 1965;
Fitch, 1971), maximum-likelihood (Felsenstein, 1981), or Bayesian methods (e.g., MrBayes;
Huelsenbeck & Ronquist, 2001; see Holder & Lewis, 2003, for a review) work directly with
character data, but usually require an evaluation of super exponentially many tree
topologies; therefore one reverts to heuristics. There seems to be no universally
accepted best method, especially for large datasets.
All of the above-mentioned methods aim to estimate a fully resolved tree from scratch.
In some cases, this results in more computational work than needs to be done, or even
than the data can robustly support, as usually evidenced by low bootstrap values in the
basal branching pattern of the tree. In several analyses, however, we are mainly interested
in the basal branching pattern of known or clearly separated and fully supported
subclades. In other words, given families of closely related sequences, what is the
topology showing the relationships between these families? The problem arises for many
large phylogenetic datasets. Mller, Rahmann, Dandekar, and Wolf (2004) recently
published a case study on the Chlorophyceae (green algae). Further prominent large
phylogenetic datasets in question are the given families of related sequences of, e.g., the
Viridiplantae, the Metazoa, the eukaryotes, or even of all the species classified within the
tree of life.
As the field of phylogenetics involves both the biological and the mathematical
community, the language can sometimes be confusing. For the purpose of this article,
we use the terms subtree and subclade synonymously; each subtree defines a
monophyletic group of taxa, i.e., a grouping in which all species share a common ancestor,
and all species derived from that common ancestor are included. We usually work with
unrooted trees, i.e., connected undirected acyclic graphs, in which every internal node
has exactly three neighboring nodes. The taxa are the leaves of the tree.
106 Rahmann, Mller, Dandekar, & Wolf
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
It is well known that there exist super-exponentially many unrooted tree topologies for
N taxa, the precise number is (2N-5)!! := (2N-5) x (2N-7) x ... x 3; see the book by Felsenstein
(2003) for more details. For the purposes of illustration, we shall assume that there are
five taxa A, B, C and D, and E, for which there are 15 unrooted tree topologies. If we already
know that A and B form a subclade S, the number of possibilities reduces to 3 unrooted
trees (e.g., C and S could be neighbors, along with the other neighboring pair D and E;
we denote this tree by CS|DE. The other two trees would be DS|CE and ES|CD). So even
a little additional knowledge leads to a significant reduction in the number of possible
tree topologies, an important aspect when using cubic-time tree reconstruction algo-
rithms. As an additional benefit, we observe below that the resulting basal trees tend to
be more robust, i.e., are supported by higher bootstrap values, than they would be when
reconstructed from scratch.
Seen from a broader perspective, the knowledge of a single split (i.e., a single edge of the
tree that separates the taxa into two disjoint nonempty subsets) allows us to divide the
large problem of reconstructing the tree on n taxa into two smaller problems, namely
reconstructing trees on n
1
and n
2
taxa, respectively, with n
1
+n
2
=n.
The Challenge: Integrating Additional Knowledge
Interest in large phylogenies, e.g., the whole tree of life, or only the ancestral
relationships among Eukaryotes, has been and still is increasing, but none of the
standard methods mentioned above is well suited for datasets of this scale, at least for
reconstructing a fully resolved tree from scratch. However, we often do have access to
additional information that allows the definition of subclades. Thus, we face the
challenge of integrating additional knowledge efficiently into the established tree
reconstruction algorithms. A solution to the problem can be formulated in several ways;
we shall now compare some of them to give the reader an overview.
The first possibility consists of keeping the whole set of taxa, but to restrict the set of
allowable tree topologies to those that are consistent with the known monophyletic
groups. Such approaches work best when used with character-based methods, e.g., in
a maximum likelihood or in a Bayesian framework, since the constraints can be integrated
in the heuristic topology search that has to be performed anyway. The PAUP* package
(Swofford, 2003) implements such constraints, for example. This type of method works
less efficiently with distance-based methods, since we do not save any distance
computations.
A second possibility consists of solving many small problems, each with a reduced set
of taxa, and to assemble the resulting (sub-)trees into a consistent supertree. There are
two scenarios.
1. The subtrees are overlapping, i.e., they may share taxa, but propose inconsistent
neighbor relationships, and the problem becomes to find a supertree that is
maximally consistent with the set of subtrees for some measure of consistency.
Such general supertree methods are reviewed by Bininda-Emonds (2004); the
TREE-PUZZLE software (Schmidt & von Haeseler, 2003) that assembles a phylo-
Efficient and Robust Analysis of Large Phylogenetic Datasets 107
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
genetic tree from its quartets (subtrees with four taxa) is an example. These methods
do not require previous knowledge of monophyletic groups, but can benefit from
it.
2. In the second scenario, all subtrees are disjoint and form a partition of the whole
set of taxa. Such approaches are also called divide-and-conquer approaches; here
knowledge of the partition of the taxa is required. The problem reduces to (a)
representing the subtree by a new type of characters (it is not always represented
as a simple molecular sequence), and (b) to find a method that allows tree
reconstruction on this new set of characters.
To solve problem (a), it has been proposed to use consensus sequences, in other words,
a majority vote of all sequences in the subtree, or the most parsimonious ancestral
sequence, which is the sequence that requires the smallest total number of mutations in
the subtree, or to derive the most likely ancestor in a likelihood setting, or to use a
probabilistic model for the ancestral sequence. To solve problem (b) in the case of more
complex characters than molecular sequences, one has the possibility to extend existing
character-based methods or to first compute distances between the subtrees and then
use an existing distance-based method.
The approach taken in this chapter consists of replacing the taxa in a known subclade
by a single supertaxon, which is represented by a probabilistic sequence, also called
a sequence profile (Gribskov, McLachlan, & Eisenberg, 1987). Then evolutionary
distances can be estimated by a generalization of the standard maximum likelihood
distance estimator this is explained in detail below and represents our solution to
problem (b) above. Subsequently, any distance-based reconstruction method can be
used. In our implementation, we have chosen the Neighbor Joining algorithm; hence we
call the resulting methop profile neighbor-joining (PNJ).
Naturally, other combinations of approaches would also work. Currently there is
relatively little practical experience with supertree methods of the PNJ type, but we expect
to see more research on the topic in the future.
The remainder of this chapter is organized as follows: After discussing the ideas behind
PNJ in detail, as well as an automatic iterative variant, we present a freely available
implementation, ProfDist (Friedrich, Wolf, Dandekar, & Mller, 2005), and as a biological
application, a reconstructed tree of the eukaryotes. A discussion concludes this chapter.
Profile Neighbor-Joining
Profiles
A sequence profile is a stochastic model of a sequence family. It can also be pictured
as a fuzzy or smeared-out sequence. Formally, a profile consists of a number of
probability distribution vectors, one for each position. So each position k specifies its
108 Rahmann, Mller, Dandekar, & Wolf
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
own nucleotide distribution
k
= (
k,A
,
k,C
,
k,G
,
k,T
). Nucleotide sequences are thus
special profiles (singleton families), where each
k
is given by one of the unit vectors A
= (1, 0, 0, 0), C = (0, 1, 0, 0), and so on.
There exist different methods for estimating profiles from sequence families. For example,
given a phylogenetic tree of sequence families, we could estimate the most likely profile
at the root, i.e., the profile that maximizes the joint probability of all descendant sequences
under a given evolutionary model. This idea is used, for example, in Felsensteins
maximum likelihood tree reconstruction method (Felsenstein, 1981). In the context of this
chapter, however, we are more interested in the center of gravity of the sequence family
distribution. Therefore, we use the position-specific relative character frequencies over
all sequence family members. This results in a robust estimate that is independent of
estimated subclade topologies; in fact, it corresponds to a star tree topology in the
subclade, giving the same weight to all sequences. Naturally, sequence weighting
schemes (e.g., see Krogh & Mitchison, 1995) can be used to obtain a different behavior
and simulate different subclade topologies. Our choice is motivated by the attempt to find
a middle road between simplicity, robustness, and accuracy; this will be discussed
further below.
The profiles can now be interpreted as the taxas characters, and any method that
supports characters with a continuous value range could directly be used to build a
tree. For example, Felsensteins maximum likelihood (ML) method easily accommodates
profiles. For efficiency reasons, however, we consider it advantageous to use a distance-
based method and thus require a way to estimate distances between sequence profiles.
Evolutionary Markov Processes
Maximum likelihood methods for distance estimation (also called correction formulas,
because they map or correct the number of observed mutations to an evolutionary
distance measured in units of percent of expected mutations) rely on a model of molecular
substitution. Substitutions are modeled by an evolutionary Markov process (EMP)
acting independently on each site of the sequence. Different models have been proposed
and estimated for nucleotide and amino acid sequences (Lanave, Preparata, Saccone, &
Serio, 1984; Mller & Vingron, 2000).
An EMP is uniquely described by its starting distribution
0
and its rate matrix Q, called
the substitution model. In the nucleotide case,
0
is a row vector of nucleotide frequen-
cies summing to 1, and Q is a 4x4 matrix, where for ij, the entry Q
ij
specifies the rate at
which nucleotide i mutates into nucleotide j. We assume that all rates are positive, i.e.,
that all substitutions have nonzero probability in any time interval. The diagonal entry
Q
ii
is set to the negative sum of the other rates in row i, so each row sums to zero. We
assume that Q has a unique stationary distribution , i.e., a distribution that remains
unchanged under the evolutionary dynamics defined by Q; it is the only probability row
vector that satisfies Q = 0. Further, we assume that the process starts in its stationary
distribution, so
0
=. We also need to fix a time unit: The rates are calibrated in such a
way that per 100 time units, one substitution is expected to occur:
Efficient and Robust Analysis of Large Phylogenetic Datasets 109
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.

i

j i
Q
ij
=
i

i
Q
ii
= 1/100.
This condition can always be satisfied by re-scaling Q by a constant factor, if necessary.
Depending on its parameterization, the rate matrix is called Jukes-Cantor model (1
parameter, all rates are equal; Jukes & Cantor, 1969), Kimura model (2 parameters, with
different rates for transitions and transversions; Kimura, 1980), or general time reversible
(GTR) model (6 free parameters; Lanave et al., 1984). The parameters (such as the
transition-transversion ratio in the Kimura model) must be estimated from the data. Since
our main aim is to reconstruct the basal branching topology, we can use sequence data
from all subclades to estimate a single overall rate matrix; different methods have been
described in the literature (e.g., Mller & Vingron, 2000) or integrated into phylogenetic
software (e.g., MrBayes; Huelsenbeck & Ronquist, 2001).
Distance Estimation with EMPs
We symbolically write ij for the event that character i has been substituted by
nucleotide j after some unknown time t. The probability of this substitution is given by
the (i,j) entry of the time-t transition matrix P
t
, which is related to the rate matrix Q via the
matrix exponential P
t
= exp(tQ). The entry P
t
ij
is equal to the conditional probability that
what is now character i will have become character j after time t.
To estimate the evolutionary distance between two sequences, we first compute a
pairwise alignment (or take a 2-dimensional projection of a multiple alignment) and count
the number of all nucleotide pair types in the alignment. Let N
ij
be the number of observed
events ij. In general, the higher the off-diagonal entries (mismatch counts) are in
comparison to the diagonal entries (match counts), the larger is the evolutionary
distance. As we are working with unrooted trees where we initially cannot infer the
direction of time, we work with time-reversible models. In particular, we count substitu-
tions in both directions, so seeing i aligned with j, we add a counter to both N
ij
and N
ji
,
thus obtaining a symmetric count matrix.
A well-founded framework is given by the maximum likelihood principle. If the distance
is t time units, the joint probability of all events is given by
i,j
(P
t
ij
)
Nij
. We seek the value
t maximizing this probability, or equivalently, the log-likelihood function,
L(t) =
i,j
N
ij
log(P
t
ij
).
Note that L(t) is a sum of log-probabilities log(P
t
ij
), weighted by their observed counts
N
ij
, the counts summing up to the total number 2n of sites in the alignment (the total
number is 2n instead of n because we count reversibly). For general models Q, the solution
of this one-dimensional maximization problem cannot be given in closed form, but is
easily obtained by numerical methods.
We now generalize this estimator from sequences to profiles. A site need not contribute
an integer count of 1 to a single substitution category, but fractional counts summing
110 Rahmann, Mller, Dandekar, & Wolf
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
to 1 can be spread out over all 16 categories. Assume that the profiles at a particular site
of the two families are given by = (
A
,
C
,
G
,
T
) and = (
A
,
C
,
G
,
T
), respectively.
Intuitively, if the sequence families consist of m
1
and m
2
sequences and we independently
draw a sequence from each family, we observe the nucleotide pair (i, j) on average
m
1

i
m
2

j
times, corresponding to a relative frequency of
i

j
. The product distribution
mirrors the conditionally independent drawing from within the sequence families and
does not imply that the families might be unrelated. Thus the total counts N
ij
for event
ij are given by the sum over all n sites. Again, we enforce symmetry by counting again
with exchanged and :
N
ij
=
n
k=1

(
k,i

k,j
+
k,j

k,i
)
Except for the computation of N
ij
, the likelihood function L(t) remains unchanged, and
so does the numerical maximization over t. In this way, we obtain a distance matrix of
evolutionary times between each pair of (super-)taxa, to which the neighbor-joining
method can then be applied.
As is evident from the description above, all of the sequence profiles do need to be of
the same length. This is a general requirement for molecular sequence data in
phylogenetics; usually one computes a multiple sequence alignment of all sequences
and for distance evaluation considers only columns without gap characters (some
methods treat a gap like an ordinary fifth character besides the nucleotides). Of course,
if multiple sequence alignments are computed separately for each subclade and then
converted into profiles, the resulting profiles can be of different length. In such a case,
we may apply evolutionary models that do take gaps into account to derive a maximum
likelihood distance estimator, such as the Thorne-Kishino-Felsenstein model (see Steel
& Hein, 2000). In our experiments, we have so far not used this option, but instead worked
with a single multiple alignment of all taxa sequences, which automatically results in all
profiles being of the same length.
Bootstrapping
In phylogenetics, bootstrapping is a widely accepted method to assess the robustness
of an inferred tree topology. The topology is robust if it remains invariant after slight
perturbations of the input data (i.e., the alignments and distances). Note that robustness
is in no obvious way related to accuracy (see Chapter 14.9 in Ewens & Grant, 2001, for
more careful distinctions). The following procedure has become the standard one
(Felsenstein, 1985; Hillis & Bull, 1993; Efron, Halloran, & Holmes, 1996): From the initial
sequence alignment A, several so-called bootstrap replicates A*
1
,..., A*
B
, with B 1000
to 10000, are created by randomly resampling entire columns of A. That is, each A* has
the same length as A, but contains each column of A zero times, once, or several times.
Each A* is now considered a new input alignment to the tree reconstruction algorithm.
From the B estimated bootstrap trees a so-called consensus tree T is derived (e.g., a
majority rule consensus tree; see, e.g., Felsenstein, 2003).
Efficient and Robust Analysis of Large Phylogenetic Datasets 111
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Each edge of T is then annotated with its bootstrap value, i.e., the percentage of bootstrap
trees that share the same edge. Edges with bootstrap values above 75% are generally
considered as reliable. If low bootstrap values are encountered, it probably makes sense
to aim for a less resolved tree, i.e., to allow multifurcating inner nodes.
We observe that when we estimate a tree directly from all taxa, some bootstrap values
in the basal branching pattern tend to drop, possibly because of a single taxon jumping
from one family into a different one in several bootstrap replicates, even if it is initially
obvious (e.g., from additional data) where the taxon belongs to within the phylogeny. For
the same reason, we may expect that using family profiles would make the basal branching
pattern more robust; and this expectation is confirmed by our experiments.
Iterative Profile Neighbor-Joining
The basic PNJ method requires defining monophyletic groups in the initial set of taxa
using any kind of additional knowledge; the sequences within a group are then converted
into profiles, and finally distances between pairs of profiles can be computed. It would
be even more advantageous, however, to have a fully automatic approach. We therefore
apply the PNJ method iteratively. In the first step, all taxas sequences are used to
produce a tree. Then the algorithm finds all maximal subtrees, in which all bootstrap
values exceed a pre-defined threshold (as mentioned above, a good choice is 75%). The
sequences of each so defined subtree are then converted into a profile, and the next
iteration starts. This procedure continues until no further profiles are created. To save
time in the first iteration, almost identical sequences are initially grouped together and
immediately converted into a profile; this behavior can be tuned by specifying a so-called
identity threshold.
Software
A user-friendly software package called ProfDist (Friedrich et al., 2005) implements the
PNJ and iterative PNJ methods on nucleotide sequence data. Windows, Linux, and
Macintosh versions of ProfDist are available at http://profdist.bioapps.biozentrum.uni-
wuerzburg.de and come with a graphical user interface.
Subclades for profile formation can be chosen manually or automatically based on
bootstrap confidence values. ProfDist implements the BIONJ variant (Gascuel, 1997) of
the neighbor-joining algorithm and uses the iterative version of the PNJ method to
automatically build a tree if no subclades are chosen manually. Standard tree viewers,
such as TREEVIEW (Page, 1996) or ATV (Zmasek & Eddy, 2001), can be used to visualize
the output.
The main feature of ProfDist is its efficient and one-stop-shop implementation of the PNJ
method. The different steps (drawing of bootstrap samples, distance computation, tree
estimation via BIONJ, consensus tree formation) can be started separately using the
buttons in the ProfDist main window (see Figure 1).
112 Rahmann, Mller, Dandekar, & Wolf
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
ProfDist supports different file formats for both trees and sequences (e.g., NEWICK,
FASTA, EMBL and PHYLIP). The essential parameters of ProfDist are (1) the number of
bootstrap replicates, (2) the distance estimation method, possibly supplemented with a
user-defined substitution model, and (4) the PNJ agglomeration procedure. ProfDist
implements the JC (Jukes & Cantor, 1969), K2P (Kimura, 1980), GTR (Lanave et al., 1984)
distance estimation methods, and the Log-Det transformation (Barry & Hartigan, 1987).
For the iterative PNJ variant, there are two further parameters: (1) the minimal bootstrap
value that an estimated monophyletic group has to achieve before it is considered
trustworthy and represented as a profile in the next iteration, and (2) an identity threshold
for immediate profile formation in the first iteration.
Nucleotide ProfDist is very time and space efficient: Nucleotide alignments are handled
in an encoded diff-like format that makes it easy to count the number of nucleotide
substitutions between any pair of sequences in a multiple alignment.
A Tree of the Eukaryotes
Based on 18S rRNA
From the European ribosomal rRNA database (Wuyts, Perriere, & Van de Peer, 2004), we
obtained all available eukaryotic 18S SSU (small subunit) rRNA sequences in aligned
Figure 1. ProfDist user interface: Sequences can be loaded in FASTA or EMBL format;
trees are output in NEWICK format

Efficient and Robust Analysis of Large Phylogenetic Datasets 113
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
format. Using the Jukes-Cantor substitution model, an initial identity threshold of 90%,
and a bootstrap confidence threshold of 50%, we ran ProfDist with the iterative version
of the PNJ algorithm, after we had generated profiles on predefined monophyletic groups
(Metazoa, Fungi, Viridiplantae, Alveolates, Stramenopilates, Red algae, Slim molds,
Euglenids and Diplomonads). The resulting tree is shown in Figure 2. It corresponds to
trees computed from SSU rRNA with other methods (van de Peer, Baldauf, Doolittle, &
Meyer, 2000), but has been computed within only a few minutes and has high bootstrap
support values.
Figure 2. A tree of the eukaryotes computed with ProfDist, based on SSU rRNA data:
The upper part shows the unrooted tree; the lower part shows a rooted tree with
bootstrap support values (numbers in square brackets denote the number of taxa in the
respective subtree)
114 Rahmann, Mller, Dandekar, & Wolf
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Summary
We have presented the profile neighbor-joining (PNJ) approach, in which known or
previously inferred subclades are converted into sequence profiles. As discussed in the
introduction, other approaches are possible: For example, we might use one representa-
tive sequence from each subclade, or estimate the most likely sequence in the subclade
root. The profile-based approach appears preferable because it integrates information
from all family members; each member sequence is weighted equally, i.e., the subtree
topology is ignored in this approach. This could be changed through sequence weight-
ing schemes, but we may question whether this is desirable: We have to take into account
the possibility that our sequence alignment contains fast evolving species or strong
variation over sites. Then it could happen that one or more sequences in the subtree are
connected accidentally to the rest by a Long Branch Attraction (LBA) phenomenon. This
would result in a very strong signal of this sequence in the weighted profile generated
from these sequences. In this sense, our approach is more robust. Also, knowledge of
the intra-subclade topologies is not necessary to construct the profile. However, this
robustness and simplicity come at a price: We must expect a drop in accuracy, i.e., in
simulated experiments, the true basal branching pattern may be reconstructed less often
if subclade topologies differ extremely from star trees, because of the resulting bias in
the profiles. Interestingly, though, we could not observe such an accuracy drop in our
simulations (Mller et al., 2004), indicating that even this simple profile construction
method works well in practice. Further studies, especially comparing the relative
efficiency of different tree estimators, would be an interesting research topic for the
future.
Profile distances could be viewed more or less as averaged sequence distances, but there
is one difference between profile distances and average distances: Average distances
are a mean of maximum-likelihood distances of sequences between sequences of both
groups, whereas the profile distance is a maximum-likelihood estimation between the
mean of the sequences in both groups. In this sense, using profile distances results
in a more computationally efficient estimator and leads to a more robust formulation,
compared to consensus methods.
As is well known, if the distances between a set of objects (taxa or subclades) are additive
(i.e., if they fulfill the four-point condition, see Felsenstein (2003) for an in-depth
discussion), there exists a unique unrooted tree representing these distance, and the NJ
method will recover it exactly. Even if the distance data is not tree-like (i.e., far from being
additive), the NJ method always reconstructs a tree, but the meaning of this tree is not
obvious. Such trees will tend to have relatively low bootstrap support since their
topology is not well supported by the data.
We observe that trees constructed with PNJ tend to have much better bootstrap support
than trees constructed by other methods. However, this by itself is not necessarily of
biological significance. The central question is whether the PNJ method is more accurate
than other methods, i.e., whether it reconstructs the true tree topology more often. While
this question is difficult to answer for real large datasets, where the true topology is
Efficient and Robust Analysis of Large Phylogenetic Datasets 115
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
usually unknown, it can be answered for simulated datasets: Based on the Robinson-
Foulds distance measure (Robinson & Foulds, 1981) between true and estimated
topology, Mller et al. (2004) have shown that PNJ indeed outperforms NJ in terms of
accuracy.
Many questions remain unanswered about PNJ and related methods at this point. Is it
statistically consistent like maximum likelihood methods in the case of tree-like data, or
how do we need to modify the profile computation procedure to make it consistent? How
long do the sequences need to be to recover the true tree with high probability, in other
words, what is the convergence rate? Can it be boosted by general principles like the disc-
covering method (DCM; Huson, Nettles, & Warnow, 1999)? How will improving the
convergence rate affect the robustness against single outlier (e.g., fast-evolving)
sequences or errors in the sequence alignments? These are questions of general interest
that are increasingly investigated by the scientific community.
In the future, we plan to extend the PNJ software to handle protein alignment data, to
automatically estimate the substitution model from the data (up to now it has to be
specified by the user), and to include the possibility of rate variation over sites into the
distance estimators. We cordially invite the reader to try out the latest version at http:/
/profdist.bioapps.biozentrum.uni-wuerzburg.de.
References
Barry, D., & Hartigan, J. (1987). Asynchronous distance between homologous DNA
sequences. Biometrics, 43, 261-276.
Bininda-Emonds, O. R. P. (2004). The evolution of supertrees. Trends in Ecology and
Evolution, 19(6), 315-322.
Bruno, W. J., Socci, N. D., & Halpern, A. L. (2000). Weighted neighbor joining: A
likelihood-based approach to distance-based phylogeny reconstruction. Molecu-
lar Biology and Evolution, 17, 189-197.
Camin, J., & Sokal, R. (1965). A method for deducing branching sequences in phylogeny.
Evolution, 19, 311-326.
Desper, R., & Gascuel, O. (2002). Fast and accurate phylogeny reconstruction algorithms
based on the minimum-evolution principle. Journal of Computational Biology, 19,
687-705.
Efron, B., Halloran, E., & Holmes, S. (1996). Bootstrap confidence levels for phylogenetic
trees. PNAS, 93, 7085-7090.
Ewens, W. J., & Grant, G. R. (2001). Statistical Methods in Bioinformatics. New York:
Springer-Verlag
Felsenstein, J. (1981). Evolutionary trees from DNA sequences: A maximum likelihood
approach. Journal of Molecular Evolution, 17, 368-376.
116 Rahmann, Mller, Dandekar, & Wolf
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Felsenstein, J. (1985). Confidence limits on phylogenies: An approach using the boot-
strap. Evolution, 39, 783-791.
Felsenstein, J. (2003). Inferring Phylogenies. Sunderland, MA: Sinauer Associates.
Fitch, W.M. (1971). Toward defining the course of evolution: Minimum change for a
specified tree topology. Systematic Zoology, 20, 406-416
Friedrich, J., Wolf, M., Dandekar, T., & Mller, T. (2005): ProfDist: A tool for the
construction of large phylogenetic trees based on profile distances. Bioinformatics,
21, 2108-2109.
Gascuel, O. (1997). BIONJ: An improved version of the NJ algorithm based on a simple
model of sequence data. Molecular Biology and Evolution, 14, 685-695.
Gribskov, M., McLachlan, A., & Eisenberg, D. (1987). Profile analysis: Detection of
distantly related proteins. PNAS, 84, 4355-4358.
Hillis, D., & Bull, J. (1993). An empirical test of bootstrapping as a method for assessing
confidence in phylogenetic analysis. Systematic Biology, 42, 182-192.
Holder, M., & Lewis, P. (2003). Phylogeny estimation: Traditional and Bayesian ap-
proaches. Nature Reviews, 4, 275-284.
Huelsenbeck, J., & Ronquist, F. (2001). MRBAYES: Bayesian inference of phylogeny.
Bioinformatics, 17, 754-755.
Huson, D. H., Nettles, S. M., & Warnow, T. J. (1999). Disk-covering, a fast-converging
method for phylogenetic tree reconstruction. Journal of Computational Biology,
6, 369-386.
Jukes, T., & Cantor, C. R. (1969). Evolution of protein molecules. In H. Munro (Ed.),
Mammalian protein metabolism (pp. 21-132). New York: Academic Press.
Kimura, M. (1980). A simple method for estimating evolutionary rate of base substitu-
tions through comparative studies of nucleotide sequences. Journal of Molecular
Evolution, 16, 111-120.
Krogh, A., & Mitchison, G. (1995). Maximum entropy weighting of aligned sequences of
proteins or DNA. In C. Rawlings, D. Clark, R. Altman, L. Hunter, T. Lengauer, & S.
Wodak (Eds.), Proceedings of the Third International Conference on Intelligent
Systems for Molecular Biology (pp. 215-221). Menlo Park, CA: AAAI Press.
Lanave, C., Preparata, G., Saccone, C., & Serio, G. (1984). A new method for calculating
evolutionary substitution rates. Journal of Molecular Evolution, 20, 86-93.
Mller, T., Rahmann, S., Dandekar, T., & Wolf, M. (2004). Accurate and robust phylogeny
estimation based on profile distances: a study of the Chlorophyceae (Chlorophyta).
BMC Evolutionary Biology, 4(20).
Mller, T., & Vingron, M. (2000). Modeling amino acid replacement. Journal of Compu-
tational Biology, 7, 761-776.
Page, R. D. M. (1996). TREEVIEW: An application to display phylogenetic trees on
personal computers. Computer Applications in the Biosciences, 12, 357-358.
Robinson, D., & Foulds, L. (1981). Comparison of phylogenetic trees. Mathematical
Biosciences, 53, 131-147.
Efficient and Robust Analysis of Large Phylogenetic Datasets 117
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Saitou, N., & Nei, M. (1987). The neighbor-joining method: A new method for reconstruct-
ing phylogenetic trees. Journal of Molecular Evolution, 4, 406-425.
Schmidt, H. A., & von Haeseler, A. (2003). Maximum-likelihood analysis using TREE-
PUZZLE. In A. D. Baxevanis, D. B. Davison, R. D. M. Page, G. Stormo, & L. Stein
(Eds.), Current protocols in bioinformatics (Unit 6.6). New York: Wiley and Sons.
Steel, M., & Hein, J. (2001). Applying the Thorne-Kishino-Felsenstein model to sequence
evolution on a star-shaped tree. Applied Mathematics Letters, 14, 679-684.
Swofford, D. L. (2003). PAUP*: Phylogenetic Analysis Using Parsimony (*and other
methods), Version 4. Sunderland, MA: Sinauer Associates.
van de Peer, Y., Baldauf, S. L., Doolittle, W. F., & Meyer, A. (2000). An updated and
comprehensive rRNA phylogeny of (crown) eukaryotes based on rate-calibrated
evolutionary distances. Journal of Molecular Evolution, 51, 565-76.
Wuyts, J., Perriere, G., & Van de Peer, Y. (2004). The European ribosomal RNA database.
Nucleic Acids Research, 32, 101-103. Retrieved February, 2006, from http://
www.psb.ugent.be/rRNA
Zmasek, C., & Eddy, S. (2001). ATV: display and manipulation of annotated phylogenetic
trees. Bioinformatics, 17, 383-384.
118 Akutsu
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Chapter VII
Algorithmic Aspects of
Protein Threading
Tatsuya Akutsu, Kyoto University, Japan
Abstract
This chapter provides an overview of computational problems and techniques for
protein threading. Protein threading is one of the most powerful approaches to protein
structure prediction, where protein structure prediction is to infer three-dimensional
(3-D) protein structure for a given protein sequence. Protein threading can be modeled
as an optimization problem. Optimal solutions can be obtained in polynomial time
using simple dynamic programming algorithms if profile type score functions are
employed. However, this problem is computationally hard (NP-hard) if score functions
include pairwise interaction preferences between amino acid residues. Therefore,
various algorithms have been developed for finding optimal or near-optimal solutions.
This chapter explains the ideas employed in these algorithms. This chapter also gives
brief explanations of related problems: protein threading with constraints, comparison
of RNA secondary structures and protein structure alignment.
Algorithmic Aspects of Protein Threading 119
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Introduction
Inference and mining of functions of genes is one of the main topics in bioinformatics.
Protein structure prediction provides useful information for that purpose because it is
known that there exists close relationship between structure and function of a protein,
where protein structure prediction is a problem of inferring three-dimensional structure
of a given protein sequence. Computational inference of protein structure is important
since determination of three-dimensional structure of a protein is much harder than
determination of its sequence.
There exist various kinds of approaches for protein structure prediction (Clote &
Backofen, 2000; Lattman, 2001; Lattman, 2003). Ab initio approach tries to infer structure
of a protein based on the basic principles (e.g., energy minimization) in physics. In this
approach, such techniques as molecular dynamics and Monte Carlo simulations have
been employed. Homology modeling approach tries to infer structure of a protein using
structure of a homologous protein (i.e., a protein whose structure is already known and
whose sequence is very similar to the target protein sequence). In this approach,
backbone structure of a protein is first computed from structure of a homologous protein
and then the whole structure is computed by using molecular dynamics and/or some
optimization methods. Secondary structure prediction approach does not aim to infer
three-dimensional structure. Instead, it tries to infer which structural class (, , others)
each residue belongs to. Such information is believed to be useful for inference of three-
dimensional structure and/or function of a protein. In secondary structure prediction
approach, various machine learning methods have been employed, which include neural
networks and support vector machines.
Protein threading is another major approach for protein structure prediction. In this
approach, given an amino acid sequence and a set of protein structures (structural
templates), a structure into which the given sequence is most likely to fold is computed.
In order to test whether or not a sequence is likely to fold into a structure, an alignment
(i.e., correspondence) between spatial positions of a 3-D structure and amino acids of
a sequence is computed using a suitable score function. That is, an alignment which
minimizes the total score (corresponding to the potential energy) is computed. This
minimization problem is called the protein threading problem. Though there exists some
similarity between protein threading and homology modeling, these are usually consid-
ered to be different: alignment between a target sequence and a template structure is
computed in protein threading whereas alignment between two sequences is computed
in homology modeling.
Many studies have been done on protein threading. Most of them focus on improvement
of practical performances using heuristic and/or statistical techniques, and few atten-
tions had been paid to algorithmic aspects of protein threading. Since protein threading
is NP-hard in general, heuristic algorithms have been widely employed, which do not
necessarily guarantee optimal solutions. However, recent studies (Xu, Xu, Crawford, &
Einstein, 2000; Xu, Li, Kim, & Xu, 2003) suggested that it is possible to compute optimal
solutions in reasonable CPU time for most proteins and computation of optimal threadings
is useful for improving practical performances of protein threading. Furthermore, there
exist several important problems in bioinformatics, which are closely related to protein
120 Akutsu
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
threading. Therefore, in this chapter, we overview algorithmic aspects of protein
threading and related problems.
This chapter is organized as follows. First, we formally define the protein threading
problem and show NP-hardness of protein threading. Then, after briefly reviewing
heuristic methods, we describe three exact approaches for computing optimal threading:
branch-and-bound approach (Lathrop & Smith, 1996), divide-and-conquer approach
(Xu, Xu, & Uberbacher, 1998) and linear programming approach (Xu et al., 2003). Next,
we briefly explain a variant of protein threading (protein threading with constraints) and
related problems (comparison of RNA secondary structures and protein structure
alignment). Finally, we conclude with future directions.
Protein Threading Alignment Between
Sequence and Structure
As mentioned in the introduction, protein threading is a problem of computing an
alignment between a target sequence and a template structure. First we define a usual
alignment for two sequences. Let be the set of amino acids (i.e., ||=20). Let s =s
1
s
m
and t=t
1
t
n
be strings over . An alignment between s and t is obtained by inserting gap
symbols (-) into or at either end of s and t such that the resulting sequences s and t
are of the same length l, where it is not allowed for each i that both s
i
and t
i
are gap
symbols.
For exampl e, consi der t wo sequences s= LDVQWAVDEGDKVV, t =
DVQWSVEKRHGDKLVLT. Then, the followings are examples of alignments:
AL
1
AL
2
AL
3

LDVQWAVDE---GDKVV-- LDVQWAVDE---GDK-VV- LDVQWAVDE---GDKVV
-DVQWSV-EKRHGDKLVLT -DVQWSV-EKRHGDKLVLT VDQWSVEKRHGDKLVLT
Let f(x,y) be a score function from

to the set of real numbers that satisfies f(x,y)=f(y,x).
We extend f(x,y) to include gap symbols by defining f(x,-) = f(-,y)=-d for all x, y in , where
-d (d>0) is a penalty per gap. The score of alignment (s,t) is defined by:
. ) ' , ' ( ) ' , ' (
1

=
=
l
i
i i
t s f t s score
An optimal alignment is an alignment having the maximum score. Suppose that f(x,x)=10
and f(x,y)=-10 for all xy, and d=10. Then, the scores of AL
1
, AL
2
and AL
3
are 10, 10 and
Algorithmic Aspects of Protein Threading 121
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
-150, respectively. In this case, both AL
1
and AL
2
are optimal alignments whereas AL
3
is
a non-optimal alignment.
In this example, we assumed linear gap cost. Though affine gap cost should be used in
practice, we use linear gap cost in this chapter for simplicity. However, all the results in
this chapter can be extended for affine gap cost.
It is well-known that optimal alignments can be computed in O(mn) time using a simple
dynamic programming algorithm. The following procedure computes the score of an
optimal alignment:
), ) , ( ] 1 , 1 [ , ] 1 , [ , ] , 1 [ max( ] , [
, ] , 0 [
, ] 0 , [
j i
t s f j i D d j i D d j i D j i D
j j D
i i D
+

where D[i,j] corresponds to the optimal score between s

1
s
2
s
i
and t
1
t
2
t
j
. An optimal
alignment can also be obtained from this matrix by using the traceback technique (Clote
& Backofen, 2000) .
Next, we define the protein threading problem (see Figure.1). Recall that in protein
threading, we compute an alignment between a target sequence and a template structure.
Let s =s
1
s
m
be a target sequence over . Let t = t
1
t
n
be a template protein structure,
where t
i
is the i-th residue in t. We can regard t as as a sequence of C(or C) atoms of
the protein structure. Then, a threading between s and t is an alignment between s and t.
Various score functions have been used in protein threading. One of the simplest score
functions is a profile. A profile f
t
is defined for a template structure t, and is a function
Figure 1. Threading between a target sequence s and a template structure t. Shaded
residues are aligned with gap symbols. In this case, s=MTLC-VERLD-E and t= t
1
t
2
-
t
3
t
4
t
5
t
6
- - t
7
t
8
t
9
.

t
2
t
9
t
3
t
4 t
8
t
7
t
5
t
6
t
1
122 Akutsu
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
from } , ,..., { }) { (
1

n
t t to the set of reals. Then, the score of a threading (s,t) is
defined by:
. ) ' , ' ( ) ' , ' (
1

=
=
l
i
i i t
t s f t s score
It should be noted that
i
s' is aligned to
i
t' where
i
t' denotes some residue in t or a gap.
For a profile type score function, an optimal threading can be computed in O(mn) time
as in the case of sequence alignment. Various types of profile-type score functions have
been proposed. The 3-D profile proposed by Bowie, Luthy, & Eisenberg (1991) is one of
the earliest profiles. A well-known sequence search software PSI-BLAST (Altschul et al.,
1997) employs a kind of profiles. Sequence alignment using an HMM (Hidden Markov
Model) can also be considered as alignment with profiles (Durbin,Eddy, Krogh, &
Mitchison , 1998). Recently, profile-profile alignment is regarded as a powerful method
for protein threading, in which a kind of profile is used also for s instead of a sequence.
An optimal alignment can still be computed in O(mn) time in this case. Edgar and Sjolander
(2004) compared 23 types of score functions for profile-profile alignment.
In the above, score functions do not include pairwise intersection preferences. However,
it is well-known that pairwise interactions between amino acid residues play a very
important role in the process of protein folding (i.e., folding into a three-dimensional
protein structure). Therefore, various types of score functions have been proposed in
which pairwise interaction preferences are taken into account. In this case, the score
function basically consists of two terms: f
t
(x,y) and g
t
(x,y,t
i
,t
j
), where g
t
(x,y,t
i
,t
j
) is a
function from } , , { } , , {
1 1 n n
t t t t K K to the set of reals. In a simplest case, g
t
can be
a contact potential defined by

>
=
, otherwise , ) , (
, ) , ( if , 0
) , , , (
0
y x g
t t dist
t t y x g
j i
j i t
where dist(t
i
,t
j
) is the Euclidean distance between C atoms of t
i
and t
j
, is a threshold,
and g
0
(x,y) is a function from

to the set of reals. Then, the score of a threading (s,t)
is defined by
, ) ' , ' , ' , ' ( ) ' , ' ( ) ' , ' (
1

< =
+ =
j i
j i j i t
l
i
i i t
t t s s g t s f t s score
where the sum in the second term is taken over all pairs of i,j such that none of s'
i
, s'
j
, t'
i
,
t'
j
is a gap symbol. It is worthy to note that alignment problems (including threading with
profiles) are usually defined as maximization problems whereas threading problems with
pair score functions are usually defined as minimization problems. In this chapter, we
basically follow this tradition.
Algorithmic Aspects of Protein Threading 123
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Hardness of Threading with
Pair Score Functions
As previously shown, protein threading with profiles can be solved in O(mn) time. On
the other hand, it is known that protein threading with pair score functions is NP-hard
(Lathrop, 1994). In the following, we present a simple proof based on Akutsu and Miyano
(1999).
Theorem 1.
Protein threading with pair score functions is NP-hard.
Proof. We use a polynomial time reduction from MAX CUT, where MAX CUT is
a well-known NP-hard problem (Garey & Johnson, 1979). Recall that MAX CUT is,
given an undirected graph G(V,E), to find a subset V of V maximizing the cardinality
of the cut (i.e., the number of edges between V and V-V). From G(V,E), we
construct an instance of protein threading in the following way (see also Fig. 2)
,where we only consider two types of amino acids (Alanine (A) and Leucine (L)).
Let V={v
1
,v
2
,,v
n
}. Then, we let t=t
1
t
n
(we identify v
i
with t
i
) and let s be n
repetitions of substring AL. Next, we define f
t
(x,y)=0 for all x, y, and we define
g
t
(x,y,t
i
,t
j
) by

<
=
otherwise. , 0
, and , } , { , if , 1
) , , , (
y x E v v j i
t t y x g
j i
j i t
Then, each threading corresponds to a cut by letting V={v
i
| A is assigned to t
i
}.
Moreover, the score of the threading corresponds to the cardinality of the
Figure 2. Reduction from MAXCUT to protein threading. A set of vertices aligned to
A (Alanine) corresponds to V and a set of vertices aligned to L (Leucine) corresponds
to V-V.

v
2
v
3
v
4
v
1
v
5
v
6
124 Akutsu
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
cut where the sign of the score should be inverted. Since the reduction can be
clearly done in polynomial time, protein threading with pair score functions is
NP-hard. Q.E.D.
Akutsu and Miyano (1999) extended this proof for showing strong hardness results on
finding approximate solutions. They also developed approximation algorithms for
special cases.
Algorithms for Threading
with Pair Score Functions
Although NP-hardness of protein threading was proven, several approaches can be
considered. One approach is to develop exact algorithms that guarantee optimal solu-
tions using such techniques as branch-and-bound. Though such an approach is
desirable, all of existing algorithms are somewhat complicated, to be explained later.
Another approach is to develop heuristic algorithms. Though optimal solutions are not
necessarily guaranteed in this approach, some of existing algorithms are very simple and
have reasonable performances. Since the second approach is much simpler, we begin
with the second one.
Heuristic Algorithms
Here, we briefly overview two well-known heuristic algorithms for protein threading with
pair score functions: frozen approximation (Godzik & Skolnick, 1992) and double dynamic
programming (Jones, Taylor, & Thornton, 1992).
The idea of frozen approximation is very simple. It replaces g
t
(x,y,t
i
,t
j
) by g
t
(x,l(t
j
),t
i
,t
j
)
where l(t
j
) is the amino acid type of the j-th residue in the template structure. That is, we
use amino acid type of the template structure for y instead of using amino acid type of
the target sequence. If we use g
t
(x,l(t
j
),t
i
,t
j
) in place of g
t
(x,y,t
i
,t
j
), the optimal solution can
be computed by using a dynamic programming algorithm similar to that for protein
threading with profiles. It should be noted that this optimal solution is different from that
of the original threading problem, and there is no theoretical guarantee on the accuracy
of the solution obtained by using frozen approximation. However, frozen approximation
has been adopted by many protein threading programs because it is very easy to
implement.
The double dynamic programming technique was originally developed for computing
structure alignment between two protein structures. It uses two levels of dynamic
programming: upper level dynamic programming and lower level dynamic programming.
Lower level dynamic programming is performed to compute the score F[i,j] which
indicates how likely it is that the pair (s
i
,t
j
) is on an optimal alignment. Then, upper level
dynamic programming is performed as in sequence alignment using F[i,j] as a score
function.
Algorithmic Aspects of Protein Threading 125
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Exact Algorithms
Though protein threading is NP-hard in general, the sizes of real proteins is limited
(usually each protein consists of less than several hundred residues) and thus there exist
chances that practically efficient exact algorithms can be developed. Indeed, several
algorithms have been developed which can compute optimal threadings for proteins
consisting of up to several hundred residues. Here, we briefly review important three of
such algorithms: a branch-and-bound algorithm (Lathrop & Smith, 1996), PROSPECT (Xu
et al., 2000), and RAPTOR (Xu et al., 2003).
Before presenting algorithms, we simplify the threading problem (see also Figure. 3). It
is known that gaps are seldom inserted in the core regions of proteins, where a core region
is either an helix or a -strand (Lathrop & Smith, 1996). Therefore, we assume that gaps
are not inserted into core regions (either in the target sequence or in the template
structure). Let C
1
,,C
M
be the core regions of the template structure. Since it is not
difficult to modify the algorithms so that terms of f
t
(x,y) can be taken into account, we
assume that the score consists of terms given by g
t
(x,y,t
i
,t
j
). Moreover, we consider
pairwise interaction terms among residues only in core regions because it is considered
that pairwise interactions among core regions are much more important than the other
pairwise interactions. Let p
i
denote the position of s to which the first position of C
i
is
aligned. Let c
i
denote the length of C
i
. It is easy to pre-compute the score between any
pair of core regions because no gaps are inserted into any core region. Thus, we denote
by g(i,j,p
i
,p
j
) the score between C
i
and C
j
when the first positions of C
i
and C
j
are aligned
to the p
i
-th and p
j
-th positions of s respectively. That is:
, ) , , , ( ) , , , (
' 0 ' 0
' ' ' '
< <
+ + + +
=
i j
j i j i
c i c j
j q i q j p i p t j i
t t s s g p p j i g
where q
i
and q
j
are the beginning positions of cores C
i
and C
j
, respectively.
Figure 3. In exact algorithms, we assume that gaps are not inserted into core regions.
The first residue position of core C
i
is aligned to the p
i
-th residue in a target sequence.
C
1
C
3
C
2
p
1
p
2
p
3
126 Akutsu
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Since we do not consider scores for non-core regions and we assume that gaps are not
inserted into core regions, it is enough to represent a threading by means of the first
positions of cores. Therefore, each threading is represented by M-tuple (p
1
,,p
M
). It
should be noted that core regions must not overlap in a feasible threading. That is,
1 +
+
i i i
p c p must hold for all i<M. Then, we re-formulate the threading as: given a target
sequence s=s
1
s
m
, core lengths c
1
, , c
M
, and a score function g(i,j,p
i
,p
j
), to find a feasible
threading (p
1
,,p
M
) minimizing the score defined by:
1
( , , ) ( , , , ).
M i j
i j
score p p g i j p p
<
=

K
Lathrop and Smith (1996) first proposed an efficient and exact algorithm based on a
branch-and-bound t echni que. They defi ned t he set of al l owed ranges
([b
1
,d
1
],[b
2
,d
2
],...,[b
M
,d
M
]), which means that b
i
p
i
d
i
must hold for each i. Their algorithm
begins with the set of ranges corresponding to all possible threadings and iteratively
subdivides each set of ranges into smaller sets of ranges. For a simplest example,
P=([b
1
,d
1
],[b
2
,d
2
],...,[b
M
,d
M
]) can be subdivided into three sets of ranges:
]), , [ , ], , [ ], , 1 [ ], , [ , ], , ([
]), , [ , ], , [ ], , [ ], , [ , ], , ([
]), , [ , ], , [ ], 1 , [ ], , [ , ], , ([
1 1 1 1 1 1 3
1 1 1 1 1 1 2
1 1 1 1 1 1 1
M M i i i i i i
M M i i i i i i
M M i i i i i i
d b d b d e d b d b P
d b d b e e d b d b P
d b d b e b d b d b P
K K
K K
K K
+ +
+ +
+ +
+ =
=
=
where choice of i and e
i
is done based on some heuristic criteria. It should be noted that
union of three sets of ranges is equal to the original set of ranges. If we repeat this
subdivision process, we finally obtain a set of all individual threadings. However, an
exponential number of threadings would be generated. Therefore, we should adopt some
technique to limit the subdivision process. For that purpose, Lathrop and Smith proposed
a few kinds of lower bounds of the scores attained by any threading from P. For example,
the following is the simplest one:
1 1
1
( , , ) ( , , ) ,
min ( , , ) min ( , , , ) min ( , , , )
M M i i j j
M i j
p p P p p P b x d b y d
i j i j
score p p g i j p p g i j x y

< <
=

K K
K
Let LB(P) denote the rightmost term in the above inequality. Let S
opt
denote the minimum
score among threadings found so far. Then, we need not search for subdivisions of P
if LB(P) is greater than S
opt
because any threading from P can not attain the score lower
than S
opt
. Therefore, we can reduce the search space. In order to improve the efficiency,
Lathrop and Smith proposed stronger bounds and succeeded to compute optimal
threadings for many medium-size proteins in reasonable CPU time.
Algorithmic Aspects of Protein Threading 127
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Xu et al. (1998) developed the PROSPECT algorithm/system based on a divide-and-
conquer strategy. PROSPECT repeatedly subdivides the template structure into sub-
structures until each substructure contains one core region, where each (sub)structure
t is divided into two smaller substructures t and t and one non-core region t. Then,
PROPECT finds optimal threadings for each substructure from smaller ones to larger
ones. In order to compute optimal threadings for t, PROSPECT combines optimal
threadings for t and t. PROSPECT also utilizes a property that only a few cores strongly
interact with each core. Thus, PROSPECT considers interactions for pairs of strongly
interacting cores, where this simplification is also employed by RAPTOR. We denote
by E the set of such pairs. For example:
)} , ( ), , ( ), , ( ), , ( ), , {(
5 4 5 3 4 2 3 2 2 1
C C C C C C C C C C E =
in Figure 4. We explain the key idea of PROSPECT using this figure. Suppose that we
need to compute optimal threadings between template structure t and all possible
subsequences . We call a pair of aligned positions a link. For example, link o
1
is (p
2
,p
4
)
and o
2
is (p
3
,p
5
). Suppose that optimal threadings for t are already computed for all
combinations of k
1
, k
2
, o
1
, o
2
. Suppose also that optimal threadings for t are already
computed for all combinations of k
3
, k
4
, o
1
, o
2
. Then, an optimal threading between t and
1 4
k k
s s K can be obtained by taking the minimum of the sum of threadings for t and t where
the minimum is taken over all combinations of k
2
, k
3
, o
1
, o
2
. Though PROSPECT takes
exponential time in general, it works efficiently if the number of links between t and t
is small for each substructure t.
Xu et al. (2003) developed a novel algorithm/system named RAPTOR by formulating
protein threading as an integer program (IP). Let a variable x
i,j
mean that core C
i
is aligned
to the j-th position of the target sequence s (if x
i,j
=1). Let a variable y
(i,l),(j,k)
mean that
cores C
i
and C
j
are aligned to the l-th and k-th positions of the target sequence s,
respectively. Then, the objective function of the threading problem can be represented
as a linear combination of y
(i,l),(j,k)
s:
Figure 4. A schematic of the divide-and-conquer procedure of PROSPECT

template structure t
t
O
1
C
1
C
2
C
3
C
4
C
5
t
t
k
1
(=P
1
) P
2
P
3
k
2
P
5 k
4
k
3
(=P
4
)
sequence S
128 Akutsu
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
( , )( , )
( , ) [ ] [ , , ]
min ( , , , )
i l j k
Ci Cj E l D i k R i j l
y g i j l k


where D[i] denotes all valid target sequence positions that C
i
could be aligned to, and
R[i,j,l] denotes all valid alignment positions of C
j
given C
i
is aligned to the l-th position
of s. In order to guarantee that a set of variables represents to a valid (feasible) threading,
we should add constraints. For that purpose, it is shown in Xu et al. (2003) that the
following linear constraints are enough:
,
[ ]
,
( , )( , )
( , )( , ) ,
[ , , ]
( , )( , ) ,
[ , , ]
1, 1, 2, , ,
0, [ ], 1, 2, , ,
0, [ ], [ ], , 1, 2, , ,
, ( , ) ,
, ( , ) .
i j
j D i
i j
i l j k
i l j k i l i j
k R i j l
i l j k j k i j
l R j i k
x i M
x j D i i M
y l D i k D j i j M
y x C C E
y x C C E

= =
=
=
=
=

K
K
K
For example, the first constraint means that each core can be aligned to a unique sequence
position. The last two constraints mean that if two interacted cores C
i
and C
j
are aligned
to two sequence positions l and k respectively, then the interaction score between these
two sequence segments should be counted in the threading score. Since integer
programming is known to be NP-hard, Xu et al. (2003) relaxed this integer program to a
linear program (LP) by removing integral constraints on all variables. In that case, each
variable may take non-integer value. Therefore, they employed some branch-and-bound
procedure for cases where non-integral solutions are obtained by an LP-solver. However,
they obtained a surprising result that the relaxed linear programs generated integral
solutions directly in most cases.
Threading with Constraints
As seen above, extensive studies have been done for protein threading. However,
predictive accuracy is not yet satisfactory for practical use. Recently, some advances
were made by the significant utilization of distance restraints, which can be obtained from
some biological experiments. Xu et al. (2000) showed that (partial) information obtained
from NMR experiments is useful to improve the accuracy of the protein threading method.
For that purpose, they modified PROSPECT so that constraints given from NMR
experiments were taken into account. Young et al. (2000) developed a novel experimental
method to aid in construction of a homology model by using chemical cross-linking and
time-of-flight (TOF) mass spectrometry to identify LYS-LYS cross-links. They also
suggested that distance restraints on Lysine residues are useful to improve accuracy of
structure prediction (fold recognition) based on protein threading. Some other heuristic
approaches have also been proposed (Albrecht, Hanisch, Zimmer, & Lengauer, 2002; Li.
Algorithmic Aspects of Protein Threading 129
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Zhang, & Skolnick, 2004; Meiler & Baker, 2003). Based on these, Akutsu, Hayashida,
Tomita, Suzuki, and Horimoto (2004) considered the protein threading with profiles and
constraints. In this section, we briefly explain their approach.
In order to cope with constraints, we modify protein threading with profiles as follows
(see Figure 5). Let A
s
be the set of pairs of residue positions in s for each of which the
distance is known (i.e., the distance is known for each pair ( , )
s
i j A where i < j). We define
a function IC(i,j,i,j) , where ( , )
s
i j A and i,j ( i<j) denote positions of residues in the
template structure. Then, IC(i,j,i,j) =1 if the distance between the i-th and j-th residues
and the distance between the i-th and j-th residues satisfy some constraints (e.g., the
difference between these distances is less than some threshold). Otherwise, IC(i,j,i,j)
=0. It should be noted that IC(i,j,i,j) is not defined if ( , )
s
i j A . Then, profile threading
with strict constraints is defined as a protein threading with profiles under the following
conditions: (i) any residue included in A
s
must not be aligned with a gap symbol, (ii) for
any pairs ( , )
s
i j A , IC(i,j,i,j) =0 must hold, where the i-th and j-th residues of t are
aligned with the i-th and j-th residues of s in alignment (s,t), respectively. These
conditions mean that distance constraints must be satisfied for all known distances in
the target sequence.
However, it may not be possible in some cases to satisfy all distance constraints.
Therefore, we consider profile threading with non-strict constraints. In this problem, an
optimal (maximum) threading is computed under the following conditions: (i) any residue
included in A
s
must not be aligned with a gap symbol, (ii)
( , )
( , , ', ')
s
i j A
IC i j i j

is the minimum.
Though profile threading can be solved in O(mn) time, profile threading with (both strict
and non-strict) constraints is proven to be NP-hard (Akutsu et al., 2004) . They developed
two exact algorithms for profile threading with strict constraints: CLIQUETHREAD and
BBDPTHREAD. CLIQUETHREAD reduces the threading problem to the maximum edge-
weight clique problem whereas BBDPTHREAD combines dynamic programming and
Figure 5. Protein threading with constraints. IC(i,j,i,j)=0 if |dist(s
i
,s
j
) dist(t
i
,t
j
)| is less
than a threshold, otherwise IC(i,j,i,j)=1.

dist(t i, t j)
t i
dist(si, sj)
t j
S i
S j
130 Akutsu
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
branch-and-bound techniques. Though BBDPTHREAD is faster than CLIQUETHREAD,
variants of CLIQUETHREAD can be applied to profile threading with non-strict con-
straints.
Related Problems
There exist several problems in bioinformatics that are closely related to protein
threading. Among these, comparison of RNA secondary structures and comparison of
protein structures are important. As shown below, these two problems can be defined
in similar ways as in protein threading with pair score functions. Indeed, the proofs of
general hardness results and approximation algorithms for special cases in these three
problems are similar to each other (Akutsu & Miyano, 1999; Goldman, Istrail, &
Papadimitriou, 1999; Lin, Chen, Jiang, & Wen, 2002). These similarities suggest that a
technique developed in one problem may be applied to the other problems. Here, we
briefly explain relations between these two problems and protein threading.
First, we consider a comparison of RNA secondary structures. An RNA secondary
structure can be considered as a combination of a sequence and a graph like a tree
structure. Therefore, comparison of RNA secondary structures can be defined as graph-
theoretic problems. Though various formulations have been proposed (Allali & Sagot,
2005; Hoechsmann, Voss, & Giegerich, 2004; Jiang, Lin, Ma, & Zhang, 2004; Zhang, 2004),
we consider here the longest arc-preserving common subsequence problem (LAPCS)
because it has been well-studied (Lin et al., 2002) and has a close relationship with protein
threading (see Figure 6).
An RNA sequence is a string over ={A,U,G,C}. In LAPCS, a set of arcs is associated
with each sequence. Let s =s
1
s
m
be an RNA sequence. Then, an arc set A
s
is defined as
a set of pairs of positions (i,j) such that 1 i j m < . LAPCS can be defined as a problem
Figure 6. Example of LAPCS (longest arc-preserving common sequence); arcs shown
by bold lines are preserved in this alignment

Algorithmic Aspects of Protein Threading 131
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
of finding an optimal (maximum) alignment between two RNA secondary structures (s,
A
s
) and (t, A
t
) under the following conditions: (i) f(x,y)=1 if x=y, otherwise f(x,y)=0, (ii) any
position pairs (i,j) in s and (i,j) in t that are aligned to the same positions must satisfy
( , ) iff. ( ', ')
s t
i j A i j A .
Although this definition looks different from the original definition given in Lin et al.,
these two are equivalent since the length of the longest common subsequence is equal
to the score of the optimal alignment. Here, we re-formulate LAPCS as a threading-like
problem. We define a function g
1
(i,j,i,j) by:
1
0, if ( ( , ) iff. ( ', ') ),
( , , ', ')
, otherwise.
s t
i j A i j A
g i j i j

=

Then, we find an alignment with the maximum score where the score of an alignment is
given by:
1
1 ( , ') ,( , ') ,
( ', ') ( ' , ' ) ( , , ', '),
l
i i
i i i AL j j AL i j
score s t f s t g i j i j
= <
= +

where AL i i ) ' , ( means that the i-th nucleotide of s is aligned to the i-th nucleotide of
t in (s,t). Readers can see that this score is similar to that in protein threading with pair
score functions where the sign of the score should be inverted. If we allow general f and
g
1
, the above formulation can include protein threading with pair score functions. As in
protein threading, LAPCS is proven to be NP-hard. Lin et al. considered the following
restrictions: (1) no two arcs share an endpoint (i.e., each i appears at most once in A
s
),
(2) no two arcs cross each other (i.e., ) iff. ( ) , ( ), , (
4 2 3 4 1 3 4 3 2 1
i i i i i i A i i i i
s
), (3) no
two arcs nest (i.e., ) iff. ) , ( ), , (
3 2 3 1 4 3 2 1
i i i i A i i i i
s
), (4) there are no arcs at all. Based
on these, they defined the following five levels: UNLIMITED: no restrictions; CROSS-
ING: restriction 1; NESTED: restrictions 1 and 2; CHAIN: restrictions 1,2, and 3; and
PLAIN: restriction 4. They used the notation LAPCS(L1,L2) to represent the LAPCS
problem where A
s
is of level L1 and A
t
is of level L2. It should be noted that
LAPCS(PLAIN,PLAIN) corresponds to the longest common subsequence problem
(without arcs). Various complexity results have been obtained depending on subclasses
of LAPCS based on this notation. They and other people obtained theoretical results
on various subclasses of LAPCS. For example, LAPCS(NESTED,NESTED) is NP-hard
whereas LAPCS(CHAIN,NESTED) can be solved in polynomial time. For details, see Lin
et al. and other references listed there. It should be noted that it is possible to obtain
polynomial time algorithms for RNA secondary structure comparison using dynamic
programming if we employ some other formulations (Allali & Sagot, 2005; Hoechsmann
et al., 2004; Zhang, 2004). However, it is unclear which formulation is adequate for
practice. Thus, rigorous comparison of existing algorithms should be done in the future.
132 Akutsu
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Next, we consider comparison of protein structures. Comparison of protein structures is
important because protein structures are closely related to protein functions and it is
known that there are many protein pairs with structural similarity but without sequence
similarity. Therefore, various kinds of approaches have been proposed to compare
protein structures. Protein structure alignment is one of the well-studied approaches.
Though there exist various definitions on protein structure alignment, we consider here
the contact map overlap (CMO) problem (Goldman et al., 1999; Caprara, Carr, Istrail,
Lancia, & Walenz, 2004) because it has close relationships with protein threading and
LAPCS (see Figure 7).
As in LAPCS, an arc set A
s
is associated with each protein sequence s. Each arc (i,j) means
that the distance between the i-th and j-th residues is small. Then, CMO is, given (s, A
s
)
and (t, A
t
), to find an optimal alignment which maximizes the score given by:

, ) ' , ' , , ( ) ' , ' (
, ) ' , ( , ) ' , (
2
<
=
j i AL j j AL i i
j i j i g t s score
where g
2
is defined by:

=
otherwise. , 0
, ) ' , ' ( and ) , ( if , 1
) ' , ' , , (
2
t s
A j i A j i
j i j i g
From this definition, readers can see close relationships between CMO and other
problems. As in other cases, CMO is proven to be NP-hard (Goldman et al., 1999).
However, practical algorithms, in which optimal solutions are guaranteed, are developed
using branch-and-bound and Lagrangian relaxation techniques (Caprara et al., 2004).
Figure 7. Example of CMO (contact map overlap); each arc means that the distance
between two residues connected by the arc is less than some threshold, arcs shown by
bold lines are common in two structures (i.e., the score of this alignment is 5)

Algorithmic Aspects of Protein Threading 133
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Summary
In this chapter, we overviewed algorithms for computing optimal solutions for protein
threading and related problems. It is shown that various kinds of optimization techniques
have been applied to these problems in non-trivial manners. Owing to these develop-
ments, it has become possible to compute optimal threadings in reasonable CPU time for
most proteins. However, quick computation is required if we apply protein threading to
annotation of genome sequences. Therefore, further improvements should be done. As
seen in the latter part of this chapter, it is interesting that several important problems in
bioinformatics have similar problem structures. Therefore, techniques developed for one
of these problems may also be applied to other problems.
Computation of optimal solutions is important not only from an algorithmic viewpoint but
also from a practical viewpoint. Indeed, PROSPECT and/or RAPTOR showed good
performances in CASP competitions, which are the world wide competitions for protein
structure prediction. These results suggest that computation of optimal solutions is
important to get good prediction results in practice. Information about both PROSPECT
and RAPTOR is available from http://www.bioinformaticssolutions.com/. It should also
be noted that special issues (Lattman, 2001, 2003) in a protein journal and the Web pages
(http://predictioncenter.llnl.gov/) for CASP are good sources to know state-of-the-art
techniques, software, and benchmark data for protein structure prediction and protein
threading.
In this chapter, we did not explain details of score functions. However, score functions
greatly affect the quality of threading results. Thus, developing a good score function
is important for improving the performance of protein threading. Indeed, a lot of studies
have been done for deriving good score functions from training data based on statistical
and/or heuristic techniques. RAPTOR also adopted SVM to tune score functions.
However, a few studies have been done from an algorithmic viewpoint (Akutsu &
Yagiura, 1998). Therefore, more algorithmic studies should be done for deriving good
score functions.
References
Akutsu, T., Hayashida, M., Tomita, E., Suzuki, J., & Horimoto, K. (2004). Protein threading
with profiles and constraints. Proceedings of 4th

IEEE International Symposium
of BioInformatics and BioEngineering (pp. 537-544). Los Alamitos: IEEE.
Akutsu, T., & Miyano, S. (1999). On the approximation of protein threading. Theoretical
Computer Science, 210, 261-275.
Akutsu, T., & Yagiura, M. (1998). On the complexity of deriving score functions from
examples for problems in molecular biology. Lecture Notes in Computer Science,
1143, 832-843. Berlin: Springer.
134 Akutsu
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Albrecht, M., Hanisch, D., Zimmer, R., & Lengauer, T. (2002). Improving fold recognition
of protein threading by experimental distance constraints. In Silico Biology, 2,
0030.
Allali, J., & Sagot, M.-F. (2005). A new distance for high level RNA secondary structure
comparison. IEEE/ACM Trans. Computational Biology and Bioinformatics, 2, 3-
14.
Altschul, S. F., Madden T. L., Schaffer A. A., Zhang, J., Zhang, Z., Miller, W., & Lipman,
D. J. (1997). Gapped BLAST and PSI-BLAST: A new generation of protein database
search programs. Nucleic Acids Research, 25, 3389-3402.
Bowie, J. U., Luthy, R., & Eisenberg, D. (1991). A method to identify protein sequences
that fold into a known three-dimensional structure. Science, 253, 164-179.
Caprara, A., Carr, R., Istrail, S., Lancia, G., & Walenz, B. (2004). 1001 Optimal PDB structure
alignments: Integer programming methods for finding the maximum contact map
overlap. Journal of Computational Biology, 11(1), 27-52.
Clote, P., & Backofen, R. (2000). Computational molecular biology: An introduction.
New York: Wiley.
Durbin, R., Eddy, S. R., Krogh, A., & Mitchison, G. (1998). Biological sequence analysis:
Probabilistic models of proteins and nucleic acids. Cambridge, UK: Cambridge
University Press.
Edgar, R. C., & Sjolander, K. (2004). A comparison of scoring functions for protein
sequence profile alignment. Bioinformatics, 20, 1301-1308.
Garey, M. R., & Johnson, D. S. (1979). Computers and intractability: A guide to the
theory of NP-completeness. New York: Freeman.
Goldman, D., Istrail, S., & Papadimitriou, C. H. (1999). Algorithmic aspects of protein
structure similarity. Proceedings of the 40th IEEE Symposium on Foundations of
Computer Science (pp. 512-522). Los Alamitos: IEEE.
Godzik, A., & Skolnick, J. (1992). Sequence-structure matching in globular proteins:
Application to supersecondary and tertiary structure determination. Proceedings
of National Academy of Sciences USA, 89, (pp. 12098-12102).
Hoeshmann, M., Voss, B., & Giegerich, R. (2004). Pure multiple RNA secondary structure
alignments: A progressive profile approach. IEEE/ACM Trans. Computational
Biology and Bioinformatics, 1, 53-62.
Jiang, T., Lin, G-H., Ma, B., & Zhang, K. (2004). The longest common subsequence
problem for arc-annotated sequences. Journal of Discrete Algorithms, 2, 257-270.
Jones, D. T., Taylor, W. R., & Thornton, J. M. (1992). A new approach to protein fold
recognition. Nature, 358, 86-89.
Lathrop, R. H. (1994). The protein threading problem with sequence amino acid interac-
tion preferences is NP-complete. Protein Engineering, 7, 1059-1068.
Lathrop, R. H., & Smith, T. F. (1996). Global optimum protein threading with gapped
alignment and empirical pair score functions. Journal of Molecular Biology, 255,
641-665.
Algorithmic Aspects of Protein Threading 135
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Lattman, E. E. (2001). CASP4 editorial. Proteins: Structure, function, and genetics,
45(S5), 1-1.
Lattman, E. E. (2003). Fifth meeting on the critical assessment of techniques for protein
structure prediction. Proteins: Structure, Function, and Genetics, 53(S6), 333-
333.
Li, W., Zhang, Y., & Skolnick, J. (2004). Application of sparse NMR restraints to large-
scale protein structure prediction. Biophysical Journal, 87, 1241-1248.
Lin, G-H., Chen, Z-Z, Jiang, T., & Wen, J. (2002). The longest common subsequence
problem for sequences with nested arc annotations. Journal of Computer and
System Sciences, 65, 465-480.
Meiler, J., & Baker, D. (2003). Rapid protein fold determination using unassigned NMR
data. Proceedings of National Academy of Sciences USA, 100, (pp. 15404-15409).
Xu, J., Li, M., Kim, D., & Xu, Y. (2003). RAPTOR: Optimal protein threading by linear
programming. Journal of Bioinformatics and Computational Biology, 1, 95-118.
Xu, Y., Xu, D., Crawford, O. H., & Einstein, J. R. (2000). A computational method for NMR-
constrained protein threading. Journal of Computational Biology, 7, 449-467.
Xu, Y., Xu, D., Uberbacher, E. C. (1998). An efficient computational method for globally
optimal threading. Journal of Computational Biology, 5, 597-614.
Young, M. M. et al. (2000). High throughput protein fold identification by using
experimental constraints derived from intermolecular cross-links and mass spec-
trometry. Proceedings of the National Academy of Sciences USA, 97, 5802-5806.
Zhang, K. (2004). RNA structure comparison and alignment. In Wang, J. T-L. et al. (Eds.),
Data mining in bioinformatics (pp. 59-81). Berlin: Springer.
136 Kelemen & Liang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Chapter VIII
Pattern Differentiations
and Formulations for
Heterogeneous
Genomic Data through
Hybrid Approaches
Arpad Kelemen, The State University of New York at Buffalo, USA
& Niagara University, USA
Yulan Liang, The State University of New York at Buffalo, USA
Abstract
Pattern differentiations and formulations are two main research tracks for heterogeneous
genomic data pattern analysis. In this chapter, we develop hybrid methods to tackle
the major challenges of power and reproducibility of the dynamic differential gene
temporal patterns. The significant differentially expressed genes are selected not only
from significant statistical analysis of microarrays but also supergenes resulting from
singular value decomposition for extracting the gene components which can maximize
the total predictor variability. Furthermore, hybrid clustering methods are developed
based on resulting profiles from several clustering methods. We demonstrate the
developed hybrid analysis through an application to a time course gene expression
data from interferon--1a treated multiple sclerosis patients. The resulting integrated-
condensed clusters and overrepresented gene lists demonstrate that the hybrid methods
can successfully be applied. The post analysis includes function analysis and pathway
discovery to validate the findings of the hybrid methods.
Pattern Differentiations and Formulations for Heterogeneous Genomic Data 137
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Introduction
Progress in mapping the human genome and developments in microarray technologies
have provided considerable amount of information for delineating the roles of genes in
disease states. Since complex diseases typically involve multiple intercorrelated genetic
and environmental factors that interact in a hierarchical fashion and the clinical charac-
teristics of diseases are determined by a network of interrelated biological traits,
microarrays hold tremendous latent information but their analysis is still a bottleneck.
Pattern analysis can be useful for discovering the knowledge on gene array data related
to certain diseases (Neal et al., 2000; Slonim, 2002). The associations between patterns
and their causes are the bricks from which the wall of biological knowledge and medical
decisions are built.
Pattern differentiations and pattern formulations are two major tracks of patterns
analysis. Pattern differentiation of gene expressions is the first step to identify potential
relevant genes in biological processes. The coordinated/temporal gene arrays are widely
used for pattern formulation in order to study the common functionalities, co-regula-
tions, and pathways that ultimately are responsible for the observed patterns. The
identification of groups of genes with similar temporal patterns of expression is usually
a critical step in the analysis of kinetic data because it provides insights into the gene-
gene interactions and thereby facilitates the testing and development of mechanistic
models for the regulation of the underlying biological processes. These temporal pattern
analyses provide clues for genes that are related in their expression through linkage in
a common developmental pathway.
There are several critical challenges in the pattern analyses. One is in the pattern
differentiations, the notorious large p small n problem (West, 2000). The large number
of irrelevant and redundant genes with high level noise measurements and uncertainty
severely degrade both classification and prediction accuracy. The solution for the large
p problem is through affine transformation and feature selection. Affine transformation
such as principal component analysis (PCA) or singular value decomposition (SVD) has
advantage of simplicity and it may remove non-discriminating and irrelevant features
(i.e., genes) by extracting eigenfeatures corresponding to the large eigenvalues (Alter,
Brown, & Botstein, 2000; Yeung and Ruzzo, 2001; Wall, Dyck, & Brettin, 2001). Yet it is
very difficult to identify important genes with these methods and the inherent linear
nature is their

prominent disadvantage.
Feature selection consists of two strategies: screening and wrappers. In the screening
approaches, all genes are analyzed and tested individually to see whether they have
higher expression level in one class than in the other (Baldi & Long 2001; Tusher,
Tibshirani, & Chu, 2001; Storey & Tibshirani 2003; Hastiel et al., 2000). The disadvantage
of screening processes is that they are non-invertible and can cause multiple testing and
model selection problems (Westfall & Young, 1993; Benjamini & Hochberg, 2002).
In wrapper methods, genes are tested not independently, but as ensembles, and
according to their performance in the classification model (Golub et al., 1999; Khan et al.,
2001; Wuju & Momiao, 2002). Since the number of feature subsets increases exponen-
tially with the dimensions of the feature space, wrappers are computationally intractable
for high-dimensional gene data.
138 Kelemen & Liang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
The second critical challenge in the patterns analysis is that there are some limitations
of most of the existing pattern formulation methods such as clustering methods (Eisen,
Spellman, Brown, & Botstein, 1998; Tamayo et al., 1999; Yeung et al., 2003). First, they
may not detect genes with different functions showing similar profiles by chance or
stochastic fluctuation from high level noise. Second, most of these approaches have not
incorporated the temporal information, such as the features of significant nonlinear trend
and the presence of sudden bursts of amplitude at irregular time intervals and the short-
term dynamics of the data. Third, these methods may not detect indirectly correlated
genes, which show weak correlations due to a time delay or nonlinear association. Last,
most of them can not determine the number of patterns a priori and the final partition of
the data requires evaluation.
In this chapter we develop hybrid methods and several interrelated integrative proce-
dures for overcoming some limitations of existing approaches, increasing the power and
the reproducibility of differentially expressed gene temporal pattern profiles from
heterogeneous genomic data. These procedures include prior analysis with screening
methods for mining the significant differential gene profiles, pattern formulations with
different clustering methods, and post analysis for pathway and function study. The
differentially expressed gene profiles were firstly selected not only from q-value com-
puted from Significant Analysis of Microarray (SAM) but also the supergenes resulting
from Singular Value Decomposition (SVD) methods (Tusher et al., 2001; Alter et al., 2001).
Hybrid pattern formulation methods will be further developed based on the resulting
profiles from different clustering methods, in our case including Hierarchical clustering,
k-means, dynamic Bayesian clustering and singular value decomposition (SVDMAN)
(Eisen et al., 1998; Tamayo et al., 1999, Ramoni, Sebastiani, & Kohane, 2002; Wall et al.,
2001). Most existing clustering methods are based on resampling procedures and
statistical validity testing (Kerr & Churchill, 2001; Dudoit & Fridlyand, 2002; Tibshirani
et al., 2001; McShane et al., 2002). For example, Kerr and Churchill have proposed using
bootstrap techniques and confidence interval to validate the reproducibility of the
clustering results. McShane et al. developed hypothesis testing approaches and sug-
gested using two-step approaches to evaluate the sample clustering. We propose hybrid
clustering approaches (the clustering of the clusters themselves), which not only allow
us visual examination of the degrees of heterogeneity between methods but also help to
improve the clustering power and reproducibility.
Since different clustering algorithms optimize different objective functions and criteria,
hybrid clustering will provide the agreement of various methods, help to detect true
patterns and differentiations not by chance or stochastic fluctuations, and provide an
Integrated-Condensed (I/C) overrepresented gene list with low false positive rate. The
follow-up post analysis such as function analysis and pathway discovery for construct-
ing the association of the informative genes is performed to validate the findings of the
methods (Hosack, Hosack, Sherman, Lane, & Lempicki, 2003; Glynn et al., 2003). We also
demonstrate the developed hybrid temporal analysis through an application to a time
course of lymphocyte gene expression data from interferon--1a treated multiple sclero-
sis (MS) patients (Weinstock-Guttman et al., 2003).
Pattern Differentiations and Formulations for Heterogeneous Genomic Data 139
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Multiple Sclerosis Microarray Data Set
Multiple sclerosis (MS) is an autoimmune disease in which the bodys immune system
attacks the brain and spinal cord, resulting in neurological disabilities. Multiple sclerosis
is an inflammatory-demyelinating disease of the central nervous system that affects over
1 million patients worldwide. It is a complex, variable disease that causes physical and
cognitive disabilities: nearly 50% of the patients diagnosed with MS are unable to walk
after 15 years. The etiology and pathogenesis of MS remains poorly understood (Jacobs
et al., 1996). Recombinant human IFN- is one of the most commonly prescribed forms
of therapy for relapsing multiple sclerosis patients on the basis of several clinical trials
(Steinman & Zamvil, 2003; Weinstock-Guttman et al., 2003). The effects of IFN-
treatment are complex, and its pharmaco-dynamics at the genomic level in humans are
poorly understood. Patients with relapsing MS respond better to IFN- treatment than
patients with progressive disease. However, this better response is not homogeneous
among relapsing patients. Studies were conducted to characterize the dynamics of the
gene expression induced by IFN--1a treatment in MS patients and to examine the
molecular mechanisms potentially capable of causing heterogeneity in response to
therapy (Ramanathan et al., 2001).
Fourteen patients with active relapsing remitting MS were recruited for the study. These
patients had not previously received IFN- and were clinically stable for the preceding
four weeks. Peripheral blood samples were obtained just before treatment and at 1, 2, 4,
8, 24, 48, 120, and 168 hrs. and three months after a 30 mg dose of i.m. IFN--1a was
administered. The GeneFillers GF211 DNA arrays (Research Genetics, Huntsville, AL)
were used in this study. Each filler contained multiple positive control spots and
housekeeping genes. Each housekeeping gene was spotted in duplicate, and there were
over 4000 known genes in the GF211 array (see ftp://ftp.resgen.com/pub/genefilters/).
Fourteen patients with 4324 genes were measured at each time point, except the first
patient who was not measured at three months. Because treatment can often cause
changes in the expression of housekeeping genes, the global normalization option was
selected, i.e., each of the filters were normalized using the intensity from all spots on the
GF211 GeneFilter DNA array and for intensity ranges. In addition, it was discovered that
266 genes (henceforth referred to as unknowns) did not have titles but they all had clone
numbers.
Methods and Results
Pattern Differentiations
In order to identify the significant differences in expression levels between two different
conditions (time points) significant analysis of microarray (SAM) is performed first.
SAMs have utilized the multi-class grouping option to facilitate comparison of gene
expression levels across the entire time course of gene expression. There are several other
140 Kelemen & Liang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
major advantages using SAM algorithms to find a list of the most significant genes (i.e.,
the genes whose expression changed significantly in response to treatment). First,
SAMs use q-value instead of p-value to control the false discovery rate where multiple
testing problems and dependences among genes are taken into account. Second, they
utilize the permutation methods in which no assumptions are made regarding the null
distribution. The tuning parameter, known as delta (), is chosen based on an
acceptable false positive rate. Since SAM does not provide any direct means of
normalizing/centering the data, these steps are performed prior to inputting the data into
SAM.
Due to the fact that complex diseases are contributed by many small effects of multiple
genes, principal component analysis (PCA) is also used to characterize the data
(including the variations and structures of the data) and to look for combination effects
(supergenes) that have major contributions to the given disease. To deal with the
computationally inhibitive large p, small n problem, we used singular value decompo-
sition analysis of microarray (SVDMAN), which uses singular value decomposition as
a computational shortcut of PCA. The general idea is that the genes by array space is
transformed through a translation matrix into a reduced eigengene x eigenarray space to
capture the inherent dependency and high variations of the data. One advantage of
SVDMAN is that it can be further used for pattern formulations. Figure 1 provides the
Venn diagram showing the relative proportions and intersection of the output of
significant genes from the SAM and SVD outputs.
Pattern Formulations
The hierarchical clustering with mean centered, and the uncentered Pearson correlation
coefficient (PCC) similarity metric, k-means, Bayesian cluster analysis of gene expression
(CAGED), and SVDMAN were applied to differentially expressed genes selected by SAM
and SVD methods. Hierarchical clustering and k-means methods have visualization
Figure 1. Venn Diagram showing the relative proportions and intersection of the
output of significant genes from the SAM and SVD outputs
285
117 232
4324 genes
SAM-402 SVD-349
Pattern Differentiations and Formulations for Heterogeneous Genomic Data 141
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
advantages and straightforward interpretations and are model free. The advantage of
hierarchical clustering is that it places genes into non-distinct clusters and a biologist
can look at the structure of the tree and judge which genes were clustered in a meaningful
manner. For the hierarchical clustering, the data was hierarchically clustered using the
usual uncentered PCC similarity metric and the centroid linkage clustering. At the risk
of a sounding flippant, k-means clustering has the advantage of placing genes into (a
predetermined number of) distinct clusters. That is, the biologists can force the genes
into a predetermined number of clusters (10 clusters in this case). However, it is
subjective for the choice of the number of clusters; both methods have no measures for
the stochastic fluctuations and they disregard the dynamic features of the observations
and temporal information. CAGED and SVDMAN methods are both model based and they
utilize the significance of the similarity to identify the best number of clusters from the
data (Ramoni et al., 2002). CAGED methods are also referred to as Bayesian model-based
approaches that account for the dependences of the temporal information using
autoregressive models and use agglomerative procedures to search for the most
probable set of clusters. The CAGED method has two major advantages over other
methods. First, it determines the number of clusters based on its internal modeling
algorithm (see Figure 2). Second, it takes into account that the series is a time series and
therefore the nature of a given point must take into consideration the point(s) before it.
That is, if we would rearrange the order of the series similarly for all genes (e.g. switched
Figure 2. Part of the Dendogram from the model based clustering of the MS data using
the CAGED 1.0 program; the unknown samples (numbers) can be seen (cluster 11)

142 Kelemen & Liang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Figure 3(a). Graphs of the Right Singular Vectors from the SVDMAN program run on
the MS data; the first four principal components zero through three are shown (note:
Principal component zero appears flat because of the scale)
Figure 3(b). Eigenexpression Fractions of the ten principle components (in order),
which capture the most variations of the data
the one hour and seven day values), the hierarchical and k-means clustering results
would be the same, but the results from the CAGED analysis would change. The
advantage of SVD is that it can capture most of the variation and the distribution of the
data and provide supergenes that are functionally not separable (see Figure 3(a) and
Figure 3(b)).
Pattern Differentiations and Formulations for Heterogeneous Genomic Data 143
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Hybrid Methods for Pattern Differentiation and
Formulations
Resampling techniques, such as permutation and bootstrap techniques, can improve
statistical significance and the reproducibility of the results. However, the genes
observed to be differentially expressed from a given method should be confirmed with
other independent methods on independent samples. For instance, the wrapper methods
can be utilized in the classification model and use the known classes information and
classification error to verify whether the best predictors have been obtained. In this
chapter we develop another strategy, called hybrid methods to improve the significance
and reproducibility of the results. Due to the fact that different methods apply different
algorithms and criteria, the hybrid multiple criteria could help to improve the quality of
the resulting informative gene profiles. For hybrid pattern differentiation, we applied
SAMs and SVDMANs results, in which controlling the false positive rate and capturing
the majority of the variations of the data can be achieved. This also improved the
identifications of the novel genes that are truly responsible for specific diseases and with
low false-positive rate and good precision while maintaining sensitivity to relatively
subtle changes.

- 0. 3
- 0. 2
- 0. 1
0
0 .1
0 .2
0 .3
0 .4
0 .5
0 .6
P r e 1 ho ur 2 ho ur 4 ho ur 8 ho ur 2 4 h ou r 4 8 h ou r 1 2 0 ho u r 1 6 8 ho u r 3 m on t h s - 0. 3
- 0. 2
- 0. 1
0
0 .1
0 .2
0 .3
0 .4
0 .5
P r e 1 ho ur 2 ho ur 4 ho ur 8 ho ur 2 4 h ou r 4 8 h ou r 1 2 0 ho u r 1 6 8 ho u r 3 mo n t h s - 0. 2
- 0. 1
0
0 .1
0 .2
0 .3
0 .4
P r e 1 ho ur 2 ho ur 4 ho ur 8 ho ur 2 4 h ou r 4 8 h ou r 1 2 0 ho u r 1 6 8 ho u r 3 m on t h s
- 0. 2
- 0. 1
0
0 .1
0 .2
0 .3
0 .4
P r e 1 ho ur 2 ho ur 4 ho ur 8 ho ur 2 4 h ou r 4 8 h ou r 1 2 0 ho u r 1 6 8 ho u r 3 mo n t h s - 0. 3
- 0. 2
- 0. 1
0
0 .1
0 .2
0 .3
0 .4
0 .5
P r e 1 ho ur 2 ho ur 4 ho ur 8 ho ur 2 4 h ou r 4 8 h ou r 1 2 0 ho u r 1 6 8 ho u r 3 mo n t h s - 1. 4
- 1. 2
- 1
- 0. 8
- 0. 6
- 0. 4
- 0. 2
0
0 .2
0 .4
0 .6
P r e 1 ho ur 2 ho ur 4 ho ur 8 ho ur 2 4 h ou r 4 8 h ou r 1 2 0 ho u r 1 6 8 ho u r 3 m on t h s
- 1. 4
- 1. 2
- 1
- 0. 8
- 0. 6
- 0. 4
- 0. 2
0
0 .2
0 .4
0 .6
P r e 1 ho ur 2 ho ur 4 ho ur 8 ho ur 2 4 h ou r 4 8 h ou r 1 2 0 ho u r 1 6 8 ho u r 3 m on t h s - 0. 6
- 0. 5
- 0. 4
- 0. 3
- 0. 2
- 0. 1
0
0 .1
0 .2
0 .3
0 .4
P r e 1 ho ur 2 ho ur 4 ho ur 8 ho ur 2 4 h ou r 4 8 h ou r 1 2 0 ho u r 1 6 8 ho u r 3 m on t h s - 1.2
- 1
- 0. 8
- 0. 6
- 0. 4
- 0. 2
0
0 .2
0 .4
P r e 1 ho ur 2 ho ur 4 ho ur 8 ho ur 2 4 h ou r 4 8 h ou r 1 2 0 ho u r 1 6 8 ho u r 3 mo n t h s
-1. 2
-1
-0. 8
-0. 6
-0. 4
-0. 2
0
0.2
0.4
Pr e 1 hour 2 hour 4 hour 8 hour 24 hour 48 hour 120 hour 168 hour 3 mont hs
- 0 . 2
- 0 . 1
0
0 . 1
0 . 2
0 . 3
0 . 4
P r e 1 h o ur 2 h o ur 4 h o ur 8 h o ur 2 4 h ou r 4 8 h ou r 1 2 0 ho u r 1 6 8 ho u r 3 m o n t h s - 0 . 2
- 0 . 1
0
0 . 1
0 . 2
0 . 3
0 . 4
P r e 1 h o u r 2 h o u r 4 h o u r 8 h o u r 2 4 h o u r 4 8 h o u r 1 2 0 h o u r 1 6 8 h o u r 3 m o n t h s
- 0 . 1 5
- 0 . 1
- 0 . 0 5
0
0 . 0 5
0 . 1
0 . 1 5
0 . 2
0 . 2 5
0 . 3
0 . 3 5
P r e 1 h o ur 2 h o ur 4 h o ur 8 h o ur 2 4 h ou r 4 8 h ou r 1 2 0 ho u r 1 6 8 ho u r 3 m o n t h s - 0 . 2
- 0 . 1
0
0 . 1
0 . 2
0 . 3
0 . 4
P r e 1 h o ur 2 h o ur 4 h o ur 8 h o ur 2 4 h ou r 4 8 h ou r 1 2 0 ho u r 1 6 8 ho u r 3 m o n t h s - 0 . 3
- 0 . 2
- 0 . 1
0
0 . 1
0 . 2
0 . 3
0 . 4
0 . 5
0 . 6
P r e 1 h o u r 2 h o u r 4 h o u r 8 h o u r 2 4 h o u r 4 8 h o u r 1 2 0 h o u r 1 6 8 h o u r 3 m o n t h s
- 1 . 2
- 1
- 0 . 8
- 0 . 6
- 0 . 4
- 0 . 2
0
0 . 2
0 . 4
0 . 6
P r e 1 h o u r 2 h o u r 4 h o u r 8 h o u r 2 4 h o u r 4 8 h o u r 1 2 0 h o u r 1 6 8 h o u r 3 m o n t h s - 0 . 6
- 0 . 5
- 0 . 4
- 0 . 3
- 0 . 2
- 0 . 1
0
0 . 1
0 . 2
0 . 3
P r e 1 h o ur 2 h o ur 4 h o ur 8 h o ur 2 4 h ou r 4 8 h ou r 1 2 0 ho u r 1 6 8 ho u r 3 m o n t h s - 0 . 3
- 0 . 2
- 0 . 1
0
0 . 1
0 . 2
0 . 3
0 . 4
0 . 5
P r e 1 h o u r 2 h o u r 4 h o u r 8 h o u r 2 4 h o u r 4 8 h o u r 1 2 0 h o u r 1 6 8 h o u r 3 m o n t h s
- 0 . 2
- 0 . 1 5
- 0 . 1
- 0 . 0 5
0
0 . 0 5
0 . 1
0 . 1 5
0 . 2
0 . 2 5
0 . 3
0 . 3 5
P r e 1 h o u r 2 h o u r 4 h o u r 8 h o u r 2 4 h o u r 4 8 h o u r 1 2 0 h o u r 1 6 8 h o u r 3 m o n t h s
0
0 . 0 5
0 . 1
0 . 1 5
0 . 2
0 . 2 5
0 . 3
0 . 3 5
0 . 4
0 . 4 5
P r e 1 h o u r 2 h o u r 4 h o u r 8 h o u r 2 4 h o u r 4 8 h o u r 1 2 0 h o u r 1 6 8 h o u r 3 m o n t h s 0
0 . 0 5
0 . 1
0 . 1 5
0 . 2
0 . 2 5
0 . 3
0 . 3 5
P r e 1 h o u r 2 h o u r 4 h o u r 8 h o u r 2 4 h o u r 4 8 h o u r 1 2 0 h o u r 1 6 8 h o u r 3 m o n t h s - 0 . 3
- 0 . 2 5
- 0 . 2
- 0 . 1 5
- 0 . 1
- 0 . 0 5
0
P r e 1 h o u r 2 h o u r 4 h o u r 8 h o u r 2 4 h o u r 4 8 h o u r 1 2 0 h o u r 1 6 8 h o u r 3 m o n t h s
- 0 . 2
- 0 . 1 8
- 0 . 1 6
- 0 . 1 4
- 0 . 1 2
- 0 . 1
- 0 . 0 8
- 0 . 0 6
- 0 . 0 4
- 0 . 0 2
0
P r e 1 h o u r 2 h o u r 4 h o u r 8 h o u r 2 4 h o u r 4 8 h o u r 1 2 0 h o u r 1 6 8 h o u r 3 m o n t h s - 0 . 4
- 0 . 3 5
- 0 . 3
- 0 . 2 5
- 0 . 2
- 0 . 1 5
- 0 . 1
- 0 . 0 5
0
P r e 1 h o u r 2 h o u r 4 h o u r 8 h o u r 2 4 h o u r 4 8 h o u r 1 2 0 h o u r 1 6 8 h o u r 3 m o n t h s - 0 . 6
- 0 . 5
- 0 . 4
- 0 . 3
- 0 . 2
- 0 . 1
0
P r e 1 h o u r 2 h o u r 4 h o u r 8 h o u r 2 4 h o u r 4 8 h o u r 1 2 0 h o u r 1 6 8 h o u r 3 m o n t h s
0
0 . 1
0 . 2
0 . 3
0 . 4
0 . 5
0 . 6
0 . 7
0 . 8
0 . 9
P r e 1 h o u r 2 h o u r 4 h o u r 8 h o u r 2 4 h o u r 4 8 h o u r 1 2 0 h o u r 1 6 8 h o u r 3 m o n t h s 0
0 . 1
0 . 2
0 . 3
0 . 4
0 . 5
0 . 6
0 . 7
0 . 8
P r e 1 h o u r 2 h o u r 4 h o u r 8 h o u r 2 4 h o u r 4 8 h o u r 1 2 0 h o u r 1 6 8 h o u r 3 m o n t h s 0
0 . 1
0 . 2
0 . 3
0 . 4
0 . 5
0 . 6
P r e 1 h o u r 2 h o u r 4 h o u r 8 h o u r 2 4 h o u r 4 8 h o u r 1 2 0 h o u r 1 6 8 h o u r 3 m o n t h s
- 0 . 8
- 0 . 7
- 0 . 6
- 0 . 5
- 0 . 4
- 0 . 3
- 0 . 2
- 0 . 1
0
P r e 1 h o u r 2 h o u r 4 h o u r 8 h o u r 2 4 h o u r 4 8 h o u r 1 2 0 h o u r 1 6 8 h o u r 3 m o n t h s 0
0 . 1
0 . 2
0 . 3
0 . 4
0 . 5
0 . 6
0 . 7
0 . 8
0 . 9
1
P r e 1 h o u r 2 h o u r 4 h o u r 8 h o u r 2 4 h o u r 4 8 h o u r 1 2 0 h o u r 1 6 8 h o u r 3 m o n t h s0
0 . 2
0 . 4
0 . 6
0 . 8
1
1 . 2
1 . 4
P r e 1 h o u r 2 h o u r 4 h o u r 8 h o u r 2 4 h o u r 4 8 h o u r 1 2 0 h o u r 1 6 8 h o u r 3 m o n t h s
0
0 . 2
0 . 4
0 . 6
0 . 8
1
1 . 2
1 . 4
1 . 6
P r e 1 h o u r 2 h o u r 4 h o u r 8 h o u r 2 4 h o u r 4 8 h o u r 1 2 0 h o u r 1 6 8 h o u r 3 m o n t h s- 2
- 1 . 8
- 1 . 6
- 1 . 4
- 1 . 2
- 1
- 0 . 8
- 0 . 6
- 0 . 4
- 0 . 2
0
P r e 1 h o u r 2 h o u r 4 h o u r 8 h o u r 2 4 h o u r 4 8 h o u r 1 2 0 h o u r 1 6 8 h o u r 3 m o n t h s
-1 8 0
-1 7 5
-1 7 0
-1 6 5
-1 6 0
-1 5 5
P r e 1 h ou r 2 h ou r 4 h ou r 8 h ou r 2 4 hour 4 8 hour 1 20 h ou r 1 68 h ou r 3 mon th s
-15
-10
-5
0
5
10
15
20
25
30
35
P r e 1 hour 2 hour 4 hour 8 hour 24 hour 48 hour 120 hour 168 hour 3 mont hs
-1 5
-1 0
-5
0
5
1 0
1 5
P r e 1 h ou r 2 h ou r 4 h ou r 8 h ou r 2 4 hour 4 8 hour 1 20 h ou r 1 68 h ou r 3 mon th s
-1 5
-1 0
-5
0
5
1 0
P r e 1 h ou r 2 h ou r 4 h ou r 8 h ou r 2 4 hour 4 8 hour 1 20 h ou r 1 68 h ou r 3 mon th s
-8
-6
-4
-2
0
2
4
P r e 1 h ou r 2 h ou r 4 h ou r 8 h ou r 2 4 hour 4 8 hour 1 20 h ou r 1 68 h ou r 3 mon th s
-4
-3
-2
-1
0
1
2
3
4
5
P r e 1 h ou r 2 h ou r 4 h ou r 8 h ou r 2 4 hour 4 8 hour 1 20 h ou r 1 68 h ou r 3 mon th s
-4
-3
-2
-1
0
1
2
3
4
P r e 1 h ou r 2 h ou r 4 h ou r 8 h ou r 2 4 hour 4 8 hour 1 20 h ou r 1 68 h ou r 3 mon th s
-4
-3
-2
-1
0
1
2
3
P r e 1 h ou r 2 h ou r 4 h ou r 8 h ou r 2 4 hour 4 8 hour 1 20 h ou r 1 68 h ou r 3 mon th s
-2
-1
0
1
2
3
4
P r e 1 h ou r 2 h ou r 4 h ou r 8 h ou r 2 4 hour 4 8 hour 1 20 h ou r 1 68 h ou r 3 mon th s
Figure 4. Differential patterns from various existing clustering methods for MS disease:
top left: Hierarchical clustering, profiles of HN1-5 (hierarchical negatively expressed
clusters 1 to 5) and HP1-4 (hierarchical positively expressed clusters 1 to 4);top right:
k-means (K0-K9); bottom left: Dynamic clustering from CAGED (C1-C14); bottom
right: Singular value decomposition from SVDMAN, the nine principal components
zero through eight are shown (S0-S8). x-axis: time, y-axis: gene expression level
144 Kelemen & Liang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
The determined numbers of clusters for each method and cluster profiles for our data were
as follows: 9 for hierarchical: Profiles of HN1-5 (hierarchical negatively expressed
clusters 1 to 5) and HP1-4 (hierarchical positively expressed clusters 1 to 4), 10 for k-
means (K0-K9), 14 for CAGED (C1-C14), and 9 for SVDMAN (S0-S8) as shown in Figure
4. Accordingly, hybrid clustering (clustering of the clusters) from four clustering
methods (hierarchical, k-means, CAGED, and SVDMAN) was developed and applied to
our data sets to compare the results of different methods and to improve the power of
reproducibility. That is, after the four different methods were performed, the individual
profiles had to be matched up from one method to the next. Instead of doing this by
hand, it dawned on the authors of this chapter that a better way to do this is by clustering
the clusters themselves. The profiles of the clusters from all four clustering methods
(hierarchical, k-means, CAGED, and SVDMAN) were compiled into one spreadsheet and
the spreadsheet was loaded into Cluster 3.0 (Eisen, 1998). The data was then hierarchi-
cally reclustered using the usual uncentered PCC similarity metric and the centroid
linkage clustering. Tight clusters were handpicked and composite graphs from the mean
centered data are illustrated in Figure 5. The resulting composite C/Cs (Clustering of the
Clusters) demonstrate that this was performed successfully (see Figure 5 and Figure 6).
Figure 5. Hybrid clustering (clustering of the resulting profiles from various methods);
the global view window from JavaTreeView has been sized down to zero for this
illustration

Pattern Differentiations and Formulations for Heterogeneous Genomic Data 145
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Comparison of the Four Clustering Methods via Hybrid
Clustering (C/C)
The results from the hierarchical, k-means, and CAGED cluster analyses came up with
similar profiles, as evidenced by the inclusion of all the three methods in four of the six
C/Cs and the inclusion of two of the three methods in the other two C/Cs. SVD only found
one profile through C/C (S1 in C/C-3). The CAGED program usually had multiple members
in the handpicked clusters from the clustering of the clusters. That is, CAGED had a
tendency to be divisive breaking clusters excessively into subclusters. The SVD
Figure 6. Hybrid clusters (C/C): clustering after clustering (hierarchical, k-means,
Bayesian with CAGED, and SVD); overall general functional characteristics: left:
down regulated (Top to bottom: negatively expressed C/C-1 proliferation; C/C-2
phosphorylation; C/C-3 signal receptors); right: positively expressed up regulated
(top to bottom: C/C-4 stress; C/C-5 DNA replication, membrane; C/C-6 protein
biosynthesis and transport)

-1.4
-1.2
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
Pre 1 hour 2 hour 4 hour 8 hour 24 hour 48 hour 120 hour 168 hour 3 months
hierar pos2
K-Means0
CAGED12
CAGED7
CAGED9
-0.6
-0.5
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0.4
Pre 1 hour 2 hour 4 hour 8 hour 24 hour 48 hour 120 hour 168 hour 3 months
hierar pos3
CAGED1
K-Means7
-1.4
-1.2
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
Pre 1 hour 2 hour 4 hour 8 hour 24 hour 48 hour 120 hour 168 hour 3 months
hierar pos1
K-Means6
CAGED13
CAGED11
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
Pre 1hour 2hour 4hour 8hour 24hour 48hour 120hour 168hour 3months
hierar neg1
K-Means5
CAGED5
CAGED10
CAGED6
-0.2
-0.1
0
0.1
0.2
0.3
0.4
Pre 1hour 2hour 4hour 8hour 24hour 48hour 120hour 168hour 3months
hierar neg3
K-Means2
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Pre 1hour 2hour 4hour 8hour 24hour 48hour 120hour 168hour 3months
svd1
K-Means1
CAGED3
CAGED4
146 Kelemen & Liang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
analysis resulted in a list of genes that did not coincide well with the list of significant
genes from the SAM output. The SVD analysis resulted in a list of 351 genes distributed
across 10 groups (the total length of the list was 929). Most genes appeared in multiple
groups. The average gene appeared in 2.65 groups (929 total length of list/351 unique
genes). The SVD analysis did, however, result in profiles that were markedly different
from one another. That is, the trajectories of a given SVD profile were partial inversions
in some way of the other SVD profiles. This can be seen visually by the fact that in the
clustering of the profiles the SVDMAN clusters were all at the end of long branches
indicating that the SVDMAN analysis correctly extracted the principal components.
Post Analysis: Function and Pathway Analysis of Genes
Resulting from Hybrid Methods
The functions of genes and the functional relationships between genes that were
statistically significant from hybrid clustering were further investigated using the
C/C Cluster Gene Category EASE score
1 HN3 development 0.042
1 HN3 cell proliferation 0.071
1 K2 development 0.089
1 K2 cellular process 0.148
1 K2 coenzyme and prosthetic group biosynthesis 0.156
1 K2 coenzyme and prosthetic group metabolism 0.156
2 S1 organelle organization and biogenesis 0.047
2 HN1 protein modification 0.009
2 K5 protein modification 0.017
2 C5 signal transduction 0.020
2 HN1 phosphate metabolism 0.029
2 HN1 Phosphorus metabolism 0.029
2 C5 intracellular signaling cascade 0.032
2 HN1 protein amino acid phosphorylation 0.033
2 C5 cell surface receptor linked signal transduction 0.036
2 C5 cell communication 0.042
2 HN1 phosphorylation 0.047
2 C10 cell communication 0.049
3 S1 induction of apoptosis by extracellular signals 0.004
3 K1 G-protein coupled receptor protein signaling pathway 0.009
3 S1 EGF receptor signaling pathway 0.010
3 C3 vesicle-mediated transport 0.014
3 S1 cell organization and biogenesis 0.015
3 S1 cytoskeleton organization and biogenesis 0.022
3 S1 cytoplasm organization and biogenesis 0.033
3 K1 cell communication 0.039
3 K1 second-messenger-mediated signaling 0.043
3 C3 DNA metabolism 0.045
3 S1 nitric oxide mediated signal transduction 0.046

Table 1(a). EASE Scores and overrepresented gene categories of negatively expressed
hybrid clusters (C/C)
Pattern Differentiations and Formulations for Heterogeneous Genomic Data 147
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Expression Analysis Systematic Explorer (EASE) software. A variant one-tailed Fisher
exact probability referred to as EASE score is a statistical measure of the
overrepresentation of a class of genes within the genes resulting from SAM and
SVDMAN. The gene list was sorted using the accession numbers (kind of gene ID), and
the accession numbers for each hybrid cluster were run through EASE to obtain the
themes of the overrepresented genes (see Tables 1(a) and 1(b)). The GenBank Human
Supplemental was selected in the 2. Input Genes Either Load pulldown menu in the
EASE program. The rows with the lowest EASE scores were copied and pasted to a new
spreadsheet for each hybrid cluster (cluster of clusters) (C/C-1 through C/C-6). The
cutoffs for the EASE scores for each cluster of clusters were given as follows: C/C-1<0.15,
C/C-2<0.05, C/C-3<0.05, C/C-4<0.20, C/C-5<0.20, and C/C-6<0.20. Each Excel sheet of
lowest EASE scores was sorted by the EASE scores. The GO cellular and GO
molecular rows in the EASE were deleted, leaving only the GO biological rows (except
Table 1(b). EASE Scores and overrepresented gene categories of positively expressed
hybrid clusters (C/C)
C/C Cluster Gene Category EASE score
4 C1 Oxygen and reactive oxygen species metabolism 0.070
4 K7 response to oxidative stress 0.076
4 K7 physiological process 0.111
4 K7 Oxygen and reactive oxygen species metabolism 0.113
4 HP3 physiological process 0.157
4 K7 response to external stimulus 0.164
4 K7 response to stress 0.179
5 K6 DNA dependent DNA replication 0.104
5 K6 S phase of mitotic cell cycle 0.128
5 K6 DNA replication 0.128
5 K6 DNA replication and chromosome cycle 0.175
5-c C11 Membrane 0.101
5-c C11 Integral to membrane 0.110
5-c C11 Unlocalized 0.158
5-c K6 Integral to membrane 0.161
5-c C11 plasma membrane 0.165
5-c K6 Membrane 0.166
6 C12 protein metabolism 0.047
6 HP2 protein biosynthesis 0.054
6 C12 protein transport 0.085
6 C12 intracellular protein transport 0.085
6 HP2 macromolecule biosynthesis 0.131
6 C12 physiological process 0.132
6 C12 protein biosynthesis 0.144
6 C12 intracellular transport 0.157
6 C12 M phase of mitotic cell cycle 0.168
6 C12 Mitosis 0.168
6 K0 macromolecule biosynthesis 0.178
6 C12 Metabolism 0.194
6 HP2 Biosynthesis 0.199

148 Kelemen & Liang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
for C/C-5 only the GO molecular rows were deleted, leaving both the biological and
cellular rows (this was done because the cellular rows seemed to have a general theme
also)).
Binding of IFN- to the cell receptor eventually results in various cross-phosphoryla-
tions and other sequential phosphorylations. The peculiar result of the EASE analysis
is that the genes that tended to be associated with phosphorylation and signal receptors
(C/C-2 and C/C-3 resp.) were in the down regulated clusters the opposite of what would
be expected. Perhaps somewhere in the reduction of the information from EASE, the
meaning was lost. This data and the summary of the data for each cluster and cluster of
clusters (C/C) are given in Tables 1(a) and 1(b). The results in Tables 1(a) and 1(b) were
combined and condensed for each C/C (two times over). The results of these two waves
of combination and condensation are given in Table 2 as Combined-Condensed
Overrepresented Gene Categories and Overall General Characteristic, respectively.
DAVID-KEGG Analysis
DAVID (Database for Annotation, Visualization and Integrated Discovery) is an on-line
tool that allows rapid biological interpretation of gene lists and performs theme discovery
and annotation. All the pathways of all the clusters contained in each hybrid cluster were
compiled and submitted to EASE; then the obtained output was sent trough DAVID via
the accession numbers of genes. The resulting list of pathways was then reduced by
eliminating the less promising pathways (see Tables 2, 3). For example, if a given cluster
produced a list of pathways, which contained 3, 2, and 1 gene(s), the pathways containing
only 1 gene were deleted. If the list contained only pathways that contained one gene,
C/C
Combined-Condensed Overrepresented Gene Categories
Overall EASE
General
Characteristic
1
development, cell proliferation, cellular process, coenzyme and
prosthetic group biosynthesis/metabolism PROLIFERATION
2
organelle organization and biogenesis, protein modification, signal
transduction, phosphate/ phosphorous metabolism, intracellular
signaling cascade, protein amino acid phosphorylation, cell
surface receptor linked signal transduction, cell communication
PHOSPHORY-
LATION
3
induction of apoptosis by extracellular signals, G-protein coupled
receptor protein / EGF receptor signaling pathways, vesicle
mediated transport, cell/cytoskeleton/cytoplasm organization and
biogenesis, cell communication, second-messenger mediated
signaling, DNA metabolism, nitric oxide mediated signal
transduction
SIGNAL RECEPTORS
4
oxygen and reactive oxygen species metabolism, response to
oxidative stress/external stimulus/stress, physiological process
STRESS
5
DNA dependent DNA replication, S phase of mitotic cell cycle,
DNA replication and chromosome cycle
DNA REPLICATION
5-c integral to membrane, unlocalized MEMBRANE
6
protein metabolism/biosynthesis/transport, intracellular protein
transport, macromolecule/protein biosynthesis, physiological
process, M phase of mitotic cell cycle
PROTEIN
BIOSYNTHESIS AND
TRANSPORT

Table 2. Overall general characteristics of hybrid clusters (C/Cs)
Pattern Differentiations and Formulations for Heterogeneous Genomic Data 149
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
C/C 1 2 3
EASE General
Characteristic
Proliferation Phosphorylation Signal Receptors
DAVID-
KEGG
Pathways
Folatebiosynthesis (1)

MAPKsignaling pathway
(1)

Porphyrin and chlorophyll
metabolism (1)


Purine metabolism (3)

MAPKsignaling pathway (2)

Oxidative phosphorylation (2)

N-Glycansbiosynthesis (2)

Tryptophan metabolism (2)

Tyrosine metabolism (2)

beta-Alanine metabolism (2)

Arginine and proline
metabolism (1)

Integrin-mediated celladhesion
(1)


Purine metabolism (3)

Ribosome (3)

Proteasome (3)

Phosphatidylinositol signaling
system (3)

Glycerolipid metabolism (2)

Glutathione metabolism (2)

MAPKsignaling pathway (2)

Oxidative phosphorylation (2)

Starch and sucrosemetabolism
(2)


Table 3(a). DAVID-KEGG Pathways for negative hybrid clusters (C/Cs)
C/C 4 5 6
EASE General
Characteristic
Stress DNA replication, Membrane
Protein Biosynthesis
and Transport
DAVID-KEGG
Pathways
Glutathione
metabolism (2)

Cell cycle (1)

Nitrogen metabolism (1)

Purine metabolism (1)


Ribosome (1)

Glycosylphosphatidyl
inositol (GPI)-anchor
biosynthesis (1)


Table 3(b). DAVID-KEGG Pathways for positive hybrid clusters (C/Cs)
the list was left alone. Next, the pathways that were found in only one cluster for a given
C/C were deleted.
When the gene lists from our hybrid clusters were run in DAVID, only three or fewer
genes from a given cluster were found in a given pathway. This is not an unusually low
number, and not surprisingly some of the pathways returned in the DAVID-KEGG
analysis did coincide with the EASE results. That is, C/C-2 (phosphorylation) returned
150 Kelemen & Liang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
the pathway oxidative phosphorylation (2), C/C-3 (signal receptors) returned the
pathways phosphatidylinositol signaling system (3) and MAPK signaling pathway
(2), C/C-5 (DNA replication, membrane) returned the pathway cell cycle (1), and C/
C-6 (protein biosynthesis and transport) returned ribosome (1). One definite future
direction is determining which genes from the gene lists (accession numbers) are the ones
matching up in the pathways returned by DAVID-KEGG.
Interferon Related Genes
Figure 7 illustrates the interferon pathway of Multiple Sclerosis disease, which is part of
the generalized JAK-STAT pathway [http://stke.sciencemag.org/cgi/cm/CMP_8390].
Any gene title from the list of overrepresented genes that contained the word inter-
feron was investigated as to what C/C and what cluster(s) it was a member of, and what
the general direction of change of expression of that gene was. Results are given in Table
4. It is surprising that four of the five interferon-related genes were actually down
regulated.
Figure 7. The interferon pathway for MS disease

Pattern Differentiations and Formulations for Heterogeneous Genomic Data 151
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Web-Based Resources
A summary of the Web-based software support used in this chapter is provided in Table
5. In the table we list the names of the methods, a brief description, the corresponding
references, and the freely available Web-based software support.
Summary
Our goals include the investigation and assessment of the reproducibility among
different pattern differential and formulation methods, the optimization of the perfor-
mance of a single model, and most importantly the provision of the procedures that
Table 5. Various software and publicly available resources used in this chapter
ACCESSION
NUMBER
TITLE CLUSTER C/C
Gene Expression
Response
AA464417
interferon induced transmembrane
protein 3 (1-8U)
K1, C3 3,3 Negative
N98563
retinoic acid- and interferon-inducible
protein (58kD)
C5 2 Negative
AA827287 interferon-induced protein 35 C5 2 Negative
AA157813 interferon, alpha-inducible protein 27 K6 5 Positive
AA676598
interferon-related developmental
regulator 1
HN1, K5, C5 2,2,2 Negative

Name Type Reference Web-Based Resources
SAM
Significance Analysis
of Microarrays
Tusher et al., (2001)
http://www-stat-class.stanford.edu/SAM/SAMServlet
Cluster
Treeview
Hierarchical and
k-means, etc.
Eisen, et al. (1998)
Cluster and Treeview
Manual.
http://bonsai.ims.u-
tokyo.ac.jp/~mdehoon/software/cluster/software.htm
http://genetics.stanford.edu/~alok/TreeView/
CAGED Model
CAGED v 1.0 user
manual (2002)
http://genomethods.org/caged/about.htm

SVDMAN

Modified PCA
Alter, et al. (2000)
and
Wall et al. (2001)
http://public.lanl.gov/svdman/
EASE Function analysis Hosack, et al., (2003) http://david.niaid.nih.gov/david/ease.htm
DAVID-
KEGG
Pathway analysis Glynn et al. (2003)
http://david.niaid.nih.gov

Table 4. Interferon induced or related genes and their corresponding hybrid clusters
(C/Cs)
152 Kelemen & Liang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
integrate various methods and the derivations of their results. Through the developed
hybrid methods we can increase the power and highlight the quality of the systematic
analysis. Our results demonstrate that this idea is feasible and the resulting I/C clusters
demonstrate that this hybrid method can be successfully achieved. The most novel
innovation in this chapter is the successful clustering of the clusters themselves (hybrid
clustering) in order to compare the results of different methods and the improvement of
the power of reproducibility.
The biologist is faced with a dizzying array of choices when it comes to choosing the
proper algorithm for pattern discovery. One approach is to utilize various algorithms and
combine the results by hybrid methods such as clustering the clusters as was done in
this chapter. Afterwards, Web-based resources such as EASE and DAVID-KEGG can be
used to determine the tendency of the biological function of the genes in a discovered
hybrid pattern as was also performed in this chapter. This may aid the determination of
the function of unknown genes based on their inclusion in a given cluster.
This chapter discusses a promising way to find consistent, reproducible, nearly optimal
solutions for genomic data analysis. It is clearly the way future scientists will have to
investigate by putting together most available methods and all of the research results
in order to make progress in a diverse, massively data intensive field, such as genomics
or bioinformatics. Undoubtedly, such analysis is better equipped to discover new,
biologically meaningful associations than separate semi-heuristic selection and appli-
cation from a pool of specialized algorithms, each offering unique features, abilities. Also,
questions can be addressed in a more informative way both during and after the analysis
and could easily lead to insights beyond what is provided by the calculation of a single
method and can significantly contribute to the generalization of study results and lead
to the potential identification of the most promising or urgent research question. It may
permit a more accurate calculation of some parameters such as the minimum sample size
required for future studies. It can provide a more precise estimate of a treatment effect,
and may explain heterogeneity between the results of separate studies. Results show that
our analysis show lower false positive rate, higher power, and more stable results than
most existing approaches.
References
Alter, O., Brown, P. O., & Botstein, D. (2000). Singular value decomposition for genome-
wide expression data processing and modeling. Proceedings of the National
Academy of Science, 97, (pp. 10101-10106).
Baldi, P., & Long, A. D. (2001). A bayesian framework for the analysis of microarray
expression data: Regularized t-test and statistical inferences of gene changes.
Bioinformatics, 17(6), 509-519.
Benjamini, Y., & Hochberg, Y. (1995). Controlling false discovery rate: A practical and
powerful approach to multiple testing. Journal of Royal Statistical Society, B 57,
289-300.
Pattern Differentiations and Formulations for Heterogeneous Genomic Data 153
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Dudoit, S., & Fridlyand, J. (2002). A prediction-based resampling method to estimate the
number of clusters in a dataset. Genome Biology, 3, RESEARCH0036.
Eisen, M. B., Spellman, P. T., Brown, P. O., & Botstein, D. (1998). Cluster analysis and
display of genome-wide expression patterns. Proceedings of the National Acad-
emy of Science, 95, (pp. 14863-14868).
Glynn, D., Sherman, B. T., Hosack, D. A., Yang, J., Gao, W., Lane, H. C., et al. (2003).
DAVID: Database for annotation, visualization, and integrated discovery. Genome
Biology, 4, R60.
Golub, T. R., Slonim, D. K., Tamayo, P., Huard, C., Gaasenbeek, M., Mesirov, J. P., et al.
(1999). Molecular classification of cancer: Class discovery and class prediction by
gene expression monitoring. Science, 286, 531-537.
Hastiel, T., Tibshirani, R., Eisen, M. B., Alizadeh, A., Levy, R, Staudt, L., et al. (2000). Gene
shaving as a method for identifying distinct sets of genes with similar expression
patterns. Genome Biology, 1(2), research0003.1-0003.21.
Hosack, D., Hosack, A., Jr., G. D., Sherman, B. T., Lane, H. C., & Lempicki, R. A. (2003).
Identifying biological themes within lists of genes with EASE. Genome Biology,
4, R70.
IFNB Multiple Sclerosis Study Group. (1993). Interferon b-1b is effective in relapsing-
remitting multiple sclerosis. I. Clinical results of a multicenter, randomized double-
blind, placebo-controlled trail. Neurology, 43, 655.
Jacobs, L. D., Cookfair, D. L., et. al. (1996). Intramuscular interferon b-1a for disease
progression in relapsing multiple sclerosis: The Multiple Sclerosis Collaborative
Research Group (MSCRG). Annual Neurology, 39, 285.
Kerr, M., & Churchill, G. (2001). Bootstrapping cluster analysis: Assessing the reliability
of conclusions from microarray experiments. Proceedings of the National Acad-
emy of Science, 97, (pp. 8961-8965).
Khan, J., Wei, J. S., Ringner, M., Saal, L. H., Ladanyi, M., Westermann, F., et al. (2001).
Classification and diagnostic prediction of cancers using gene expression profiling
and artificial neural networks. Nature Medicine, 7(6), 673-679.
McShane, L. M., Radmacher, M. D., Friedlin, B., Yu, R., Li, M. C., & Simon, R. (2002).
Methods for assessing reproducibility of clustering patterns observed in analyses
of microarray data. Bioinformatics, 18, 1462-1469.
Neal, S. H., Madhusmita, M., Holter, N. S., Mitra, M., Maritan, A., Cieplak, M., et al. (2000).
Fundamental patterns underlying gene expression profiles: Simplicity from
complexicity. Proceedings of the National Academy of Science, 97, 8409-8414.
Ramanathan, M., Weinstock-Guttman, B., Nguyen, L. T., Badgett, D., Miller, C., Patrick,
K., et al. (2001). Gene expression revealed using DNA arrays: The pattern in multiple
sclerosis patients compared with normal subjects. Journal of Neuroimmunology,
116(2), 213-219.
Ramoni, M., Sebastiani, P., & Kohane, I. (2002). Cluster analysis of gene expression
dynamics. Proceedings of the National Academy of Science, 99, 9121- 9126.
Slonim, D. K. (2002). From patterns to pathways: Gene expression data analysis comes
of age. Nature genetics supplement, 32, 502.
154 Kelemen & Liang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Steinman, L., & Zamvil, S. (2003). Transcriptional analysis of targets in multiple sclerosis,
Nature Review Immunology, June 3(6), 483-92.
Storey, J. D., & Tibshirani, R. (2003). Statistical significance for genomewide studies.
Proceedings of the National Academy of Science, 100(16), 9440-9445.
Tamayo, T., Slonim, D., Mesirov, J., Zhu, Q., Kitareewan, S., Dmitrovsky, E., Lander, E.
S., & Golub, T. R. (1999). Interpreting patterns of gene expression with self-
organizing maps: Methods and application to hematopoietic differentiation. Pro-
ceedings of the National Academy of Science, 96, (pp. 2907-2912).
Tibshirani, R., Walter, G., & Hastie, T. (2001). Estimating the number of clusters in a data
set via the gap statistic. Journal of Royal Statistical Society, B63, (pp. 411-423).
Tusher, V. G., Tibshirani, R., & Chu, G. (2001). Significance analysis of microarrays
applied to the ioniaing radiation response. Proceedings of the National Academy
of Science, 9, 5116-5121.
Wall, M. E., Dyck, P. A., & Brettin, T. S. (2001). SVDMAN - Singular value decomposition
analysis of microarray data. Bioinformatics, 17, 566-568.
Weinstock-Guttman, et al. (2003). Genomic effects of IFN-beta in multiple sclerosis
patients. Journal of Immunology, 171, 2694-702.
West, M. (2000). Bayesian regression analysis in the Large p, Small n paradigm
(Technical Report 00-22). Institute of Statistics and Decision Sciences, Duke
University, CSE-2000-08-01.
Westfall, P., & Young, S. (1993). Resampling-based multiple testing. Wiley-Interscience.
Wuju, L., & Momiao, X. T. (2002). Class: Tumor classification system based on gene
expression profile. Bioinformatics, 18, 325-326.
Yeung, K. Y., & Ruzzo, W. L. (2001). Principal component analysis for clustering gene
expression data. Bioinformatics, 17, 763-774.
Parameterless Clustering Techniques for Gene Expression Analysis 155
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Chapter IX
Parameterless
Clustering Techniques
for Gene Expression
Analysis
Vincent S. Tseng, National Cheng Kung University, Taiwan
Ching-Pin Kao, National Cheng Kung University, Taiwan
Abstract
In recent years, clustering analysis has even become a valuable and useful tool for in-
silico analysis of microarray or gene expression data. Although a number of clustering
methods have been proposed, they are confronted with difficulties in meeting the
requirements of automation, high quality, and high efficiency at the same time. In this
chapter, we discuss the issue of parameterless clustering technique for gene expression
analysis. We introduce two novel, parameterless and efficient clustering methods that
fit for analysis of gene expression data. The unique feature of our methods is they
incorporate the validation techniques into the clustering process so that high quality
results can be obtained. Through experimental evaluation, these methods are shown
to outperform other clustering methods greatly in terms of clustering quality, efficiency,
and automation on both of synthetic and real data sets.
156 Tseng & Kao
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Introduction
Clustering analysis has been applied in a wide variety of fields such as biology, medicine,
psychology, economics, sociology, and astrophysics. The main goal of clustering
analysis is to partition a given set of objects into homogeneous groups based on their
features such that objects within a group are more similar to each other and more different
from those in other groups (Chen, Han, & Yu, 1996). Clustering groups the genes into
biologically relevant clusters with similar expression patterns. The genes clustered
together tend to be functionally related, hence clustering can reveal the co-expression
of genes which were previously uncharacterized. In recent years, clustering analysis has
even become a valuable and useful tool for in-silico analysis of microarray or gene
expression data (Eisen et al., 1998; Alon et al., 1999; Ben-Dor & Yakhini, 1999; Tamayo
et al., 1999). For example, Eisen et al. (1998) applied a variant of hierarchical clustering
to identify groups of co-expressed yeast genes. Alon et al. (1999) used a two-way
clustering technique to detect groups of correlated genes and tissues. Self-organizing
maps were used by Tamayo et al. (1999) to identify clusters in the yeast cell cycle data
set and human hematopoietic differentiation data set.
Thoroughly and extensive overviews of clustering algorithms are given by Aldenderfer
and Blashfield (1984) and Jain and Dubes (1998). Although a number of clustering
methods have been studied in the literature (Carpenter & Grossberg, 1987; Kohonen,
1990; Zhang, Ramakrishnan, & Livny, 1996; Chen, Han, & Yu, 1996; Guha, Rastogi, &
Shim, 1998; Guha, Rastogi, & Shim, 1999), they are not satisfactory in terms of: (1)
automation, (2) quality, and (3) efficiency. First, most clustering algorithms request users
to specify some parameters. In real applications, however, it is hard for biologists to
determine the suitable parameters manually. Thus an automated clustering method is
required. Second, most clustering algorithms aim to produce the clustering results based
on the input parameters and their own criterions. Hence, they are incapable of producing
optimal clustering result. Third, the existing clustering algorithms may not perform well
when the optimal or near-optimal clustering result is enforced from the universal
criterions.
On the other hand, a variety of clustering validation measures are applied to evaluate the
validity of the clustering results, the suitability of parameters, or the reliability of
clustering algorithms. A good overview of clustering validation can be found in the book
Cluster Analysis (Aldenderfer & Blashfield, 1984), in which numerous validation index
are considered, like DB-index (Davies & Bouldin, 1979; Aldenderfer & Blashfield, 1984),
Simple matching coefficient (Aldenderfer & Blashfield, 1984; Jain & Dubes, 1998),
Jaccard coefficient (Aldenderfer & Blashfield, 1984; Jain & Dubes, 1998), Huberts
statistic (Aldenderfer & Blashfield, 1984; Jain & Dubes, 1998), etc. There also exist
several other measures, like ANOVA (Kerr & Churchill, 2001), FOM (Yeung, Haynor, &
Ruzzo, 2001), VCV (Hathaway & Bezdek, 2003). Nevertheless, the roles of them are placed
only on the phase of post-validation. The study on how to integrate validation
techniques with clustering methods tightly for improving the clustering quality has been
absent.
Parameterless Clustering Techniques for Gene Expression Analysis 157
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
In this chapter, we will describe new clustering methods that are integrated with
validation techniques. We introduce two novel, parameterless and efficient clustering
methods that are fit for analysis of gene expression data. The methods determine the best
grouping of genes on-the-fly without any user-input parameter.
Background
The related work of clustering analysis can be divided into three types (Aldenderfer &
Blashfield, 1984). The first type is similarity measurements that will affect the final
clustering results directly. Euclidean distance and correlation coefficient are most
popular similarity measures. Surely this is an important task for clustering analysis, but
it is not the focus of this chapter. The second type is clustering methods that are the core
of clustering analysis and have received extensive attentions. The third type is validation
techniques that are applied to evaluate the validity of the clustering results, the
suitability of parameters, or the reliability of clustering algorithms. In addition, all
validation techniques may be roughly divided into two main types: scalar measurements
and intensity image methods (Hathaway & Bezdek, 2003). Clustering methods and
validation techniques are introduced briefly in the following section.
Clustering Methods
The most well-known clustering method is probably the k-means algorithm (Aldenderfer
& Blashfield, 1984; Jain & Dubes, 1998), which is a partitioning method. The k-means
algorithm partitions the dataset into k groups, based primarily on the distance between
data items, where k is a parameter specified by the user. Hierarchical methods, for example,
UPGMA (Rohlf, 1970), BIRCH (Zhang, Ramakrishnan, & Livny, 1996), CURE (Guha,
Rastogi, & Shim, 1998), and ROCK (Guha, Rastogi, & Shim, 1999), are another popular
kind of clustering methods. Agglomerative or divisive are used to represent distance
matrices as ultrametric trees. In recent years, artificial neural networks, such as self-
organizing maps (SOM) (Kohonen, 1990), competitive learning networks (Rumelhart &
Zipser, 1985), and adaptive resonance theory (ART) networks (Carpenter & Grossberg,
1987), have also been used for clustering.
Cluster affinity search technique (CAST) (Ben-Dor & Yakhini, 1999) takes as input a
parameter called the affinity threshold t, where 0 < t < 1, and tries to guarantee that the
average similarity in each generated cluster is higher than the threshold t. CAST
generates one cluster at a time and selects the object with the most neighbors as a seed
for the current cluster. It adds un-clustered objects with high affinity (the average
similarity between the object and the cluster is greater than t) to the current cluster and
removes objects with low affinity from the current cluster iteratively. The main advantage
of CAST is that it can detect outliers more effectively.
158 Tseng & Kao
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Scalar Measurements
Jain and Dubes (1998) divided the cluster validation procedure into two main parts:
external and internal criterion analysis. External indices are used to measure the extent
to which cluster labels (a clustering result) match externally supplied class labels (a
partition that is known a priori). There are many statistical measures that assess the
agreement between an external criterion and a clustering result. For example, Milligan,
Soon, and Sokol (1983) and Milligan and Cooper (1986) evaluated the performance of
different clustering algorithms and different statistical measures of agreement for both
synthetic and real data. The problem of external criterion analysis is that reliable external
criteria are rarely available when gene expression data are analyzed. In contrast, internal
indices are used to measure the goodness of a clustering structure without respect to
external information. For example, compactness and isolation of clusters are possible
measures of goodness of fit. A measure called the Figure of Merit (FOM) was used by
Yeung, Haynor, and Ruzzo (2001) to evaluate the quality of clustering performed on a
number of real datasets.
We introduce an internal index, namely Huberts (gamma) statistic (Aldenderfer &
Blashfield, 1984; Jain & Dubes, 1998) and two external indices, namely Simple matching
coefficient and Jaccard coefficient (Aldenderfer & Blashfield, 1984; Jain & Dubes, 1998).
The definition of Huberts statistic is as follows:
Let X=[X(i, j)] and Y=[Y(i, j)] be two n n proximity matrices on the same n genes. From
the viewpoint of correlation coefficient, X(i, j) indicates the observed correlation
coefficient of genes i and j and Y(i, j) is defined as follows:

otherwise. 0
cluster, same in the clustered are and genes if 1
) , (
j i
j i Y
(1)
The Huberts statistic represents the point serial correlation between the matrices X
and Y, and is defined as follows when the two matrices are symmetric:

+
(
(
,
\
,
,
(
j

(
(
,
\
,
,
(
j


1
1 1
) , ( ) , ( 1
n
i
n
i j
Y X
Y j i Y X j i X
M
, (2)
where M = n (n - 1) / 2 is the number of entries in the double sum, and
X
and
Y
denote
the sample standard deviations while X and Y denote the sample means of the entries
of matrices X and Y.
From the viewpoint of distance, X(i, j) indicates the observed distance of genes i and j
and Y(i, j) is defined as (1) by exchange the 1 and 0. The value of is between [-1,
1] and a higher value of represents the better clustering quality.
Parameterless Clustering Techniques for Gene Expression Analysis 159
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Figure 1 illustrates the concept of Huberts statistic. For matrices X and Yas described
above, the solid lines represent the original coordinate and the dotted lines represent the
coordinate centered by ( , X Y). A point A
i,j
that falls in quadrant I of dotted coordinate
indicate that genes i and j have higher similarity and are clustered in the same cluster.
A point B
i,j
that falls in quadrant II indicate that genes i and j have lower similarity but
are clustered in the same cluster. In contrast, a point C
i,j
that falls in quadrant III of dotted
coordinate indicate that genes i and j have lower similarity and are clustered in different
clusters. A point D
i,j
that falls in quadrant IV indicate that genes i and j have higher
similarity but are clustered in different clusters. In brief, the more points that fall in
quadrants I and III, the better the quality of clustering results are. Therefore, the point
serial correlation between the matrices X and Y can be used to measure the quality of
clustering results.
The definitions of Simple matching coefficient and Jaccard coefficient (Aldenderfer &
Blashfield, 1984; Jain & Dubes, 1998) are as follows:
Let U=[U(i, j)] and V=[V(i, j)] be two n n binary matrices on the same n genes. The matrices
indicate two distinct clustering results and are defined as follows:
Figure 1. The concept of Huberts statistic
Table 1. Association table
Matrix U
1 0
1 a b
Matrix V
0 c d

160 Tseng & Kao
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.

=
otherwise 0
cluster, same in the clustered are and genes if 1
) , (
j i
j i G , G=U, V.
(3)
Let us consider the contingency table of the matrices U and V as shown in Table 1. Where
a indicates the number of entries on which U and V both have values 1 and b indicates
the number of entries on which U have values 1 and V have values 0, and so on.
The Simple matching coefficient represents the point serial correlation between the
matrices X and Y, and is defined as follows when the two matrices are symmetric:
The Simple matching coefficient is defined by:
d c b a
d a
S
+ + +
+
= .

(4)
that is total number of matching entries divided by total number of entries. The Jaccard
coefficient is defined by:
c b a
d a
S
+ +
+
= . (5)
which is similar to the Simple matching coefficient, only with negative matches (d) are
ignored. The Simple matching coefficient usually varies over a smaller range than the
Jaccard coefficient because negative matches are often a dominant factor.
Intensity Image Methods
One main problem of scalar validity measures is that representing the accuracy of
different clustering methods by a single real number may lose much information.
Accordingly, intensity image methods use all of the information generated by the
clustering methods to produce an nn gray scale intensity image. It converts the
similarity (or dissimilarity) into gray scale and reorders the data of similarity matrix to
display the clustering result. For instance, VCV (Hathaway & Bezdek, 2003) approach
retains and organizes the information that is lost through the massive aggregation of
information by scalar measures.
Parameterless Clustering Techniques for Gene Expression Analysis 161
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Smart-Cast and Correlation Search
Technique (CST)
In this section, we will present the ideas and algorithms of our novel and parameterless
clustering methods, namely Smart-CAST (Tseng & Kao, 2005) and Correlation Search
Technique (CST) (Tseng & Kao, 2005). Before the methods can be applied, we have to
produce a similarity matrix S based on the original dataset. The matrix S stores the degree
of similarity between each pair of genes in the dataset, and the range of the degree is [0,
1]. This much reduces the computation overhead incurred by some clustering algorithms
that calculate the similarities dynamically. You may obtain the similarity by using any
similarity measurements (e.g., Euclidean distance, Pearsons correlation coefficient, etc.)
for various purposes. Then, Smart-CAST and CST can automatically cluster the genes
according to the similarity matrix S without any user input parameters.
Tseng & Chen (2002) evaluated a number of validity indexes for measuring the quality
of clustering results. They concluded from the experimental results that Huberts
(gamma) statistic (Aldenderfer & Blashfield, 1984; Jain & Dubes, 1998) might be the best
index for both partition-based clustering methods and density-based methods for both
low-similarity and high-similarity data. So we use the validation index, namely Huberts
statistic, as the key measure of clustering quality during the clustering process.
Smart-CAST
Preview of Smart-CAST
The main ideas of the Smart-CAST method are as follows. In the first step, a density-and-
affinity based algorithm is applied as the basic clustering algorithm. Along with a
specified input parameter, the basic clustering algorithm utilizes the similarity matrix S
Figure 2. Huberts statistic vs. values of t

0
0.1
0.2
0.3
0.4
0.5
0.6
5 15 25 35 45 55 65 75 85 95
t (%)

s
t
a
t
i
s
t
i
c
162 Tseng & Kao
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
to conduct the clustering task. Thus, a clustering result will be produced by the basic
clustering algorithm based on the given input parameter. A good candidate for the base
clustering algorithm is CAST (Ben-Dor & Yakhini, 1999), which needs only one input
parameter, called the affinity threshold t, where 0 < t < 1. The reasons why we prefer to
use CAST as the base clustering method are as follows:
1. CAST has the capability of isolating outliers.
2. Unlike other algorithms that are unstable (e.g., k-means), CAST is a stable
algorithm.
3. The execution time of CAST is shorter than most of the other clustering algorithms.
In the second step, a validation test is performed to evaluate the quality of the clustering
result produced in step two. We adopt Huberts statistic (Aldenderfer & Blashfield,
1984; Jain & Dubes, 1998) to measure the quality of clustering, because Tseng & Chen
(2002) concluded from the experimental results that Huberts might be the best index.
With the above steps, it is clear that high quality clustering can be achieved by applying
a number of different values of the affinity threshold t as input parameters to the CAST
algorithm, by calculating the Huberts statistic of each clustering result, and by
choosing the one with the highest value of the Huberts statistic as the output. In this
way, a local-optimal clustering result can be obtained by users automatically. An example
is shown in Figure 2, where the X axis represents the values of the affinity threshold t input
to CAST and the Y axis shows the obtained Huberts statistic for each of the clustering
results. The peak in the curve corresponds to the best clustering result, which has a
statistic value of around 0.52 when t is set to be 0.25.
This approach is feasible in that firstly, CAST executes very quickly since the similarity
matrix of gene expressions is obtained in advance; secondly, the Huberts statistic for
each clustering result can be calculated easily. However, one problem encountered is
how to determine suitable values of the affinity threshold t. The easiest way is to
increment the value of the affinity threshold t with a fixed interval. For example, we may
increase the value of t from 0.05 to 0.95 in increments of 0.05. We call this approach CAST-
FI (Fixed Increment) in the following. The main disadvantage of CAST-FI is that many
iterations of computations are required. Therefore, a new method is introduced in the next
section to reduce the computation overhead.
Computation Reduction of Smart-CST
The idea behind the method is to reduce the amount of computation by eliminating
unnecessary executions of clustering so as to obtain a nearly-optimal clustering result
instead of the optimal one. That is, we try to execute CAST as few times as possible.
Parameterless Clustering Techniques for Gene Expression Analysis 163
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Therefore, we need to narrow down the range of the parameter affinity threshold t
effectively. The method works as follows:
1. Initially, a testing range R for setting the affinity threshold t is set to be [0, 1]. We
divide R equally into m parts with the points P
1
, P
2
,, P
m-1
, where P
1
< P
2
< < P
m-
1
, m 3. Then, the value of each P
i
is taken as the affinity threshold t for executing
CAST, and the G statistic of the clustering result for each of P
i
is calculated. We
call this process a run.
2. When a clustering run is completed, the clustering at point P
b
that produces the
highest G statistic is considered to be the best clustering. The testing range R is
then replaced by the range [P
b-1
, P
b+1
] that contains the point P
b
.
3. The above process is repeated until the testing range R is smaller than a threshold
or the difference between the maximal and minimal quality values is smaller than
another threshold .
4. The best quality clustering result obtained through the tested process is output
as the answer.
In this way, we can obtain a clustering result that has nearly-optimal clustering quality
with much less computation. In the next section, through empirical evaluation, we shall
evaluate how good the generated clustering result is and to what extent the amount of
computation could be reduced by our approach.
Correlation Search Technique (CST)
Preview of CST
The main idea behind CST is to integrate clustering method with validation technique so
that it can cluster the genes quickly and automatically. Since validation technique is
embedded, CST can obtain a near-optimal clustering result.
The validation index we used is Huberts statistic (Aldenderfer & Blashfield, 1984; Jain
& Dubes, 1998). Since the computing of statistic may spend considerable time, we
simplify it to enhance the performance of our approach. We expand Huberts statistic
as follows:
2
1
1 1
1
1 1
2
2
1
1 1
1
1 1
2
1
1 1
1
1 1
1
1 1
) , ( ) , ( ) , ( ) , (
) , ( ) , ( ) , ( ) , (

= + =

= + =

= + =

= + =

= + =

= + =

= + =
n
i
n
i j
n
i
n
i j
n
i
n
i j
n
i
n
i j
n
i
n
i j
n
i
n
i j
n
i
n
i j
j i Y j i Y M j i X j i X M
j i Y j i X j i Y j i X M
.
(6)
164 Tseng & Kao
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
While we decide which clustering result is better, the measurement of quality of
clustering result can be simplified as follows:
2
1
1 1
1
1 1
1
1 1
1
1 1
1
1 1
) , ( ) , (
) , ( ) , ( ) , ( ) , (

= + =

= + =

= + =

= + =

= + =
n
i
n
i j
n
i
n
i j
n
i
n
i j
n
i
n
i j
n
i
n
i j
j i Y j i Y M
j i Y j i X j i Y j i X M
.

(7)
We call (7) as simplistic Huberts statistic. For more details, readers are referred to the
literature by Tseng and Kao (2005).
CST Algorithm
The input of CST is a symmetric similarity matrix X, where X(i, j) [0, 1]. CST is a greedy
algorithm that constructs clusters one at a time, and the currently constructed cluster is
denoted by C
open
. Each cluster is started by a seed and is constructed incrementally by
adding (or removing) elements to (or from) C
open
one at a time. The temporary clustering
result of each addition (or removal) of x is computed by simplistic Huberts statistic,
i.e. (7), and is denoted by
add
(x) (or
remove
(x)). In addition, the currently maximum of
simplistic Huberts statistic is denoted by
max
. We say that an element x has high
positive correlation if
add
(x)
max
, and x has high negative correlation if
remove
(x)

max
. CST takes turns between adding high positive correlation elements to C
open
, and
removing high negative correlation elements from it. When C
open
is stabilized by addition
and removal procedure, this cluster is complete and next one is started.
The addition and removal procedure strengthen the quality of C
open
gradually. Moreover,
the removal procedure exterminates the cluster members that were inaccurately added at
early clustering stages. In addition, a heuristics is added to CST for selecting an element
with the maximum number of neighbors to start a new cluster.
To save computation time, we simplify the simplistic Huberts statistic, i.e. (7), further.
While we decide which element will be added to (or removed from) C
open
, the measurement
" of effect of each added (or removed) element is:

= + =
=
1
1 1
) , ( ) , (
n
i
n
i j
j i Y j i X
(8)
Parameterless Clustering Techniques for Gene Expression Analysis 165
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
More details on the CST algorithm can be found in the literature by Tseng & Kao (2005).
A pseudo-code of CST is shown in Figure 3. The subroutine MaxValidaty() computes
the possible maximum of measurement yet, i.e. (7), while a certain element is added (or
removed). In the addition stage, it is equal to:
( )
2
open open
open
|) | ( |) | ( *
|) | ( * }) | ) ( max{ ( *
C S C S M
C S S U u u a S M
Y Y
Y X XY
+ +
+ +


(9)
Figure 3. A pseudo-code of CST
Input: An n-by-n similarity matrix :

0. Initialization:
M = n (n - 1) / 2
S
X
=

= + =
1
1 1
) , (
n
i
n
i j
j i X
S
Y
= 0
S
XY
= 0
C = /* The collection of closed clusters */
U = {1, 2, , n} /* Elements not yet assigned to any cluster */

max
= 0
1. while (U ? ) do
C
open
=
a () = 0
1.1. SEED: Pick an element u U with most neighbors
U = U - {u} /* Remove u from U */
For all i U set a(i) = X(u, i) /* Update the affinity*/
C
open
= {u} /* Insert u into C
open
*/
1.2. ADD: while MaxValidaty( ) ?
max
do
Pick an element u U with maximum a()
U = U - {u} /* Remove u from U */
S
Y
= S
Y
+ |C
open
|
S
XY
= S
XY
+ a(u)
For all i U U C
open
set a(i) = a(i) + X(u, i) /* Update the affinity*/
C
open
= C
open
U {u} /* Insert u into C
open
*/

max
= MaxValidaty( )
1.3. REMOVE: while MaxValidaty( ) >
max
do
Pick an element v C
open
with minimum a()
C
open
= C
open
- {v} /* Remove v from C
open
*/
S
Y
= S
Y
- |C
open
|
S
XY
= S
XY
- a(u)
For all i U U C
open
set a(i) = a(i) - X(u, i) /* Update the affinity*/
U = U U {v} /* Insert v into U */

max
= MaxValidaty()
1.4. Repeat steps ADD and REMOVE as long as there are no elements been removed.
1.5. C = C U {C
open
}

end
2. Done, return the collection of cluster, C.

166 Tseng & Kao

Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
For the removal stage, it becomes:
( )
2
open open
open open
) 1 | | ( ) 1 | | ( *
) 1 | | ( * }) | ) ( min{ ( *
+ +
+


C S C S M
C S S C v v a S M
Y Y
Y X XY

(10)
Figure 4. Profiles of the four seed sets: (a) Dataset I, (b) Dataset II, (c) Dataset III, (d)
Dataset IV
-3
-2
-1
0
1
2
3
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
C1 C2 C3

-3
-2
-1
0
1
2
3
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
C1 C2 C3 C4 C5 C6

(a)
(b)

-3
-2
-1
0
1
2
3
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
C1 C2 C3 C4 C5

-3
-2
-1
0
1
2
3
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
C1 C2 C3 C4 C5
C6 C7 C8 C9 C10

(c)
(d)
(a) (b)
(c) (d)
Parameterless Clustering Techniques for Gene Expression Analysis 167
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Empirical Evaluation
We process a series of experiments to evaluate the performance and accuracy of the
Smart-CAST and CST methods. Accordingly, we tested the methods on four synthetic
datasets generated by our correlation-inclined cluster dataset generator. Meanwhile, a
real dataset, namely the yeast cell-cycle dataset by Spellman et al. (1998), was also used
for testing the methods. In the synthetic data generator, the users can set up some
parameters for generating various kinds of gene expression datasets with variations in
terms of the number of clusters, the number of genes in each cluster, the number of
outliers (or noise), etc. Firstly, a number of seed genes are produced by the generator or
input by the user. All the seed genes must have same dimensions (number of conditions)
and very low similarity mutually so as to ensure that the generated clusters will not
overlap. Second, the generator produces clusters one by one. It randomly produces
genes in each cluster, and all produced genes are similar to the seed gene of this cluster
in the viewpoint of correlation. Third, the outliers are produced randomly and added into
the dataset. It should be noted that the outliers may fall in some clusters. In this way, all
genes in the same cluster will have very high similarity and they will have high
dissimilarity with genes in other clusters.
We first generate four seed sets with size three, six, five, and ten, respectively. The seeds
in each set have 15 dimensions and their correlation coefficient is less than 0.1. The
profiles of these four seed sets are shown in Figure 4. Then, we load these four seed sets
into our dataset generator to generate four synthetic gene expression datasets for
testing, respectively, namely Dataset I, Dataset II, Dataset III and Dataset IV. The Dataset
I contains three gene clusters of sizes 900, 700, and 500, with additional 400 outliers. The
Dataset II contains six gene clusters of sizes 500, 450, 400, 300, 250, and 200, with
additional 400 outliers. The Dataset III contains five gene clusters of sizes 1200, 1000,
800, 700, and 500, with 800 additional outliers. The Dataset IV contains ten gene clusters
of sizes 650, 550, 550, 500, 450, 400, 350, 300, 250, and 200, with 800 additional outliers.
Table 2 shows the cluster structures of these four datasets.
For the real dataset, the yeast cell-cycle dataset contains 72 expressions (includes alpha
factor, cdc15, cdc28, and elutriation). Spellman et al. (1998) used hierarchical clustering
and analysis of the 5' regions of the genes to identify 8 groups, namely CLN2 (76 genes),
Table 2. Cluster structure of datasets
Dataset I Dataset II Dataset III Dataset IV
# Clusters 3 6 5 10
Cluster Size
900, 700,
500.
500, 450, 400,
300, 250, 200.
1200, 1000,
800, 700, 500.
650, 550, 550, 500, 450,
400, 350, 300, 250, 200.
# Outliers 400 400 800 800
# Total
Patterns
2500 2500 5000 5000

168 Tseng & Kao
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Y (31 genes), Histone (9 genes), MET (20 genes), CLB2 (35 genes), MCM (34 genes), SIC1
(27 genes), and MAT (13 genes). Since several groups are similar in terms of the
expression profiles, the 8 groups are further grouped into 4 main clusters according to
the hierarchical structure of Spellmans analysis, namely G1 (CLN2 and Y groups), S/G2
(Histone and MET groups), M (CLB2 and MCM groups), and M/G1 (SCI1 and MAT
groups). Since we would like to cluster the genes with main considerations on the
expression profiles, we take the G1, S/G2, M, and M/G1 groups as the standard for the
clustering results.
We compare the Smart-CAST and CST methods with the well-known clustering method,
namely k-means (Aldenderfer & Blashfield, 1984; Jain & Dubes, 1998), and the CAST-FI
methods we introduced before. For k-means, the value of k is varied from 2 to 21 in step
of 1, and from 2 to 41 in step of 1, respectively. For CAST-FI, the value of affinity threshold
t is varied from 0.05 to 0.95 in fixed increment of 0.05. Furthermore, the parameters m,
and were set to default values of 4, 0.01 and 0.01, respectively The quality of clustering
results was measured by using Huberts statistic, Simple matching coefficient
(Aldenderfer & Blashfield, 1984; Jain & Dubes, 1998), and Jaccard coefficient (Aldenderfer
& Blashfield, 1984; Jain & Dubes, 1998). Simple matching coefficient and Jaccard
coefficient can evaluate the quality of clustering results in pre-clustering dataset. We
also use intensity image to exhibit the visualized results of the clustering methods. Since
correlation coefficient is widely adopted in most studies on gene expression analysis,
we evaluate only correlation coefficient for the similarity measure.
Simulated Datasets
Table 3 presents the total execution time and the best clustering quality of the tested
methods on Dataset I. The notation M indicates the number of main clusters, and we
consider clusters whose sizes are smaller than 50 and 5 as outliers for simulated datasets
and the yeast cell-cycle dataset, respectively.
It is observed that CST, Smart-CAST and CAST-FI outperform k-means substantially in
both of execution time and clustering quality, i.e., Huberts statistic, Simple matching
coefficient, and Jaccard coefficient. In particular, CST performs 396 times to 1638 times
faster on Dataset I than k-means with k ranged from 2 to 21 and 2 to 41, respectively. In
Table 3. Experimental results (Dataset I)
Methods Time (s) # Clusters Statistic
Matching
Coefficient
Jaccard
Coefficient
CST < 1 65 (M=3) 0.800 0.981 0.926
Smart-CAST 12 85 (M=3) 0.800 0.986 0.944
CAST-FI 54 91 (M=3) 0.799 0.986 0.945
k-means (k=2-21) 396 6 (M=6) 0.456 0.825 0.427
k-means (k=2-41) 1638 6 (M=6) 0.456 0.825 0.427
Parameterless Clustering Techniques for Gene Expression Analysis 169
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
addition, the results also show that the clustering quality generated by CST is very close
to that of Smart-CAST and CAST-FI in all of Huberts statistic, Simple matching
coefficient, and Jaccard coefficient. It means that the clustering quality of our approach
is as good as Smart-CAST and CAST-FI even though the computation time of our
approach is reduced substantially.
Table 3 also shows that k-means produced six main clusters as the best clustering result
for Dataset I, while the sizes of clusters are 580, 547, 457, 450, 352, and 114. This is true
no matter k is varied from 2 to 21 and from 2 to 41 and does not match the original cluster
structure. In contrast, CST produces 65 clusters as the best clustering result. In
particular, it is clear that three main clusters are generated, with sizes 912, 717, and 503.
This tightly matches the original cluster structure.
The intensity images of the original cluster structure and the clustering results for
Dataset I are shown in Figure 5. From Figure 5(a), we easily observe that there are three
Figure 5. Intensity images on Dataset I: (a) The prior cluster structure, (b) the actual
input to the algorithms, (c) clustering results of CST, (d) clustering results of CAST

(a)
(b)


(c)
(d)
(a) (b)
(c) (d)
170 Tseng & Kao
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
main clusters and quite a number of outliers in Dataset I. It is observed that Figure 5(c)
and Figure 5(d) are very similar to Figure 5(a), meaning that the clustering results of CST
and Smart-CAST are very similar to the original cluster structure in Dataset I.
It is obvious that the observations on Dataset II, Dataset III and Dataset IV are similar
to that on Dataset I. For more detail, readers are referred to the literature by Tseng & Kao
(2005).
The Yeast Cell-Cycle Dataset
Table 4 presents the total execution time and the best clustering quality of the tested
methods on the real dataset, i.e., the yeast cell-cycle dataset as described previously. The
CST method produced four main clusters with one outlier, while other methods generated
five main clusters. It is observed that CST, Smart-CAST and CAST-FI outperform k-
Table 4. Experimental results (Yeast dataset)
Methods Time (s) # Clusters Statistic
Matching
Coefficient
Jaccard
Coefficient
CST < 1 5 (M=4) 0.703 0.983 0.941
Smart-CAST < 1 5 (M=5) 0.706 0.978 0.922
CAST-FI 3 5 (M=5) 0.705 0.973 0.908
k-means (k=2-21) 75 5 (M=5) 0.683 0.914 0.724
k-means (k=2-41) 325 5 (M=5) 0.683 0.914 0.724
Figure 6. Intensity images of the yeast dataset: (a) The original cluster structure, (b)
Clustering results of CST

(a) (b)
Parameterless Clustering Techniques for Gene Expression Analysis 171
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
means substantially in execution time and clustering quality. In particular, CST delivers
the best execution time and clustering quality. Specifically, CST performs 75 to 325 times
faster than k-means with k ranged as [2, 21] and [2, 41], respectively.
To investigate the clustering quality, Figure 6 shows the intensity images for the original
cluster structure of the dataset and the clustering results. From Figure 6(a), we can easily
observe that there are four main clusters in the yeast dataset. It is observed that Figure
6(b) is very similar to Figure 6(a), meaning that the clustering result generated by CST
is very close to the original cluster structure. In fact, CST produced four main clusters,
namely G1, S/G2, M and M/G1 clusters, with one outlier named SRD1. This matches the
actual structure of this yeast dataset, as described previously. In contrast, Smart-CAST,
CAST-FI, and k-means produced five main clusters as the best clustering results, which
is kind of deviated from the actual structure of the dataset. Hence, it is shown that CST
outperforms other methods substantially in terms of execution time and clustering
quality on the real yeast cell-cycle dataset.
Summary
Clustering analysis is a valuable and useful technique for in-silico analysis of microarray
or gene expression data, but most clustering methods incur problems in the aspects of
automation, quality, and efficiency. In this chapter, we explore the effective integration
between clustering methods and validation techniques. We introduce two novel,
parameterless, and efficient clustering methods, namely Smart-CAST and CST, that are
fit for analysis of gene expression data.
The primary characteristics of the introduced methods are in as follows: First, the Smart-
CAST and CST methods are parameterless clustering algorithms. Secondly, the quality
of clustering results generated by the introduced methods is nearly optimal. Through
performance evaluation on synthetic and real gene expression datasets, Smart-CAST and
CST methods are shown to achieve higher efficiency and clustering quality than other
methods without requesting the users to input any parameter. Therefore, the methods
can provide high degree of automation, efficiency and clustering quality, which are
lacked in other clustering methods for gene expression mining.
In the future, the following issues may be further explored:
1. Reduction of memory requirements: Smart-CAST and CST need memory to store
the similarity matrix. Hence, a memory-efficient and memory reduction scheme is
needed, especially when the number of genes in the microarry is large.
2. Applications of CST on more real microarray datasets: Everyone can apply Smart-
CAST and CST on real microarray datasets to evaluate the validity of the clustering
results, with the aim to promote related bioinformatics research like disease marker
discovery.
172 Tseng & Kao
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
References
Aldenderfer, M. S., & Blashfield, R. K. (1984). Cluster analysis. Newbury Park, CA: Sage
Publications.
Alon, U., Barkai, N., Notterman, D. A., Gish, K., Ybarra, S., Mack, D., & Levine, A. J. (1999).
Broad patterns of gene expression revealed by clustering analysis of tumor and
normal colon tissues probed by oligonucleotide arrays. Proceedings of National
Academy of Science, 96(12), 6745-6750.
Ben-Dor, A., & Yakhini, Z. (1999). Clustering gene expression patterns. Journal of
Computational Biology, 6(3-4), 281-297.
Carpenter, G. A., & Grossberg, S. (1987). A massive parallel architecture for a self-
organizing neural pattern recognition machine. Computer Vision, Graphics, and
Image Processing, 37, 54-115.
Chen, M. S., Han, J., & Yu, P. S. (1996). Data mining: An overview from a database
perspective. IEEE Transactions on Knowledge and Data Engineering, 8(6), 866-
883.
Davies, D. L., & Bouldin, D. W. (1979). A cluster separation measure. IEEE Transactions
on Pattern Analysis and Machine Intelligence, 1(2), 224-227.
Eisen, M. B., Spellman, P. T., Brown, P. O., & Botstein, D. (1998). Clustering analysis and
display of genome wide expression patterns. Proceedings of the National Acad-
emy of Sciences, (Vol. 95, pp. 14863-14868).
Guha, S., Rastogi, R., & Shim, K. (1998). CURE: An efficient clustering algorithm for large
databases. Proceedings of the 1998 ACM SIGMOD International Conference on
Management of Data, (pp. 73-84).
Guha, S., Rastogi, R., & Shim, K. (1999). ROCK: A robust clustering algorithm for
categorical attributes. Proceedings of the 15th International Conference on Data
Engineering, Australia, (pp. 512-521).
Hathaway, R. J., & Bezdek, J. C. (2003). Visual cluster validity for prototype generator
clustering models. Pattern Recognition Letters, 24(9-10), 1563-1569.
Jain, A. K., & Dubes, R. C. (1998). Algorithms for clustering data. Englewood Cliffs, NJ:
Prentice Hall.
Kerr, M. K., & Churchill, G. A. (2001). Churchill. Bootstrapping cluster analysis: Assess-
ing the reliability of conclusions from microarray experiments. Proceedings of
National Academy of Science, 98(16), 8961-8965.
Kohonen, T. (1990). The self-organizing map. Proceedings of the IEEE, 78(9), 1464-1479.
Milligan, G. W., & Cooper, M. C. (1986). A study of the comparability of external criteria
for hierarchical cluster analysis. Multivariate Behavioral Research, 21, 441-458.
Milligan, G. W., Soon, S. C., & Sokol, L. M. (1983). The effect of cluster size, dimension-
ality and the number of clusters on recovery of true cluster structure. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 5, 40-47.
Parameterless Clustering Techniques for Gene Expression Analysis 173
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Rohlf, F. J. (1970). Adaptive hierarchical clustering schemes. Systematic Zoology, 19, 58-
82.
Rumelhart, D. E., & Zipser, D. (1985). Feature discovery by competitive learning.
Cognitive Science, 9, 75-112.
Spellman, P. T., Sherlock, G., Zhang, M. Q., Iyer, V. R., Anders, K., Eisen, M. B., et al.
(1998). Comprehensive identification of cell cycle-regulated genes of the yeast
Saccharomyces Cerevisiae by Microarray Hybridization. Molecular Biology of the
Cell, 9(12), 3273-3297.
Tamayo, P., Slonim, D., Mesirov, J., Zhu, Q., Kitareewan, S., Dmitrovsky, E., et al. (1999).
Interpreting patterns of gene expression with self-organizing maps: Methods and
application to hematopoietic differentiation. Proceedings of National Academy of
Science, 96(6), 2907-2912.
Tseng, S. M., & Chen, L. J. (2002). An empirical study of the validity of gene expression
clustering. Proceedings of the 2002 International Conference on Mathematics
and Engineering Techniques in Medicine and Biological Sciences (METMBS02),
USA.
Tseng, S. M., & Kao, C. P. (2005). Mining and validating gene expression patterns: An
integrated approach and applications. Informatica, 27, 21-27.
Tseng, V. S. (n.d.). Parameter-less clustering method. IEEE/ACM Transactions on
Computational Biology and Bioinformatics (TCBB).
Yeung, K. Y., Haynor, D. R., & Ruzzo, W. L. (2001). Validating clustering for gene
expression data. Bioinformatics, 17(4), 309-318.
Zhang, T., Ramakrishnan, R., & Livny, M. (1996). BIRCH: An efficient data clustering
method for very large databases. Proceedings of the 1998 ACM SIGMOD Inter-
national Conference on Management of Data, (pp. 103-114).
174 Zhang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Chapter X
Joint Discriminatory
Gene Selection for
Molecular Classification
of Cancer
Junying Zhang, Xidian University, China
Abstract
This chapter introduces gene selection approaches in microarray data analysis for two
purposes: cancer classification and tissue heterogeneity correction and hence is
divided into two respective parts. In the first part, we search for jointly discriminatory
genes which are most responsible to classification of tissue samples for diagnosis. In
the second part, we study tissue heterogeneity correction techniques, in which
independent component analysis is applied to tissue samples with the expression levels
of only selected genes, the genes which are functionally independent and/or jointly
discriminatory; we also employ non-negative matrix factorization (NMF) to
computationally decompose molecular signatures based on the fact that the expression
values in microarray profiling are non-negative. Throughout the chapter, a real world
gene expression profile data was used for experiments, which consists of 88 tissue
samples of 2308 effective gene expressions obtained from 88 patients of 4 different
neuroblastoma and non-hodgkin lymphoma cell tumors.
Gene Selection for Molecular Classification of Cancer 175
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Introduction
Spotted cDNA microarray, along with recent advances in machine learning and pattern
recognition methods, promise cost effective and powerful new tools for the large-scale
analysis of gene expressions. Still, the biological and medical research advances present
more challenging topics to science and engineering. The most interesting ones among
them include discriminatory gene selection, tissue heterogeneity correction, tissue and
or gene clustering and tissue classification for cancer diagnosis.
As a common feature in microarray profiling, gene expression profiles represent a
composite of more than one distinct source (i.e., the observed signal intensity will consist
of the weighted sum of activities of the various sources). For example, molecular analysis
of cells in their native tissue environment provides the most accurate picture of the in
vivo disease state. However, tissues are complicated three-dimensional structures and
the cell subpopulation of interest (e.g., malignant cells in tumor) may constitute only a
small fraction of the total tissue volume. More specifically, in the case of solid tumors,
the related issue is called partial volume effect, in other words, the heterogeneity within
the tumor samples caused by stromal contamination. Blind application of the microarray
profiling could result in extracting signatures reflecting the proportion of stromal
contamination in the sample, rather than underlying tumor biology. Such artifacts
would be real, reproducible, and potentially misleading, but would not be of biological
or clinical interest. As a result, the overlap of source signals can severely decrease the
sensitivity and specificity for the measurement of molecular signatures associated with
different disease processes. Despite its critical importance to almost all the follow-up
analysis steps, this issue (i.e., heterogeneity correction) is often less emphasized or at
least has not been rigorously addressed as compared to the overwhelming interest and
effort in pheno/geno-clustering and class prediction.
As a counterpart to heterogeneity correction, feature selection, the task of selecting
(hopefully) the best subset of the input feature set, is becoming more and more important,
especially since the improvement of biology and the advent of spotted cDNA microarray,
which permit scientists to screen thousands of genes simultaneously to determine
whether those genes are active, hyperactive or silent in normal or cancerous tissue.
Gene selection is one of the most important inseparable step from the other steps in
microarray data analysis, and is featured with multiple purposes. At first, the genes with
their expression levels not stably expressed, in other words, the genes we refer to as the
unstable expressed genes (UEG), should be removed; also, in carrying out comparisons
of expression data using measurements from multiple samples or arrays, the question of
normalizing data arises, where the control gene set, i.e., the designated subset expected
to be unchanged over most circumstances, what we call the constantly expressed genes
(CEG), should be obtained for normalization; differentially expressed genes are those
which result in differences in expression levels due to tissue samples which come from
tissues of different diseases or tumors/tumor levels, which may functionally tell the
biological/clinical cause of the diseases; finally, jointly discriminatory genes (JDG) are
those we should select from gene set via some discriminatory/separability measurement
such that the best classification/prediction performance can be reached when cancer
classification and/or cancer diagnosis is expected.
176 Zhang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
On the other hand, gene selection is strongly encouraged from tissue clustering/
classification point of view. Since the number of genes is huge compared with the number
of tissue samples, tissue clustering and tissue classification are such tasks which will
undergo severely the curse of dimensionality if it were fulfilled directly in the huge
dimensional gene space in which much less number of samples exist (in most cases, this
is unrealistic due to the super-high dimensionality of the space). Gene selection, as a tool
to reduce dimensionality of the space, is widely applicable. Among the most widely cited
reasons to reduce data dimensionality are (1) to eliminate redundancy, (2) to reduce the
computational time and cost of performing the analysis and (3) to reduce misclassifications.
Perhaps the most important reason is the ability of a reduced data dimensionality to
reduce the misclassification rate. This rate is reduced because as the number of
measurements in a data set increases, the misclassification rate first decreases and then
increases. With high data dimensionality, the classifier will have to allocate computa-
tional resources to deal with the noise dimensions, leaving fewer available to encode the
knowledge embedded in the informative dimensions. Limiting dimensionality also
reduces the impact of the curse of dimensionality. Thus, by excluding genes that clearly
show little likelihood of correlation with phenotype/outcome, the precision of analyses
can be improved and the unnecessary variation associated with clusters that are not
relevant to the goal of the analysis is reduced.
This chapter is devoted to some work done in microarray data analysis, relating to jointly
discriminatory gene selection, and tissue heterogeneity correction, as well as cancer
classification/prediction. The chapter is divided into two parts, focused on gene
selection for cancer diagnosis, and heterogeneity correction respectively. The first part
looks for jointly discriminatory genes which are most responsible to classification of
tissue samples for diagnosis. In the second part, we study heterogeneity correction
techniques, in which independent component analysis (ICA) is applied to tissue samples
with the expression levels of only selected genes, the genes which are functionally
independent, and/or jointly discriminatory. Also, we employ non-negative matrix factor-
ization (NMF) to computationally decompose molecular signatures based on the fact that
the expression values in microarray profiling are non-negative. The real-world data used
for experiment in this chapter is NCI data (Khan et al., 2001). The data set consists of 88
tissue samples of 2308 effective gene expressions obtained from 88 patients of four
different neuroblastoma and non-hodgkin lymphoma cell tumors.
Problem of Finding
Joint Discriminatory Genes for
Cancer Diagnosis
The advent of DNA microarray technology has brought to data analysis broad patterns
of gene expression simultaneously recorded in a single experiment. In the past few
months, many data sets have become publicly available on the Internet. These data sets
present multiple challenges (Anil, Robert, & Mar, 2000), including a huge number of gene
Gene Selection for Molecular Classification of Cancer 177
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
expression values per experiment (several thousands to tens of thousands), and a
relatively small number of experiments (a few dozen) belonging to several disease
classes. The gene expression data set is generally expressed in a matrix, where each row
indicates an experiment (a tissue sample), with its elements being the expression levels
of different genes. Generally speaking, the number of genes is over 2000, while the number
of samples is comparatively small, say, less than 100.
Gene selection becomes more and more important especially at present for DNA
microarray data set since the number of dimensions (genes) becomes higher and higher
compared with a very limited number of samples (experiments) in gene space (G-space)
where the curse of dimensionality problem emerges (Trunk, 1979). In fact, the data set
itself has its intrinsic dimensionality. Gene selection is an approach of selecting a
subgroup of genes based on some criterion, such as discriminative power of the samples
belonging to different classes. There are three main reasons to keep the dimensionality
of the pattern representation (i.e., the number of features/genes) as small as possible:
biologists interest, measurement cost, and classification accuracy. A limited yet salient
feature set simplifies both the pattern representation and the classifiers that are built on
the selected representation. Consequently, the resulting classifier will be faster and will
use less memory. Also, a small number of features/genes can alleviate the curse of
dimensionality when the number of training samples is limited. Most importantly, on the
other hand, seeking a subset of genes, the genes that have the greatest discriminatory
power of different diseases, is one of the main interests of biologists.
A serious curse of dimensionality emerges in the case of DNA microarray profiles
(Toronen, 2004). Let m
0
be the dimensions of the space, N be the number of tissue samples
in the space, and the tissue samples belong to K tumor classes. If the distribution of the
data set is known in this space, the larger the ratio of m
0
to is, N the less the Bayesian
decision error would be, which will approach zero when the number of dimensions of the
space is infinite; however, when the distribution of the data set in this space is unknown,
which is the case for real DNA microarray data set, the Bayesian decision error would
converge to its possible maximum value as m0 approaches infinity, the phenomenon
referred to as the curse of dimensionality (Haykin, 1999; Choi, 2002). Notice that an
acceptable number of samples in the space should be more than 10 times the number of
dimensions of the space, a good practice to follow in classifier design, for avoiding the
curse of dimensionality if the distribution of the data is unknown (Haykin, 1999). For the
case when each dimension corresponds to a gene and each sample corresponds to an
experiment for the spotted cDNA microarray dataset, the space is called gene space, or
simply G-space. From this viewpoint, a serious curse of dimensionality occurs in the
spotted cDNA microarray dataset, where in general the number of dimensions (genes)
is larger than 2000, and ever increasing as the development of advanced biochip
techniques, while the number of samples obtained for clinical experiments is very limited
and comparatively small, say, less than 100. In fact, the data has its intrinsic dimensions:
when the number of training samples is limited and fixed, too many dimensions will not
benefit classification but bring more misclassification. On the other hand, what the
biologists are interested in is to select dimensions (genes) which are responsible for the
classification of cancer diseases, then a biochip of much smaller scale could be produced,
which includes only the expression profile of the cancer-related genes. This will induce
much less disgnostic cost for cancer disgnosis.
178 Zhang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Recently many researchers do research work on gene selection and gene expression
pattern clustering. Hierarchical growing neural network was presented for clustering
gene expression patterns (Herrero, Valencia, & Dopazo, 2001), support vector machine
(SVM) method was used in selecting genes (Guyon, Weston, Barnhill, & Vapnik, 2002),
and statistical PCA method was presented for biomarker identification of the diseases
(Xiong, Fang, & Zhao, 2001).
The next two sections will be focused on two gene selection methods. One, referred to
as class-space method (C-space method), transforms the high-dimensional sample space
into a low-dimensional class space to tackle the curse of dimensionality, and the
discriminatory genes are selected via principal component analysis (PCA) in the class
space. Another, referred to as contribution space method, is based on the fact that a small
number of samples, under some trivial condition, is generally linearly separable in a high-
dimensional space especially when the number of samples is much less than that of the
dimensions of the sample space, which is just the situation of the microarray profiles. This
characteristic encourages us to simply design a linear support vector machine (SVM) to
classify tissue samples which belong to pair-wise classes, since a simple linear SVM is
sufficient as a classifier with the greatest generalization potential, while the margin vector
of the SVM is directly applicable to select genes which are responsible to the classifi-
cation of these two classes. The method is then extended to gene selection for the case
of more than two cancer classes. Diagnostic genes for real DNA microarray data set were
selected and compared with the ones selected with SNR method, TSM method and SVM/
MLP-CV method.
Finding Discriminatory Genes
for Molecular Classification of
Cancer in Class Space
In this section, an effective gene selection method is presented based on the transfor-
mation of gene space (G-space) into its dual space, referred to as class space (C-space)
for tackling the curse of dimensionality. The genes are selected so that the largest
discriminative power is attained in separating samples from belonging to different
classes with the help of principal component analysis (PCA). Our experimental results
on real DNA microarray data set are evaluated with Fisher criterion, weighted pairwise
Fisher criterion and leave-one-out cross validation, showing the effectiveness and
efficiency of the proposed method.
Space Transformation: From Gene Space to Class Space
Assume that the gene expression data set at hand is expressed in the matrix of:
Gene Selection for Molecular Classification of Cancer 179
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
1 2
11 21 1
1
12 22 2 2
1 2

g
g
g
s
s s g s
N
N
N
N
N N N N
g g g
g g g
s
g g g s
s
g g g








L
L
L
M
M M M M
L
,
where each row corresponds to the gene expression levels over all genes in a specific
experiment (sample), and each column corresponds to gene expression levels of a specific
gene over all the experiments; the number of dimensions N
g
of the G-space is much larger
than the number of samples N
g
in the space; also, each sample belongs to one of K classes.
For avoiding the curse of dimensionality of the G-space, we make a space transformation,
from the gene space (G-space) to class space (C-space). Construct a class space (C-
space) each dimension of which corresponds to a class, and each sample of which is a
gene in G-space. Since each gene is expressed as:
,1 1 ,2 2 ,
,
s s
i i i i N N
g g s g s g s = + + + L (1)
it can also be expressed in C-space as,
,1 1 ,2 2 ,
,
i i i i K K
g g S g S g S = + + + L (2)
Figure 1. Demonstration for absolute/relative contribution of genes to separation of
two different deceases


y
g
n
g

o
45

2
g

1
g
l

1
S

2
S

180 Zhang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
where Sj is the jth class axis, and,
1 2
2 2 2 1/ 2
, , , ,
( )
m
i j i k i k i k
g g g g
= + + + L
(3)
if the samples belonging to the ith class are
1 2
, , ,
, , , .
m
i k i k i k
s s s L . Evidently,
,
, i j
g
indicates
virtually the expression level of gene gi over all samples belonging to the jth class
(expression level is always positive). Since the number of dimensions is and the number
of samples is N
g
in C-space, curse of dimensionality does not occur in C-space.
Now let us consider discriminatory gene selection problem. A simple example is given
first where there are only two genes ga and gb, and only two tissue classes, normal and
abnormal, for introducing the main principle of selecting genes. Assume that the gene
expression level is (g
a
, g
b
) = (3, 3) for normal tissue, and is (g
a
, g
b
) = (30, 3) for abnormal
tissue, i.e.,
normal 3 3
abnormal 30 3
a b
g g

Accordingly, the expression level of g
a
changing from 3 to 30 is the reason of change from
normal tissue to abnormal tissue which can be used for the diagnosis of the disease, while
that of g
b
without any change in its expression level indicates that the expression level
of g
b
is not responsible to the classification and diagnosis of normal tissue and abnormal
tissue. From this viewpoint, gene g
a
rather than g
b
should be selected to distinguish
normal tissue and abnormal tissue, i.e., gene ga has its contribution to the separation of
normal tissue and abnormal tissue, while the contribution of gene g
b
to this separation
is 0. Figure 1 shows the two genes in C-space for two class situation, where none of the
genes on the 45-degree line l have contribution to separation of these two classes: their
contribution values are all 0s.
It is obvious that the genes which are closer to one have smaller contribution values,
while those which are closer to the normal class axis S1/abnormal class axis S2 have larger
contribution values, and the two genes which are symmetric against l should have same
contribution value. Now we divide contribution value of gene g to the separation of these
two classes into absolute contribution and relative contribution, denoted by c
a
(g) and
c
r
(g) respectively, and use their multiplication to indicate the contribution of gene to
this separation, i.e.,
( ) ( ) ( ).
r a
c g c g c g =

(4)
For a gene g=(g
y
,g
n
) in C-space, where g
y
and g
n
are the projections of in normal and
abnormal class axis respectively, the absolute contribution value of g to separation of
normal and abnormal classes is defined as the distance from g to the 45-degree line l, i.e.,
Gene Selection for Molecular Classification of Cancer 181
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
c
a
(g)=d(g,l). In Figure 1, g
1
and g
2
have the same absolute contribution value since they
have the same distance to the line l. However, their closeness to the normal class axis/
abnormal class axis is different. Thus they should have different relative contribution
values. The relative contribution value of gene g, C
r
(g) , is defined as the angle between
vector g and the line l, since it reflects the ratio of expression levels of normal to abnormal
situation.
The line l separates the C-space into two areas, which normal/abnormal class axis is
inside, referred to as normal/abnormal class area. The genes that are in normal class area
have contributions to separating normal class to abnormal class, while those that are in
abnormal class area have contributions to separating abnormal class to normal class.
Finding Joint Discriminatory Genes in Class Space
Figure 2 demonstrates three different gene distribution situations in 2 dimensional C-
space. For Figure 2a, the genes which have no contributions should be on 45-degree line
l
1
, while for Figure 2b and c, such genes should be on lines l
2
and l
3
respectively according
to the distribution of the majority of the data set. Here we call and quasi 45-degree lines.
In higher dimensional C-space, they would be quasi 45-degree hyperplanes, or simply
qusi-hyperplanes.
Figure 2. Possible quasi-hyperplanes in two-class situation


2
S

1
S

2
S

1
S

2
l 1
l

3
l

2
S

1
S
(a) (b)
(c)
Figure 3. Derivation of quasi 45-degree hyperplane equations


2
S

3
S

1
S

1
l

3
l

2
l
3
rd
class area
1
st
class area
2
nd
class area
182 Zhang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
We perform PCA on gene data set in C-space, find the principal component direction, and
then find the quasi-hyperplane for separation of the data in one class and the data in all
other classes. For doing this, the following four steps are necessary:
Step 1. Centering by,
(1)
g g g = , (5)
where g is the mean of all of the genes in C-space.
Step 2. Performing PCA by,
(2) (1) T
g Q g = ,
(6)
where Q = (q
1
, q
2
, , q
K
) is a matrix whose columns are the eigenvectors of the covariance
matrix of
(1)
g . .
Step 3. Whitening by,
(3) 1/ 2 (2)
g D g

= , (7)
where D is the diagonal matrix with its diagonal elements being the eigenvalues of the
covariance matrix of g
(1)
.
Step 4. Finding the 45-degree hyperplanes in space.
In fact, the 45-degree hyperplanes in space are just the quasi-hyperplanes in C-space.
For the case of K = 3 shown in Figure 3, the equations of the 45-degree hyperplanes in
g
(3)
space, i.e., the quasi-hyperplanes in C-space, are:

(3) (3) (3) (3)
1 1 2 1 3
(3) (3) (3) (3)
2 2 1 2 3
(3) (3) (3) (3)
3 3 1 3 2
: , ,
: , ,
: , .
l g g g g
l g g g g
l g g g g
Gene Selection for Molecular Classification of Cancer 183
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
These equations can also be written in the matrix form of:

(3)
1
(3)
2
(3)
3
1 1 0
: 0,
1 0 1
1 1 0
: 0,
0 1 1
1 0 1
: 0.
0 1 1
l g
l g
l g
Similarly, the equations of K - 1 quasi-hyperplanes in K dimensional C-space, i.e., the 45-
degree hyperplanes in g
(3)
space for separating the ith class and all other classes, can
be expressed in the matrix form of:
(3)
( ) [0, 0, , 0]
T
G i g = L
,
, (8)
where


1 1
1 1
( ) 1
1
1 1
i
G i i





=

O M
M
O

and all of the other elements in G(i) are 0s. Combining equations (5)-(8), we get all the
quasi-hyperplanes in C-space as:
1/ 2
( ) ( ) 0, 1, 2, , .
T
G i D Q g g i K

= = L (9)
For each i, i = 1, 2, , K, the ith class area is referred to the area surrounded by the quasi-
hyperpl anes shown i n equat i on (8) (i . e. , t he area whi ch sat i s-
fies
1/ 2
( ) ( )
T
G i D Q g g

[0, 0, , 0]
T
L . ).
184 Zhang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
In order to calculate the contribution of each gene to separation, it is necessary to know
which class area the gene belongs to, and then this gene has a contribution to separate
this class from all other classes. Otherwise, it does not have this contribution. The class
area that gene
g
belongs to can be obtained by finding i, i = 1, 2, , K, such:
G i D Q g g
T
( ) ( )
/

1 2
[0, 0,
...
, 0]
T
. (10)
Then gene g has contribution to separate ith class from all other classes. The key point
K - 1 for calculating this contribution is to calculate the closeness between vector g and
the quasi-hyperplanes presented by equation (8) in C-space. The closer they are, the less
contribution it has, while no genes on quasi hyperplanes have contribution to separa-
tion. Here relative and absolute contribution of a gene g are defined, but both are
dimensional vectors. The ith element of the relative contribution and the absolute
contribution of gene g means the relative contribution and the absolute contribution of
gene g to separating the ith class from all other classes respectively. The absolute
contribution vector is defined as a vector the ith element of which is the distance between
g g and l
i
, and the relative contribution vector of g is defined as a vector the ith element
of which is the angle between the g(i) and g(i), where is the projection point of g on the
ith quasi-hyperplane.
The contribution of gene g to separation is the sum of contributions of gene g to
separating the ith class to all other classes for i = 1, 2, , K, that is:
( ) ( ) ( )
T
r a
c g c g c g =
.
. (11)
Gene selection is performed by ranking all the genes according to their contributions to
separation, and selecting the top M genes, where M is the number of genes one wants
to select. Another way for gene selection, referred to as within-class ranking, is to rank
genes according to their contributions to separation in each class area and select M
1
, M
2
,
Figure 4. Gene selection performance vs. the number of genes selected by c-space
ranking
Gene Selection for Molecular Classification of Cancer 185
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
, M
K
genes in the K class areas respectively, where M
1
, M
2
, , M
K
are predetermined
numbers of genes one wants to select in the corresponding class areas.
Experiments and Results
We performed gene selection on the real NCI microarray data set (Khan et al., 2001). For
this dataset, the dimensionality of the C-space is K = 4, and there are altogether 2308
sample points (genes) in the C-space. Our experiments include three parts: (1) calculate
the linear separability criterion of the data on the selected gene subspace according to
different gene contribution ranking method; (2) study the linear separability of the data
on selected gene subspace with different gene selection methods (SNR method, three-
step method (TSM) and C-space methods); (3) study the nonlinear separability of the
data on the selected gene subspace by understanding the relation between the number
of genes selected with C-space method and the nonlinear separability of the data in the
selected gene subspace.
Experimental Results on Different Contribution Ranking Methods
For NCI data, gene selection is performed in a K = 4 dimensional C-space, where there
are altogether 2308 genes (samples). Figure 4 presents the number of selected genes
versus the linear separability criterion of data in the selected gene subspace, where M
1
= M
2
= M
3
= M
4
= 150 for within-class ranking, and M = 1200 for all-gene ranking and
select top M. In our experiments, the linear separability is measured with Fisher criterion
and weighted Fisher criterion. Shown in Figure 4 are two very consistent curves, which
get to the largest when M = 60 for both, indicating that for NCI data, both the ranking
methods get approximately the same linear separability criterion for the same number of
selected genes.
However, notice that the selected subgroup of 60 genes with Fisher criterion and that
with weighted Fisher criterion are not the same. In our following experiments, only all-
gene ranking is used. In order not to lose genes which may play an important role in
biology, the predetermined number of genes selected in our following experiments is 150,
which is much larger than 60, and also larger than the number of samples.
Comparison with Other Gene Selection Methods
Finding 150 genes from 2308 genes which can guarantee the best separability of the data
in this selected gene subspace, is an NP complete problem (Pudil & Novovicova, 1998).
Thus, some methods have been developed for finding the suboptimal gene subgroup.
A simple method biologists often used for gene selection, SNR method (Dudoit,
Fridlyand, & Speed, 2002) , performs gene selection directly in gene space, and consider
the separability of the data in each gene expression level independently, i.e., to select
genes with the top largest SNR. That is, for gene
i
, the SNR of the data set in this gene
is calculated by:
186 Zhang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
2
2
( )
k ki i
k
i
k ki
k
p
SNR
p

,
,
where
i

is the mean of the ith gene expression levels of the data,
ki

,
2
ki

, and
2
ki

are
respectively the mean and variance of the ith gene expression levels of the data which
belong to the kth class, and pk is the proportion of the kth class data to all the data.
The three-step method (TSM) (Wang, Zhang, Wang, Clarke, & Khan, 2001) divides gene
selection process into three steps with three parameters (K
1
, K
2
, K
3
). K
1
, K
2
, K
3
indicate
the number of the genes removed in the first, the second and the third step respectively.
The first step is just SNR method. The second step is to find a gene subspace from the
remained K - 1 dimensional gene subspace by calculating the within-scatter matrix and
between-scatter matrix:
,
( )( ) ,
w k k
k
T
b k k k
k
S p C
S p
=
=

as well as the eigenvalue

m
and eigenvector
m
W of matrix
1
w b
S S

, , where
k
and
k
C is the
means and the covariance matrix of the samples belonging to the kth class, and the
means of overall training samples. Rank the remained K - 1 genes according to:
Table 1. Separability criterion of the projected data from selected gene subspace into
top three principal axes
Projection method DCAF projection DCAW projection PCA projection
S
e
p
a
r
a
b
i
l
i
t
y

m
e
a
s
u
r
e

F
i
s
h
e
r

W
e
i
g
h
t
e
d

f
i
s
h
e
r

F
i
s
h
e
r

W
e
i
g
h
t
e
d

f
i
s
h
e
r

F
i
s
h
e
r

W
e
i
g
h
t
e
d

f
i
s
h
e
r

N
u
m
b
e
r

o
f

g
e
n
e
s

s
e
l
e
c
t
e
d

C
o
m
p
u
t
a
t
i
o
n

t
i
m
e

C-space method 266.8972 463.8205 269.2144 463.3149 1.6933 0.032578 150 10 s
TSM
(2138,10,10)
285.6568 475.5676 286.1407 480.8895 2.6211 0.028189 150 20 hr
TSM
(2138,15,5)
186.0106 370.8758 186.5364 384.5828 0.83254 0.013029 150 9 hr
TSM
(2138,20,0)
115.0260 287.3228 115.3815 291.6633 1.8997
0.009975
1
150 30 min
SNR method 147.2021 327.4194 150.3970 342.1500 5.6257 0.048028 150 15 min

Gene Selection for Molecular Classification of Cancer 187
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
1
1
n K
j m mj
m
c W

=
=
,
and remove the last K
2
genes. The third step is to remove K
3
genes, one at a time, by
removing the one which induces the greatest decreasing of Fisher criterion/weighted
Fisher criterion value of the data on the unremoved gene subset.
For TSM, its first step, just SNR, is very simple and time saving, but the separability of
the data on the selected gene subspace may not be satisfied. Hence, a small K
1
should
be selected, such that important genes could not be lost. The second step of TSM is to
consider the scatter of data on all genes, while the third step to consider the separability
of the whole data. Therefore, to get better gene selection result, K
2
, K
3
should be larger,
but the computation for step 2 and step 3 will increase very seriously. In our three
experiments with (K
1
, K
2
, K
3
) being (2138, 10, 10), (2138, 15, 5), (2138, 20, 0) respectively,
it took 20 hours to find 150 genes with TSM of (2138,10,10) while it took only 10 seconds
to find 150 genes with our C-space method.
In order to visualize the separability of the data in the selected gene subspace, PCA/
DCAF/DCAW are used respectively (PCA principal component analysis (Haykin,
1999), DCAF/DCAW discriminate component analysis based on Fisher (Wang et al.,
2001)/weighted Fisher criterion (Loog, Duin, & R.P.W., 2001). That is, first, calculate,
0 0
1
1
1
1 1
PCA: ,
DCAF: ,
DCAW: ( )( )( )
p
F w b
K K
T
w w i j ij i j i j
i j i
J C
J S S
J S p p w

= = +
=
=
=



TSM(2138,10,10)
(DCAW,286.1407)
TSM(2138.15.5)
(DCAW,186.5364)

TSM(2138,20,0)
(DCAW,115.3815)
5
5
5
140
0
0
5
c-space method
(DCAW,269.2144)
TSM(2138,10,10)
(DCAW,286.1407)
SNR method
(DCAW,150.397)
39
21
18
87
6
33
72
(a) (b)
Figure 5. Overlap situation of gene subsets selected with different methods. (a) Overlap
for three TSM selected gene subsets; (b) overlap for three different gene selection
methods
188 Zhang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
where C is the covariance matrix of the data in gene space, ( )
ij
w =
2
(2 )
2 2
ij
ij
erf

is the
weight added to the class i and class j between-class scatter matrix,
ij

is the Mahalanobis
distance between centers of class i and class j. Then, get the eigenvalue and eigenvector
of J, and project the data into the top three principal eigenvector axes of the matrix J. Table
1 shows the linear separability Fisher criterion/weighted Fisher criterion of the projected
data from the selected 150 gene subspaces with the above three gene selection methods.
Notice that the subsets of the genes selected with the above methods are not the same
and have some overlaps.
It is seen from Table 1 that TSM is decreasing in its Fisher criterion value in the order
of (2138, 10, 10), (2138, 15, 5), (2138, 20, 0), which is in consistency with our analysis. The Fisher
criterion value does not decrease in the same order for PCA method since PCA is an
unsupervised learning which can guarantee that the reconstruction structure error is minimum
while DCAF/DCAW is a supervised learning. It is also seen from Table 1 that the Fisher criterion
value of C-space method is only a little bit less than that of (2138, 10, 10) TSM, while the
computation time decreased greatly from 20 hours for (2138, 10, 10) TSM to 10 seconds for C-
space method, indicating that the C-space method is greatly effective and efficient.
It is also seen from Table 1 that DCAW is superior in finding the projection axes which
preserve separability of the data better compared with DCAF.
The overlap situation of the selected gene subsets with different methods was studied
and is shown in Figure 5, where each number indicates the number of genes in the
corresponding subset. It is shown from Figure 5(a) that among 150 genes selected with
(2138, 10, 10) TSM, 140 genes are also selected with both (2138, 15, 5) TSM and (2138,
20, 0) TSM, but the separability performances decrease greatly from 286.1407 to 186.5364
to 115.3815. This seems to indicate that the remaining 10 genes are significant for the
separability of the data. However, all of these 10 genes are not included in the 150 genes
selected by C-space method. This indicates that the separability of the data is obtained
by the combination of genes, rather than obtained by genes independently. This can also

0 2 4 6 8 10 12 14 16 18 20
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
no. of selected features
r
a
t
e
misclassification rate
Figure 6. The number of genes selected vs. misclassification rate with a L-3-4 MLP,
L=2~20
Gene Selection for Molecular Classification of Cancer 189
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
be seen from the overlaps of the gene subsets obtained by (2138, 10, 10) TSM, C-space
method and SNR method shown in Figure 5(b) where the overlap of genes selected with
TSM and with C-space method is only 39+18 in size but the separability of data by
selecting genes with C-space method is already 269.2144, much better than SNR method
where the separability is only 150.397.
The Number of Selected Genes vs. Nonlinear Classification
The number of selected genes is studied for nonlinear separability performance. L genes
were selected with the C-space method, and a multiplayer perceptron (MLP) network (L-
3-4) with L inputs, three hidden neurons, and four output neurons was trained for
classification of samples belonging to four classes. The hidden and output neurons are
activated with sigmoidal function and linear activation function respectively. Figure 6
shows the number of genes selected vs. misclassification rate of the trained network,
where each point in Figure 6 is the one with the lowest misclassification rate among 30
trials with random initiation of the network parameters for avoiding getting stuck into
local minimum of the training process. It is shown from Figure 6 that only nine genes rather
than 150 genes are enough if a nonlinear classifier is used.
The generalization performance of the MLP is studied with leave-one-out cross valida-
tion approach for the 60 genes selected with the C-space method. The process was: (1)
60 genes were selected with C-space method; (2) DCAW was used to project the data with
the selected 60 dimensions (genes) into top three principal axes (a linear projection); (3)
a 3-3-4 MLP was trained (a nonlinear classifier). Each sample among 64 samples was used
as a testing sample, while all other samples were used as training samples, to train a MLP.
Hence, there were altogether 64 MLPs trained and tested. Our experimental result shows
that the misclassification rate of all of these 64 MLPs for testing samples is zero. This
indicates that these 64 genes selected with C-space method is very effective and efficient.
Notice that this method makes a quasi 45 degree hyperplane found with PCA method be
a reference for overall genes, and calculates contribution of each gene for classification
independently. This makes the method effective in both the separation performance of
data on the selected genes and in computation complexity.
Finding Discriminatory Genes
for Molecular Classification of
Cancer in Contribution Space
Noticing the fact that a small number of samples, under some trivial condition, is generally
linearly separable in a high-dimensional space especially when the number of samples
is much less than that of the dimensions of the sample space, which is just the situation
of the microarray profiles, and also the fact that support vector machine (SVM) is such
a classifier designed for binary classification that brings the least empirical risk and best
190 Zhang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
generalization performance when only a limited number of training samples are used for
designing it, we simply design a linear support vector machine (SVM) to classify tissue
samples which belong to pair-wise classes, with the margin vector of the SVM directly
used to select genes which are responsible to the classification of these two classes. The
method is then extended to gene selection for the case of more than two cancer classes.
Contribution Space and Gene Selection Method
As is known, in gene space (G-space), the number of dimensions (genes) m
0
is much larger
than the number of samples n which belong to k disease classes. This makes it practical
that the samples in G-space be linearly separable. Thus, for the case of only two classes
(cancer/normal), a linear SVM can be trained as a linear classifier for classification of two
class samples in G-space. The ith element g
i
of the margin vector M=[g
1
,g
2
,L, g
m0
] tells
the margin value defined by the SVM classifier in the ith gene axis, where m
0
is the total
number of genes in microarray profiles. Notice that the larger the g
i
is, the ith gene is more
possible to be selected since the SVM results in larger margin for separation of samples
in two classes compared with the other genes. For multiclass situation, say, k class
situation, the binary class-pair includes ,(1,2), (1,3), (1,4),,(k - 1,k), and k (k - 1)/2 SVMs
can be trained as classifiers for classification of each pair of samples belonging to two
different classes respectively. The resultant margin vectors can be obtained and
expressed as follows:
0
0
0
1 2
12 12 12 12
1 2
1 1 1 1
1 2
( 1) ( 1) ( 1) ( 1)
m
m
K K K K
m
K K K K K K K K
M g g g
M M g g g
g g g M





= =




L
M M M M M
L
M M M M M
L
.
. (12)
Figure 7. Absolute contribution and relative contribution of a gene for overall
classification

Gene Selection for Molecular Classification of Cancer 191
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
We define a contribution space where each axis
ij
C ( , 1, 2, , , ) i j k i j = L corresponds to
the margin component derived from the SVM which classifies the ith class and the jth
class. Evidently, the contribution space has k(k - 1)/2 dimensions, and each gene is a
sample or a vector in the space. The ith column vector in matrix M defines a contribution
vector of gene i, for i=1,2,...m
0
. Firstly, take a look at a simple gene selection problem:
select only one gene i from three genes through the following contribution matrix:
1 2 3
12
13
23
10 8 20
9 8 3
10 8 2
g g g
C
C
C
.
. (13)
Even though gene 3 contributes the classification of class 1 and class 2 greatly (the
contribution value is 20) compared with all of other genes for this classification (the
contribution value is only 10 and 8 respectively), its contributions to classification of
class 1 and 3, and of class 2 and 3 are very small (the values are only 3 and 2 respectively).
In contrast, gene 1 has moderately large contributions to classification of all the class-
pairs. Hence, gene 1 would be selected as a gene selection result. Shown in Figure 7 is
a simple 3-dimensional contribution space for illustration. A is a vector with same
components, say, 1s, in each axis of the space, corresponding to the largest normalized
contribution to classification of overall class-pairs. O is a zero vector with zero compo-
nents in each axis of the space, corresponding to no contribution to overall class-pairs.
A novel idea for gene selection is to calculate the overall classification contribution of
each gene by comparing the corresponding column of contribution matrix M and the
reference gene A, rank the overall classification contribution of all the genes in an
descendent order, and select the first r genes as gene selection result.
We measure the overall classification contribution (OCC) of the ith gene g
i
by two
contribution issues: absolute contribution c
a
(g
i
) and relative contribution. The absolute
contribution c
s
(g) is defined as the length of the projected vector which is obtained by
projection of the ith column vector in M contribution matrix to the reference vector A,
and c
r
(g
i
) is defined as the cosine function value of the angle between the ith column
vector in and the reference vector A. The final OCC of the ith gene g
i
is defined as a
combination of absolute contribution and the relative contribution as follows:
( ) ( ) ( )
i i i
a r
OCC g c g c g = . (14)
Notice that the reference vector A does not affect the gene selection result as long as the
elements in A are all the same. Thus, we have our gene selection algorithm as follows.
Step 1. Train a linear SVM for each class pair, class i and class j, and obtain the margin
vector M
jj
of the SVM, for i,j=1,2,L,k,ij. Normalize the margin vectors M
jj
such that each
vector has a standard norm of 1. Then the contribution matrix M

defined in equation 12
is obtained.
192 Zhang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.

-6 -4 -2 0 2 4
-5
-4
-3
-2
-1
0
1
2
3
4
projection of gene subspace to top 2 wDCA space
Fisher:25.6394; wFisher:54.6125
c1
c2
c3
c4
-5
0
5
-5
0
5
-4
-2
0
2
4
Fisher:27.5023; wFisher:68.3611
projection of gene subspace to top 3 wDCA space
c1
c2
c3
c4

-6 -4 -2 0 2 4
-5
-4
-3
-2
-1
0
1
2
3
projection of gene subspace to top 2 wDCA space
Fisher:29.9794; wFisher:59.266
c1
c2
c3
c4
-5
0
5
-5
0
5
-6
-4
-2
0
2
4
Fisher:37.4817; wFisher:102.5749
projection of gene subspace to top 3 wDCA space
c1
c2
c3
c4
Figure 8a. the scatter plot of samples expressed with gene expression levels of 20
selected genes obtained by (a) SVM/MLP-CV method and
Figure 8b. CS method and projected onto top two(left) and top three(right) DCA space
for visualization
Step 2. Let the reference vector in contribution space to be ones in its elements. Calculate
the OCC of each gene in contribution space by computing the absolute contribution and
the relative contribution, and combining them with equation 14.
Step 3. Sequence the genes according to their OCCs in descendent order, and let the
number of selected genes r=1;
Step 4. Select the first r genes from the gene sequence as the selected gene subset G(r).
Both Leave-one-out and leave-4-out cross validations (for leave-four-out, we leave a
sample out from each class as training data to avoiding too large computation) are applied
and the largest misclassification rate (r), for testing samples is used for the evaluation
of the generalization performance of the linear classifier structure. Notice that we still use
SVM based method for cross validation, i.e., train a linear SVM for classification of
samples in each class-pair but expressed with only expression levels of the selected
genes. The largest misclassification rate (r) for testing samples is calculated.
Step 5. If (r)0 and r + 1 r go to step 4; otherwise, G(r) is the final selected gene subset.
Gene Selection for Molecular Classification of Cancer 193
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Experiment on Real Microarray Data
For NCI data, we constructed contribution space, which is of six dimensions, and there
are altogether 2308 genes in the space for NCI data. We used the proposed contribution
space based gene selection method (CS method) for gene selection. In order to compare
the separability performance of the samples in the selected gene subgroup with the
proposed method, we introduce another gene selection method briefly, called SVM/MLP
cross validation method (SVM/MLP-CV). By noticing that the dimension of the gene
space is too high while there are only a comparatively small number of training samples,
for the situation of the samples belonging to only two classes, class i and class j, support
vector machine (SVM) is used respectively for gene selection directly in gene space. The
absolute value of the elements in the margin vector obtained from SVM is ordered
decreasingly, and the first genes (determined by leave-one-out and leave-four-out cross
validation with zero misclassification) are selected according to the first r
jj
ordered
absolute margin elements since it will bring the largest generalization power for classi-
fication of samples belonging to class i and class j in the selected gene subspace. After
the genes are selected for samples belonging to each pair of classes, the selected genes
are merged into a larger group of gene subset with, say, r genes. Then, sequential
backward selection is used for gene subgroup selection until the largest misclassification
rate for leave-one-out and leave-four-out cross validation with an MLP of r inputs (notice
that by sequential backward selection, the number r is decreasing during gene selection
process), two hidden neurons and four output neurons (each corresponds to a type of
disease) is larger than zero. In this way, the final gene selection result is obtained.
In our experiment, Fisher criterion and weighted pairwise Fisher criterion were employed
for evaluation of the separability performance of the samples in the selected gene
subspace. Table 2 shows the selected gene indices, and the Fisher and weighted Fisher
criterion values obtained from our experiment on this data set, when the number of
selected genes is 20. It is seen from the Fisher and weighted Fisher criterion values in
Table 1 that the separability of the samples in the selected gene subspace is much larger
Table 2. Comparison of SVM/MLP-CV method and the CS method for selecting 20 genes
and the separability of samples in the selected gene subspace and their projection to
the top 2/top 3 DCA/wDCA directions
Separability evaluation
Method
Selected gene
indices
20-D
selected
gene space
Project to
Top 2 DCA
Project to
Top 3 DCA
Project to
Top 2 wDCA
Project to Top
3 wDCA
SVM/M
LP-CV
135,151,187,246,
276,469,509,523,
545,1093,1105,13
89,1557,1645,175
0,1915,1954,1955
,2046,2050
31.8404
78.8581
25.7403
55.9319
27.5091
68.4060
25.6394
54.6125
27.5023
68.3611
CS
1,107,151,187,24
6,276,469,509,52
3,545,572,742,84
2,1084,1093,1389
,1601,1915,1954,
1955
40.9699
109.1131
29.6184
60.1432
36.6256
99.3981
29.9794
59.2660
37.4817
102.5749

194 Zhang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
when the genes are selected with the proposed CS method, compared with the previously
proposed SVM/MLP-CV method. Notice that it was verified that in Table 2, the
misclassification rates for leave-one-out and leave-4-out cross validation are all zeros
when the group of genes was selected with both methods.
We also projected the samples with gene expression levels of only selected genes onto
top 2/3 DCA subspace (i.e., top 2/3 discriminate component directions with discriminate
component analysis (DCA)), and top 2/3 wDCA subspace (i.e., top 2/3 weighted
discriminate component directions with weighted discriminate component analysis
(wDCA)) for visualization of the separability of the samples in the selected gene subset.
Figure 15(a) shows the scatter plot of the samples in the selected gene subspace obtained
by SVM/MLP-CV method projected onto top 2/3 wDCA subspace, and Figure 15 (b)
shows that of the samples in the gene subspace obtained by CS method projected onto
top 2/3 wDCA subspace. By comparison of Figure 15(a) and Figure 15 (b), it is seen that
the data belonging to different classes are more separable for the gene subset selected
by CS method than that for the gene subset selected by SVM/MLP-CV method. This is
because the CS method takes into account the contribution of a gene to classification
ability on separating all class-pairs, rather than on separating only some specific pair of
classes. This makes it possible not to loss the genes which have relatively small
contribution values for separating some pairs of classes, but have great contribution
values for overall classification.
Problem of Correcting Tissue
Heterogenity Effect in Molecular
Microarray Profiles
As a common feature in microarray profiling, gene expression profiles represent a
composite of more than one distinct sources (i.e., the observed signal intensity will
consist of the weighted sum of activities of the various sources). Such a heterogeneity
effect would be real, reproducible, and potentially misleading, but would not be of
biological or clinical interest. This is why heterogeneity correction should be performed
to obtain microarray profiling which reflects the real biological molecular signature of the
underlying tissues.
Various methods have been proposed to correct tissue heterogeneity in microarray
studies. One of the main approaches is laser capture microdissection (LCM). The method
uses specialized technology to separate and isolate the cancer cells (or other tissue cell
subpopulations of interest) directly from the tissue. The molecular signatures in the
isolated subpopulations can then be analyzed individually. The disadvantage of LCM
is at the level of resources and expertise. Although LCM obtains pure tissue cell
subpopulations, the amount of material can be quite small and has to be analyzed with
appropriately sensitive analysis systems.
Independent component analysis (ICA) (Hyvarinen et al., 2000 , 2001; Chung et al., 2005;
Choudrey & Roberts, 2003; Deniz et al., 2003; Lu et al., 2000; Lotlikar & Kothari, 2000;
Gene Selection for Molecular Classification of Cancer 195
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Ristaniemi & Joutsensalo, 2000) is a newly-developed method that has been applied to
blind source separation (BSS) for decomposition of composite signals. It aims at
recovering the unobservable independent sources (or signals) from multiple-observed
data masked by linear or nonlinear mixing. Most existing algorithms for linear mixing
models stem from the theory of ICA. One of the basic assumptions for ICA model is the
statistical independence between components. However, the dependent components
are the often situation in the real world. Relaxing the very assumption of independence,
thus explicitly formulating new data models, is an approach in this point. Three recently
developed methods in this category are multidimensional ICA (MICA) (Hyvarinen et al.,
2001; Cardoso, 1998), independent subspace analysis (ISA) (Hyvarinen et al., 2000, 2001)
and topographic ICA (TICA) (Hyvarinen et al., 2001). In multidimensional ICA, it is
assumed that only certain sets (subspaces) of the components are mutually independent.
ISA is a closely related method where a particular distribution structure inside such
subspaces is defined. TICA, on the other hand, attempts to utilize the dependence of the
estimated independent components to define a topographic order, where the closer
components in topography are more dependent to each other.
Notice that ICA can only work well with a strict restriction that each component should
be statistically independent and non-gaussian distributed. This restriction does not hold
in many applications, especially in real microarray profiles. Of course, if independence
does not hold for some of component pairs, any connection to ICA would be lost, and
the model would not be useful in those applications where ICA has proved useful. Notice
that ICA can decompose the statistically independent signals, while the most genes in
microarray profiles function dependently. If the independent genes are selected at first
and ICA is applied to the microarray data with the expression values of only these
selected genes (this is what is referred to as partially independent component analysis,
PICA), the microarray profiling which reflects the real biological molecular signature of
the underlying tissues can still be obtained. However, how to select such genes still
remains an open problem due to its relation to up to infinitely-high-order statistical
properties. Notice that computational decomposition of molecular signatures for hetero-
geneity correction and molecular classification of cancer tissue for cancer diagnosis are
similar in their first step, or, gene selection. For computational decomposition of
molecular signatures to correct heterogeneity with partially independent component
analysis, we need to select genes which are of most functional independence; for
molecular classification of cancer tissues, we need to select genes that the samples with
expression levels of these genes are most separable. A novel computational heteroge-
neity correction method is developed in which informative genes are firstly selected with
the largest separability criterion, and then ICA is conducted on tissue samples with the
expression levels of only the selected genes for blind decomposition of the molecular
signatures in the microarray profiles.
By noticing that the expression values in gene microarray profiling are only non-
negative, a novel idea for heterogeneity correction is: it is a non-negative matrix
factorization (NMF) problem without enforcement of statistical characteristics on sources
and NMF technique is applied to heterogeneity correction of microarray profiles.
Simulations and experiments on real microarray data show effectiveness of the proposed
method.
196 Zhang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Heterogenity Correction with
Independent Component Analysis
Basic ICA Model
In basic ICA model, it is assumed that each observation is a linear mixture of independent
components s
i
, i.e.:
1 1 2 2
1, 2, ,
i i i in n
x a s a s a s i n = + + + = L L
,
,
or
1 1
2 2
n n
x s
x s
x s



=




A
M M

, or ,
= x As , (15)
where s is an n-dimensional random component vector, x is an n-dimensional random
observation vector, and A is an n by n mixing matrix. The statistically independence of
the components is one of the most critical assumptions in ICA model (the nongaussianity
of the components is another one). That means that the joint probability density function
(pdf) of the components and the marginal pdfs satisfy:
1 2
1
( , , , ) ( )
n
n i
i
p s s s p s
=
=

L .
. (16)
However, this restriction does not hold in many cases in the real world. Several people
talking to each other with a same topic is an example; The mixtures of several images with
a same or similar background is another example; Of course, if independence would not
hold for most component pairs, any connection to ICA would be lost, and the model
would not be very useful in those applications where ICA has proved useful.
Methodology of PICA
In contrast to basic ICA model, PICA model assumes that the components expressed by
s, i.e., s
1
,s
2
,L, s
n
may be statistically dependent, the dependence may be linear, or
nonlinear.
At the beginning, let us review the mechanism of independent component analysis on
basic ICA model. From the very starting point, the components are statistically
independent, while based on the Central Limit Theorem, the distribution of a sum
(observations) of independent random variables (components) tends toward a Gaussian
Gene Selection for Molecular Classification of Cancer 197
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
distribution, under certain conditions. Therefore, the procedure that ICA searches for
estimates of the components is to find directions, such that the projections of observa-
tions on each direction are distributed with most nongaussian distribution.
It is clear that ICA algorithm can help find the precise estimate (statistically precise) of
the components if the components are completely independent (except two ambiguities
of ICA, i.e., scale and order of the components). However, it will mislead direction finding
if the components are dependent but ICA algorithm is superimposed on the observations
which are mixtures from the dependent components. This results in estimates of virtual
components (the estimates of components), the components which are as independent
as possible. Then, what is the relation between the virtual components and the real
components (dependent)?
Let s' and A' be the virtual component vector and virtual mixing matrix obtained by
applying ICA directly on observations. We have:
s'=(A')
-1
As (17)
meaning that the virtual components are still linear mixture of real components. Compared
with observations, x, which are still linear mixtures of the real components, ordinarily
speaking, the virtual components are rough estimates of real components since they are
separation results by virtue of ICA.
Denote the component vector s as
s= [S
1,
S
2
, L, S
m
] (18)
where each column is a realization of random vector . We assume there are m realizations
in all. In order to find the precise estimates of the real components, a set of realization
indices I, I {1,2,...m}, should be found, such that
(a) s
1
={S
i
,iI} are independent, or say, the joint pdf and the marginal pdfs of random
variables over the realizations of is
1 2
1
( , , , ) ( )
n
I I nI iI
i
p s s s p s
=
=

L
(19)
(b) {s
1
, S
j
}are dependent for jI .
In other words, if we find the largest independent part of the real dependent components,
ICA model will be precisely satisfied on this part of the real components. Then ICA could
be performed on this part of the observations and this part of the real components
together with the real mixing matrix could be estimated precisely. Finally, precise
estimates of the real components can be obtained by s=A
-1
x.
198 Zhang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
The great difficulty for the resolution of this problem is that we want to find the largest
independent part of the real components, but the real components are unknown and are
the ones which remain to be estimated.
ICA is one of the approaches for blind separation of signals, while classification is an
approach to separate patterns. If each signal or random variable is viewed as a pattern
which represents that signal or random variable, it seems that there is some relation
between separation of signals and the pattern expression of the signals. This will be
studied in our next section.
Feature Analysis on Component Patterns
Now we denote the real components as:
[ ]
1 11 12 1
2 21 22 2
1 2
1 2
, , ,
m
m
m
n n n nm
s s s s
s s s s
S S S
s s s s



= = =



s
L
L
L
M L L L L
L
, (20)
where s
i
is the ith row of s indicating the ith component, S
j
(j=1,2,L,...m) is the jth column
of s indicating the jth realization of random vector s, and s
ij
is the jth realization of ithe
th components. Thus, each component s
i
is expressed by
1 2
( , , , )
i i i im
s s s s = L . . (21)
This means that it is s
i
rather than sjsince its pattern corresponds to
1 2
( , , , )
i i im
s s s L . .
Therefore, each component s
i
can be thought of as a pattern defined by a vector in m-
dimensional feature space and its jth element s
ij
is the jth feature of this component. The
feature which can be induced from another or other features should be removed for more
compact representation of the components and for dimension reduction (Tenenbaum,
Silva, & Langford, 2000; Pudil & Novovicova, 1998). Of course, the only features which
are independent to each other are not enough for representation of the components
except they are the largest independent features to wholly express the features of the
components.
Feature selection becomes more and more important since the number of dimensions (the
number of features) of the data set met in the real world becomes greater and greater
compared with the number of samples in the feature space (Haykin, 1999). Otherwise, the
curse of dimensionality problem emerges. The curse is that as the dimensionality of data
increases, they become dramatically more and more sparse in a multidimensional space
(Cardoso , 1998). One example is microarray data set, where there are thousands of genes
(dimensions) and only a few samples (experiments) which belong to several classes
Gene Selection for Molecular Classification of Cancer 199
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
(disease categories). Only a few components but each have so many realizations (each
realization corresponds to a dimension of the feature space that component falls in)
shows us another example.
Pattern recognition is basically a classification problem combined with dimensionality
reduction of pattern features that serve as the input to the pattern classifier. This
reduction is a preprocessing of the data for obtaining a smaller set of representative
features and retaining the optimal salient characteristics of the data. This will result in
not only decreases of the processing time but also leads to more compactness of the
model for classification task and better generalization of the designed classifier. The
principle of feature selection for classification is mainly based on maximizing the
separability of patterns in the reduced dimensional feature subspace.
For statistical independence of feature i and feature j, we can have the following two
understandings:
(a) They are independent over projections of all data samples onto the subspace
spanned by feature i and feature j (class unrelated).
(b) They are independent over projections of all class centers onto the subspace
spanned by feature i and feature j (class related).
The feature selection for pattern classification should be the latter (class related). In our
analysis, each real component is viewed as a class center which is the center of several
virtual samples in the class, and the representation of the component, s
i
=(s
i1
,s
i2
,...,s
im
),
is the feature pattern of that class.
In order to understand the statistical relation of feature and feature , to know if they are
independent or not, a large number of class centers is needed since only a large enough
class centers can generate a distribution. However, in our ICA situation, the number of
components is much less than the number of features which represent those components.
The situation here is exactly the same with microarray data situation where the number
of samples in feature space is too small to generate a distribution in the feature space for
understanding the statistical relation of each pair of features. To tackle this problem, a
component space is defined, where each axis of the space corresponds to a component,
and each sample in this space corresponds to a feature.
The dependence of components is defined by the distribution of the features (feature
scatter plot) in component space and the dependence of the features is defined by the
distribution of component vectors in feature space (even though the number of compo-
nents is small compared with the number of dimensions of the feature space). ICA is an
approach to tackle the blind separation of components, which has an assumption that
the components should be independent to each other, but we assume the components
are not independent to each other. Feature selection aims at finding a feature subset to
maximize the separable power of the classes. We have made a connection of component
and its feature representation in our above study. Thus, we can study the independence
of the components in its feature subspace and the separability of the components in this
subspace.
200 Zhang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Assume that I
( {1, 2, , }) m L be features which are selected by some feature selection
method, and span an m
1
-dimensional feature subspace. It is reasonable that we use
Euclidian distance between n
1 2
, s s , i.e., [d(s
1
, s
2
)]
2
= (s
1j
s
2j
)
2
, to measure the separability
of components in feature space. We have

2 2
1 2 1 2
2 2
1 2
2 2
1 2 1 2 1 2
[ ( , )] ( )
( )
( ) ( , )
j j
j
d s s s s
m E s s
m s s p s s ds ds
=

=

(22)
and
E(s
1I
s
2I
)
2
= E(s
1I
2
) + E(s
2I
2
) 2E(s
1I
s
2I
) (23)
where the first two terms are the separability contributed by features I with an assumption
that the projections of the components s
1
,s
2
, i.e., s
1I
,s
2I
, are independent in component
space, and the last term is the separability contributed by the correlation of components
s
1I
,s
2I
in the same space. Obviously, due to the possibility of E(s
1I
,s
2I
) either greater or less
o b s e r v a t i o n s
o b s e r v a t i o n s
d e m i x e d w i t h I C A
d e m i x e d w i t h I C A
o b s e r v a t i o n s ( s . i . )
o b s e r v a t i o n s ( s . i . )
d e m i x e d w i t h P I C A ( s . i . )
d e m i x e d w i t h P I C A ( s . i . )
d e m i x e d w i t h P I C A
d e m i x e d w i t h P I C A
s o u r c e s
s o u r c e s

(a) (b) (c) (d) (e) (f)
Figure 9. Blind separation with ICA and PICA of the two dependent sources, where one
source is a transpose of the other source: (a) Observations, (b) dexmixed with ICA, (c)
observations (s.i.), (d) demixed with PICA (s.i.), (e) demixed with PICA, (f) sources
j
Gene Selection for Molecular Classification of Cancer 201
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
than 0, the separability of s
1I
,s
2I
can either be better or worse than that when no correlation
exists between s
1I
,s
2I
.
Assume that I
1
/I
2
are two separate feature subsets such that
1 1
1 2
,
I I
s s /
2 12
1 2
,
I I
s s aare indepen-
dent in component space, and assume
2 1
I I . . We have
2 2 1 1
2 2
1 2 1 2
[( ) ] [( ) ]
I I I I
E s s E s s > . (24)
Figure 10. Blind separation of tire and coin images with ICA and PICA: (a) Observations,
(b) results, (c) observations (s.i.),(d) ICA results (s.i.), (e) PICA results, (f) sources


Figure 11. Scatter plots of the sources(a), observations(b), ICA estimated sources(c)
and PICA estimated sources(d) as well as the scatter plots of the observations(e)/ICA
estimated sources(f) with only the selected indices
0 1 2 3
-4
-2
0
2
source
-2 0 2
-2
0
2
4
observation
-1 0 1 2 3
-4
-2
0
2
ICA result
-2 0 2
-2
0
2
4
observation(s.i.)
-1 0 1 2
-3
-2
-1
0
1
2
ICA result(s.i.)
0 1 2 3
-4
-2
0
2
PICA result

(a) (b) (c) (d) (e) (f)
202 Zhang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Therefore, the largest independent part of the components (in component space) results
in the greatest separability of the component patterns (in feature space).
All in one, under the assumption that we have a compact representation of component
features, it is coincident to find the largest independent part of the components (in
component space) and to select a feature subset (in feature space) which can maximize
the separability of the component patterns in feature subspace defined by the selected
features.
Figure 12. Comparison of the mixing matrix between the one estimated via ICA, PICA
as well as the real one

0.5
1
30
210
60
240
90
270
120
300
150
330
180 0
real mixing matrix
0.5
1
30
210
60
240
90
270
120
300
150
330
180 0
result from ICA
0.5
1
30
210
60
240
90
270
120
300
150
330
180 0
result from PICA

Figure 13. Heterogeneity correction result (c) from the microarray profiles (b) and its
comparison with the ground truth (a)
Gene Selection for Molecular Classification of Cancer 203
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
PICA Algorithm
C-space method is a feature selection method we developed for gene selection, which
has been studied in the third section. It converts the feature selection problem in feature
space into contribution calculation of each feature to classification task in class space
and the top M features in their contributions are selected. M is an integer parameter for
our C-space feature selection method. Since selecting the features that make the
component patterns to be best separable in the feature subspace defined by these
features is coincident with finding the greatest independent part of the components in
feature space, we apply our C-space method for PICA. The PICA algorithm is as follows:
Step1. perform ICA on observations. A very rough estimate of the real component vector,
i.e., , and a very rough estimate of the real mixing matrix , i.e., s' are obtained, due to the
dependence of the real components A; A' each row of s' is the feature pattern of the
corresponding component.
Step 2. calculate the contribution of each feature for separation of all the component
patterns by applying C-space method, and rank the features with descending order
according to their contributions; Select top M features as the feature subset I.
Step 3. perform ICA on rough estimates of the components s', i.e., s'
I
. Then the estimate
of mixing matrix A
2
is obtained.
Step 4. the final estimate of the real mixing matrix is A'A
2
, and the final estimate of the
real component vector is (A'A
2
)
-1
x.
In the above procedure, ICA is performed twice, first to blindly separate observations
into rough estimates of the real components, and then to blindly separate observations
which are represented with only compact features into precise estimates of the real
Figure 14. Comparison of the scatter plots between the ground truth and the
heterogeneity correction result obtained by PICA for the number of genes selected
equals to64, 12(a), 128, 12 (b), and 2308, 12(c)
(a) (b) (c)
204 Zhang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
components for those compact features. With this method, the mixing matrix and the
components are finally estimated precisely.
The features which will be fed to the second ICA calculation are selected with the
principle of selecting features with the greatest separability of the component patterns
in feature subspace defined by the selected features. This is compatible to finding the
greatest independent part of the components. Therefore, we can think of the features
selected to be the independent part of the components (we do not declare any proof on
this point). Then we can make an alternative for step 3 and step 4 with the following steps:
Step 3. perform ICA on the projections of the observations x in the feature subspace,
i.e., x
I
, and obtain precise estimate of the real mixing matrix, A.
Step 4. the final estimate of the real component vector is A
-1
x.
Simulations and Heterogeneity Correction Experiments
A large number of experiments for testing and comparison are performed, including
simulation experiment and application of PICA to the resolution of partial volume
correction for real microarray data set. In all of our simulation experiments, each element
in mixing matrix is a uniformly distributed random variable ranged in [0,1] and mixing matrix
has independent elements. The symbol s.i. in our following figures indicates the
indices of selected features by C-space feature selection method. In the following, we
exemplify only two simulation results because of limitation of the space.
Our aim is to separate two 242242 images, coin image and its transpose, from two
mixtures of them. Evidently, the component sources are dependent again. The PICA is
performed with the number of features selected to be , i.e., 1000 out of 242242=58564
features are selected. Much better estimation performance for components is again
obtained by comparison of the closeness between Figure 9 (e) and 9 (f) and the closeness
between Figure 9 (b) and Figure 9 (f)
For separating the mixture of tire image and coin image (both are in the size of 205232),
feature selection parameter M is set to M = 3000, i.e., the mixture images with 30000 out
of 205232=47560 selected features are fed to PICA algorithm. The performance of
component estimates from PICA and its comparison to that from ICA can be seen in Figure
10 and Figure 11. The scatter plot of the features in component space sketched in Figure
6.3 also shows the much better performance of PICA than that of ICA on estimates of the
components. In practice, the correlation coefficient of the pure tire and coin image is -
0.1843, while the correlation coefficient of them with only selected features becomes only
0.00092433. This means that C-space feature selection with an appropriate choice of
parameter M do functions of decorrelation, which leads to components represented in
the corresponding feature subspace to be much more independent, while independence
is a prerequisite of applying basic ICA for getting better recoveries.
9(c) and 9 (d), and 10 (c) and 10 (d) help us to have a better understanding on how PICA
works. The observations with only selected features (pixels) and the estimates of
Gene Selection for Molecular Classification of Cancer 205
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
components (source images) with these features (pixels) are shown in Figure 9(c) and 9
(d), and 10 (c) and 10 (d), where the pixels with the lowest intensities (black pixels) are unselected
ones. It is seen that the features (pixels) which are selected are nearly the overlap pixels of the
objects in two source images, while the features (pixels) which could be though of as common
background of the source images are not selected by the feature selection algorithm.
The effectiveness of our PICA method was also tested on NCI data for partial volume
correction (PVC) problem. The Source and mixed signals are given in Figure 13 respec-
tively. The original 2308 genes were ranked according to their separable power to the two
classes and the top 64/128 features (genes) are selected for ICA analysis. The original
expression levels (reflected by intensities of the array image) of the top 64/128 features,
mixed expression levels and the separation result are shown in Figure 13 (a) and (b) and
(c) respectively (the upper part corresponds to the situation of 64 genes, the lower part
corresponds to the situation of 128 genes). By comparison of Figure 13 (a) and Figure
13 (c), the blind source separation by PICA recovered the original array almost perfectly.
Many other independent trials using other gene sets reached a similar result.
The scatter plots of genes in component space are shown in Figure 14 for the number of
selected genes being M = 64,128,2308 from top to down and for the case when all of the
M = 2308 genes (the bottom scatter plot in Figure 14) are selected. Our experiment
indicates that the scatter plot that the estimates of the sources are almost the real sources
for M 256. This indicates that the statistically independent genes were retained
for M 256. That the estimates of the sources are very far from the real sources for too
large M, say for M 256, indicates that if too large number of genes are retained for PICA,
some dependent part of the sources are also retained, which will mislead direction
searching in the following-up ICA analysis. Therefore, the number of features selected
is an important parameter and problem dependent in application of PICA to real world
problems.
Heterogeneity Correction with
Non-Matrix Factorization (NMF)
In this section, we propose a tissue heterogeneity correction approach based on non-
negative matrix factorization (NMF) by noticing that the microarray profiles are non-
negative in gene expression levels. NMF is a newly developed approach to factorize a
non-negative matrix into two non-negative matrices, and has found wide applications in
feature extraction (Lee & Seung, 1999; Guilamet & Vitria, 2001, 2003; Li et al., 2004; Liu
& Zheng, 2004). Assume that V is a n x m non-negative matrix showing m samples in n
dimensional space. NMF is to factorize V into WH where W and H are n x r and r x m non-
negative matrix respectively such that some cost function is the minimum. The two typical
cost functions are
1
[ log( ) ( ) ]
iu iu iu
i u
F V WH WH =

and
2
2
( ( ) )
iu iu
i u
F V WH =

. The
non-negative constraints of NMF model lead to a part-based representation because
they allow only additive, not subtractive, combinations. NMF is already successful in
learning facial parts and semantic topics (Lee & Seung, 1999). Also, Lee and Seung
206 Zhang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
proved the convergence of NMF algorithms for the above two cost functions respec-
tively (Daniel & Seung, 2001).
Matrix Factorization and Blind Source Separation with
Non-Negative Matrix Factorization
BSS is a very active topic recently in signal processing and neural network fields
(Hyvarinen et al., 2001; Hoyer & Hyvrinen, 2000). It is an approach to recover the sources
from their combinations (observations) without any understanding of how the sources
are mixtured. For a linear model, the observations are a linear combination of the sources,
i.e., X = AS, where S is an r n matrix indicating r source signals in n dimensional space,
X is an m n matrix showing observations in dimensional space, and A is an m r mixing
matrix. Therefore, BSS is a matrix factorization, to factorize observation matrix V into
mixing matrix A and source matrix S.
An NMF problem is: given a non-negative n m matrix V, find non-negative n r and
matrix factors W and H such that the difference measure between V and WH is the minimum
according to some cost function, i.e.:
Figure 15. Dependent source separation with NMF: (a) Real sources; (b) observations;
(c) recovered sources with ICA method; (d) recovered sources with NMF; (e) cost
function vs. iterations for NMF

(a) (b) (c) (d)

(e)
Fig. 7.1 dependent source separation with NMF, (a) real sources; (b) observations;
(a) (b) (c) (d)
(e)
Gene Selection for Molecular Classification of Cancer 207
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
V WH (25)
NMF is a method to obtain a representation of data using non-negative constraints.
These constraints lead to a part-based representation because they allow only additive,
not subtractive, combinations of the original data. For the ith column of equation (25),
i.e., , where v
i
and hi is the ith column of V and H, the ith datum (observation) is a non-
negative linear combination of the columns of (W
1
,W
2
,...,W
r
), while the combinatorial
coefficients are the elements of . Therefore, the columns of , i.e.,
i i
v Wh = ,, can be viewed
as the basis of the data V when V is optimally estimated by its factors.
If (W
1
,W
2
,...,W
r
) are considered to be r sources/basis of the observation space spanned
from the columns of V, notice that NMF does not constrain any statistical properties on
the sources. However, we will show that they have some association: they should be of
Figure 16. Partial Volume Correction (PVC) with ICA and NMF: (a) BSS results with
ICA and NMF, and (b) the scatter plots of the recovered sources

(a)

(b)
Fig. 7.2 Partial Volume Correction (PVC) with ICA and NMF, (a) BSS
208 Zhang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
linear independence (as the basis of a space) which will induce some degree statistical
independence (i.e., statistical dependence) in statistical meanings. Denote w
i
to be a
random variable where each element in W
i
is its realization. Notice that there may be two
parts of correlation between observations. One is from the model, where the observations
are linear combinations of sources, which leads to first-order statistical original depen-
dence between observations, and another, which comes from the original dependent
sources themselves, may exist according to the sources. It is clear that the task of BSS
is to decorrelate the observations such that the recovered sources has the lowest origin-
based correlation coefficient:
2 2
( )
( , )
( ) ( )
T
i j i j
i j
i j
i j
E ww W W
w w
W W
E w E w
= =

(26)
rather than the center-based one
2 2
[( )( )]
( , )
[( ) ] [( ) ]
i i j j
i j
i i j j
E w w w w
w w
E w w E w w

=


Consider the case that W
i
and W
j
are orthogonal, which is an extreme case of linear
independence between W
i
and W
j
. The correlation coefficient ( , )
i j
w w becomes 0,
indicating that W
i
and W
j
are first-order origin-based statistically independent. In contrast
to linear independence between W
i
and W
j
, let us consider the case whereW
i
and W
j
are
linearly dependent, i.e.,
i j
W kW = , where is a constant. We have ( , ) 1
i j
w w = , , indicating
that W
i
and W
j
are totally first-order origin-based statistically dependent.
In fact, the linear independence between the columns of themselves corresponds to some
degree first-order origin-based statistical independence (i.e., statistical dependence),
i.e., the correlation coefficient
( , ) [0,1)
i j
W W
. This becomes our fundamental under-
standing that NMF can help to find the basis of the data space which are some degree
first-order origin-based statistically independent (i.e., statistical dependent). From this
understanding, NMF can be used for decorrelating observations to recover the sources
that are not necessarily statistically independent and/or Gaussian distributed. This is the
reason why our experiments with NMF in the fourth section worked well for non-negative
BSS even when there is statistically linear dependence between the sources. Therefore,
when NMF is used for non-negative BSS, the ICA transpose model, i.e.:
T T T
X S A =
(27)
is used, and the transpose of the factored W will be the recovered source matrix S.
Gene Selection for Molecular Classification of Cancer 209
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
In our following experiments, the NMF algorithm presented in Lee and Seung (1999), i.e.,
( )
( )
i
ia ia a
i
ia
ia
ja
j
i
a a ia
i i
V
W W H
WH
W
W
W
V
H H W
WH

(28)
is used. The convergence of the algorithm is proved in Daniel and Seung (2001).
Blind Source Separation and Heterogeneity Correction
Experiments
We did a lot of simulation experiments on non-negative BSS with NMF. The BSS
simulation experiments are divided into three groups: statistically independent / depen-
dent sources, statistically independent Gaussian distributed sources, and linear bar
separation problem (dependent sources) respectively. In the following, only the linear
bar separation is illustrated because of the limitation of the space, even though good
results were obtained for the other group of simulation experiments. Notice that in our
simulation, each element of mixing matrix is a random number that follows [0,1] uniform
distribution, and fastICA algorithm (Foldiak, 1990) is used when ICA method is applied
for BSS. Then we will illustrate our NMF method for heterogeneity correction of the real
world microarray data.
The linear bar problem (Foldiak, 1990) is a blind separation of bars from their combina-
tions. Eight feature images (sources) sized 4 4 including four vertical and four horizontal
bar images are randomly mixtured to form 20 observation images. It is very clear that there
exist highly statistical correlation between sources. For example, the background of the
images is the same for all of the sources. Therefore, ICA does not work well. Figure 15
demonstrates our result of the recovered sources from NMF method compared with that
from ICA method. It is seen from Figure 15 that only seven sources rather than eight were
found with ICA method, and the recovered seven sources are very far from the real
sources, while the recovered eight sources from NMF are very close to the real sources,
indicating that NMF is effective in separating sources where the sources are highly
dependent to each other.
The effectiveness of our NMF method was also tested on NCI data for heterogeneity
correction. Two Sources (samples from two diseases respectively), two mixed signals,
recovered sources from ICA and the ones from NMF are given in Figure 16 (a)
respectively. Notice that the true sources are determined, in our present case, by
210 Zhang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
separately profiling the pure cell lines that provide the ground truth of the gene
expression profiles from each cell populations. In our clinical case, we use laser-capture
microdissection (LCM) technique to separate cell populations from real biopsy samples.
By comparison of the second and the third column in Figure 16 (a), the blind source
separation by NMF method recovered the original array very well. Many other indepen-
dent trials using other gene sets reached a similar result.
Our experiments also showed that the NMF algorithm sometimes does not converge to
desired solutions. For example, we mixed three independent Gaussian distributed signals
with a linear model and got three observation signals. However, the real sources are much
less likely to be well-recovered by the application of NMF algorithm to these observa-
tions. To our knowledge, it seems that there are two main reasons for NMF to converge
to undesired solutions. One is that the basis of a space may not be unique theoretically,
and therefore separate runs of NMF may lead to different results. Another reason may
come from the algorithm itself that the cost function sometimes gets stock into local
minimum during its iteration, the same situation as that in ICA algorithm. From the scatter
plots shown in Figure 16 (b), it is clear by comparison of the third column (the scatter plot
of the recovered sources) and the fourth column (the scatter plot of the real sources) that
the NMF decorrelates the origin-based correlation between observations to recover
sources without consideration of whether such correlation in observations comes from
the linear combination of the sources or from the sources themselves. This also tells us
that the NMF method presented here for non-negative BSS still has room to be improved.
Summary
In this chapter we present material dealing with the jointly discriminatory gene selection
for molecular signature decomposition. It is appropriate that we now reflect over that
material and ask: how useful is jointly discriminatory gene selection? The answer to this
question, of course, depends on the application of interest.
If the main objective is to achieve high classification rate for tissue sample classification
and good generalization performance for cancer prediction, the use of jointly discrimi-
natory gene selection offers a useful learning procedure, and a compact and intrinsic
representation for good data compression while preserving as much information about
the inputs as possible for classification and prediction of cancer. Firstly, it can both
compress the microarray data and preserve the separable information for classification
of tissue samples. Secondly, since the separable information for classification of tissue
samples has its own intrinsic dimensions which is less than the dimensionality of the
sample space due to the huge dimensionality of the space and a relatively small number
of tissue samples in the space, a representation of a tissue sample based on the selected
jointly discriminatory genes is preferable and more compact for classification, which will
result in higher classification rate, compared to their original microarray gene profiles of
the tissue samples. The more compact and intrinsic representation of the tissue samples
by their expression levels of the selected jointly discriminatory genes may make the
classifier to be more effective and efficient, i.e., the classifier easier to design both in
structure and parameters, and with better generalization potentials.
Gene Selection for Molecular Classification of Cancer 211
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Heterogeneity correction of gene microarray profiles is another challenging problem that
we tackled in this chapter, which is also related to jointly discriminatory gene selection.
For heterogeneity correction to get a molecular decomposition of the microarray profiles
at hand, independent component analysis is not applicable due to the fact that the
sources are mostly statistically dependent, which disabbays the independence assump-
tion of the basic ICA model. In order for ICA to be still applicable for heterogeneity
correction of gene microarray profiles, functionally independent genes needs to be
selected. Due to the difficulty of finding functionally independent genes, jointly
discriminatory genes are selected instead. By noticing that jointly discriminatory genes
are in some sense, the genes which are most independent in the function of classification
of cancer, a novel idea, first selecting jointly discriminatory genes and then performing
ICA on microarray profiles with the expression levels of only selected genes, is
presented. Accordingly, separation of tissue samples and heterogeneity correction of
tissue samples via jointly discriminatory gene selection are compactly correlated in some
technical point of view.
References
Anil, K., Robert, P. R., & Mar, J. (2000). Statistical pattern recognition: A Review. IEEE
Trans. on Pattern Analysis and Machine Intelligence, 22(1), 4-37.
Cardoso, J. F. (1998). Multidimensional independent component analysis, In Proc. IEEE
Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP98). Seattle, WA.
Choi, K. (2002). Input feature selection for classification problems. IEEE Trans. on
Neural Networks, 13(1), 143-159.
Choudrey, R. & Roberts, S. (2003). Variational mixture of bayesian independent compo-
nent analyzers. Neural Computation, 15, 213-252.
Chung, S. H., Park, C. S., & Park, K. S. (2005). Application of independent component
analysis (ICA) method to the Raman spectra processing. Proceedings of SPIE,
5702, 168-172.
Daniel, D. L., & Seung, H. S. (2001). Algorithms for non-negative matrix factorization.
Advances in Neural Information Processing Systems, 13, 556-562.
Dniz, O., Castrilln, M., & Hernndez, M. (2003). Face recognition using independent
component analysis and support vector machines. Lecture Notes in Computer
Science, 2091, 59.
Dudoit, S., Fridlyand, J., & Speed, T. P. (2002). Comparison of discrimination methods
for the classification of tumors using gene expression data. J. Am. Stat. Assoc,
97(457), 77-87.
Foldiak, P. (1990). Forming sparse representations by local anti-Hebbian learning.
Biological Cybernetics, 64, 165-170.
Guillamet, D., Vitri, J. (2003). Evaluation of distance metrics for recognition based on
non-negative matrix factorization. Pattern Recognition Letters, 24(9-10), 1599-
1605.
212 Zhang
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Guillamet, D., & Vitria, J. (2001). Unsupervised learning of part-based representations.
Lecture Notes in Computer Science, 2124, 700.
Guillamet, D., & Vitri, J. (2003). Evaluation of distance metrics for recognition based on
non-negative matrix factorization. Pattern Recognition Letters, 24(9-10), 1599-
1605.
Guillamet, M. B., & Vitria, J. (2001). Weighted Non-negative matrix factorization for local
representations. Proceedings of 2001 Computer Vision and Pattern Recognition.
Retrieved from http://citeseer.nj.nec.com/context/2095821/0
Guyon, I., Weston, J., Barnhill, S., & Vapnik, V. (2002). Gene selection for cancer
classification using support vector machines. Machine Learning, 46(3), 389-422.
Haykin, S. (1999). Neural Networks: A comprehensive foundation. Prentice Hall Inc.
Herrero, J., Valencia, A., & Dopazo, J. (2001). A hierarchical unsupervised growing neural
network for clustering gene expression patterns. Bioinformatics, 17(2), 126-136.
Hoyer, P. O., & Hyvrinen, A. (2000). Independent component analysis applied to feature
extraction from color and stereo images. Network: Computation in Neural Systems,
11(3), 191-210.
Hyvrinen, A. (1999). Sparse Code Shrinkage: Denoising of nongaussian data by
maximum likelihood estimation. Neural Computation, 11, 1739-1768.
Hyvrinen, A., Hoyer, P. O., & Inki, M. (2001). Topographic independent component
analysis. Neural computation, 13, 1527-1558.
Hyvrinen, A., Karhunen, J., & Oja, E. (2001). Independent component analysis. New
York: John Wiley.
Hyvrinen, A., & Hoyer, P. O. (2000). Emergence of phase and shift invariant features by
decomposition of natural images into independent feature subspaces. Neural
Computation, 12(7), 1705-1720.
Hyvrinen, A., & Oja, E. (2000). Independent Component Analysis: Algorithms and
applications. Neural Networks, 13, 411-430.
Khan, J., Wei, J. S., Ringner, M., Saal, L. H., Lananyi, M., Westermann, F., et al. (2001).
Classification and diagnostic prediction of cancers using gene expression profiling
and artificial neural networks. Nature Medicine, 7(6), 673-679.
Lee, D., & Seung, H. S.(1999). Learning the parts of objects by non-negative matrix
factorization. Nature, 401, 788.
Li, Y. Q., Cichocki, A., & Amari, S (2004). Analysis of sparse representation and blind
source separation. Neural Compupation, 16, 1193-1234.
Liu, W. X., Zheng, N. N. (2004). Non-negative matrix factorization based methods for
object recognition. Pattern Recognition Letters, 25(8), 893-897.
Loog, M., Duin, & R. P. W. (2001). Multiclass Linear fimension teduction by weighted
pairwise fisher criteria. IEEE trans. on Pattern Analysis and Machine Intelligence,
23(7), 762-766.
Lotlikar, R., & Kothari, R. (2000). Fractional-step dimensionality reduction. IEEE Trans.
on Pattern Analysis and Machine Intelligence, 22(8), 623-627.
Gene Selection for Molecular Classification of Cancer 213
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Lu, C. S., Huang, S. K., Sze, C. J., & Liao, H. Y. M. (2000). Cocktail watermarking for digital
image protection. IEEE Trans. on Multimedia, 2(4), 209-224.
Pudil, P., & Novovicova, J. (1998). Novel methods for subset selection with respect to
problem knowledge. IEEE Intelligent Systems, 13(2), 66-74
Ristaniemi, T. (2000). Synchronization and blind signal processing in CDMA systems.
Unpublished doctoral dissertation, University of Jyvaskyla, Jyvaskyla, Finland.
Ristaniemi, R., & Joutsensalo, J. (2000). Advanced ICA-based receivers for DS-CDMA
systems. Proceedings of IEEE Int. Conference on Personal, Indoor, and Mobile
Radio Communications (PIMRC00). London.
Tenenbaum, J. B., Silva, V. D., & Langford, J. C. (2000). A global geometricframework for
nonlinear dimensionality reduction. Science, 290(22), 2319-2325.
Toronen, P. (2004). Selection of informative clusters from hierarchical cluster Tree with
Gene Classes. BMC Bioinformatics, 5(1), 32.
Trunk, G. V. (1979). A Problem of Dimensionality: A simple example. IEEE Trans. on
Pattern Analysis and Machine Intelligence, 1(3), 306-307.
Wang, Y., Zhang, J., Huang, K., Khan, J., & Szabo, Z. (2002). Proc. IEEE Intl. Symp.
Biomed. Imaging (pp.457-460). Washington, DC.
Wang, Y., Zhang, J., Wang, Z., Clarke, R., & Khan, J. (2001). Gene selection by machine
learning in microarray Studies (Tech. Rep. No. CUA01-018). United States,
Washington, DC.: The Catholic University of America.
Xiong, M., Fang, X., & Zhao, J. (2001). Biomarker identification by feature wrappers.
Genome Research (see www.genome.org), 11, 1878-188.
Zhang, J., Wei, L., & Wang, Y. (2003). Computational decomposition of molecular
signatures based on blind source separation of non-negative dependent sources
with NMF. In 2003 IEEE 13th Workshop on Neural Networks for Signal Process-
ing (NNSP03), Toulouse, France (pp. 409- 418).
214 Kido
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Chapter XI
A Haplotype
Analysis System for
Genes Discovery of
Common Diseases
Takashi Kido, HuBit Genomix, Inc., Japan
Abstract
This chapter introduces computational methods for detecting complex disease loci
with haplotype analysis. It argues that the haplotype analysis, which plays a major role
in the study of population genetics, can be computationally modeled and systematically
implemented as a means for detecting causative genes of complex diseases. In this
chapter, the author provides a review of issues on haplotype analysis and proposes the
analysis system which integrates a comprehensive spectrum of functions on haplotype
analysis for supporting disease association studies. The explanation of the system and
some real examples of the haplotype analysis will not only provide researchers with
better understanding of current theory and practice of genetic association studies, but
also present a computational perspective on the gene discovery research for the
common diseases.
A Haplotype Analysis System for Genes Discovery of Common Diseases 215
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Introduction
In recent years, much attention has been focusing on finding causative genes for
common diseases in human genetics. (Badano & Katsanis, 2002; Daly, 2001; Fan &
Knapp, 2003; Gabriel et al., 2002) These findings of causative genes of common diseases
including diabetes, hypertension, heart disease, cancer, and mental illness are expected
to be opening doors for realizing new diagnoses and drug discoveries. A promising
approach for the gene discovery on common diseases is to statistically examine genetic
association between the risk of common diseases and DNA variations in human
populations. While single nucleotide polymorphisms (SNPs), the most common genetic
variation, are widely used for this genetic association study, haplotypes, the combina-
tion of closely linked SNPs on a chromosome, has been shown to have pivotal roles in
the study of the genetic basis of disease (Clark, 2004; Niu, 2004; Schaid, 2004). The main
purpose of this report is to provide a comprehensive review of haplotype analysis in
genetic association studies on complex, common diseases and provide the computa-
tional framework which enables us to carry out successful high-throughput genome-
wide association study.
In addition to the review of the recent developments of haplotype analysis, the author
presents the design, implementation, and application of a haplotype analysis system for
supporting genome-wide association study. While there are some useful tools or
programs available for haplotype analysis (Kitamura et al., 2002; Niu, Zin, Zu, & Liu, 2002;
Sham & Curtis, 1995; Stephens, Smith, & Donnelly, 2001), little work has been reported
for a comprehensive analysis pipleline for large-scale and high-throughput SNPs
screening which fully integrate these functions. HAPSCORE (Zhang, Rowe, Struewing,
& Buetow, 2002) is one of the few examples of those pipeline systems; however, it does
not include some analysis functions such as automatic linkage disequilibrium (LD) block
partitioning and disease association analysis tools. In this report, the author presents
a system, LDMiner (Higashi et al., 2003), which represents the pioneer pipeline system
that integrates a comprehensive spectrum of functions related to haplotype analysis.
This report introduces the details of LDMiner and shows some examples of haplotype
analysis with LDMiner, which helps to explain the theory and practice on population-
based association study for common diseases.
Background
Genetic Variations and Common Diseases
The progress on human genome science is opening doors for the discovery of new
diagnostics, preventive strategies, and drug therapies for common complex diseases
including diabetes, hypertension, heart disease, cancer, and mental illness. Analysis of
human genome primarily focuses on variations in the human DNA sequence, since these
differences can affect the potential risk of disease outbreaks or the effectiveness of a drug
treatment of the diseases.
216 Kido
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
A common method for determining the genetic differences between individuals is to find
single nucleotide polymorphisms (SNPs). A SNP is defined as a DNA sequence variation
referring to an alteration of a single nucleotide (A, T, G, C). SNPs represent the most
common genetic variations. In fact, there are millions of SNPs in the human genome
(Kruglyak & Nickerson, 2001), and it is estimated that there will be on average one SNP
every 1,000 base pairs. SNPs are caused when nucleotides replicate imperfectly or mutate.
Although most of these SNPs have no ostentatious impacts on the survival of the
species, certain SNPs may confer beneficial effects allowing species to evolve and to
adapt to new environments more successfully, while certain others may be detrimental.
These SNPs are passed on from generation to generation. After hundreds of years, some
SNPs become established in the population.
A number of instances are known for which a particular nucleotide at a SNP locus (i.e.,
a particular SNP allele) is associated with an individuals propensity to develop a disease.
For example, it has been reported that functional SNPs in the lymphotoxin-alpha gene
were associated with susceptibility to myocardial infection by means of a large-scale
case-control association study using 92,788 SNP markers (Ozaki et al., 2002). There have
also been a number of reports that show some SNPs in certain genes can determine
whether a drug can treat a disease more effectively in individual with certain genotypes
compared to those who do not carry such SNPs. For example, Cummins et al. (2004)
reported that there is a strong association in Han Chinese between a genetic marker, the
human leukocyte antigen HLA-B*1502, and Stevens-Johnson syndrome induced by
carbamazepine, a drug commonly prescribed for the treatment of seizures. These SNPs
are often linked to the causative genes, but may not be themselves proved to be
causative. These are often called surrogate markers for the disease. These disease-
associated SNPs are expected to be validated by a number of investigations including
biological experiments and epidemiological studies in the next several years.
Why Study Haplotypes?
A haplotype is a set of closely linked alleles (SNPs) inherited as a unit. For example, lets
assume that there is a gene containing three SNPs. We now represent two alleles of each
SNP with A and B (A is minor allele). The number of possible combinations of the three
SNPs is eight (i.e. A-A-A, A-A-B, A-B-A, A-B-B, B-A-A, B-A-B, B-B-A, B-B-B).
However, the total number of common (i.e. frequency > 5%) haplotypes in the population
usually converge to a number less than eight, for example, three (A-A-A, A-B-A, B-B-
B). Also, the number of common haplotypes varies depending on the chromosome
regions. The recent empirical findings in human genetic studies show that genomic
sequence is comprised of parts; cold spots (LD blocks) having much less variations than
other regions and hot spots having more variations (Cardon & Abecasis, 2003). Some
recombination hot spots tend to include a large number of haplotypes, while cold spots
(or LD blocks) include only small number of haplotypes.
Haplotypes play important rules for searching the causative sites associated with several
common diseases. As Clark (2004) discussed the role of haplotypes in candidate gene
studies, there are three primary reasons for considering the haplotype configuration,
which are listed as follows.
A Haplotype Analysis System for Genes Discovery of Common Diseases 217
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Biological function: Haplotypes may be defined as functional units of genes; the
protein products of the candidate genes may depend on particular combinations
of amino acid. For example, ApoE is a protein whose function is influenced by a
pair of polymorphic amino acids (Fullerton et al., 2000). The two-site haplotypes
best describe the functional differences of carious ApoE protein isoforms rather
than those of individual SNPs in the gene.
Tracing past evolutionary history: The variation in populations is inherently
structured into haplotypes. The haplotypes cover the informative small segments
on ancient ancestral chromosome that may harbor a disease locus. Therefore the
information on haplotype structure is useful for localizing and for tracing the past
historical events on causative sites. For example, we can trace the evolutionary
events by classifying haplotypes into clusters on the basis of the idea of popula-
tion genetic theory, which gives us a useful insight on the origin of the disease.
(Kido et al., 2003; Templeton, 1995)
Statistical power advantage: The haplotype-based analysis, which combines the
information of adjacent SNPs into multi-locus haplotypes, may increase statistical
power of conventional SNP-based analysis to detect disease associated sites.
Bader (2001) showed that haplotype-based association tests can have greater
power than SNP-based association tests in the case when the disease locus has
multiple disease-causing alleles.
Linkage Disequilibrium and Haplotype Analysis
The case-control study, which compares a group of people with disease to a similar group
of people without disease, is commonly used in population-based association study
design. If we have a strong hypothesis about the special causative alleles, causal alleles
can be directly evaluated. For common complex diseases, however, strong hypotheses
about the specific causal alleles are generally not available, so the indirect association
approach has become widespread. In indirect association studies, we expect that the
actual causative allele is not genotyped but might be located near another marker that
is genotyped.
LD is an important concept for indirect association study, which is defined as the lack
of independence among different polymorphisms. That is, alleles A and B are said to be
in LD if the frequency of pairs of alleles AB is not equal to the product of the frequencies
of A and B alleles. In LD analysis, we assume that causative alleles at tightly linked
markers will remain in significant LD for extended periods of time from ancestors. Various
statistical metrics can be used to summarize the extent of pairwise LD between two linked
markers, such as D and r
2
.
One of the important questions for indirect association studies is how much LD is
needed to detect a disease allele using its nearby genetic markers? Krina and Cardon
(2004) discussed the complex interplay among factors that influence allelic association.
They outlined the four parameters that affect an odds ratio test of association with a
single SNP: (1) the odds ratio of true disease-causing SNP, (2) LD between markers and
the causal SNP, (3) the marker allele frequency, and (4) the disease allele frequency.
218 Kido
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Ideally a disease allele can be detected provided that the extent of LD between the marker
and the causative allele is high and the marker allele frequency matches with the allele
frequency of the disease allele.
Since haplotype analysis captures regional LD information of SNPs, characterization of
LD patterns is important. Characterization of LD patterns across the human genome is
at present an area of highly active research. International HapMap project gathers
information on LD structure of variation in human populations (The International
HapMap Consortium, 2003). This will facilitate genome-wide association analysis and
the search for the genetic determinations of complex diseases.
Computational Algorithms for Haplotype Analysis
Defining LD Blocks
The recent finding in human genetics shows that human genome is broken up into regions
of weak LD (called hot spots) and strong LD (called cold spots or LD blocks). Strong LD
within a genomic region implies that most of the variation of that region can be captured
by just a few informative SNPs, called tag SNPs. It is hoped that the regions of strong
LD will facilitate the discovery of genes related to common diseases through association
studies.
There are a variety of proposed empiric definitions for haplotype blocks. For example,
Gabriel et al. (2002) developed a descriptive approach to comparing block boundaries
across different populations. They examined adjacent pairs of SNPs in at least two
populations, and labeled SNP pairs as concordant or discordant, based on the strength
of linkage disequilibrium between two populations. They then calculated the percentage
of concordant pairs among all pairs to define the block boundaries. Liu et al. (2004)
propose several similarity measures to compare haplotype boundaries across popula-
tions. They found that haplotype block boundaries vary among populations and the
definition of haplotype blocks is highly dependent on the density of SNPs and the
method used to find blocks. The understanding of the similarities and differences of
haplotype blocks between different populations will be important for future genome
association studies.
Inferring Haplotype Phases
In most cases, haplotypes are not read directly when multiple human SNPs are genotyped,
but must be inferred from unphased genotype data. Although a spectrum of molecular
haplotyping methods such as single-molecule dilution, long-range allele-specific PCR
were developed, these methods are not widely used because of their low-throughput
performance and important technical problems remain unresolved. Therefore a number
of algorithms for inferring haplotypes from unphased genotype data have been devel-
oped.
A Haplotype Analysis System for Genes Discovery of Common Diseases 219
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Niu et al. (2004) provides a good review of recent computational methods of population-
based haplotype inference methods. Clarks algorithm is based on the principle of
maximum parsimony which resolves the haplotypes by starting with identifying all
unambiguous haplotypes (all homozygotes and single-site heterozygotes) and repeat-
edly adding a new haplotype to the resolved haplotype. (Clark, 1990) Although Clarks
algorithm is a relatively straight-forward procedure, the algorithm does not give unique
solutions, because the phasing results are dependent on the order of genotypes that
need to be phased. Expectation-maximization (EM) algorithm estimates population
haplotype probabilities based on maximum likelihood, finding the values of the haplotype
probabilities which optimize the probability of the observed data, based on the assump-
tion of Hardy-Weinberg equilibrium (HWE). (Excoffer & Slatkin, 1995) Actually, EM
algorithm is based on solid statistical theory and are quite effective as simulation studies
demonstrate that its performance is not strongly affected by the departures from HWE.
(Niu et al., 2004). Some disadvantages of EM algorithm, however, are that the iteration
may lead to locally optimal point when there are many distinct haplotypes, and it cannot
handle a large number of loci. Stephens et al. (2001) proposed a coalescence-based
Markov-chain Monte Carlo (MCMC) approach: a pseudo-Gibbs sampler (PGS) for
reconstructing haplotypes from genotype data. This algorithm is implemented as
program named PHASE. Niu et al. introduced a divide-conquer-combine algorithm:
partition-ligation (PL), which can handle a large number of loci. This algorithm is
implemented as programs named Haplotyper (Bayesian implementation) and PLEM (EM
implementation).
Although there is an ongoing debate in regard to whether the best haplotype phasing
algorithm exits, different algorithms perform differently in different populations with their
own strengths and limitations. Niu et al. demonstrates that different algorithms have
different degrees of sensitivity to various extends of population diversities and genotyping
error rates using empirical simulations based on real genome data sets.
Statistical Advantages and Methods for Haplotype-Based
Disease Association
Haplotype analyses can sometimes provide greater power than single-marker analyses
for genetic disease associations, due to the ancestral structure captured in the distribu-
tion f haplotypes. Whether haplotype analyses outperform single locus analyses varies
depending on the assumptions about the number of disease causing SNPs, the amount
of LD among SNPs from the marker and disease causing SNPs. Schaid (2004) reviews the
recent works on the statistical advantages of haplotypes. Bader (2001) reports that when
the set of measured SNPs includes causative SNPs, single-locus tests are more powerful
than haplotype-based tests when the number of causative SNPs is less than the number
of haplotypes. In contrast, haplotype analyses can be more powerful than single-locus
analyses when the SNPs are in LD with a causative diallelic locus. (Akey & Xiong, 2001)
Both single-locus analyses and haplotype analyses lose power when there are multiple
alleles at a causative locus, but comparatively speaking haplotype analyses lose less
power. In this situation, the power advantage for haplotype analyses is greatest when
220 Kido
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
the marker alleles are not in strong LD with each other, yet in strong LD with the causative
alleles.
Schaid provides a comprehensive review on how to evaluate the association of haplotypes
with human traits. When haplotypes are directly observed, then statistical methods can
be applied to compare the frequencies of haplotypes between cases and controls. If the
haplotype- phases are unknown, haplotype frequencies for the cases and controls can
be estimated by EM algorithm. One of the common methods for haplotype-based disease
association is to use likelihood ratio statistic test (LRT) which tests the equality of
haplotype frequencies between cases and controls (Zhao & Sham, 2002). Although LRT
is a practical method, one of the disadvantages of LRT is that it may not be adequate when
there are many haplotypes, it lacks adjustment for environmental covariates. In order to
solve this problem, Schaid introduces regression models for haplotypes in which
haplotypes are treated as categorical covariates with other environmental covariates.
Schaid also reviews recent developments of statistical regression methods for haplotype
analyses.
A Haplotype System Analysis
LDMiner: The Pipeline System for SNP Analysis
We have developed a system named LDMiner that supports disease association studies
to detect genes that may cause complex diseases (Higashi et al., 2003). The main functions
of LDMiner are as follows:
1. LDMiner is a total analysis system for SNPs, LD blocks haplotypes, and disease
association.
2. LDMiner automatically combines SNPs located in the vicinity with respect to each
other on a chromosomal region and constructs LD block, and accordingly haplotypes
within the block using the EM algorithm (Kitamura et al., 2002).
3. LDMiner automatically assesses the association of haplotypes and diplotypes
between cases and controls. By combining SNPs together into LD blocks, we can
improve the statistical power for association study.
4. LDMiner visualizes the analyzed results along with a genome viewer. The main
viewer displays the genomic structure and is linked to another main viewer showing
the in-depth analysis result. These viewers allow the user to easily check and make
interpretations of the results.
LD analysis and some advanced association studies can be efficiently performed using
LDMiner with handy tools for eliminating the inadequate data and so on. Consequently,
the number of SNPs the system can analyze is about 30 to 50 times higher than by the
standard manual procedures per unit of time (Higashi et al., 2003).
A Haplotype Analysis System for Genes Discovery of Common Diseases 221
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
The system design of the LDMiner analysis pipeline is shown in Figure 1. In phase 1, we
import genotype and medical data with SNP annotation for the analysis. In phases 2 and
3, we examine the extents of LDs among SNPs within the same genes or the same
chromosomal regions as well as to estimate haplotypes and diplotypes using EM
algorithm. In phase 4, LD blocks are automatically constructed by a sliding-window
approach in searching for strong LD regions as discretized LDblocks. In phase 5, we
perform several association tests for either individual SNPs or haplotypes within LD
blocks. In phase 6, all association results are automatically annotated with SNPs and LD
blocks information in order to easily retrieve and compare. These analysis results are
visualized by linking the genome browser which maps LD status into genome structure.
Algorithms
Haplotyping
Given the genotypes of the individuals in a sample, the frequencies of the haplotypes
in the population can be estimated. There are primarily three categories of algorithms for
the inference of haplotype phases of individual genotype data; these categories are
Clarks algorithm (Clark, 1990), the EM algorithm (Kitamura et al., 2002) and Bayesian
Figure 1. Analysis pipeline with LDMiner

222 Kido
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
algorithm (Stephens et al., 2001). Niu et al. reports the performance comparison of these
algorithms and conclude that EM algorithm is shown to be accurate and practical.
Furthermore they proposed PL algorithm, which is an extended version of a stochastic
sampling algorithm incorporating the idea of divide-conquer-combine, implemented
the program called HAPLOTYPER which is applicable to a large number of SNPs.
Although our system can incorporate several haplotyping programs as module, we
mainly use EM algorithm implemented as LDSupport programs by Kamatani et al. (2002)
and then determine the posterior probability distribution of diplotype configuration
(diplotype distribution).
Automatic LD Block Extraction
We define the LD block based on the idea that SNPs within each LD block are tightly
linked, which means that most of the pairwise LD measurement of two SNPs within the
LD block is high, for example D 0.8 Because our major purpose of LD block extraction
is to improve the performance of haplotype association analysis, the number of SNPs in
the LD block should be limited in order to keep down the degree of freedom of the
association study. For example, based on our experience, three to five SNP markers are
appropriate for our association study when the majority of the pairwise LD values across
all SNPs within the same LD block is strong enough. Otherwise, if we select too many
SNPs for the LD block, the power of the multi-locus association test becomes smaller
because of the inflation of the number of haplotypes as well as the fact that the posterior
probability of the diplotype determination becomes smaller. For these reasons, we
propose the window-sliding method for LD block construction where the maximum
number of SNPs in the LD block is always lower than predefined window size. One of the
Figure 2. Window-sliding method for LD block extraction

1 1 1
0.947
0.912
0.948 0.978 0.975
0.946 0.952
0.925
A Haplotype Analysis System for Genes Discovery of Common Diseases 223
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
examples of the window-sliding method is shown in Figure 2. In this example, SNP2, 3,
4, 5 were extracted to be within a single LD block because all pairwise LDs among the four
SNPs is strong enough (D > 0.9). Although D is the quite popular measure of LD,
appropriate LD measures and parameters should be selected depending on the situation.
LDminer provides a variety of LD measures and parameters for flexibly defining the LD
block. We implement this method for both of gene-base and chromosome-base LD block
extraction.
Disease Association for Multi-Locus Haplotypes
After building LD blocks, our system automatically performs both single-locus tests and
haplotype /displotype association tests. In haplotype / diplotype association tests, we
have options to select several multi-locus association tests such as Log Likelihood Test
(LRT) (Zhao & Sham, 2002), Monte Carlo tests. (Sham & Curtis, 1995)
LRT is the nonparametric tests for homogeneity in haplotype / diplotype frequencies
between cases and controls. For example, the fast EH program (FEHP) developed by Zhao
and Sham can be used for this purpose. In their statistics, the maximum log-likelihood was
calculated for cases alone, for controls alone, and for cases and controls pooled
al t oget her. Denot i ng t hese maxi mum l og-l i kel i hoods as
case
L ln ,
control
L ln
and
control case
L
+
ln , LRT statistic was defined as 2(
case
L ln +
control
L ln -
control case
L
+
ln ),
which is asymptotically chi-squared with the degree of freedom of, where is the number
of haplotypes / diplotypes. Then we obtain asymptotic p-values as well as empirical p-
values by the use of permutation procedures.
In Monte Carlo tests, some programs are well known such as CLUMP by Sham and Curtis
(1995), which outputs the results of four different association tests from T1 to T4. In order
to estimate the odds ratio (OR: a measure of the degree of association which estimates
the relative risk of the disease in case-control study) and 95%CI for each haplotype/
diplotype, we perform the Fishers exact test for each haplotype by constructing a 2
2 table which consists of the number of haplotype holders /others in the column and the
number of case/control in the row.
Viewer for Genome Structure and Tools for Analysis Interpretation
LDMiner also has a supplicated vitalization tool (Higashi et al., 2003). The main viewer
displays the genomic structure and is linked to another main viewer showing the in-depth
analysis result. As real causative polymorphisms are likely to be located in the vicinity
of the analyzed polymorphisms rather than themselves, it is important to examine the
surrounding genes and polymorphisms. For this purpose, we also proposed the algo-
rithm on haplotype classification for detecting disease-associated sites (Kido et al.,
2003). These viewers and tools allow the user to easily check and make an interpretation
of the results.
224 Kido
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Examples of Haplotype Analyses
Experimental Setting: Genotype Data Used
To evaluate our methods, we performed a simulation study based on observed SNPs from
genes around NPHS1 in chromosome 19 regions with a Japanese population of 375
independent individuals, a Caucasian population of 49 individuals and an African-
American population of 49 individuals. This enables us to base our study on real patterns
of LD within those genes. Some genes were genotyped in both exonic and intronic
regions. We chose a subset of 10 SNPs mainly from NPHS1 genes denoted by 1 to 10 as
shown in Figure 3.
Experimental Setting for the Simulation Study for
Association Study
In order to examine whether the multi-locus association tests of LD0blocks have more
statistical power than that of single-locus marker and how LD status influences the power
of association, we evaluated it with simulated data sets. We designed a disease model
for typical common disease described in report by Jannot, Essioux, Reese, & Clerget-
Darpoux (2003) and phenotypic trait distributions were simulated under the samples of
Japanese 375 (case 148, control 227) individuals, keeping our SNP typing information. In
this model, we assumed that SNP 7, which allele frequency is 0.068, is the causative with
prevalence rate of 0.08. We then compared the multi-locus haplotype tests with single-
locus tests.
Single-Locus Analysis
The results of the single-locus analysis of allelic and genotype association are shown
in Figure 4. As we designed, the causative SNP7 shows significant difference between
Figure 3. Selected SNPs and detected LD block for Caucasian, African American and
Japanese

A Haplotype Analysis System for Genes Discovery of Common Diseases 225
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Figure 4. P-value distribution of the single-locus analysis and LD mapping
Figure 5. P-value distribution of the haplotype analysis and comparison with those
of single-locus analysis

0. 000001
0. 00001
0. 0001
0. 001
0. 01
0. 1
1
1 2 3 4 5 6 7 8 9 10
Locus
P
-
v
a
l
u
e
A : B ( A l l e l i c ) A A + A B: B B ( A D o m i n a n t ) A A : A A + A B ( A R e c e s si v e )

HAP3
HAP5
Single-locus analysis
Disease locus
226 Kido
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
case and control allele frequencies with a p-value is less than 0.00001. SNP9 also has a
low p-value in allelic association. This is because LD between SNP7 and SNP9 is strongly
high (|D|=0.978) and the discrepancy between SNP7 and SNP9 frequencies is very small
(SNP7 minor allele frequency [MAF]=0.068, and SNP9 MAF = 0.064). LD theory
introduced by Zondervan and Cardon (2004) estimates that the expected OR of locus 9
is 2.06 and this fits with our simulation result. However, the SNP6, SNP8, SNP10 are also
in the strong LD with locus, their p-values are not significant. This can be explained by
the discrepancy between the disease SNP (locus 7) and marker SNP frequency is not
small. The MAF of SNP6 is 0.135, that of SNP8 is 0.120 and that of SNP10 is 0.056. This
case shows that common variant marker which has high allele frequency can not detect
the effect of row variant causative SNP even though they are in the same LD block.
In LD mapping, the red part shows the strong LD (D is higher than 0.8) and the blue part
shows weak LD. Read parts can be extracted as LD block.
Haplotype Analysis
Selecting SNPs for Haplotype Analysis
First we selected 5 SNPs - SNP6, SNP7 (causative), SNP8, SNP9, SNP10 that are in the
strong LD block as show in Figure 4. The average of |D| in this LD block is 0.964. We then
select 3 SNPs - SNP6, SNP8 SNP10. This is because, minor allele frequencies of SNP7 and
SNP9 are small (less than 0.1) and they are usually considered as inappropriate SNP
markers. Here, an interesting question for us is whether haplotype analysis consisting
of three SNPs can detect the causative region or not, although all three SNPs are not
positive in single-locus analysis.
Results of Haplotype Association: Single-Locus Analysis vs. Haplotype
Analysis
The p-value distributions of the haplotype association and comparisons with those of
single-locus analysis are shown in Figure 5. The HAP5 consisting of five SNPs has a
significant overall p-value by FEHP program; empirical p-value is less than 0.001.
However it is not higher than those of single-locus analysis at SNP7 and SNP9.
Interestingly, HAP3 consisting of three SNPs has also significant overall p-value, which
is less than 0.01. (The empirical p-value by FEHP program with 20000 permutations is
0.0022, and the asymptotic p-value is 0.0007. In Figure 3, we plot the empirical p-value
of FEHP.) Furthermore this p-value is higher than those of single-locus analysis at SNP6,
SNP8 and SNP10. This is a good example to show the advantage of haplotype analysis;
although three SNPs in HAP3 are not positive in single-locus analysis, haplotype
analysis can detect the significant difference in the HAP3 block.
Table 1 shows the results of haplotype analysis with three SNP markers (HAP3). Only
six haplotypes are estimated by EM algorithm, major haplotype BBB dominates 85.8 %
over the population. For the major haplotype BBB, the difference of frequencies
A Haplotype Analysis System for Genes Discovery of Common Diseases 227
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
between case and control is not significant with a p-value obtained by Fishers exact test
of 0.085. The frequency of the minor haplotype AAB - 6.5 %, is significant (p-value is
less than 0.0001). It works as a risk haplotype, such that its frequency among cases (11%)
is much higher than in controls (3%). The Fishers p-value of AAB is less than 0.0001
and the OR is 3.75 and its 95% confidence interval is 1.99 7.03. Because the posterior
probability for determining diplotype is almost 100% (represented by blue bar) or more
than 99% (represented by green bar) in most haplotypes, we can validate that haplotype
estimation is mostly reasonable.
Diplotype Analysis
We also perform three-locus diplotype association as well as haplotype association
study. This is done by determining individuals diplotype after estimating haplotype by
EM algorithm (Kitamura et al., 2002). We only choose them from which posterior
probability is more than 80%, otherwise we rule out from our analysis. The results are
shown in Table 2. Minor diplotype [1:2] which represents the combination of haplotype1
(BBB) and haplotype2 (AAB) works as risk diplotype (Fishers p-value is 0.005,
OR=2.65). The diplotype also works as risk diplotype in our analysis (Fishers p-value
is 0.008, OR=17), although it is in the rare cases.
Summary
The main purpose of this chapter was to review the roles of haplotype analysis for
association study on complex common diseases and provide the computational frame-
work that enables to realize high throughput genome-wide association study. The
motivation, background and computational methods for finding causative genes by use
Table 1. The results of haplotype analysis with 3 SNP markers

228 Kido
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
of haplotype and LD information have been reviewed. A Haplotype Analysis System,
LDMiner, has been introduced with some real examples of association studies. Although
there have been significant advances in this area, further developments of computational
methods for the genome-wide association studies and the detection of complex gene-
gene interactions have been pursued. (Meng, Zaykin, Zu, Wanger, & Ehm, 2003; Ritchie
et al., 2001; Toivonen et al., 2000) It is hoped that this report will entice further research
and developments in bioinformatics for finding causative genes.
Acknowledgment
We would like to thank Dr. Tianhua Niu at Harvard University for useful discussions and
comments.
References
Akey J., & Xiong, M. (2001). Haplotype vs single marker linkage disequilibrium tests:
What do we gain? Eur J Hu Genet, 9, 291-300.
Table 2. The result of diplotype analysis with 3 SNP markers

A Haplotype Analysis System for Genes Discovery of Common Diseases 229
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Badano, J. L, & Katsanis, N. (2002). Beyond Mendel: An evolving view of human genetic
disease transmission. Nature Genetics, 3, 779-789.
Bader, J. S. (2001). The relative power of SNPs and haplotype as genetic markers for
association tests. Pharmacogenomics, 2(1), 11-24
Cardon, L. R, & Abecasis, G.R. (2003). Using haplotype blocks to map human complex
trait loci. Trends Genet., 19, 135-140.
Clark, A. G. (1990). Inference of haplotypes from PCR-amplified samples of diploid
populations. Mol. Biol. Evol., 7, 111-122.
Clark, A. G. (2004). The role of haplotypes in candidate gene studies. Genetic Epidemi-
ology, 27, 321-333.
Cummins, J. M., Rago, C., Kohli, M., Kinzer, K. W., Lengauer, C., & Vogelstein, B. (2004).
A marker for Stevens-Hohnson Syndrome. Nature, 428,486.
Excoffier, L., & Slatkin, M. (1995). Maximum-likelihood estimation of molecular haplotype
frequencies in a diploid population. Mol. Biol. Evol., 12, 921-927.
Fan, R., & Knapp, M. (2003). Genome association studies of complex diseases by case-
control designs. Am .J .Hum. Genet., 72,850-868.
Fullerton, S. M., Clark A. G, Weiss, K. M, Nickerson, D. A., Taylor, S. L., Stengard,
J. H., et al. (2000). Apolipoprotein E variation at the sequence haplotype level:
Implications for the origin and maintenance of a major human polymorphism. Am.
J. Hum. Genet., 67, 881-900.
Gabriel, S. B., Schaffner, S. F., Nguyen, H., Moore, J. M., Roy, J., Blumenstiel, B., et
al. (2002). The structure of haplotype blocks in the human genome. Science, 296,
(pp. 2225-2229).
Higashi, Y., Higuchi, H., Kido, T., Matsumine, H., Baba, M., Morimoto, T., et al. (2003).
SNP analysis system for detecting complex disease associated sites. The Proceed-
ing of IEEE Computational Systems Bioinformatics (CSB2003), 450-451.
Hung, S. L., Chung, W. H., Liou, L. B., Cgy, C. C., Lin, M., Huang, H. P., et al. (2005).
HLA-B*5801 allele as a genetic marker for severe cutaneous adverse reactions
caused by allopurinol. PNAS, 102,11, 4134-4139.
International HapMap Project. (n.d.). Retrieved from http://www.hapmap.org/
The International HapMap Consortium. (2003) The International HapMap Project.
Nature, 18, 426(6968),789-96.
Jannot, A. S., Essioux, L., Reese, M. G., & Clerget-Darpoux, F. (2003). Improved use of
SNP information to detect the role of genes. Genetic Epidemiology, 25, 158-167.
Kido, T., Baba, M., Matsumine, H., Higashi, Y., Higuchi, H., & Muramatsu, M. (2003).
Haplotype pattern mining & classification for detecting disease associated site.
The proceeding of IEEE Computational Systems Bioinfomratics (CSB2003), (pp.
452-453).
Kitamura, Y., Moriguchi, M., Kaneko, H., Morisaki, H., Morisaki, T., Toyama, K., &
Kamatani, N. (2002). Determination of probability distribution of diplotype con-
figuration (diplotype distribution) for each subject from genotypic data using the
EM algorithm. Annals of Human Genetics, 66, 183-193.
230 Kido
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Kruglyak, L., & Nickerson, D. A. (2001 March). Variation is the spice of life. Nature
Genetics., 27(3), 234-6.
Liu, N., Sawyer, S. L., Mukherjee, N., Pakstis, A. J., Kidd, J. R., Kidd, K. K., et al. (2004).
Haplotype block structures show significant variation among populations. Ge-
netic Epidemiology, 27, 385-400.
Meng, Z., Zaykin, D. V., Zu, C. F., Wanger, M., & Ehm, M. G. (2003). Selection of genetic
markers for association analysis, using linkage disequilibrium and haplotypes. Am.
J .Hum. Genet., 73, 115-130.
Niu, T. (2004). Algorithms for Inferring Haplotypes. Genetic Epidemiology, 27, 334-347.
Niu, T., Zin, Z. S., Zu, X., & Liu, J. S. (2002). Bayesian haplotype inference for multiple
linked single-nucleotide polymorphisms, Am. J. Hum. Genet., 70, 157-169.
Ritchie, M. D, Hahn, L. W., Roodi, N., Dupont, W. D., Parl, F. F., & Moore, J. H. (2001).
Multifactor dimensionality reduction reveals high-order interactions among estro-
gen metabolism genes in sporadic breast cancer. Annals of Human Genetics, 69,
138-147.
Schaid, D. J. (2004). Evaluating associations of haplotypes with traits. Genetic Epide-
miology, 27, 348-364.
Sham, P. C., & Curtis, D. (1995). Monte Carlo tests for associations between disease and
alleles at highly polymorphic loci. Annals of Human Genetics, 59, 97-105.
Stephens, M., Smith, N. J., & Donnelly, P. (2001). A new statistical method for haplotype
reconstruction from population data. Am. J. Hum . Genet., 68, 978-989.
Toivonen, H. T. Onkamo, P., Vasko, K., Ollikainen, V., Sevon, P., Mannila, H., Herr,
M., & Kere, J. (2000). Data mining applied to linkage disequilibrium mapping. Am.
J. Hum. Genet., 67, 133-145.
Zhang, J., Rowe, W. L., Struewing, J. P., & Buetow, K. H. (2002). HapScope: A software
system for automated and visual analysis of functionally annotated the allelic
association. Nature Genetics, 5, 89-100.
Zhao, J. H. (2004). Haplotype block structures show significant variation among popu-
lations. Genetic Epidemiology, 27, 385-400.
Zhao, J. H., & Sham, P.C. (2002). Faster haplotype frequency estimation using unrelated
subjects. Hum Hered., 53, 36-41.
A Bayesian Framework for Improving Clustering Accuracy 231
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Chapter XII
A Bayesian Framework
for Improving
Clustering Accuracy
of Protein Sequences
Based on
Association Rules
Peng-Yeng Yin, National Chi Nan University, Taiwan
Shyong-Jian Shyu, Ming Chuan University, Taiwan
Guan-Shieng Huang, National Chi Nan University, Taiwan
Shuang-Te Liao, Ming Chuan University, Taiwan
Abstract
With the advent of new sequencing technology for biological data, the number of
sequenced proteins stored in public databases has become an explosion. The structural,
functional, and phylogenetic analyses of proteins would benefit from exploring
databases by using data mining techniques. Clustering algorithms can assign proteins
into clusters such that proteins in the same cluster are more similar in homology than
those in different clusters. This procedure not only simplifies the analysis task but also
enhances the accuracy of the results. Most of the existing protein-clustering algorithms
compute the similarity between proteins based on one-to-one pairwise sequence
232 Yin, Shyu, Huang, & Liao
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
alignment instead of multiple sequences alignment; the latter is prohibited due to
expensive computation. Hence the accuracy of the clustering result is deteriorated.
Further, the traditional clustering methods are ad-hoc and the resulting clustering
often converges to local optima. This chapter presents a Bayesian framework for
improving clustering accuracy of protein sequences based on association rules. The
experimental results manifest that the proposed framework can significantly improve
the performance of traditional clustering methods.
Introduction
One of the central problems of bioinformatics is to predict structural, functional, and
phylogenetic features of proteins. A protein can be viewed as a sequence of amino acids
with 20 letters (which is called the primary structure). The explosive growth of protein
databases has made it possible to cluster proteins with similar properties into a family
in order to understand their structural, functional, and phylogenetic relationships. For
example, there are 181,821 protein sequences in the Swiss-Prot database (release 47.1)
and 1,748,002 sequences in its supplement TrEMBL database (release 30.1) up to May
24, 2005. According to the secondary structural content and organization, proteins were
originally classified into four classes: , , +, and / (Levitt & Chothia, 1976). Several
others (including multi-domain, membrane and cell surface, and small proteins) have been
added in the SCOP database (Lo Conte, Brenner, Hubbard, Chothia, & Murzin, 2002).
Family is a group of proteins that share more than 50% identities when aligned, the SCOP
database (release 1.67) reports 2630 families.
Pairwise comparisons between sequences provide good predictions of the biological
similarity for related sequences. Alignment algorithms such as the Smith-Waterman
algorithm and the Needleman-Wunsch algorithm and their variants are proved to be
useful. Substitution matrices like PAMs and BLOSUMs are designed so that one can
detect the similarity even between distant sequences. However, the statistical tests for
distant homologous sequences are not usually significant (Hubbard, Lesk, & Tramontano,
1996). Pairwise alignment fails to represent shared similarities among three or more
sequences because it leaves the problem of how to represent the similarities between the
first and the third sequences after the first two sequences have been aligned. It is
suggested in many literatures that multiple sequence alignment should be a better choice.
While this sounds reasonable, it causes some problems we address here. The most critical
issue is the time efficiency. The natural extension of the dynamic programming algorithm
from the pairwise alignment to the multiple alignment requires exponential time (Carrillo
& Lipman, 1988), and many problems related to finding the multiple alignment are known
to be NP-hard (Wang & Jiang, 1994). The second issue is that calculating a distance matrix
by pairwise-alignment algorithm is fundamental. ClustalW (Thompson, Higgins, &
Gibson, 1994) is one of the most popular softwares for multiple-alignment problems. It
implements the so-called progressive method, a heuristic that combines the sub-
alignments into a big one under the guidance of a phylogenetic tree. In fact, the tree is
built from a pre-computed distance matrix using pairwise alignment.
A Bayesian Framework for Improving Clustering Accuracy 233
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Many protein clustering techniques exist for sorting the proteins but the resulting
clustering could be of low accuracy due to two reasons. First, these clustering techniques
are conducted according to homology similarity, thus a preprocessing of sequence
alignment should be applied to construct a homology proximity matrix (or similarity
matrix). As we have mentioned, applying multiple sequence alignment among all proteins
in a large data set is prohibited because of expensive computation. Instead, an all-
against-all pairwise alignment is adopted for saving computation time but it may cause
deterioration in accuracy. Second, most of the traditional clustering techniques, such as
hierarchical merging, iterative partitioning, and graph-based clustering, often converge
to local optima and are not established on statistical inference basis (Jain, Murty, &
Flynn, 1999).
This chapter proposes a Bayesian framework for improving clustering accuracy of
protein sequences based on association rules. With the initial clustering result obtained
by using a traditional method based on the distance matrix, the strong association rules
of protein subsequences for each cluster can be generated. These rules satisfying both
minimum support and minimum confidence can serve as features to assign proteins to
new clusters. We call the process to extract features from clusters the alignment-less
alignment. Instead of merely comparing similarity from two protein sequences, these
features capture important characteristics for a whole class from the majority, but ignore
minor exceptions. These exceptions exist due to two reasons: the feature itself or the
sequence itself. For the first reason, the feature being selected could be inappropriate
and thus causes exceptions. Or, there does not exist a perfect feature that coincides for
the whole class. The second reason is more important. The sequence causing the
exception may be pre-classified into a wrong cluster; therefore, it should be re-assigned
to the correct one. The Bayes classifier can provide optimal protein classifications by
using the a priori feature information through statistical inference. As such, the
accuracy of the protein clustering is improved.
The rest of this chapter is organized as follows. The background reviews existing
methods relevant to protein clustering and the motivations of this chapter. The third
section presents the ideas and the theory of the proposed method. The fourth section
gives the experimental results with a dataset of protein sequences. The final section
concludes this chapter.
Background
Related Works
Many clustering techniques for protein sequences have been proposed. Among them,
three main kinds of approaches exist, namely the hierarchical merging, iterative partition-
ing, and graph-based clustering. All of these methods use a pre-computed similarity
matrix obtained by performing pairwise alignments on every pair of proteins. In the
following we briefly review these approaches.
234 Yin, Shyu, Huang, & Liao
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Hierarchical Merging: The hierarchical merging clustering (Yona, Linial, & Linial,
1999, 2000; Sasson, Linial, & Linial, 2002) starts with a partition that takes each
protein as a separate cluster, and then iteratively merges the two clusters that have
the highest similarity from all pairs of current clusters. The similarity between two
clusters is derived from the average similarity between the corresponding members.
Thus, the hierarchical merging procedure forms a sequence of nested clusterings
in which the number of clusters decreases as the number of iterations increases.
A clustering result can be obtained by specifying an appropriate cutting-off
threshold, in other words, the iterative merging procedure progresses until the
maximal similarity score between any two clusters is less than the cutting-off
threshold. The algorithm for hierarchical merging clustering is summarized in
Figure 1.
Iterative Partitioning: The iterative partitioning method (Jain & Dubes, 1988;
Guralnik & Karypis, 2001; Sugiyama & Kotani, 2002) starts with an initial partition
of k clusters. The initial partition can be obtained by arbitrarily specifying k
proteins as cluster centers then assigning each protein to the closest cluster whose
center has the most similarity with this protein. The next partition is obtained by
computing the average similarity between each protein and all members of each
cluster. The partitioning process is iterated until no protein changes its assignment
between successive iterations. Although the iterative partitioning method has
some variants like ISODATA (Ball & Hall, 1964) and K-means (McQueen, 1967),
their general principles can be described as shown in Figure 2. A post-processing
stage can be added to refine the clusters obtained from the iterative partitioning
by splitting or merging the clusters based on intra-cluster and inter-cluster
similarity scores (Wise, 2002).
Graph-based Clustering: These methods (Tatusov, Koonin, & Lipman, 1997;
Enright & Ouzounis, 2000; Bolten, Schliep, Schneckener, Schomburg, & Schrader,
2001) represent each protein sequence as a graph vertex, and every pair of these
vertices are connected by an edge with a label on it. The label denotes the similarity
score between the two proteins represented by the vertices connecting to the
corresponding edge. A partition of the graph can be generated by cutting off the
edges whose labels are less than a specified similarity threshold, and each
connected component of vertices corresponds to a cluster of proteins since the
Figure 1. Summary of hierarchical merging clustering
create.cluster(p);
Repeat
Find clusters x, y such that similarity(x, y) is maximal;
If similarity(x, y)>cutting_off_threshold then merge.cluster(x, y);
Otherwise, terminate;

A Bayesian Framework for Improving Clustering Accuracy 235
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
similarity scores between proteins from the same component are higher than those
between proteins from different components. The general idea of the graph-based
clustering is outlined in Figure 3. A post-refining process can be conducted by
using the graph-based clustering result as the input to a cluster-merging algorithm
which iteratively merges the nearest neighboring clusters if the relative entropy
decreases with the merging of the clusters (Abascal & Valencia, 2002).
Motivations
The protein clustering result obtained by using the above mentioned approaches could
be of low accuracy. This is partly due to the reason that these clustering methods use
only pairwise sequence alignment information and partly because these clustering
techniques often converge to local optima. Since it is computationally prohibitive to
derive multiple sequence alignment information, an alternative is to calculate the
statistics from the sequences directly by using data mining techniques. More precisely,
the association rules between the amino acids in the protein sequences are mined within
each cluster. Then the rules satisfying minimum confidence can serve as salient features
to identify each cluster. Moreover, matching each protein sequence with the association
rule provides a good estimate of the a priori probability that the protein satisfies the rule.
The statistical inference can compensate the accuracy inadequacy of the clustering
Figure 2. Summary of iterative partitioning method
Figure 3. Summary of graph-based clustering
Select k proteins as initial cluster centers
Repeat
Compute the average similarity between each protein and each cluster;
Generate a new partition by assigning each protein to its closest cluster;
Until no proteins change assignment between successive iterations

Represent each input sequence as a vertex, and every pair of vertices are connected by
an edge.
Label each edge with the similarity score between the two connected vertices.
Create a partition of the graph by cutting off the edges whose similarity scores are less
than a specified threshold.
236 Yin, Shyu, Huang, & Liao
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
result due to the pairwise alignment and the local clustering technique. Therefore, we
propose to improve the clustering accuracy by using the Bayes classifier with the
conditional probabilities of association rules with each cluster and the alignment scores
between protein sequences and these rules.
Methods
With the assistance of association rule mining and Bayes classifier, our system improves
the clustering accuracy of traditional protein-clustering methods. The system overview
is shown in Figure 4. First, a traditional protein-clustering approach (either one of the
hierarchical merging, iterative partitioning, or graph-based clustering methods) is
performed to obtain an initial clustering result of the input protein sequences. In general,
the traditional protein-clustering approach consists of three steps as shown in the upper
grey box: (1) perform the local alignment (such as BLAST with scoring matrix of BLOSUM
62) between each pair of protein sequences, (2) construct a distance matrix from the raw
distance scores (such as the E-values produced by BLAST) of the local alignment, and
(3) apply the clustering method with the distance matrix to get the clustering result.
Second, the refining process (as shown in the lower grey box) is performed such that the
clustering result is improved. The refining process also consists of three steps: (1)
generate strong association rules which satisfy minimum confidence for each cluster, (2)
perform sequence rule matching between each protein sequence and each association
rule based on local alignment, and (3) adapt cluster membership of each protein sequence
by using Bayes classifier and its matching score. The proposed refining process
improves the clustering result based on statistical inference. The association rules used
are statistically confident and the reassignment of protein sequences using Bayes
classifier satisfies the maximum a posteriori criterion.
We now present the refining process in details.
Sequence Association Rule Generation
Association rule mining has been intensively used for finding correlation relationships
among a large set of items and has delivered many successful applications, such as
catalog design, cross marketing, and loss-leader analysis (Han & Kamber, 2001).
Association rule mining finds significant associations or correlation relationships
among a large repository of data items. These relationships, represented as rules, can
assist the users to make their decisions. Traditionally, association rule mining works with
unordered itemset, that is, the order of the items appearing in the itemset does not matter
(Agrawal, Imielinski, & Swami, 1993). However, the order of the amino acids appearing
in a protein sequence reserves important phylogeny information among homologies and
should be taken into account. The sequence Apriori algorithm (Agrawal & Srikant, 1994,
1995) which adapts the classical association rule mining algorithm for sets of sequences
can be used for this purpose.
A Bayesian Framework for Improving Clustering Accuracy 237
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Given a cluster C of protein sequences, the algorithm finds sets of sequences that have
support above a given minimum support. The support of a sequence X with respect to
C, denoted by support
C
(X), is the number of sequences in C that contain X as subse-
quence. A sequence is called frequent if its support is greater than the given minimum
support. To expedite the search for all frequent sequences, the algorithm enforces an
iterative procedure based on the a priori property which states that any subsequence of
a frequent sequence must be also frequent. It starts with the set of frequent 1-sequences
which are the frequent sequences of length 1 and uses this set to find the set of frequent
2-sequences, then the set of frequent 2-sequences is used to find the set of frequent 3-
sequences, and so on, until no more longer frequent sequences can be found. As the
number of frequent sequences can be extremely large for a cluster of protein sequences
having length of hundreds of amino acids, we retain the frequent sequences that are not
contained in longer ones and restrict the search for frequent sequences of length 5 to
10. The reduced set of frequent sequences is sufficient for deriving similarity statistics
in homology since longer frequent sequences of length more than 10 usually have lower
support and a number of short frequent sequences within the specified range of lengths
that the longer ones contain as subsequences can still be reserved.
With the set of found frequent sequences, we generate sequence association rules as
follows. For a given frequent sequence X, it can be divided into two disjoint subse-
quences A and B with their position information attached. For example, let X be abcde,
Figure 4. System overview

Protein Sequences
Distance Matrix
Scoring Matrix
(BLOSUM62)
Initial Clustering
Result
Improved Clustering
Result
Pairwise Alignment
Using BLAST
Clustering
Association Rules Generation
Sequence Rule Matching
Cluster Membership Adapting
Using Bayes Classifier
Traditional
Clustering
Approach
Refining
Process
238 Yin, Shyu, Huang, & Liao
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
one of the possible divisions could be A = a_c_e and B = _b_d_. A sequence
association rule is of the form A B where both A and B contain at least one amino
acid. A is called the rule antecedent and B is called the rule consequent. A sequence
association rule is strong if it has a confidence value above a given minimum confidence.
The confidence of a rule A B with respect to protein cluster C, denoted
by ( ) B A
C
confidence , is defined as:

( )
( )
( ) A
B A
B A
C
C
C
support
support
confidence

=
, (1)
where B A denotes the supersequence that is properly divided into A and B. We
generate all strong sequence association rules for each cluster of protein sequences.
Sequence Association Rule Matching
In order to determine the cluster membership of each protein, we need to propose a
measure which estimates the possibility that the evolution of a protein follows a
particular association rule. Here we propose a rule matching scheme which is analogous
to local sequence alignment without gaps. Given a sequence t and a strong association
rule r, we compute the alignment score between t and r according to a substitution matrix
with the constraint that no gap is allowed and the rule antecedent cannot be substituted
in order to detect most similarities. For example, suppose we use the substitution matrix
as shown in Figure 5 (a). Let the sequence be adabdacd and the association rule be
da_d _ _d_, there are two possible alignments as illustrated in Figure 5 (b) and the
best alignment score without gaps is 11.
Figure 5. An illustrative example of sequence association rule matching: (a) A
substitution matrix, (b) Two possible alignments between the sequence and the
association rule
a b c d
a 2 -1 2 0
b -1 3 -2 1
c 2 -2 2 -1
d 0 1 -1 4

Sequence: adabdacd Score
Association rule: dadd 4+2+1+4=11
Association rule: dadd 4+2 -1+4= 9
(a) (b)
A Bayesian Framework for Improving Clustering Accuracy 239
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Assume that we obtain a set of strong association rules { }
n
r r r ,..., ,
2 1
= from the
procedure of sequence association rule generation, the probability that the evolution of
a protein t follows a particular association rule r
i
can be estimated by:

=
=
n
h
h
i
i
t r w
t r w
t r p
1
) , (
) , (
) | (

, (2)
where w(r
i
, t) is the alignment score between r
i
and t. As such, we can use the probabilities
( ) ) | ( ..., ), | ( ), | (
2 1
t r p t r p t r p
n

as the feature values of protein t, and determine its cluster
membership by using the Bayes classifier.
Cluster Membership Adapting by Using Bayes Classifier
The Bayes classifier is one of the most important techniques used in data mining for
classification. It predicts the cluster membership probabilities based on statistical
inference. Studies have shown that the Bayes classifier is comparable in performance
with decision tree and neural network classifiers (Han & Kamber, 2001). Herein we
propose to predict the cluster membership of each protein by using the Bayes classifier
with association rules.
Let the initial clustering result consists of k clusters, denoted by { }
k
C C C ,..., ,
2 1
= , from
which a set of n strong association rules is derived. The a priori probability p(C
i
) that
a protein belongs to cluster C
i
can be calculated by counting the ratio of C
i
in size to the
whole set of proteins. The condition probability p(r
i
C
j
) that a protein satisfies associa-
tion rule r
i
given that this protein is initially assigned to cluster C
j
is estimated by the
average probability p(r
i
t) for any t C
j
. The conditional probability p(r
i
t) that the
evolution of protein t follows association rule r
i
is estimated by using Equation (2). We
then use the nave Bayes classifier to assign protein t to the most probable cluster C
Bayes
given by:
( )

=

=

n
i
j i j i
C
Bayes
C p t r p C r p C
j 1
) ( ) | ( ) | ( max arg
. (3)
In theory, the nave Bayes classifier makes classification with the minimum error rate.
240 Yin, Shyu, Huang, & Liao
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Experimental Results
We validate our method by using protein sequences selected from SCOP database
(release 1.50, Murzin, Brenner, Hubbard, & Chothia, 1995; Lo Conte et al., 2002) which
is a protein classification created manually. SCOP provides a hierarchy of known protein
folds and their detailed structure information. We randomly select 1189 protein se-
quences from SCOP and these sequences compose 388 protein clusters according to
manual annotations on structure domains. The mean length of these protein sequences
is 188 amino acids. Our method is implemented in C++ programming language and the
experiments are conducted on a personal computer with a 1.8 GHz CPU and 512 MB RAM.
Accuracy Measures
To define the accuracy measures of protein classification, some notations are first
introduced. For every pair of protein sequences, the predicted classification and the
annotated classification have four possible combinations. TP (true positive) is the
number of pairs predicted in the same domain given that they are in the same SCOP
domain, TN (true negative) is the number of pairs predicted with different domains given
that they are in different SCOP domains, FP (false positive) is the number of different
SCOP-domain pairs that are predicted in the same domain, and FN (false negative) is the
number of SCOP-domain pairs that are predicted in different domains. Two accuracy
measures, namely sensitivity (S
n
) and specificity (S
p
), are defined as follows:
n
TP
S
TP FN
=
+
(4)
and
p
TP
S
TP FP
=
+
. (5)
Sensitivity is the proportion of SCOP-domain pairs that have been correctly identified,
and specificity is the proportion of pairs predicted in the same domain that are actually
SCOP-domain pairs. Sensitivity and specificity cannot be used alone since perfect
sensitivity can be obtained if all the pairs are predicted in the same domain, and specificity
is not defined if all the pairs are predicted in different domains. To compare the
performance between two competing methods, a sensitivity vs. specificity curve is
usually used for evaluation. Or alternatively, the mean of sensitivity and specificity
values can be used as a unified measure.
A Bayesian Framework for Improving Clustering Accuracy 241
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Performance Evaluation
We first evaluate the clustering performance using the traditional methods, in particular,
we have implemented the hierarchical merging method (Sasson et al., 2002), the K-means
algorithm (modified from Guralnik & Karypis, 2001, by changing the feature space to pair-
wise E-value), and the graph-based clustering (Bolten et al., 2001). Various clustering
thresholds are specified to obtain the performances at different specified numbers of
Figure 6. Performances of traditional clustering methods, (a) hierarchical merging
clustering, (b) K-means clustering, (c) graph-based clustering

(a)
(b)
(c)
S
n
S
p
S
n
S
p
S
n
S
p
(c)
(b)
(a)
242 Yin, Shyu, Huang, & Liao
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Figure 7. Performance improvement for specificity vs. sensitivity, (a) hierarchical
merging clustering, (b) K-means clustering, (c) graph-based clustering
clusters. Figures 6 (a) - 6 (c) show the variations of sensitivity and specificity as the
number of clusters increases for hierarchical merging, K-means, and graph-based
clustering, respectively. These curves are intuitive since the mean size of clusters is
smaller if the partition with more clusters is obtained and, in general, the smaller the
cluster-mean size is, the lower the sensitivity is, but the higher the specificity is. If we
take the accuracy values obtained when the number of clusters is equal to the true number
(388), the hierarchical merging method with S
n
= 0.79 and S
p
= 0.73 is superior to the K-

(a)

(b)

(c)
S
p
S
p
S
p
S
n
S
n
S
n
(a)
(b)
(c)
A Bayesian Framework for Improving Clustering Accuracy 243
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
means algorithm which produces S
n
= 0.69 and S
p
= 0.72, while the graph-based clustering
has medium performance (S
n
= 0.80 and S
p
= 0.63).
Next we evaluate the improvement for clustering accuracy due to our Bayesian frame-
work. From the clustering results obtained by traditional clustering methods, we apply
the sequence association rule mining to generate strong association rules. Each protein
Figure 8. Performance improvement for the unified measure, (a) hierarchical merging
clustering, (b) K-means clustering, (c) graph-based clustering

(a)

(b)



(c)
(
S
n

+

S
p
)
/
2
(
S
n

+

S
p
)
/
2
(
S
n

+

S
p
)
/
2
(c)
(b)
(a)
244 Yin, Shyu, Huang, & Liao
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Table 1. The computation time used by each component of the proposed method
sequence is matched with these rules and updates its cluster membership by using the
Bayes classifier in order to improve the accuracy. Because the sensitivity values
obtained by hierarchical merging, K-means, and graph-based clusterings with the true
cluster number (388) are 0.79, 0.69, and 0.80, respectively, we compute the specificity
improvement for these methods within a range of sensitivity close to these values.
Figures 7 (a) - 7 (c) show the sensitivity vs. specificity curve for illustrating the
improvement achieved. It is observed that the proposed refining process can signifi-
cantly improve the specificity of the traditional methods. The average improvements in
Sp are 0.09, 0.11, and 0.06 for hierarchical merging, K-means, and graph-based clustering,
respectively. We also compute the accuracy improvement by using the unified measure
of (S
n
+ S
p
)/2 as shown in Figs. 8 (a) - 8 (c). The average improvements over various
specified numbers of clusters are 0.03, 0.04, and 0.03 for hierarchical merging, K-means,
and graph-based clustering, manifesting the robustness of the proposed framework.
Table 1 shows the incurred computation time (in seconds) by each component of the
proposed framework and the corresponding percentage to the whole for our collective
database. We observe that the distance matrix computation using pairwise sequence
alignment and the association rules generation using sequence Apriori algorithm are the
most time-consuming components, and they consume 35% and 33% of the total compu-
tation time, respectively. Further, the computation time for the refining process involving
the last two components of Table 1 is about half of the whole time needed. That is, for
our collective database, the proposed framework provides a considerable amount of
accuracy improvement (as shown in Figures 7-8) for the traditional protein-clustering
algorithms by doubling the computation time. In general cases, the time proportion for
the refining process diminishes when the number of sequences in the database increases.
Suppose there are m sequences in the database, and for simplicity, assume that their
average length is r. The time complexity for calculating a distance matrix is (m
2
r
2
), for
performing traditional clustering is (m
2
), and for generating the association rules of
fixed length is (mr). Rule matching and Bayesian classification depends on two factors:
the number of features extracted and the number of clusters classified (and remember that
each rule is within a fixed length). The former can be regarded as a constant once the
lengths of rules and the alphabet (which is 20 for amino acids) are fixed to constants.
Therefore a rough calculation asserts that it takes (mrc) where c is the number of
CPU Time (second) Percentage
Distance Matrix Computation 162 35%
Traditional Clustering 64 14%
Association Rules Generation 157 33%
Rule Matching and Bayesian Classification 87 18%
A Bayesian Framework for Improving Clustering Accuracy 245
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
clusters. As a summary, the time complexity for each component is (m
2
r
2
), (m
2
), (mr),
and (mrc), respectively. Parameter r, the average length of peptide sequences, is around
300 and can be regarded as a constant. Parameter c grows mildly as the number of
sequences increases. As a result, when the number of sequences in a database grows
larger, the percentage of the computing time on the refining process becomes smaller.
Summary
Protein sequence clustering is useful for structural, functional, and phylogenetic
analyses. As the rapid growth in the number of sequenced proteins prohibits the analysis
using multiple sequences alignment, most of the traditional protein-clustering methods
derive the similarity among sequences from pairwise sequence alignment. In this chapter,
we have proposed a Bayesian framework based on association rule mining for improving
the clustering accuracy using existing methods. A selective dataset from SCOP has been
experimented and the result manifests that the proposed framework is feasible. The main
features of the proposed framework include:
The proposed framework improves the accuracy of a given initial clustering result
which can be provided by any clustering methods. Therefore, the user is still able
to choose a particular clustering method which is suited to his/her own analysis
of the final result.
From the initial clustering result, the sequence association rules among amino acids
are mined. These rules represent important relationships for the sequences belong-
ing to the same cluster.
The association rules serve as features for classification of proteins. Using Bayes
classifier, the classification error can be minimized based on statistical inference.
Future research is encouraged in expediting the computation for distance matrix and
association rule generation in order to extend the application of the proposed framework
to large protein databases such as SWISSPROT.
References
Abascal, F., & Valencia, A. (2002). Clustering of proximal sequence space for the
identification of protein families. Bioinformatics, 18(7), 908-921.
Agrawal, R., Imielinski, T., & Swami, A. (1993). Mining association rules between sets
of items in large databases. Proceedings of ACM-SIGMOD, International Confer-
ence on Management of Data, Washington, DC (pp. 207-216).
246 Yin, Shyu, Huang, & Liao
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Agrawal, R., & Srikant, R. (1994). Algorithms for mining association rules. Proceedings
of the 20th International Conference on Very Large Databases, Santiago, Chile
(pp. 487-499).
Agrawal, R., & Srikant, R. (1995). Mining sequential patterns. Proceedings of the
International Conference on Data Engineering (ICDE), Taipei, Taiwan.
Ball, G. H., & Hall, D. J. (1964). Some fundamental concepts and synthesis procedures
for pattern recognition preprocessors. Proceedings of International Conference
on Microwaves, Circuit Theory, and Information Theory, Tokyo.
Bolten, E., Schliep, A., Schneckener, S., Schomburg, D., & Schrader, R. (2001). Clustering
protein sequencesstructure prediction by transitive homology. Bioinformatics,
17(10), 935-941.
Carrillo, H., & Lipman, D. (1988). The multiple sequence alignment problem in biology.
SIAM Journal on Applied Mathematics, 48, 1073-1082.
Enright, A. J., & Ouzounis, C. A. (2000). GeneRAGE a robust algorithm for sequence
clustering and domain detection. Bioinformatics, 16(5), 451-457.
Guralnik, V., & Karypis, G., (2001, August 26). A scalable algorithm for clustering protein
sequences. Proceedings of Workshop on Data Mining in Bioinformatics, San
Francisco (pp. 73-80). ACM Press.
Han, J., & Kamber, M., (2001). Data mining: Concepts and techniques. San Francisco:
Morgan Kaufmann Publishers.
Hubbard, T. J., Lesk, A. M., & Tramontano, A. (1996). Gathering them into the fold. Nature
Structure Biology, 4, 313.
Jain, A. K., & Dubes, R. C. (1988). Algorithms for clustering data. Englewood Cliffs, NJ:
Prentice Hall.
Jain, A. K., Murty, M. N., & Flynn, P. J. (1999). Data clustering: A review. ACM Computing
Surveys, 13(3), 264-323.
Levitt, M., & Chothia, C. (1976). Structural patterns in globular proteins. Nature, 261,
552-558.
Lo Conte, L., Brenner, S. E., Hubbard, T. J. P., Chothia, C., & Murzin, A. G. (2002). SCOP
database in 2002; refinements accommodate structural genomics. Nucleic Acids
Research, 30, 264-267.
McQueen, J.B. (1967). Some methods of classification and analysis of multivariate
observations. Proceedings of Fifth Berkeleys Symposium on Mathematical Sta-
tistics and Probability (pp. 281-297).
Murzin A. G., Brenner S. E., Hubbard T., & Chothia C. (1995). SCOP: A structure
classification of proteins database for the investigation of sequences and struc-
tures. Journal of Molecular Biology, 247, 536-540.
Sasson, O., Linial, N., & Linial, M. (2002). The metric space of proteinscomparative
study of clustering algorithms. Bioinformatics, 18(1), S14-S21.
Sugiyama, A., & Kotani, M. (2002). Analysis of gene expression data by self-organizing
maps and k-means clustering. Proceedings of IJCNN (pp. 1342-1345).
A Bayesian Framework for Improving Clustering Accuracy 247
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Tatusov, R. L., Koonin, E. V., & Lipman, D. J. (1997). A genomic perspective on protein
families. Science, 278, 631-637.
Thompson, J. D., Higgins, D. G., & Gibson, T. J. (1994). CLUSTAL W: Improving the
sensitivity of progressive multiple sequence alignment through sequence weight-
ing, position-specific gap penalties and weight matrix choice, Nucleic Acids
Research, 22, 4673-4680.
Wang, L., & Jiang, T. (1994). On the complexity of multiple sequence alignment. Journal
of Computational Biology, 1, 337-348.
Wise, M. J. (2002). The POPPs: clustering and searching using peptide probability
profiles. Bioinformatics, 18(1), S38-S45.
Yona, G., Linial, N., & Linial, M. (1999). ProtoMap: Automatic classification of protein
sequences, a hierarchy of protein families, and local maps of the protein space.
Proteins, 37, 360-378.
Yona, G., Linial, N., & Linial, M. (2000). ProtoMap: Automatic classification of protein
sequences, a hierarchy of protein families, and local maps of the protein space.
Nucleic Acids Research, 28, 49-55.
248 Park, Dam, Pan, Xu, Geist, Heffelfinger, & Samatova
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Chapter XIII
In Silico Recognition of
Protein-Protein
Interactions:
Theory and Applications
Byung-Hoon Park, Oak Ridge National Laboratory, USA
Phuongan Dam, University of Georgia, USA
Chongle Pan, University of Tennessee, USA
Ying Xu, University of Georgia, USA
Al Geist, Oak Ridge National Laboratory, USA
Grant Heffelfinger, Sandia National Laboratories, USA
Nagiza F. Samatova, Oak Ridge National Laboratory, USA
Abstract
Protein-protein interactions are fundamental to cellular processes. They are responsible
for phenomena like DNA replication, gene transcription, protein translation, regulation
of metabolic pathways, immunologic recognition, signal transduction, etc. The
identification of interacting proteins is therefore an important prerequisite step in
understanding their physiological functions. Due to the invaluable importance to
various biophysical activities, reliable computational methods to infer protein-
protein interactions from either structural or genome sequences are in heavy demand
lately. Successful predictions, for instance, will facilitate a drug design process and
the reconstruction of metabolic or regulatory networks. In this chapter, we review: (a)
In Silico Recognition of Protein-Protein Interactions: Theory and Applications 249
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
high-throughput experimental methods for identification of protein-protein
interactions, (b) existing databases of protein-protein interactions, (c) computational
approaches to predicting protein-protein interactions at both residue and protein
levels, (d) various statistical and machine learning techniques to model protein-
protein interactions, and (e) applications of protein-protein interactions in predicting
protein functions. We also discuss intrinsic drawbacks of the existing approaches and
future research directions.
Introduction
Protein-protein interactions are one of the most ubiquitous and fundamental phenomena
in all cellular activities including DNA transcription and regulation of metabolic or
signaling pathways. Uncovering the mechanisms that govern protein-protein interac-
tions, therefore, has been the subject of great interest in the post-genomic era. Emerging
genome-wide high-throughput experimental techniques (Ho et al., 2002; Ito et al., 2001;
Uetz et al., 2000; Zhu et al., 2001) are generating an increasing amount of data that provide
significant insights into the mechanisms underlying protein-protein interactions. In
parallel, new computational approaches are being introduced to infer novel interactions
or to reconstruct complex interactions such as functional modules and protein complexes
from experimental data.
In a narrow sense, a protein-protein interaction refers to a physical binding between two
or more proteins. Such a physical interaction can be categorized based on the compo-
sition of the complex, the function of the complex versus that of a monomer, the binding
affinity of subunits in the complex, the duration of the complex formation, or interactions
between specific functional groups. Accordingly, protein-protein interaction can be
homo-oligomeric (between identical subunits), such as a dimer of a transcription factor,
or hetero-oligomeric (between different subunits), such as the transcription machinery
(RNA polymerase). A physical interaction can also be classified based on the duration
of the contact. It can be transient as in the case of an enzyme-substrate contact or ligand-
receptor interaction, or it can be stable as in the case of the formation of an actin filament
or an antigen-antibody interaction. Furthermore, protein-protein interactions are often
characterized as interactions between specific functional groups. Examples include
interactions between SH
2
domain and a phosphorylated tyrosine residue or between two
PDZ domains.
In the field of bioinformatics, protein-protein interactions are understood in a much
broader sense. They not only include physical interactions but also embrace functional
associations between proteins at different times or locations. This new interpretation
includes two proteins that appear in the same pathway, although they are not physically
in contact with each other. Such functional interactions can be defined in many ways,
including proteins that are functioning in the same pathway, co-expressed under the
same conditions, co-evolved or co-localized in the same region of genomes, or even co-
present in the same literature abstracts (von Mering et al., 2005).
250 Park, Dam, Pan, Xu, Geist, Heffelfinger, & Samatova
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Computational methods typically address physical or functional interactions at the
levels of the protein and its component residues. The former is an effort to determine the
potential interaction between two proteins given their structural or sequence informa-
tion. In contrast, studies of protein interactions at the residue level focus on identifying
specific residues that interact with other residue(s) in different proteins. Computational
methods utilize various techniques to predict protein interactions at both levels,
including simple correlation analysis, sophisticated statistical assessments, and ma-
chine-learning approaches for classification.
In this chapter, we review in silico approaches to recognize protein-protein interactions
in a wide spectrum. We first introduce high-throughput experimental methods for sifting
putative physical interactions, followed by a discussion of a series of computational
methods targeted at protein and residue levels. Applications of protein interactions,
protein function and complex predictions are described within the context of protein-
protein interaction networks. We conclude the chapter with the discussion of some
future directions and needs for computational approaches to characterizing protein-
protein interactions.
Experimental Methods and Databases
of Protein-Protein Interactions
Experimental protein-protein interaction data are an invaluable resource to infer biologi-
cal phenomena. Traditionally, experimental methods for protein-protein interaction
identification are low-throughput, or focusing on individual cases. The procedures
include some degree of purification of one or more protein complexes followed by (1)
identification of the complex components using a wide-range of techniques, or (2) direct
detection of protein-protein interactions by NMR or crystallography. In practice,
purification of a protein complex largely depends on physical characteristics of the
complex as mass, ionic charge, hydrophobicity, or its affinity to a substrate such as lectin
or metal ions (Doonan, 1996). After the protein purification step, the subunits (compo-
nent proteins) are identified by their mass (mass spectrometry), their amino acid
sequence (sequencing of the peptides), or their affinity to specific antibodies generated
against them in previous studies. These low-throughput procedures are costly, time-
consuming, and do not scale.
With the advent of proteomic approaches, several high-throughput experimental
methods have emerged. Different methods can probe different types of protein interac-
tions, but all suffer from generating false positives and false negatives to some extent.
Since most protein-protein interaction databases, the foundation of many computational
approaches, are populated with data obtained through these methods, it is important to
understand their strengths and weaknesses, as well as the types of interactions they can
discover. The yeast two-hybrid assay (Fields & Song, 1989; Tucker, Gera, & Uetz, 2001;
Uetz, 2002; Uetz et al., 2000), phage display (Tong et al., 2002), affinity purification (Gavin
et al., 2002) (Aloy & Russell, 2002), and protein microarray (chip) (Zhu et al., 2001), four
of the most widely used experimental methods are depicted in Figure 1.
In Silico Recognition of Protein-Protein Interactions: Theory and Applications 251
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Protein-protein interaction data, whether produced by high-throughput methods or
retrieved from literature, are invaluable resources for computational methods. There
exists a number of publicly accessible protein interaction databases systematically
maintained to serve researchers. Built on meticulously designed relational schemas, most
provide extensive search capabilities via nice user interfaces. Retrieval of related data
in various formats is readily available, simplifying large-scale comparative analysis of
protein-protein interactions. Therefore, understanding the characteristics of these
databases is the first step in pursuing protein interaction inference.
Every year new databases dedicated for protein interactions are introduced. In particular,
a comprehensive list of protein-protein interaction databases and their URLs is pub-
lished in the first issue of Nucleic Acids Research journal each year. These protein-
protein interaction databases are currently in the category of Metabolic and Signaling
Pathways/Intermolecular interactions and signaling pathways (Galperin, 2005). From the
Figure 1. Four high-throughput experimental methods to sift protein-protein
interactions
(b) Phage Display
(a) Yeast two-hybrid


HIS3
A
DBD
DBD
C = Bait D = Prey
AD B
AD B A
DBD
HIS3
RNAP
A
DBD
AD B
Phage
P
h
a
g
e
P
h
a
g
e
Bait
Bait
Bait
L
L L
P
h
a
g
e
(b) Phage display
Positive signal detected
C
P
B
(d) Protein chip
no signal detected
Chemical linkers
Immobilized surface
B: baits P: prey C: tag for detection
P
P
Generating baits
Expression of bait proteins in the cell
Purification of protein complexes
Identification of protein complexes
A
Bait
Bait
Bait
(c) Affinity purification
P
Tag
B
A
P
252 Park, Dam, Pan, Xu, Geist, Heffelfinger, & Samatova
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
list, 29 out of 719 databases introduced in the year of 2005 are databases of protein
interactions (http://www3.oup.co.uk/nar/database/cap/). We list 11 widely used data-
bases in Table 1. Besides listed databases, some specialized databases include MHC-
Peptide Interaction Database a highly curated database for sequence-structure-
function information on MHC-peptide interactions, PDZBase a database of protein-
protein interactions involving PDZ domains, PINdb a database of protein-protein
interactions in the nucleus of human and yeast, POINT a database of predicted human
protein-protein interactome, ProNIT a protein-nucleic acid interaction database and
InterPare a database of interacting residues for all protein-protein interfaces in the PDB.
Table 1. A list of widely used protein-protein interaction databases (Note that names
of databases are shown in acronyms)

Database URL Data Source Comments
DIP
(Xenarios et al.,
2005)
http://dip.doe-mbi.ucla.edu 44,482 interactions
17,173 proteins
Experiments Vast majority of interactions are from
eukaryotes. Only Helicobacter pylori and
Escherichia coli are prokaryotes.
BIND
(Alfarano et al.,
2005)
http://bind.ca
134,886 interactions
2,525 complexes
8 pathways
Experiments
Scientific Literature
Object-oriented design. An interaction is the
basic unit, and complexes and pathways are
represented in terms of interactions.
MIPS
(Mewes et al.,
2002)
http://mips.gsf.de
11,200 interactions
1,050 complexes
Experiments Cross-reference for functional classification,
subcellular localization, and EC numbers
are available. Widely used for benchmark
data.
MINT
(Zanzoni et al.,
2002)
http://cbm.bio.uniroma2.it/
mint
43,273 interactions
18,549 proteins
Experiments
Scientific Literatures
Special emphasis on proteomes from
mammalian organisms
iPfam
(Finn, Marshall, &
Bateman, 2004)
http://www.sanger.ac.uk/S
oftware/
Pfam/iPfam/
2,828
domain interactions
PDB, Pfam Collection of domain-domain interactions
found in PDB (Pfam pairs).
InterDOM
(Ng, Zhang, &
Tan, 2003)
http://InterDom.lit.org.sg
30,037
domain interactions
Gene Fusion events
DIP, BIND, PDB
Scientific Literature
Data are integrated by assigning higher
confidence to domain interactions that are
independently derived from different data
sources and methods.
IntACT
(Hermjakob et al.,
2004)
http://www.bioinf.man.ac.
uk/ resources/
interactpr.shtml
39,715 interactions
and complexes
Scientific Literatures On-going collaboration between 9
bioinformatics centers including EBI, MPI-
MG, SIB, SDU,UB1, CNB-CSIC, HUJI,
GSK, and MINT
PDB
(Deshpande et al.,
2005)
http://www.pdb.org
29,101 molecular
structures
X-ray
crystallography or
NMR spectroscopy
Comprehensive resources on biomolecular
structures, including proteins and protein
complexes.
STRING
(Mering et al.,
2003)
http://string.embl.de/
Predicted interactions
in 444,238 genes from
110 species
Experiments
Genomic Context
Scientific literature
Predicted interactions
Database of known and predicted
interactions. Both direct and indirect
interactions are listed. Predictions in terms
of COG orthologs are also available.
PREDICTOME
(Mellor, Yanai,
Clodfelter,
Mintseris, &
DeLisi, 2002)
http://predictome.bu.edu 300,505 predicted
interactions
44 microbial
genomes
Experiments
Predicted interactions
Nice tool for visualizing the predicted
functional associations among genes and
proteins in many different organisms.
LiveDIP
(Duan, Xenarios,
& Eisenberg,
2002)
http://dip.doembi.ucla.edu/
ldip.html
35 types of chemical
modification
408 interactions
Extension of DIP Describes protein interactions by protein
states and state transitions.

In Silico Recognition of Protein-Protein Interactions: Theory and Applications 253
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Interface Site Prediction: Protein
Interaction at Residue Level
When two or more proteins physically interact, the residues on a protein surface that are
actually in contact with the residues of the other proteins are called protein interface
sites. The prediction of protein interface sites is an effort to identify residues that
constitute such contact regions from either primary sequence or tertiary structural
information. Identification of protein interface sites is of great importance in areas like
protein docking (Wodak & Mendez, 2004) that suffers from the astronomical size of the
search space of possible conformations.
Definition and Properties of Interface Sites
Protein interface sites are defined in terms of either the distances between residues or
the reduction of Accessible Surface Area (ASA). For example, if the distance between
any heavy atoms (or between C

or C

atoms) is less than a certain threshold, such as 5

(Zhou & Shan, 2001), the residue pair is classified as an interface site. In some other
cases, the threshold is set in terms of absolute or relative reduction (percentage of ASA
loss upon complexation) of ASA, such as 1
2
(Bahadur, Chakrabarti, Rodier, & Janin,
2003; Jones & Thornton, 1995), or 5% (Caffrey, Somaroo, Hughes, Mintseris, & Huang,
2004). Over the last decades, a number of attempts have been made to explain interface
sites in terms of residue propensity (Ofran & Rost, 2003b), hydrophobicity (Bahadur et
al., 2003; Jones & Thornton, 1995, , 1996; Korn & Burnett, 1991), shape complementarities
(Jones & Thornton, 1996), and in other terms. However, the underlying principles that
govern protein interface sites are not yet fully understood.
Computational Approaches to Interface Sites Prediction
Most computational approaches are based on the assumption that residue type propen-
sities, computed from multiple sequence alignment of orthologs, are not random in areas
immediately surrounding interface sites.
Since, for each residue in a known protein complex, the label (interface, surface, or core)
can be easily obtained, machine learning methods like artificial neural network or support
vector machines are widely used to predict interface sites. Specifically, for each residue,
amino acid propensities of either spatial or sequence neighbors are transformed into a
feature vector. Occasionally, additional information such as ASA is appended (Zhou &
Shan, 2001). If spatial location is considered when building the feature vector, the
resulting classifier is only valid for predicting proteins with known structures (Fariselli,
Pazos, Valencia, & Casadio, 2002; Zhou & Shan, 2001). In that sense, a classifier
constructed from sequence neighbors has wider applications (Ofran & Rost, 2003b; Yan,
Dobbs, & Honavar, 2004).
254 Park, Dam, Pan, Xu, Geist, Heffelfinger, & Samatova
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
The overall accuracy in identifying interface sites is still far from satisfactory. To boost
performance, a more sophisticated classification framework has recently been intro-
duced, where predictions from multiple SVMs are merged by a weighted scheme (Park,
Munavalli, Geist, & Samatova, 2004) or integrated by a Bayesian classifier (Yan et al.,
2004). While such efforts will help produce a slightly more accurate classifier, the
fundamental limitation of computational approaches can only be resolved when a more
appropriate set of complexes and more descriptive feature vectors becomes available.
Focusing on specific subsets of interface sites is one such effort (Ofran & Rost, 2003a,
2003b).
Computational Approaches to Predicting Protein-Protein
Interactions
While the vast amount of high-throughput experimental data gives us new opportunities
to compile rich sources of valuable genome-wide interactomes, we are still far behind in
understanding fundamental principles that govern protein-protein interactions. Further-
more, high-throughput data is considered to be error prone, and the coverage of data is
very limited to simpler organisms such as Saccharomyces cerevisiae and Hellicobacter
pylori. Computational approaches play a complementary role to experimental methods
by predicting novel interactions, assessing reliabilities of the existing interactions, and
building mathematical models that may explain protein interactions through transforming
biological intuition into mathematical formulations.
Figure 2. Genomic context-based methods are based on the assumption that functionally
related proteins are encoded by genes that are regulated or evolved together: Three
different approaches attempt to uncover such evidences through comparative genome
analysis, (a) Phylogenenetic profile captures pairs of genes based on their co-
occurrences pattern across genomes, (b) gene neighborhood identifies pairs of genes
in vicinity, and (c) gene fusion examines pairs of genes that exist individually in some
organisms, but as a fused gene in other organisms
(a) Phylogenetic Profile (b) Gene Neighborhood (c) Gene Fusion (a) Phylogenetic profile (b) Gene neighborhood (c) Gene fusion
In Silico Recognition of Protein-Protein Interactions: Theory and Applications 255
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Genomic Context-Based Methods
Genomic context-based methods are based on the assumption that functionally related
proteins are encoded by genes that are co-regulated or co-evolved. These genes are
believed to leave detectable signatures over multiple genomes that can be identified by
comparative analysis. Different methods seek to uncover such signatures in different
ways (see Figure 2). Gene fusion methods (Enright, Iliopoulos, Kyrpides, & Ouzounis,
1999; Marcotte, Pellegrini, Thompson, Yeates, & Eisenberg, 1999; Sali, 1999) search for
two genes that exist individually in some genomes, but as a fused gene in other genomes.
Typically, a non-overlapping side-by-side BLAST search is performed to identify fused
genes (Enright & Ouzounis, 2001; Huynen, Snel, Mering, & Bork, 2003; Mellor,Yanai,
Clodfelter, Mintseris, & DeLisi, 2002; Suhre & Claverie, 2004). The gene neighborhood
method considers the conserved proximity between two genes for a potential interaction
(Dandekar, Snel, Huynen, & Bork, 1998; Overbeek, Fonstein, DSouza, Pusch, & Maltsev,
1999). Whereas the above two methods consider spatial co-location in various genomes,
phylogenetic profiling (Pellegrini, Marcotte, Thompson, Eisenberg, & Yeates, 1999)
Figure 3. Two methods for predicting interacting proteins: (a) Given two proteins, a
pair of organism-aligned MSAs is built (b) from two mirror phylogenetic trees A and
B that are constructed from MSAs in (a), each pair of distances is mapped to 2-D plane.
A high correlation indicates potential interaction (c) both correlations of intra and
inter residue pairs are computed, and three distributions are compared to determine
potential interaction
org 1
Distance pairs in A
Inter Intra Intra
Intra and Inter
distributions are
compared
B
(a) Pair of organism-aligned multiple sequence alignments.
(b) Phylogenetic tree method (c) in silico two hybrid method
org 2
org n
org 2
org n
org 1
(a) Pair of organism-aligned multiple sequence alignments
(b) Phylogenetic tree method (c) in silico two hybrid method
Intra and Inter
distributions are
compared
Distance pairs in A
Intra Inter Intra
B
256 Park, Dam, Pan, Xu, Geist, Heffelfinger, & Samatova
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
examines temporal co-appearance as an evidence of potential interaction. In practice, two
proteins are predicted to interact, if their orthologs are found to have correlated
appearances (Matthews et al., 2001) over a set of genomes. In a further refinement,
phylogenetic trees (Pazos & Valencia, 2001) can be used to quantify the degree of co-
evolution rather than co-existence between two genes. For two genes, two mirror
phylogenetic trees are constructed from organism-aligned multiple sequence alignments
of associated protein families. Then all pairs of distances measured from the same path
in each tree are measured. Two genes (or proteins accordingly) are predicted to interact
if high correlation is observed between all pairs of distances. The in silico two hybrid
(i2h) method {Pazos, 2002 #129}, the distributions of correlations between inter-protein
positions and correlations between intra-protein positions are compared. Based on the
difference, the i2h predicts a potential interaction between the proteins. Detailed
procedures are shown in Figure 3.
A genomic context-based approach suggests that genomic associations between genes
reveal functional associations between them; the stronger the genomic association, the
stronger the functional association (Snel, Bork, & Huynen, 2002). For this reason,
genomic associations cover a relatively wide range of protein interactions including
physical and functional interactions. However, its applicability is somewhat limited for
several reasons; it requires sufficient number of completely sequenced genomes, it is
only applicable to prokaryotes, and not many proteins follow the principle of genomic
associations (Brun et al., 2003).
Domain-Domain Interactions
Protein domains are evolutionarily conserved segments of an amino acid sequence.
Biologically, they are often responsible for the structural formation or functional
behavior of proteins, and are frequently involved (either directly or indirectly) in
intermolecular interactions. For this reason, there have been attempts to search for pairs
of protein domains whose coexistence can account for putative protein-protein interac-
tions. In essence, most approaches focus on assessing statistical significances of
domain pairs found in high-throughput experimental data. For example, let P
i
and P
j
be
interacting proteins that include domains {a,b} and {c,d}, respectively. Then the
interaction may be explained by some of 4 possible domain interactions: (a,c), (a,d), (b,c),
and (b,d). The main objective is, through statistical assessments, to identify which
domain pairs are indeed responsible for the interaction.
Sprinzak and Margalit (2001) consider protein interactions in terms of InterPro (Mulder
et al., 2003) signature pairs whose co-occurrences are more frequent than what is
expected at random. The Protein Interaction Classification by Unlikely Profile Pair
(PICUPP) introduces a more reliable method that assesses the interaction probability of
a domain pair by bootstrapping (Park et al., 2003). Another probabilistic approach that
utilizes the scale-free property (see Section 6 for details) of protein interactions (Gomez,
Noble, & Rzhetsky, 2003) has been proposed that considers not only the number of times
the domain pair is observed, but also the number of times it is found in non-interacting
proteins. The Interaction Domain Profile-Pair (IDPP) method (Wojcik & Schachter, 2001)
represents protein interactions as interactions between domain clusters. Whereas most
In Silico Recognition of Protein-Protein Interactions: Theory and Applications 257
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
approaches aim to estimate the likelihood of each domain pair individually, a global
optimization approach that simultaneously estimates probabilities of all domain-domain
interactions using Expectation-Maximization (EM) algorithm has been proposed (Deng,
Mehta, Sun, & Chen, 2002). As an integrative approach, Ng, Zhang, and Tan (2003)
combined evidences of interacting Pfam domain pairs obtained from various data
sources. Each domain pair was given a number of odd ratio scores (observed/expected)
which were integrated using an unweighted scheme.
Classification Methods
The most commonly used classifier for predicting interacting protein pairs is the Support
Vector Machine (SVM) (Bock & Gough, 2001, 2003; Martin, Roe, & Faulon, 2005). SVM
originated from statistical learning theory (Vapnik, 1998) and has been widely accepted
in other bioinformatics areas, such as homology detection (Leslie, Eskin, Weston, &
Noble, 2003; Saigo, Vert, Ueda, & Akutsu, 2004), gene classification (Brown et al., 2000),
prediction of RNA-binding proteins (Han, Cai, Lo, Chung, & Chen, 2004), protein folding
recognition (Ding & Dubchak, 2001), prediction of subcellular localization (Park &
Kanehisa, 2003), and prediction of protein secondary structure (Hua & Sun, 2001). The
main challenge in applying SVM to predict protein-protein interactions is to employ an
appropriate encoding of protein-protein interactions in a vector space. This primarily
involves the choice of biochemical (or biological) properties that can differentiate
interacting from non-interacting pairs.
Most classification methods assume the absence of structural information, i.e. only a pair
of primary sequences is available. Therefore, physiochemical properties associated with
a primary sequence such as hydrophobicity, charge, and surface tension (Bock & Gough,
2001) are often considered. To generate a feature vector, the values for each residue in
a sequence are tabulated in a fixed order. Note, however, this type of transformation
suffers from two drawbacks. First, the transformation of proteins of various lengths into
a fixed length vector can be problematic. Second, if v
i
and v
j
denote the transformed
feature vectors for protein i and protein j, respectively, the construction of a feature
vector for protein pair (v
i
,v
j
) is not readily obvious. An early SVM-based method simply
merged the two vectors; the final feature vector for a protein pair became either v
i
+v
j
or
v
j
+v
i
, where + denotes concatenation of the two vectors. Therefore, direct mapping from
physiochemical properties to a vector space needs to be reexamined.
Despite the fact that the classification approach has been proven to be successful in
many other areas, and an early attempt applying this method to protein-protein interac-
tion prediction achieved a certain level of success (Bock & Gough, 2001), its practical
applicability is still in question mainly due to the absence of negative instances.
Technically, there exists no experimental method that can detect pairs of proteins that
do not interact at all. In practice, randomly chosen pairs of proteins substitute for
negative instances (Bock & Gough, 2003; Martin et al., 2005). Occasionally, pairs of
synthetically generated proteins are used for the same purpose (Bock & Gough, 2001).
However, their biological validity is difficult to estimate. Thus, the creation of negative
instances based on true biological principles (subcellular localization, for example) is
much anticipated.
258 Park, Dam, Pan, Xu, Geist, Heffelfinger, & Samatova
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Inference from Imperfect Data
Although the advances in high-throughput experimental methods make exhaustive
inspections for all potential interactions possible in some smaller organisms, there remain
significant discrepancies in data produced by different methods. In the Saccharomyces
cerevisiae proteome, for example, roughly 10,000 interactions are identified from four
different methods, yet the overlaps are quite limited (see Table 2). Also, the total number
of interactions currently identified is only one third of what is predicted to exist (von
Mering et al., 2002). Although the small overlap may be partially due to differences in the
types of interactions that an experimental method can discover, it mainly comes from the
lack of reproducibility of the methods as being discussed in the yeast two-hybrid system.
This unreliable characteristic of high-throughput datasets calls for (1) computational
assessment of data quality and (2) robust approaches to infer from such imperfect data.
In principle, data quality can be assessed by validation utilizing high-quality benchmark
data. However, due to the limited availability of such a reliable dataset (von Mering et
al., 2002), quality is often examined by considering additional properties found in
different information sources. For example, Deane et al. introduce two measures: the
Expression Profile Reliability Index (EPR Index) and the Paralogous Verification Method
(PVM) (Deane, Salwinski, Xenarios, & Eisenberg, 2002). The EPR assesses the quality
of protein interaction datasets by comparing mRNA expression profiles of interacting
protein pairs in the dataset to those of known interacting and non-interacting protein
pairs. The PVM predicts that two proteins potentially interact if their paralogs also
interact. Further discussions on data quality assessment can be found in (Salwinski &
Eisenberg, 2003).
Two approaches exist for inferences from imperfect data. In the first approach, the most
reliable subset of high-throughput data is sought by either taking the intersection of
multiple data sources (Bader & Hogue, 2002; von Mering et al., 2002), performing
Table 2. Intersections of protein interactions identified in different high-throughput
experimental datasets. Values in bold denote the number of interactions in each
database, and others show overlaps between databases that are referenced by
corresponding column and row. Small-scale DIP denotes the reliable small-scale
studies that are listed in DIP database (Salwinski et al., 2004). Every pair of high-
throughput datasets shows very limited agreements. However, all datasets identified
many interactions in small-scale studies. The data in this table are borrowed from
(Salwinski & Eisenberg, 2003).
Ito et al. Uetz et al. Gavin et al. Ho et al. Small-scale DIP
Ito et al. 4363 186 54 63 442
Uetz et al. 1403 54 56 415
Gavin et al. 3322 198 528
Ho et al. 3596 391
In Silico Recognition of Protein-Protein Interactions: Theory and Applications 259
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
statistical assessments (Sprinzak, Sattath, & Margalit, 2003), analyzing network topol-
ogy (Goldberg & Roth, 2003; Saito, Suzuki, & Hayashizaki, 2003), or by taking the
combination of these three (Bader, Chaudhuri, Rothberg, & Chant, 2004). The second
approach takes the opposite direction: integration of multiple data sources through
sophisticated probabilistic approaches. For example, STRING (Huynen et al., 2003)
combines predictions from multiple methods, which use different types of data, by
assigning a unique weight to each method based on its benchmark result against a
common reference set. The Probabilistic Interactome (PI) (Jansen et al., 2003) takes a step
further in merging multiple data sources. Like STRING, PI also assesses the weight of
each data source. For this, two Bayesian networks are constructed from high-throughput
experimental data and other genomic features, respectively. Then, a final Bayesian
network (a simple nave Bayes classifier) is built based on probabilities estimated from
the two networks. Alternatively, multiple data sources can be selectively integrated.
Based on structural risk minimization, Bock and Gough introduced a framework that
iteratively brings in experimental data sets for different organisms in such a way that the
prediction accuracy through cross-validation is improved (Bock & Gough, 2003). It is a
growing consensus that the integration of multiple data sources can significantly
facilitate inference of protein interactions. In the near future, more diverse data sources
are expected to be integrated along this direction (Salwinski & Eisenberg, 2003).
Characterization of
Protein Interaction Networks and
Prediction of Protein Functions
The responses of a cell to environmental conditions are governed by complex intermo-
lecular interactions. For example, in a metabolic pathway, hundreds of proteins are
interconnected through various biochemical reactions. Thanks to large-scale, high-
throughput identification of protein-protein interactions and various computational
techniques, genome-scale protein interaction networks are available, where both direct
and indirect interactions between proteins are represented. However, identifying the
functional role of a protein or cellular functional units through analysis of interaction
networks has become a major challenge in proteomics. Most approaches follow three
steps: (1) the construction of genome-wide protein-protein interaction networks in terms
of mathematical representation, i.e. a graph, where nodes are proteins, and edges are
interactions between them (Ito et al., 2001; Uetz et al., 2000); (2) the determination of graph
properties such as cliques, shortest paths, etc., and (3) functional characterization of
proteins and protein complexes based on these properties. In this section, we describe
how a protein-protein interaction network is utilized to infer biological complexes and
protein functions.
260 Park, Dam, Pan, Xu, Geist, Heffelfinger, & Samatova
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Network Properties
A number of network models have been suggested to characterize protein-protein
interaction networks. Each model lays a foundation to identify protein function and
functional modules within its framework. The most widely accepted models are scale-free
networks (Barabasi & Albert, 1999; Fraser, Hirsh, Steinmetz, Scharfe, & Feldman, 2002;
Hoffmann & Valencia, 2003; Jeong, Mason, Barabasi, & Oltvai, 2001; Wuchty, 2002) and
small-world networks (Goldberg & Roth, 2003; Watts & Strogatz, 1998). Whereas the
scale-free model delineates the global connectivity of a network, the small-world model
rather focuses on elucidating the local connectivity. The former suggests that there exist
very few highly connected nodes following the power-law distribution. It implies that
such highly connected proteins (often called hubs) participate in numerous metabolic
reactions, thus explaining the robust self-organizing behavior of a protein interaction
network. Another view, the small-world model implies that the shortest path between any
pair of proteins tends to be small, and the network is full of densely connected local
neighborhoods. The two models are not in conflict; rather, they complement each other
(Snel et al., 2002). In fact, it has recently reported that the yeast co-expression network
can be successfully modeled as a small-world, scale-free network (van Noort, Snel, &
Huynen, 2004).
An alternative view, which is based on the assumption that cellular functionality can be
divided into a set of modules, is also widely used. This modular network model (Brun et
al., 2003; Rives & Galitski, 2003) suggests that protein interaction networks consist of
several densely interconnected functional modules that are connected by few cross-
talks. However, such a modular structure implies that most nodes roughly have the equal
edge degrees, which is against the scale-free nature. There has been an attempt to resolve
this conflict by devising a hierarchical modular network, where modules constitute
higher level modules (Ravasz, Somera, Mongru, Oltvai, & Barabasi, 2002). However, the
disparity between these two models still remains.
Protein Function Prediction through Network Analysis
Protein function prediction through network analysis is based on the notion of guilt-
by-association (Oliver, 2000). It implies that a protein performs its function through
interactions with other proteins that have the same cellular function. Hence, large
complex protein interaction networks are now analyzed in detail to associate proteins of
known functions with hitherto hypothetical or uncharacterized proteins, and to extract
cellular processes in which a number of interacting proteins are involved. For example,
given a whole proteome that consists of both known and unknown proteins, functional
annotations for the unknowns can be determined by finding the best configurations that
maximize the presence of the same functional association among interacting proteins
(Vazques, Flammini, Maritan, & Vespignani, 2003).
The graph-theoretic properties of a protein interaction network facilitate functional
classifications of proteins. Following the scale-free network model, highly connected
proteins (hubs) are identified to account for viability of the cell (Jeong et al., 2001).
In Silico Recognition of Protein-Protein Interactions: Theory and Applications 261
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Biologically, such proteins are confirmed to evolve very slowly (Fraser et al., 2002). In
contrast, several studies show that clustering the interaction network can isolate a
number of functional modules, demonstrating the modular property of protein interaction
networks (Brun et al., 2003; Rives & Galitski, 2003). Such modularity is particularly
evident in networks with a high clustering coefficient that indicates how densely the
neighbors of a protein are interconnected. In summary, it has been shown that some
graph-theoretic properties are successfully mapped to cellular functional units; signal-
ing pathways, important proteins, and protein complexes are described in terms of
cliques, articulation nodes, subgraphs, etc. (Bader & Hogue, 2002; Przulj, Wigle, &
Jurisica, 2004; Watts & Strogatz, 1998).
Protein Complex Prediction
Unlike a functional module, which consists of proteins that bind to (or interact with) each
other at different times and locations to perform a certain cellular process, a protein
complex is comprised of multiple proteins (subunits) that physically interact to form a
single molecular structure. Though computational approaches to predict protein com-
plexes are typically done via docking simulations with high resolution structures
available, many recent studies demonstrate that protein complexes can also be detected
by analyzing protein-protein interaction maps. For example, Spirin and Mirny (2003)
studied highly connected subgraphs of meso-scale (5-25 proteins) clusters and con-
firmed that many of these clusters correspond to protein complexes. A probabilistic
approach also reported that a strong interaction highly correlates with a potential
membership in the same complex (Asthana, King, Gibbons, & Roth, 2004), where a strong
interaction is represented as an edge with high probability in the network.
Figure 4. Three illustrative network topologies. In scale-free networks (a), a few highly
connected nodes are linked with others. In small-world networks (b), any node can be
reached by a small number of intermediate nodes. In hierarchical modular networks
(c), a node belongs to a module, which in turn belongs to a higher module in the
hierarchy.
(a) scale-free (b) small-world (c) hierarchical modular
262 Park, Dam, Pan, Xu, Geist, Heffelfinger, & Samatova
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Computational methods for predicting protein functions and complexes inevitably deal
with imperfect data. Protein interaction networks are mostly constructed from experimen-
tal data that are susceptible to errors and noise. Although we expect more extensive and
accurate interaction data due to the advancement of recent mass spectrometry during the
last decade (Janin & Seraphin, 2003; Sobott & Robinson, 2002; Yates III, 2000), attempts
to appropriately integrate various data sources will also help encompass the coverage
and enhance the integrity of the networks (Hoffmann & Valencia, 2003; Lee, Date, Adai,
& Marcotte, 2004). For example, many signatures of protein complex membership have
been observed in other data sources including gene expression data and subcellular
localization data. Therefore, systematic integration of various genomic data with the
current protein-protein interaction maps is much anticipated (Jansen, Lan, Qian, &
Gerstein, 2002).
Summary
Ultrascale computing and high-throughput technologies, such as genomics,
transcriptomics, proteomics and metabolomics have been generating an avalanche of
data that offers new opportunities to build detailed models of molecular interaction
networks including physical, chemical, and biological processes at increasing levels of
detail. However, there still remain growing challenges of interpreting and integrating, in
a biologically meaningful way, the flux of complex and noisy data thus produced. As a
result, false-positive and false-negative ratios of current computational methods are
high. Also the coverage of such predictions, both within the same organism and among
differing organisms, is still very limited.
We foresee several directions in future computational efforts to remedy the current
limitations. First, integration of heterogeneous data of complementary methods (microarray
data, for example) into a single data mining framework will be in great demand. This is
particularly useful for systematic biological inference through well-defined data mining
techniques given the absence of strong direct evidences. For example, the growth of
identified protein complexes is outpaced by that of other experimental data. However,
given the importance of structural information for understanding how interactions are
associated with functions, an integrating effort to sift indirect inference is challenging,
yet much anticipated. Second, with a constantly increasing volume of experimental data,
the most reliable benchmark data set needs to be materialized through highly robust
statistical assessment methods. This data will greatly facilitate the development of new
computational frameworks by providing a reliable and universal test-bed. In contrast to
integration efforts, we expect more diverse and more specialized computational methods
to emerge. For example, in protein interface site prediction, mate-specific methods will
appear. Instead of trying to identify interface residues in a protein given a blind
assumption that the protein interacts with other proteins, the method will be refined to
locate interacting residues with respect to a specific partner. Finally, representing/
reconstructing protein-protein interaction networks with spatial and temporal dimen-
sions will help unveil particular functional blocks that constitute other biological
networks including metabolic/signaling pathways and regulatory networks.
In Silico Recognition of Protein-Protein Interactions: Theory and Applications 263
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Acknowledgments
This work is funded in part or in full by the US Department of Energys Genomes to Life
program (http://doegenomestolife.org/ ) under the project Carbon Sequestration in
Synechococcus sp.: From MolecularMachines to Hierarchical Modeling. The work of
N.F.S. is partially sponsored by the Laboratory Directed Research and Development
Program of Oak Ridge National Laboratory. P.D. and Y.X. are partially supported by the
National Science Foundation (#NSF/DBI-0354771, #NSF/ITR-IIS-0407204).
References
Alfarano, C., Andrade, C. E., Anthony, K., Bahroos, N., Bajec, M., Bantoft, K., et al. (2005).
The biomolecular interaction network database and related tools 2005 update.
Nucleic Acids Research, 33(suppl_1), D418-424.
Aloy, P., & Russell, R. B. (2002). The third dimension for protein interactions and
complexes. Trends in Biochemical Sciences, 27(12), 633-638.
Asthana, S., King, O. D., Gibbons, F. D., & Roth, F. P. (2004). Predicting protein complex
membership using probabilistic network reliability. Genome Res., 14(6), 1170-1175.
Bader, G., & Hogue, C. (2002). Analyzing yeast protein-protein interaction data obtained
from different sources. Nature Biotechnology, 20(10), 991-997.
Bader, J. S., Chaudhuri, A., Rothberg, J. M., & Chant, J. (2004). Gaining confidence in
high-throughput protein interaction networks. Nature Biotechnology, 22(1), 78-
85.
Bahadur, R. P., Chakrabarti, P., Rodier, F., & Janin, J. (2003). Dissecting subunit interfaces
in homodimeric proteins. Proteins, 53(3), 708-719.
Barabasi, A. L., & Albert, R. (1999). Emergence of scaling in random networks. Science,
286(5439), 509-512.
Bock, J. R., & Gough, D. A. (2001). Predicting protein-protein interactions from primary
structure. Bioinformatics, 17(5), 455-460.
Bock, J. R., & Gough, D. A. (2003). Whole-proteome interaction mining. Bioinformatics,
19(1), 125-134.
Brown, M. P. S., Grundy, W. N., Lin, D., Cristianini, N., Sugnet, C. W., Furey, T. S., et al.
(2000). Knowledge-based analysis of microarray gene expression data by using
support vector machines. PNAS, 97(1), 262-267.
Brun, C., Chevenet, F., Matin, D., Wojcik, J., Guenoche, A., & Bernard, J. (2003).
Functional classification of proteins for the prediction of cellular function from a
protein-protein interaction network. Genome Biology, 5(R6), 1465-6906.
Caffrey, D. R., Somaroo, S., Hughes, J. D., Mintseris, J., & Huang, E. S. (2004). Are protein-
protein interfaces more conserved in sequence than the rest of the protein surface?
Protein Sci, 13(1), 190-202.
264 Park, Dam, Pan, Xu, Geist, Heffelfinger, & Samatova
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Dandekar, T., Snel, B., Huynen, M., & Bork, P. (1998). Conservation of gene order: a
fingerprint of proteins that physically interact. Trends in Biochemical Sciences,
23(9), 324-328.
Deane, C. M., Salwinski, L., Xenarios, I., & Eisenberg, D. (2002). Protein interactions: Two
methods for assessment of the reliability of high-throughput observations. Mol
Cell Proteomics, M100037-MCP100200.
Deng, M., Mehta, S., Sun, F., & Chen, T. (2002). Inferring domain-domain interactions
from protein-protein-interactions. Genome Res., 12(10), 1540-1548.
Deshpande, N., Addess, K. J., Bluhm, W. F., Merino-Ott, J. C., Townsend-Merino, W.,
Zhang, Q., et al. (2005). The RCSB protein data bank: a redesigned query system
and relational database based on the mmCIF schema. Nucleic Acids Research,
33(suppl_1), D233-237.
Ding, C. H. Q., & Dubchak, I. (2001). Multi-class protein fold recognition using support
vector machines and neural networks. Bioinformatics, 17(4), 349-358.
Doonan, S. (1996). Protein purification protocols. General strategies. Methods Mol Biol,
59, 1-16.
Duan, X. J., Xenarios, I., & Eisenberg, D. (2002). Describing biological protein interac-
tions in terms of protein states and statetransitions : THE LiveDIP DATABASE.
Mol Cell Proteomics, 1(2), 104-116.
Enright, A. J., Iliopoulos, I., Kyrpides, N. C., & Ouzounis, C. A. (1999). Protein interaction
maps for complete genomes based on gene fusion events. Nature, 402, 86.
Enright, A. J., & Ouzounis, C. A. (2001). Functional associations of proteins in entire
genomes by means of exhaustive detection of gene fusions. Genome Biology, 2,
341-347.
Fariselli, P., Pazos, F., Valencia, A., & Casadio, R. (2002). Prediction of protein-protein
interaction sites in heterocomplexes with neural networks. Eur J Biochem, 269(5),
1356-1361.
Fields, S., & Song, O. (1989). A novel genetic system to detect protein-protein interac-
tions. Nature, 340(6230), 245-256.
Finn, R. D., Marshall, M., & Bateman, A. (2004). iPfam: visualisation of protein-protein
interactions in PDB at domain and amino acid resolutions. Bioinformatics, bti011.
Fraser, H. B., Hirsh, A. E., Steinmetz, L. M., Scharfe, C., & Feldman, M. W. (2002).
Evolutionary Rate in the Protein Interaction Network. Science, 296(5568), 750-752.
Galperin, M. Y. (2005). The Molecular Biology Database Collection: 2005 update. Nucleic
Acids Research, 33(suppl_1), D5-24.
Gavin, A. C., Bosche, M., Krause, R., Grandi, P., Marzioch, M., Bauer, A., et al. (2002).
Functional organization of the yeast proteome by systematic analysis of protein
complexes. Nature, 415(6868), 141-147.
Goldberg, D. S., & Roth, F. P. (2003). Assessing experimentally derived interactions in
a small world. PNAS, 100(8), 4372-4376.
Gomez, S. M., Noble, W. S., & Rzhetsky, A. (2003). Learning to predict protein-protein
interactions from protein sequences. Bioinformatics, 19(15), 1875-1881.
In Silico Recognition of Protein-Protein Interactions: Theory and Applications 265
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Han, L. Y., Cai, C. Z., Lo, S. L., Chung, M. C. M., & Chen, Y. Z. (2004). Prediction of RNA-
binding proteins from primary sequence by a support vector machine approach.
RNA, 10(3), 355-368.
Hermjakob, H., Montecchi-Palazzi, L., Lewington, C., Mudali, S., Kerrien, S., Orchard, S.,
et al. (2004). IntAct: An open source molecular interaction database. Nucleic Acids
Research, 32(90001), D452-455.
Ho, Y., Gruhler, A., Heilbut, A., Bader, G. D., Moore, L., Adams, S. L., et al. (2002).
Systematic identification of protein complexes in Saccharomyces cerevisiae by
mass spectrometry. Nature, 415(6868), 180-183.
Hoffmann, R., & Valencia, A. (2003). Protein interaction: same network, different hubs.
Trends in Genetics, 19(12), 681-683.
Hua, S., & Sun, Z. (2001). A novel method of protein secondary structure prediction with
high segment overlap measure: Support vector machine approach. Journal of
Molecular Biology, 308(2), 397-407.
Huynen, M. A., Snel, B., Mering, C. v., & Bork, P. (2003). Function prediction and protein
networks. Current Opinion in Cell Biology, 15(2), 191-198.
Ito, T., Chiba, T., Ozawa, R., Yoshida, M., Hattori, M., & Sakaki, Y. (2001). A comprehen-
sive two-hybrid analysis to explore the yeast protein interactome. PNAS, 98(8),
4569-4574.
Janin, J., & Seraphin, B. (2003). Genome-wide studies of protein-protein interaction. Curr
Opin Struct Biol, 13(3), 383-388.
Jansen, R., Lan, N., Qian, J., & Gerstein, M. (2002). Integration of genomic datasets to
predict protein complexes in yeast. Journal of Structural and Functional Genomics,
2(2), 71-81.
Jansen, R., Yu, H., Greenbaum, D., Kluger, Y., Krogan, N. J., Chung, S., et al. (2003). A
Bayesian networks approach for predicting protein-protein interactions from
genomic data. Science, 302(5644), 449-453.
Jeong, H., Mason, S. P., Barabasi, A. L., & Oltvai, Z. N. (2001). Lethality and centrality
in protein networks. Nature, 411, 41.
Jones, S., & Thornton, J. M. (1995). Protein-protein interactions: a review of protein dimer
structures. Prog Biophys Mol Biol, 63(1), 31-65.
Jones, S., & Thornton, J. M. (1996). Principles of protein-protein interactions. Proc Natl
Acad Sci USA, 93(1), 13-20.
Korn, A. P., & Burnett, R. M. (1991). Distribution and complementarity of hydropathy in
multisubunit proteins. Proteins, 9(1), 37-55.
Lee, I., Date, S. V., Adai, A. T., & Marcotte, E. M. (2004). A probabilistic functional
network of yeast genes. Science, 306(5701), 1555-1558.
Leslie, C., Eskin, E., Weston, J., & Noble, W. (2003). Mismatch string kernels for SVM protein
classification. Advances in Neural Information Processing Systems, 15, 1441-1448.
Marcotte, E. M., Pellegrini, M., Thompson, M., Yeates, T., & Eisenberg, D. (1999). A
combined algorithm for genome-wide prediction of protein function. Nature,
402(23), 25-26.
266 Park, Dam, Pan, Xu, Geist, Heffelfinger, & Samatova
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Martin, S., Roe, D., & Faulon, J. L. (2005). Predicting protein-protein interactions using
signature products. Bioinformatics, 21(2), 218-226.
Matthews, L. R., Vaglio, P., Reboul, J., Ge, H., Davis, B. P., Garrels, J., et al. (2001).
Identification of potential interaction networks using sequence-based searches for
conserved protein-protein interactions or interologs. Genome Res., 11(12), 2120-2126.
Mellor, J. C., Yanai, I., Clodfelter, K. H., Mintseris, J., & DeLisi, C. (2002). Predictome: A
database of putative functional links between proteins. Nucleic Acids Research,
30(1), 306-309.
Mering, C. v., Huynen, M., Jaeggi, D., Schmidt, S., Bork, P., & Snel, B. (2003). STRING:
A database of predicted functional associations between proteins. Nucleic Acids
Research, 31(1), 258-261.
Mewes, H. W., Frishman, D., Guldener, U., Mannhaupt, G., Mayer, K., Mokrejs, M., et
al. (2002). MIPS: A database for genomes and protein sequences. Nucleic Acids
Research, 30(1), 31-34.
Mulder, N. J., Apweiler, R., Attwood, T. K., Bairoch, A., Barrell, D., Bateman, A., et al.
(2003). The InterPro Database, 2003 brings increased coverage and new features.
Nucleic Acids Research, 31(1), 315-318.
Ng, S.-K., Zhang, Z., & Tan, S.-H. (2003). Integrative approach for computationally
inferring protein domain interactions. Bioinformatics, 19(8), 923-929.
Ofran, Y., & Rost, B. (2003a). Analysing six types of protein-protein interfaces. Journal
of Molecular Biology, 325(2), 377-387.
Ofran, Y., & Rost, B. (2003b). Predicted protein-protein interaction sites from local
sequence information. FEBS Letters, 544(1-3), 236-239.
Oliver, S. (2000). Proteomics: Guilt-by-association goes global. Nature, 403(6770), 601-
603.
Overbeek, R., Fonstein, M., DSouza, M., Pusch, G. D., & Maltsev, N. (1999). The use of
gene clusters to infer functional coupling. PNAS, 96(6), 2896-2901.
Park, B.-H., Munavalli, R., Geist, A., & Samatova, N. F. (2004, April, 2004). Analysis of
Protein Interaction Sites through Separated Data Spaces. Paper presented at the
Bioinformatics Workshop in conjunction with the Fourth SIAM International
Conference on Data Mining, Orlando, FL.
Park, B.-H., Ostrouchov, G., Gong-Xin, Y., Geist, A., Gorin, A., & Smatova, N. F. (2003).
Inference of Protein-Protein Interactions by Unlikely Profile Pair. Paper pre-
sented at the SIAM International Conference on Data Mining.
Park, K.-J., & Kanehisa, M. (2003). Prediction of protein subcellular locations by support
vector machines using compositions of amino acids and amino acid pairs.
Bioinformatics, 19(13), 1656-1663.
Pazos, F., & Valencia, A. (2001). Similarity of phylogenetic trees as indicator of protein-
protein interaction. Protein Eng., 14(9), 609-614.
Pellegrini, M., Marcotte, E. M., Thompson, M. J., Eisenberg, D., & Yeates, T. O. (1999).
Assigning protein functions by comparative genome analysis: Protein phyloge-
netic profiles. PNAS, 96(8), 4285-4288.
In Silico Recognition of Protein-Protein Interactions: Theory and Applications 267
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Przulj, N., Wigle, D. A., & Jurisica, I. (2004). Functional topology in a network of protein
interactions. Bioinformatics, 20(3), 340-348.
Ravasz, E., Somera, A. L., Mongru, D. A., Oltvai, Z. N., & Barabasi, A.L. (2002).
Hierarchical organization of modularity in metabolic networks. Science, 297(5586),
1551-1555.
Rives, A. W., & Galitski, T. (2003). Modular organization of cellular networks. PNAS,
100(3), 1128-1133.
Saigo, H., Vert, J.P., Ueda, N., & Akutsu, T. (2004). Protein homology detection using
string alignment kernels. Bioinformatics, 20(11), 1682-1689.
Saito, R., Suzuki, H., & Hayashizaki, Y. (2003). Construction of reliable protein-protein
interaction networks with a new interaction generality measure. Bioinformatics,
19(6), 756-763.
Sali, A. (1999). Functional links between proteins. Nature, 402, 23.
Salwinski, L., & Eisenberg, D. (2003). Computational methods of analysis of protein-
protein interactions. Current Opinion in Structural Biology, 13(3), 377-382.
Salwinski, L., Miller, C. S., Smith, A. J., Pettit, F. K., Bowie, J. U., & Eisenberg, D. (2004).
The database of interacting proteins: 2004 update. Nucleic Acids Research,
32(90001), D449-451.
Snel, B., Bork, P., & Huynen, M. A. (2002). The identification of functional modules from
the genomic association of genes. PNAS, 99(9), 5890-5895.
Sobott, F., & Robinson, C. V. (2002). Protein complexes gain momentum. Current Opinion
in Structural Biology, 12(6), 729-734.
Spirin, V., & Mirny, L. A. (2003). Protein complexes and functional modules in molecular
networks. PNAS, 100(21), 12123-12128.
Sprinzak, E., & Margalit, H. (2001). Correlated sequence-signatures as markers of protein-
protein interaction1. Journal of Molecular Biology, 311(4), 681-692.
Sprinzak, E., Sattath, S., & Margalit, H. (2003). How Reliable are Experimental Protein-
Protein Interaction Data? Journal of Molecular Biology, 327, 919-923.
Suhre, K., & Claverie, J.M. (2004). FusionDB: a database for in-depth analysis of
prokaryotic gene fusion events. Nucleic Acids Research, 32(90001), D273-276.
Tong, A. H. Y., Drees, B., Nardelli, G., Bader, G. D., Brannetti, B., Castagnoli, L., et al.
(2002). A Combined Experimental and Computational Strategy to Define Protein
Interaction Networks for Peptide Recognition Modules. Science, 295(5553), 321-
324.
Tucker, C. L., Gera, J. F., & Uetz, P. (2001). Towards an understanding of complex protein
networks. Trends Cell Biol, 11(3), 102-106.
Uetz, P. (2002). Two-hybrid arrays. Curr Opin Chem Biol, 6(1), 57-62.
Uetz, P., Giot, L., Cagney, G., Mansfield, T. A., Judson, R. S., Knight, J. R., et al. (2000).
A comprehensive analysis of protein-protein interactions in Saccharomyces
cerevisiae. Nature, 403(6770), 623-627.
268 Park, Dam, Pan, Xu, Geist, Heffelfinger, & Samatova
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
van Noort, V., Snel, B., & Huynen, M. A. (2004). The yeast coexpression network has a
small-world, scale-free architecture and can be explained by a simple model. EMBO
Reports, 5(3), 280-284.
Vapnik, V. (1998). Statistical learning theory. New York: Wiley Interscience.
Vazques, A., Flammini, A., Maritan, A., & Vespignani, A. (2003). Global protein function
prediction from protein-protein interaction networks. Nature Biotechnology,
21(6), 697-700.
von Mering, C., Jensen, L. J., Snel, B., Hooper, S. D., Krupp, M., Foglierini, M., et al. (2005).
STRING: known and predicted protein-protein associations, integrated and trans-
ferred across organisms. Nucleic Acids Research, 33 Database Issue, D433-437.
von Mering, C., Krause, R., Snel, B., Cornell, M., Oliver, S. G., Fields, S., et al. (2002).
Comparative assessment of large-scale data sets of protein-protein interactions.
Nature, 417(6887), 399-403.
Watts, D. J., & Strogatz, S. H. (1998). Collective dynamics of small-world networks.
Nature, 393(6684), 440-442.
Wodak, S. J., & Mendez, R. (2004). Prediction of protein-protein interactions: the CAPRI
experiment, its evaluation and implications. Current Opinion in Structural Biol-
ogy, 14(2), 242-249.
Wojcik, J., & Schachter, V. (2001). Protein-protein interaction map inference using
interacting domain profile pairs. Bioinformatics, 17(90001), 296S-305.
Wuchty, S. (2002). Interaction and domain networks of yeast. Proteomics, 2(12), 1715-
1723.
Xenarios, I., Salwinski, L., Xiaoqun, J., Higney, P., Kim, S., & Eisenberg, D. (2005). DIP
Database of Interacting Proteins. Nucleic Acids Research, 33(Database Issue).
Yan, C., Dobbs, D., & Honavar, V. (2004). A two-stage classifier for identification of
protein-protein interface residues. Bioinformatics, 20(suppl_1), i371-378.
Yates III, J. R. (2000). Mass spectrometry: From genomics to proteomics. Trends in
Genetics, 16(1), 5-8.
Zanzoni, A., Montecchi-Palazzi, L., Quondam, M., Ausiello, G., Helmer-Citterich, M., &
Cesareni, G. (2002). MINT: a Molecular INTeraction database. FEBS Letters,
513(1), 135-140.
Zhou, H.-X., & Shan, Y. (2001). Prediction of protein interaction sites from sequence
profile and residue neighbor list. PROTEINS: Strucure, Function, and Genetics,
44, 336-343.
Zhu, H., Bilgin, M., Bangham, R., Hall, D., Casamayor, A., Bertone, P., et al. (2001). Global
Analysis of Protein Activities Using Proteome Chips. Science, 293(5537), 2101-
2105.
Differential Association Rules 269
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Chapter XIV
Differential
Association Rules:
Understanding
Annotations in Protein
Interaction Networks
Christopher Besemann, North Dakota State University, USA
Anne Denton, North Dakota State University, USA
Ajay Yekkirala, North Dakota State University, USA
Ron Hutchison, The Richard Stockton College of New Jersey, USA
Marc Anderson, North Dakota State University, USA
Abstract
In this chapter, we discuss the use of differential association rules to study the
annotations of proteins in one or more interaction networks. Using this technique, we
find the differences in the annotations of interacting proteins in a network. We extend
the concept to compare annotations of interacting proteins across different definitions
of interaction networks. Both cases reveal instances of rules that explain known and
unknown characteristics of the network(s). By taking advantage of such data mining
techniques, a large number of interesting patterns can be effectively explored that
otherwise would not be.
270 Besemann, Denton, Yekkirala, Hutchison, & Anderson
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Introduction
Many high-throughput techniques have been developed to identify protein interaction
networks in organisms. One conventional definition of interactions within an organism
considers whether the protein molecules interact physically. High-throughput tech-
niques to test for such interactions include the popular yeast-two-hybrid method (Ito et
al., 2001; Uetz et al., 2000). Another common definition of interactions relies on the
concept of genetic interactions in which the lethality of two gene deletions is examined
as in (Tong et al., 2001; Tong et al., 2004). Finally, there also exist a number of computer-
driven procedures like the identification of domain-fusion interactions. Two genes have
a domain fusion interaction in one species if their homologs occur as exons of a single
gene in another species (Marcotte et al., 1999; Truong & Ikura, 2003). With so much
interaction data available, there is a clear need to develop techniques that are targeted
at understanding the role of networks in bioinformatics.
Interactions are not the only type of information available for proteins. Many genes and
proteins are studied extensively in model species, such as yeast. A rich supply of
experimental data, in the form of annotations, is available for these species and can be
studied in much the same way as data associated with objects in other disciplines. These
annotations are the items of information that belong to particular objects. A popular
technique for the study of such items is association rule mining (ARM). ARM has the
goal of discovering associations between frequently occurring sets of items. In the
prototypical case, ARM is used in the analysis of shopping transactions (Agrawal,
Imielinski, & Swami, 1993; Han & Fu, 1995); the technique has also been applied in
relational database settings (Dehaspe & Raedt, 1997; Oyama, Kitano, Satou, & Ito, 2002).
The combination of network data and protein annotation data is a special case of
relational data. In a relational database, one table is used to describe protein annotations
and one or more separate tables define the network(s). A relational framework allows
analysis beyond statistical comparison of single annotations of interacting proteins
(Schwikowski, Uetz, & Fields, 2000). Relational frameworks also support the study of
associations involving more complex attribute combinations. Within the same frame-
work, it also becomes possible to contrast different interaction definitions.
This chapter explores a novel ARM algorithm called differential ARM that enables the
discovery of differences between interacting proteins and between multiple network
types. In the Background section, we describe representative work on protein interac-
tions and introduce key definitions in association rule mining. The next section, Mining
Differential Association Rules, defines our goals and the problems to overcome, ending
with a description of the differential ARM method. Finally, in the section on Differential
Association Rule Experiments, an experiment on real-world yeast interaction data is
described.
Differential Association Rules 271
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Background
We begin by discussing related work in the area of protein interaction network analysis.
Basic notation for association rule mining is covered and notes on related references in
the area of ARM are given.
Protein Interaction Network Analysis
Several papers have analyzed yeast protein interaction networks. The majority of these
works have studied isolated networks (Wagner, 2001) or used comparisons of the
statistics of multiple networks (Maslov & Sneppen, 2002). In Schwikowski et al. (2000)
the authors did some analysis of the yeast physical protein interaction network using
counting methods to determine percentages of interactions that have certain func-
tional behavior. Most examples of interest are techniques that require some manual
intervention. In larger datasets, in which patterns of interest are harder to perceive,
automated techniques would greatly assist analysis. Tong et al. (2004) performed
statistical analysis on part of the yeast genetic interaction network. In this work, the
authors begin with observations of interesting statistics observed within interaction
networks, and conclude with clustering analysis and population genetics. Some of the
conclusions drawn relate to the density of the network and comparison of how the genetic
interaction network predicts parts of the physical interaction network. Again, the
emphasis is on a single network and simple statistical comparisons. In Ozier, Amin, and
Ideker (2003, 2003) global statistics of a physical and genetic interaction network are
compared. ARM helps move beyond these studies.
The first use of association rule mining for protein interaction graphs was documented
in Oyama et al. (2002). Oyama et al. use a relational ARM method to study yeast physical
protein interactions towards predicting protein-binding site information. The idea of the
paper is to combine protein interaction and annotation data into a single table that is
mined using traditional ARM techniques. The table has two columns, one for the left-
hand side protein and the other for the right-hand side protein of the interaction. Some
problems were observed because many interacting proteins share the same annotations.
This behavior is related to what Jensen and Neville (2002) call autocorrelation in the
context of classification. Techniques described in this chapter help address the issues;
they first appear in Besemann, Denton, Yekkirala, Hutchison, and Anderson (2004).
Association Rule Mining
General descriptions of association rule mining can be found in many papers. It is first
described in Agrawal et al. (1993). In this section, we introduce the notation for relational
ARM that is used in the later parts of this chapter.
Association rule mining is commonly defined for sets of items. We combine the concept
of sets with the relational algebra framework by choosing an extended relational model
similar to Han and Fu (1995). Attributes within this model are allowed to be set-valued,
thereby violating first normal form in database theory. For more background on relational
272 Besemann, Denton, Yekkirala, Hutchison, & Anderson
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
notation, introductory texts such as Elmasri and Navathe (2004) can be reviewed. In this
notation R(A
1
,,A
n
) represents a relation and A
i
is an attribute of the relation (e.g.,
protein relation containing annotation attributes). Classic ARM assumes that the dataset
is represented in a single relation, R(A), with the single attribute, A, being set-valued.
The members of the set-valued attribute are items. In our example, the items could be
annotations of a protein. A rule in ARM is A C for non-intersecting itemsets A and
C. The support of a rule is the percent of transactions (rows in the table) that have (A
C) and the confidence of the rule is the fraction
) sup(
) sup(
A
C A
or how often we find C
given that A is present.
The goal of ARM is to find all rules that meet thresholds of minimum support and
confidence. In the original problem, a database consisting of only a single relation was
considered. Multiple relations are considered by using relational join operations. This
approach was taken in Oyama et al. (2002). Multi-relational association rule mining has
also been performed in the context of inductive logic programming (Dehaspe & Raedt,
1997). Dehaspe and Raedt address the general task of allowing all relations in the
database to be mined using a predicate logic language such as Prolog. Association rule
mining has further been extended to structural properties of graphs. The work on graph-
based and frequent substructure ARM (Inokuchi, Washio, & Motoda, 2000; Michihiro
& Karypis, 2001) differs from relational approaches in that nodes in the graph are assumed
to have at most one label.
A difficulty in association rule mining of data associated with nodes in a network is
elimination of rules that are not considered interesting. Uninteresting rules can also be
a problem in traditional ARM. Work has been done to eliminate redundant rules using
the theory of closed sets (Zaki, 2000; Bastide, Pasquier, Taouil, Stumme, & Lakhal , 2000).
Measures other than support and confidence can be used such as in Brin, Motwani,
Ullman, and Tsur (1997). Finally, efficient methods have been developed to constrain the
desired rule output (Srikant, Vu, & Agrawal, 1997; Pei, Han, & Lakshmanan, 2004).

Node
ORF Annotations
YPR184W
YER146W
YNL287W
YBL026W
YMR207C
{<cytoplasm >}
{<cytoplasm >}
{< Sensitivity_aaaod >}
{< transcription >, < nucleus >}
{< nucleus >}
Edge
ORF0 ORF1
YPR184W YER146W
YNL287W YBL026W
YBL026W YMR207C
Figure 1. Example of a node and an edge table
Differential Association Rules 273
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Mining Differential Association Rules
A motivating scenario for the use of differential ARM can be taken from the protein
interaction setting. Consider using standard ARM to identify differences between
interacting proteins. Assume, for example, that with standard ARM proteins with
transcription as annotation are associated frequently with proteins that are localized
in the nucleus. This association rule may be due to two other rules: one that associates
transcription and nucleus within a single protein, and others that represent the
autocorrelation of transcription and/or nucleus between interacting proteins. We
would not consider this a difference between interacting proteins. Nor would we
consider the autocorrelation of transcription and/or nucleus interesting, or the fact
that transcription and nucleus are related for individual proteins. However, a rule
of the same format could still stand for a difference in other cases.
Consider the rule that proteins in the nucleus are found to interact with proteins in the
mitochondria. A single protein would not simultaneously be located in the nucleus
and in the mitochondria. We therefore assume that the rule highlights a difference
between interacting proteins and may identify an instance of compartmental crosstalk.
This rule is more interesting to a biologist than the rule relating nucleus and transcrip-
tion. It is more expressive of the properties of the respective interaction network.
We distinguish between two examples based on our biological background knowledge.
One example is an artifact while the other correctly identifies differences in interacting
proteins that shed light on the nature of the interaction in question. Two approaches
could be taken to make this distinction at the level of the ARM algorithm. We could devise
a difference criterion involving correlations between neighboring proteins and/or rules
found within individual proteins. Such an approach would not benefit from any of the
pruning that has made ARM an efficient and popular technique. The differential ARM
algorithm takes an approach that makes significant use of pruning: only those items are
considered where each item in a set is unique to only one of the interacting nodes or
proteins. The rule associating transcription and nucleus would only be evaluated
on those transcription proteins that are not themselves in the nucleus, and those
nucleus proteins, that are not themselves involved in transcription.
There are other reasons why a focus on differences is more effective for association rule
mining in networks than a standard application of ARM on joined relations. Traditionally
association rule mining is performed on sets of items with no known correlations.
Interacting proteins are known to have matching annotations in many cases (Tong et al.,
2004; Jensen & Neville, 2002). Using association rule mining on such data, in which items
are expected to be correlated may lead to output in which the known correlations
dominate all other observations either directly or indirectly. This problem is observed
when relational association rule mining is directly applied to protein networks (Oyama
et al., 2002; Besemann et al., 2004). Excluding matching items of interacting proteins is
therefore advisable in the interest of finding meaningful results. Matching annotations
can be studied by simple correlation analysis, in which co-occurrence of an annotation
in interacting proteins is tested (Tong et al. 2004). In the presence of such correlations,
association rules likely reflect nothing but similarities between interacting proteins.
274 Besemann, Denton, Yekkirala, Hutchison, & Anderson
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Differential Association Rule Notation
This section describes notation used to formalize the concept of differential ARM, and
the implementation that was used in experiments. We assume a relational framework to
discuss differences within and between interaction networks. The concept of a network
may suggest use of graph-based techniques. Graph-theory typically assumes that nodes
and edges have at most one label. Relational algebra on the other hand has the tools for
the manipulation of data associated with nodes and edges. A relational representation
of a graph with one type of node requires two relations. One relation for data associated
with nodes that we call the node relation, and a second relation that describes the
relationship between nodes that we call the edge relation.
Consider a database with a protein node relation. Tuples in the relation have a node
identifier attribute and descriptor-set attribute such as ORF and Annotations in table
Node of Figure 1. The annotations are like items in standard ARM, however in the multi-
relational setting a label on the items must identify their node of origin. Edge relations
have attributes to point to the two connected nodes as in table Edge of Figure 1.
Our goal is to mine relational basis sets constructed from the descriptors (annotations)
of connected proteins. A joined-relation basis set is the set of tuples needed to perform
basic relational ARM. Joined-relation basis sets are formed in multiple steps. Edge and
node relations are joined through a conditional join operation. The condition on the join
excludes cycles in the graph. Attribute names are changed similar to notation of the
attribute renaming function in Elmasri and Navathe (2004) such that the names are unique
between related proteins. Attributes are identified by consecutive integers to which we
refer to as origin identifiers. Origin identifiers are added to the annotations to identify
which protein contributed that annotation. The Join table in Figure 2 gives an example
of a basis set prepared as described.
The result for a 2-node setting (pair of interacting proteins) assigns 0 to the attributes
of the first protein of the edge relation, and assigns 1 to the attributes for the second
protein. This would be similar to the description from Oyama et al. (2002) with the left-
hand side protein annotations being 0 and right-hand side protein annotations being
1. Note also, that we have multiple alternatives for joining corresponding to different
Figure 2. (Left) Example of a joined table. (Right) Table after application of uniqueness
operator
TID Join Unique
1 0.cytoplasm 1.cytoplasm NULL NULL
2 0.Sensitivity_aaaod
1.transcription,
1.nucleus
0.Sensitivity_aaaod
1.transcription,
1. nucleus
3 0.transcription, 0.nucleus 1.nucleus 0.transcription NULL

Differential Association Rules 275
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
subgraph structures. As an example, four nodes can form either a straight line (0-1-2-3)
or a T-shape (0-1-2, 1-3) .
Once the joining has occurred and origin identifiers added to the annotations, we have
simple items. A simple item may be 0.nucleus that indicates the protein at index 0 of
the join has the annotation of nucleus. The Join section in Figure 2 shows an example
of the product as the result of the operations applied to Tables Node and Edge from
Figure 1.
After the joined-relation basis sets are prepared, the differential portion of the method
is applied. This takes place in the form of a uniqueness operator that ensures each
descriptor (annotation) is not repeated a tuple from the joined table. Table Unique in
Figure 2 shows the results of the unique operation on the joined portion. For example,
cytoplasm in TID 1 is not unique so it is eliminated from TID 1. The uniqueness operator
applies to all set-valued attributes of a joined-relation but alternatives are possible, such
as requiring uniqueness only across a subset of edges.
One additional problem occurs when finding patterns involving joins over multiple
tables. A pattern that expresses information about a single protein can be found in
transactions that involve single proteins, interacting protein pairs, chains of three
interacting proteins and so forth. The typical solution is to pick the simplest represen-
tation. For a pattern involving a single protein, it is the node table without performing
joins. For patterns involving two interacting proteins, we use the table that only involves
one interaction table, and so on. This situation is also identified in a report on relational
ARM (Cristofor & Simovici, 2001). We formalize this concept by calling an itemset out-
of-scope if it does not require all portions of the current join operation, i.e., every node
(and edge) from the join must be represented. From Figure 2, this means that TID 1 and
TID 3 are out-of-scope.
Summary of differential ARM algorithm:
1. Form basis sets from protein and interaction relations (join relations & set origin
identifiers)
2. Process each transaction in the basis set to remove non-unique annotations
3. Remove out-of-scope transactions and itemsets
4. Perform association rule mining over the processed dataset
Differential ARM for Network Comparison
In this section, we go beyond applying the idea of differential ARM to interacting
proteins of a single network and look at interacting proteins in multiple networks. If we
store each interaction type in an attribute of the edge relation then we can perform
differential ARM as given.
For network comparison, we use basis sets formed by joining a chain of three nodes,
involving both types of network. After the join formula is applied, the other steps of
building items and applying the uniqueness operator occur the same way as described
earlier in the chapter. Based on this definition, three nodes must interact to allow
276 Besemann, Denton, Yekkirala, Hutchison, & Anderson
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
comparison of two networks. This represents the case of node overlap when the first
network covers one node, the two networks share a second node, and the second network
covers the third node.
An example is seen in Figure 3 where the edge C, D (solid) and the edge D, E (dashed)
are compared by the algorithm, but not the edge B, C (solid) because it has no
corresponding dashed interaction for either node. Similarly, edge H, G causes either
a dashed or a solid overlap, so it is not used. When we apply the uniqueness operator
to these transactions, we are accepting only those triples that have different, non-
overlapping interaction types.
Differential Association Rule Experiments
This section gives an example of applying the differential ARM methods to a bioinformatics
dataset. First, we describe the datasets used. We then discuss the results of applying
differential ARM on a single-type interaction network and then on a multiple-type
interaction network.
Our data consists of a node relation gathered from the Comprehensive Yeast Genome
Database at MIPS (Mewes et al, 2002; CYGD, 2004). This node relation represents gene
annotation data such as function and localization. The relation contains the ORF
identifier as key and the set of annotations related to that ORF as attribute (descriptor
set).
It should be noted that the quality of annotations is still in the process of being improved.
At this point, there are many proteins with no annotations at all and some annotation
categories are only covered by a small subset of proteins. In other cases, only high-level
annotations are available. This means that the results are limited to the detail that is
provided in the databases. As annotation sources and procedures improve, so will the
results of ARM.
We used three different definitions for protein-protein interactions that are represented
as undirected edges for yeast: physical, genetic and domain fusion. The physical edge
relation was built from the ppi table at CYGD where all tuples with type label of
Figure 3. Transactions for network comparison



Transactions
C D E
C D K
G D E
G D K
J H I
A
B
C
D
E F
G
H
I J
K
Differential Association Rules 277
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
physical were used. The genetic edge relation was taken from supplemental table S1
of genetic interactions from Tong et al. (2004) where both Synthetic Sick and Syn-
thetic Lethal entries are used. Our third edge relation was the domain fusion set built
from the unfiltered results posted from Truong and Ikura (2003) and Ikura Lab (2004). The
set was filtered to reflect only ORFs contained in our node relation from MIPS in order
to improve both performance and reliability of results.
Before we discuss the association rule results we produced, we show the efficiency of
the approach. Differential ARM outperforms the standard algorithm by three orders of
magnitude in the four-node setting and one order of magnitude in the two-node setting.
The reduction in the number of rules produced is more significant. The differences
between the numbers of rules in differential vs. standard ARM demonstrate how scope
and autocorrelation greatly inflate the results. Differential ARM has more than six orders
of magnitude fewer rules to interpret. Results of trial experiments are shown in Figure
4. Those orders of magnitude determine the difference between useable and unusable
results.
More important than performance concerns are the practical benefits of rules produced
by differential ARM. We will first look at an example of a rule that is strong based on the
application of a standard ARM algorithm on joined tables but not so if only unique items
are considered. Standard ARM on joined tables returns mostly rules that show
autocorrelation of items or are out-of-scope. We look at a rule that was mentioned earlier:
{0.transcription} {1.nucleus} (support = 0.29% confidence = 28.38%).
This rule is the consequence of a single-node rule and an autocorrelation rule:
{0.transcription} {0.nucleus} (support = 0.70% confidence = 69.59%),
{0.nucleus} {1.nucleus} (support = 5.74% confidence = 29.02%).
Figure 4. Performance comparison between standard ARM and differential ARM
Figure: Processing Time Figure: Number of Rules


n-node join
1 2 3 4
1.E-02
1.E-01
1.E+00
1.E+01
1.E+02
Differential
ARM
p
r
o
c
e
s
s

t
i
m
e
n-node join
1 2 3 4
1.E+06
1.E+07
1.E+08
Differential
ARM
n
u
m
b
e
r

o
f

r
u
l
e
s
1.E+04
1.E+05
1.E+03
1.E+02
1.E+01
1.E+00
278 Besemann, Denton, Yekkirala, Hutchison, & Anderson
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Using the uniqueness operator changes the support of the original rule to 0.02% and a
confidence of 2.08%. Strong rules in our data set generally have a support around 2-4%
and confidence around 20%. Based on those numbers the rule cannot be considered
strong and ranks lower in the differential results. This shows a case where a hypothesis
may have been made based on a standard ARM result that the properties of transcription
and nucleus are typical for proteins on opposite sides of a physical interaction.
Differential ARM results do not support this hypothesis. Differential ARM shows that
nucleus is strongly associated itself and proteins co-located in the nucleus interact.
Therefore, the differential ARM results correctly exclude this rule.
The following rule was found in the physical network based on differential ARM results:
{1.mitochondria} {0.cytoplasm} (support = 1.2% confidence = 27.3%).
This rule clearly corresponds to annotations that would not be expected to hold within
a single protein but may hold between interacting ones. A protein located in the
mitochondria would not have localization cytoplasm. We do expect compartmental
crosstalk as studied in Schwikowski et al. (2000) between those two locations. The
observation confirms to us that we see rules that are sensible from a biological
perspective. We also identify patterns that have not been reported. The pattern {1.ER,
0.mitochondria} (support 0.21%) was observed. It was selected because of its compara-
tively high support and unexpected indication of proteins in the endoplasmic reticulum
(ER) physically interacting with proteins in the mitochondria.
We now look at results that derive from the network comparison formalism. Given multiple
types of protein-protein interactions we look for significant differences to aid in the
understanding of cellular function and as well as the properties and uses of the networks.
The networks themselves do not show a significant overlap, in other words, it is common
that for any given physical interaction between two proteins there will be no genetic
interaction (Tong et al., 2004).
Table 1 shows that even the statistical properties of the networks differ significantly. The
average number of interactions of proteins that show at least one interaction varies from
3.55 in the physical network to 44.5 in the domain fusion network. Comparison of
annotations across those networks has to compensate for such differences. The process
of joining relations and application of the uniqueness operator ensures that each protein
Table 1. Statistics of physical, genetic, and domain fusion networks
Statistics
Table int/orf max int #>20 #int
physical 3.55 289 73 14672
genetic 7.88 157 93 8336
domain fusion 44.6 231 305 28040

Differential Association Rules 279
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
that is considered for a physical interaction will also be considered for a genetic
interaction.
Before looking at details of individual rules, we make some general observations
regarding the number of rules we find for different combinations of networks. When
comparing physical and genetic networks we found about one order of magnitude more
strong rules relating to the physical network compared with the genetic network. Physical
interactions also produce the stronger rules when compared with domain fusion net-
works. That means that the physical network allows the most precise statements to be
made. That suggests that physical interactions reflect properties of the proteins better
than either of the other two do.
An example of a rule supported by the physical network but not the domain fusion
network:
{0.ABC trans fam sign} {1.ATP/GTP binding site} (support = 0.45% confidence
= 90%).
ORF 0 has the motif of an ABC transporter signature that implies it is an ABC transporter
coding sequence. ABC transporters have conserved ATP binding domains as the motif
in ORF 1 (Falquet et al., 2002). From the rule, we see that these two domains physically
interact. However, the domains never occur in a single gene. This supports the obser-
vation that the ATP binding domain is found in many other proteins as well (Falquet et
al., 2002) and both functions are combined through interactions at the protein level rather
than at the genetic level. This observation could warrant further studies or hypothesis
on the observed pattern. There are many other examples of rules that allow comparison
and contrast of these related networks.
Future Trends
Study of biological networks combined with annotation information is likely to become
a topic of increasing importance. Definitions of biological networks exist beyond the
ones discussed in this chapter and new ones are added frequently. Important types of
networks that have been omitted from the discussion of the chapter are metabolic,
regulatory, and signaling pathways. Network-based techniques that are aware of node
data are ideally suited as tools that are intermediate in their complexity and scope between
large-scale network analysis that ignores anything but connectivity, and small-scale
simulations of individual pathways. While data mining is applied to the general scenario
of relational databases, little has been specifically done to address the setting of a
network with node information. Much can be gained by focusing on this important
setting. Networks with node data are ubiquitous from social networks through the
structure of the World Wide Web to many areas of bioinformatics. The increasing volume
and complexity of such data will require further advances in efficiency and utility.
280 Besemann, Denton, Yekkirala, Hutchison, & Anderson
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Summary
This chapter has described issues related to data mining of protein interaction networks
in conjunction with information associated with the proteins themselves. Other tech-
niques in statistics and data mining leave significant gaps, including problems in
describing differences in the data and analyzing multiple networks. A novel concept,
differential association rule mining was introduced, that allows users to discover
associations that indicate differences in interacting proteins. The goal of this technique
is to highlight differences between items belonging to different interacting nodes or
different networks. We demonstrated that such differences would not be identified by
application of standard relational ARM techniques. Differential ARM was applied to
real-world examples of protein interactions in yeast. Within the results, we recovered
association results found in the literature as well as new associations. Overall, the
technique offers an efficient and effective method for data exploration in biological and
other networks.
References
Agrawal, R., Imielinski, T., & Swami, A. N. (1993). Mining association rules between sets
of items in large databases. Proceedings of the ACM SIGMOD International
Conference on Management of Data, Washington, DC (pp. 207- 216).
Bastide, Y., Pasquier, N., Taouil, R., Stumme, G., & Lakhal, L. (2000). Mining minimal non-
redundant association rules using frequent closed itemsets. Lecture Notes in
Computer Science, 1861(972).
Besemann, C., Denton, A., Yekkirala, A., Hutchison, R., & Anderson, M. (2004). Differ-
ential association rule mining for the study of protein-protein interaction networks.
Proceedings of the Workshop on Data Mining in Bioinformatics at KDD, Seattle,
WA.
Brin, S., Motwani, R., Ullman J. D., & Tsur, S. (1997). Dynamic itemset counting and
implication rules for market basket data. Proceedings ACM SIGMOD International
Conference on Management of Data, Tucson, AZ.
Cristofor, L., & Simovici, D. (2001). Mining association rules in entity-relationship
modeled databases. Technical report, University of Massachusetts Boston.
CYGD. (2004). Retrieved March, 2004 from http://mips.gsf.de/genre/proj/yeast/index.jsp
Dehaspe, L., & Raedt, L. D. (1997). Mining association rules in multiple relations.
Proceedings of the 7th International Workshop on Inductive Logic Program-
ming, Prague, Czech Republic (pp. 125-132).
Elmasri, R. & Navathe, S. (2004). Fundamentals of Database Systems. Boston: Pearson.
Falquet, L., Pagni, M., Bucher, P., Hulo, N., Sigrist, C. J., Hofmann, K., et al. (2002). The
prosite database, its status in 2002. Nucleic Acids Research, 30, 235-238.
Differential Association Rules 281
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Han, J., & Fu, Y. (1995). Discovery of multiple-level association rules from large
databases. Proceedings of the 21th International Conference on Very Large Data
Bases, Zurich, Switzerland.
Ikura Lab, O. C. I. (2004). Domain fusion database. Retrieved March, 2004 from http://
calcium.uhnres.utoronto.ca/pi/pubpages/download/index.htm
Inokuchi, A., Washio, T., & Motoda, H. (2000). An apriori based algorithm for mining
frequent substructures from graph data. Proceedings of the European Conference
on Principles of Data Mining and Knowledge Discovery, Lyon, France (pp. 13
23).
Ito, T., Chiba, T., Ozawa, R., Yoshida, M., Hattori, M., & Sakaki, Y. (2001). A comprehen-
sive two-hybrid analysis to explore the yeast protein interactome. Proc Natl Acad
Sci USA, 98(8), 456974.
Jensen, D., & Neville, J. (2002). Autocorrelation and Linkage Cause Bias in Evaluation
of Relational Learners. Lecture Notes in Artificial Intelligence, 2583, 101-116.
Marcotte, E. M., Pellegrini, M., Ng, H. L., Rice, D. W., Yeates, T. O., & Eisenberg, D. (1999).
Detecting protein function and protein-protein interactions from genome se-
quences. Science, 285(5428), 7513.
Maslov, S., & Sneppen, K. (2002). Specificity and stability in topology of protein
networks. Science, 296(5569), 910-3.
Mewes, H., Frishman, D., Gldener, U., Mannhaupt, G., Mayer, K., Mokrejs, M., et al.
(2002). Mips: A database for genomes and protein sequences. Nucleic Acids
Research, 30(1), 3144.
Michihiro, K., & Karypis, G. (2001). Frequent subgraph discovery. Proceedings of the
International Conference on Data Mining. San Jose, CA (pp. 313320).
Oyama, T., Kitano, K., Satou, K., & Ito, T. (2002). Extraction of knowledge on protein-
protein interaction by association rule discovery. Bioinformatics, 18(5), 70514.
Ozier, O., Amin, N., & Ideker, T. (2003). Genetic interactions: Online supplement to:
Global architecture of genetic interactions on the protein network. Retrieved
January 31, 2005, from http://web.wi.mit.edu/ideker/pub/nbt/
Ozier, O., Amin, N., & Ideker, T. (2003). Global architecture of genetic interactions on the
protein network [Letter to the editor]. Nature Biotechnology, 21(5), 490-1.
Pei, J., Han, J., & Lakshmanan, L.V. (2004). Pushing convertible constraints in frequent
itemset mining. Data Mining and Knowledge Discovery, 8, 227252.
Schwikowski, B., Uetz, P., & Fields, S. (2000). A network of protein-protein interactions
in yeast. Nature Biotechnology, 18(12), 12423.
Srikant, R., Vu, Q., & Agrawal, R. (1997). Mining association rules with item constraints.
Proceedings of the International Conference on Knowledge Discovery and Data
Mining, Newport Beach, CA, (pp. 6773).
Tong, A. H. Y., Evangelista, M., Parsons, A. B., Xu, H., Bader, G. D., Pag, N., et al. (2001).
Systematic genetic analysis with ordered arrays of yeast deletion mutants. Science,
294(5550), 23648.
282 Besemann, Denton, Yekkirala, Hutchison, & Anderson
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Tong, A. H. Y., Evangelista, M., Parsons, A. B., Xu, H., Bader, G. D., Pag, N., et al. (2004).
Global mapping of the yeast genetic interaction network. Science, 303(5695), 808
815.
Truong, K., & Ikura, M. (2003). Domain fusion analysis by applying relational algebra to
protein sequence and domain databases. BMC Bioinformatics, 4(16).
Uetz, P., Giot, L., Cagney, G., Mansfield, T. A., Judson, R. S., Knight, J. R., et al. (2000).
A comprehensive analysis of protein-protein interactions in Saccharomyces
cerevisiae. Nature, 403(6770), 6237.
Wagner, A., & Fell, D. (2001). The small world inside large metabolic networks. Proc. Roy.
Soc. London Series B, 268 (pp. 1803-1810).
Zaki, M. J. (2000). Generating non-redundant association rules. Proceedings of the ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining,
Boston (pp. 3443).
Mining BioLiterature 283
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Chapter XV
Mining BioLiterature:
Toward Automatic
Annotation of Genes
and Proteins
Francisco M. Couto, Universidade de Lisboa, Portugal
Mrio J. Silva, Universidade de Lisboa, Portugal
Abstract
This chapter introduces the use of Text Mining in scientific literature for biological
research, with a special focus on automatic gene and protein annotation. This field
became recently a major topic in Bioinformatics, motivated by the opportunity brought
by tapping the BioLiterature with automatic text processing software. The chapter
describes the main approaches adopted and analyzes systems that have been developed
for automatically annotating genes or proteins. To illustrate how text-mining tools fit
in biological databases curation processes, the chapter presents a tool that assists
protein annotation. Besides the promising advances of Text Mining of BioLiterature,
many problems need to be addressed. This chapter presents the main open problems in
using text-mining tools for automatic annotation of genes and proteins, and discusses
how a more efficient integration of existing domain knowledge can improve the
performance of these tools.
284 Couto & Silva
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Introduction
Bioinformatics aims at understanding living systems using biological information. The
facts discovered in biological research have been mainly published in the scientific
literature (BioLiterature) since the 19
th
century. Extracting knowledge from such a large
amount of unstructured information is a painful and hard task, even to an expert. A
solution could be the creation of a database where authors would deposit all the facts
published in BioLiterature in a structured form. Some generic databases, such as UniProt,
collect and distribute biological information (Apweiler et al., 2004). However, different
communities have different needs and views on specific topics, which change over time.
As a result, researchers do not look only for the facts, but also for their evidence. Before
a researcher considers a fact as relevant to his work, he checks the evidence presented
by the author, because facts are normally valid only in a specific context. This explains
why Molecular Biology knowledge continues to be mainly published in BioLiterature.
Another solution is Text Mining, which aims at automatically extracting knowledge from
natural language texts (Hearst, 1999). Text-mining systems can be used to identify the
following types of information: entities, such as genes, proteins and cellular compo-
nents; relationships, such as protein localization or protein interactions; and events,
such as experimental methods used to discover protein interactions. Bioinformatics tools
to collect more information about the concepts they analyze also use Text Mining. For
example, information automatically extracted from the BioLiterature can improve gene
expression clustering (Blaschke, Oliveros, & Valencia, 2004).
Text Mining of BioLiterature has been studied since the last decade (Andrade & Valencia,
1998). The interest in the topic has been steadily increasing, motivated by the vast
amount of publications that curators have to read in order to update biological databases,
or simply to help researchers keep up with progress in a specific area. Text Mining can
minimize these problems mainly because BioLiterature articles are quite often publicly
available. The most widely used BioLiterature repository is MEDLINE, which provides
a vast collection of abstracts and bibliographic information. For example, in 2003, about
560,000 citations have been added to MEDLINE. Reading 10 of these documents per day,
it would take around 150 years to read all the documents added in 2003. Moreover, the
number of new documents added per year increased by more than 20,000 from 2000 to
2003. Hence, text-mining systems could have a great impact in minimizing this effort by
automatically extracting information that can be used for multiple purposes and could not
possibly be organized by other means.
This chapter starts by providing broad definitions used in Text Mining and describes the
main approaches. Then, it summarizes the state-of-the-art of this field and shows how
text-mining systems can be used to automatically annotate genes or proteins. Next, the
chapter describes a tool designed for assisting protein annotation. Finally, the chapter
discusses future and emerging trends and presents concluding remarks.
Mining BioLiterature 285
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Text Mining
Text Mining aims at automatically extracting knowledge from unstructured text. Usually
the text is organized as a collection of documents, or corpus.
TextMining = NLP + DataMining
Data Mining aims at automatically extracting knowledge from structured data. (Hand,
Mannila, & Smyth, 2000). Thus, Text Mining is a special case of Data Mining, where input
data is text instead of structured data. Normally, text-mining systems generate structured
representations of the text, which are then analyzed by Data Mining tools. The simplest
representation of a text is a vector with the number of occurrences of each word in the
text (called a bag-of-words). This representation can be easily created and manipulated,
but ignores all the text structure. Text-mining systems may also use Natural Language
Processing (NLP) techniques to represent and process text more effectively. NLP is a
broad research area that aims at analyzing spoken, handwritten, printed, and electronic
text for different purposes, such as speech recognition or translation (Manning &
Schtze, 1999). The most popular NLP techniques used by text-mining systems include:
tokenization, morphology analysis, part-of-speech tagging, sense disambiguation,
parsing, and anaphora resolution.
Tokenization aims at identifying boundaries in the text to fragment it into basic units
called tokens. The first step in a text-mining system is to identify the tokens. The most
commonly used token is the word. In most languages, white-space characters can be
considered an accurate boundary to fragment the text into words. This problem is more
complex in languages without explicitly delimiters, such as Chinese (Wu & Fung, 1994).
Morphology analysis aims at grouping the words (tokens) that are variants of a common
word, and therefore are normally used with a similar meaning (Spencer, 1991). This
involves the study of the structure and formation of words. A common type of inflectional
variants results from the tense on verbs. For example, binding and binds are
inflectional variants of bind. Some other word variants result from prefixing, suffixing,
infixing, or compounding.
Part-of-speech tagging aims at labeling each word with its semantic role, such as article,
noun, verb, adjective, preposition, or pronoun (Baker, 1989). This involves the study of
the structure and formation of sentences. The tagging is a classification of words
according to their semantic role and to their relations to each other in a sentence. Sense
disambiguation selects the correct meaning of a word in a given piece of text. For example,
compound has two different senses in the expressions compound the ingredients
and chemical compound. Normally, the part-of-speech tags are used as a first step in
sense disambiguation (Wilks & Stevenson, 1997)
286 Couto & Silva
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Parsing aims at identifying the syntactic structure of a sentence (Earley, 1970). The
syntactic structure of a sequence of words is composed by a set of other syntactic
structures related to smaller sequences, except for the part-of-speech tags that are
syntactic structures directly linked to words. Normally, the syntactic structure of a
sentence is represented by a syntax tree, where leafs represent the words and internal
nodes the syntactic structures. Algorithms to identify the complete syntactic structure
of a sentence are in general inaccurate and time-consuming, given the combinatorial
explosion in long sentences. An alternative is shallow-parsing, which does not attempt
to parse complex syntactic structures. Shallow-parsing only splits sentences into
phrases, in other words, subsequences of words that represent a grammatical unit, such
as noun phrase or verb phrase. Anaphora (or co-reference) resolution aims at determin-
ing different sequences of words referring to the same entity. For example, in the sentence
The enzyme has an intense activity, thus, this protein should be used, the noun
phrases The enzyme and this protein refer to same entity.
Some of the NLP techniques described above can be implemented using algorithms also
used in Data Mining. For example, part-of-speech taggers can use Hidden Markov
Models (HMMs) to estimate the probability of a sequence of part of speech assignments
(Smith, Rindflesch, & Wilbur, 2004). Not all NLP techniques improve the performance of
a given text-mining system. As a result, designers of text-mining systems have to select
which NLP techniques would be useful to achieve their ultimate goal.
After creating a structured representation of texts, text-mining systems can use the
following approaches for extracting knowledge (Leake, 1996):
rule-based or case-based
The rule-based approach relies on rules inferred from patterns identified from the text by
an expert. The rules represent in a structured form the knowledge acquired by experts
when performing the same task. The expert analyzes a subpart of the text and identifies
common patterns in which the relevant information is expressed. These patterns are then
converted to rules to identify the relevant information in the rest of the text. The main
bottleneck of this approach is the manual process of creating rules and patterns. Besides
being time-consuming, in most cases, this manual process is unable to devise from a
subpart of the text the set of rules that encompass all possible cases.
The case-based approach relies on a predefined set of texts previously annotated by an
expert, which is used to learn a model for the rest of the text. Cases contain knowledge
in an unprocessed form, and they only describe the output expected by the users for a
limited set of examples. The expert analyzes a subpart of the text (training set) and
provides the output expected to be returned by the text-mining system for that text. The
system uses the training set to create a probabilistic model that will be applied to the rest
of the text. The main bottleneck of this approach is the selection and creation of a training
set large enough to enable the creation of a model accurate for the rest of the text.
The manual analysis of text requires less expertise in the case-based approach than in
the rule-based approach. In the rule-based approach, the expert has to identify how the
Mining BioLiterature 287
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
relevant information is expressed in addition to the expected output. However, rule-
based systems can use this expertise to achieve high precision by selecting the most
reliable rules and patterns.
State of the Art
The main problem in BioLiterature mining is coping with the lack of a standard nomen-
clature for describing biologic concepts and entities. In BioLiterature, we can often find
different terms referring to the same biological concept or entity (synonyms), or the same
term meaning different biological concepts or entities (homonyms). Genes, whose name
is a common English word, are frequent, which makes it difficult to recognize biological
entities in the text.
Recent advances in Text Mining of BioLiterature already achieved acceptable levels of
accuracy in identifying gene and protein names in the text. However, the extraction of
relationships, such as functional annotations, is still far from being solved. Recent
surveys report these advances by presenting text-mining tools that are not only run in
different corpora but also perform different tasks (Hirschman, Park, Tsujii, Wong, & Wu,
2002; Blaschke, Hirschman, & Valencia 2002; Dickman, 2003; Shatkay & Feldman, 2003).
On the other hand, recent challenging evaluations compared the performance of different
approaches in solving the same tasks using the same corpus. For example, the 2002 KDD
Cup (bio-text task) consisted on identifying which biomedical articles contained relevant
experimental results about Drosophila (fruit fly), and the genes (transcripts and proteins)
involved (Yeh, Hirschman, & Morgan, 2003). The best submission out of 32 obtained 78%
F-measure in the article decision, and 67% F-measure in the gene decision.
A similar challenging evaluation was the 2004 TREC genomics track, which consisted on
identifying relevant documents and documents with relevant experimental results about
the mouse (Hersh et al., 2004). The first task was a typical Information Retrieval task. A
list of documents and a list of topics were given as input. The goal was to identify the
relevant documents for each topic. The best submission out of 47 obtained 41%
precision. The second subtask comprised the selection of documents with relevant
experimental information. The best submission out of 59 obtained 27% F-measure. In
addition to document selection, the task also comprised automatic annotations of genes.
The best submission out of 36 obtained 56% F-measure.
Another challenging evaluation was BioCreAtIvE (Hirschman, Yeh, Blaschke, & Valencia,
2005). This evaluation comprised two tasks. The first aimed at identifying genes and
proteins in BioLiterature. The best submission out of 40 obtained 83% F-measure. The
second task addressed the automatic annotation of human proteins, and involved two
subtasks. The first subtask required the identification of the texts that provided the
evidence for extracting each annotation. From 21 submissions, the highest precision was
78% and the highest recall was 23%. The second subtask consisted on automatic
annotation of proteins. From 18 submissions, the highest precision was 34% and the
highest recall was 12%.
288 Couto & Silva
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Automatic Annotation
One of the most important applications of text-mining systems is the automatic annota-
tion of genes and proteins. A gene or protein annotation consists of a pair composed by
the gene or protein and a description of its biological role. Normally, descriptions use
terms from a common ontology. The Gene Ontology (GO-Consortium, 2004) provides a
structured controlled vocabulary that can be applied to different species (GO-Consor-
tium, 2004). GO has three different aspects: molecular function, biological process, and
cellular component. To comprehend a gene or protein activity is also important to know
the biological entities that interact with it. Thus, the annotation of a gene or protein also
involves identifying interacting chemical substances, drugs, genes and proteins.
Text-mining systems that automatically annotate genes or proteins can be categorized
according to: the mining approach taken (rule-based or case- based), the NLP techniques
applied, and the amount of manual intervention required.
Rule-Based Systems
AbXtract was one of the first text-mining systems attempting to characterize the function
of genes and proteins based on information automatically extracted from BioLiterature
(Andrade & Valencia, 1998). The system assigns relevant keywords to protein families
based on a rule comprising the frequency of the keywords in the abstracts related to the
family. In addition to using a rule-based approach, AbXtract relies in only one rule that
does not require human intervention. A similar approach is taken by the system proposed
by Prez, Perez-Iratxeta, Bork, Thode, and Andrade (2004), which annotates genes with
keywords extracted from abstracts based on mappings between different ontologies.
An example of a system based on a large number of rules is BioRAT (Corney, Buxton,
Langdon, & Jones, 2004). Given a query, BioRAT finds documents and highlights the
most relevant facts in their abstracts or full texts. However, the rules are exclusively
derived from patterns inserted by the user. Textpresso is another rule-based system that
finds documents and marks them up with terms from a built-in ontology (Mller, Kenny,
& Sternberg, 2004). The system assigns to each entry of the ontology regular expressions
that capture how the entry can be expressed in BioLiterature. Textpresso is not so
dependent on the user as BioRAT, since many of the regular expressions are automati-
cally generated to account for regular forms of verbs and nouns.
BioIE is a system that takes more advantage of NLP techniques. It extracts biological
interactions from BioLiterature and annotates them with GO terms (Kim & Park, 2004). The
system uses morphology, sense disambiguation, and rules with syntactic dependencies
to identify GO terms in the text. BioIE uses 1,312 patterns to match interactions in the
sentences, so it also requires substantial manual intervention. Koike, Niwa, and Takagi
(2005) propose a similar system that annotates gene, protein, and families with GO terms
extracted from texts. The system uses morphology, part-of-speech tagging, shallow
parsing, and simple anaphora resolution. To extract the relationships, it uses both
automatically generated and manually inserted rules.
Mining BioLiterature 289
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Case-Based Systems
A text-mining system using the case-based approach was proposed by Palakal, Stephens,
Mukhopadhyay, and Raje (2003). The system extracts relationships between biological
objects (e.g. protein, gene, cell cycle). The system uses sense disambiguation, and a
probabilistic model to find directional relationships. The model is trained using examples
of sentences expressing a relationship.
Table 1. Categorization of some recent text-mining systems designed for automatic
annotation of genes and proteins; for each system, the table indicates the mining
approach taken, the proportion of NLP techniques used and the proportion of manual
intervention needed to generate rules, patterns, or training sets
System Mining NLP Manual
Andrade and Valencia (1998) Rule-based nil nil
Prez et al. (2004) Rule-based nil nil
Corney et al. (2004) Rule-based Low High
Mller et al. (2004) Rule-based Low Medium
Kim and Park (2004) Rule-based Medium Medium
Koike et al. (2005) Rule-based High Medium
Palakal et al. (2003) Case-based Medium Low
Chiang and Yu (2003) Case-based Medium Low

Figure 1. UML use case diagram of GOAnnotator
290 Couto & Silva
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Figure 2. A list of documents related to the protein Ras GTPase-activating protein
4 provided by the GOAnnotator. The list is sorted by the similarity of the most similar
term extracted from each document. The curator can invoke the links in the Extract
column to see the extracted terms together with the evidence text. By default,
GOAnnotator uses only the abstracts of scientific documents, but the curator can
replace or add text (links in the AddText column).
MeKE is another system that extracts protein functions from BioLiterature using
sentence alignment (Chiang & Yu, 2003). MeKE also uses sense disambiguation. The
system uses a statistical classifier that identifies common patterns in examples of
sentences expressing GO annotations. The classifier uses these patterns to decide if a
given sentence expresses a GO annotation.
Discussion
The systems described above show how Text Mining can help curators in the annotation
process. Most rely on domain knowledge manually inserted by curators (see Table 1).
Domain knowledge improves precision, but it cannot be easily extended to work on other
domains and demands an extra effort to keep the knowledge updated as BioLiterature
evolves. This approach is time-consuming and makes the systems too specific to be
extended to new domains. Thus, an approach to avoid this process is much needed.
GOAnnotator
This section illustrates how text-mining can be integrated in a biological database
curation process, by describing GOAnnotator, a tool for assisting the GO annotation of

Mining BioLiterature 291
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
UniProt entries (Rebholz-Schuhmann, Kirsch, & Couto, 2005). GOAnnotator links the GO
terms present in the uncurated annotations with evidence text automatically extracted
from the documents linked to UniProt entries.
Figure 1 presents the data flow involved in the processing steps of GOAnnotator and
in its interaction with the users and external sources. Initially, the curator provides a
UniProt accession number to GOAnnotator. GOAnnotator follows the bibliographic
links found in the UniProt database and retrieves the documents. Additional documents
are retrieved from the GeneRIF database (Mitchell et al., 2003). Curators can also provide
any other text for mining. GOAnnotator then extracts from the documents GO terms similar
to the GO terms present in the uncurated annotations.
In GOAnnotator the extraction of GO terms is performed by FiGO, a tool that receives text
and returns the GO terms detected (Couto, Silva, & Coutinho, 2005). FiGO is rule-based,
does not use any NLP technique and does not require manual intervention. FiGO assigns
a confidence value to each GO term that represents the terms likelihood of being
mentioned in the text. The confidence value is the product of two parameters. The first,
called local evidence context (LEC), is used to measure the likelihood that words in the
text are part of a given GO term. The second parameter is the inverse of their frequency
in GO. GO terms are similar if they are in the same lineage or if they share a common parent
in the GO hierarchy. FiGO uses the semantic similarity measure of (Lin, 1998) to compute
the degree of similarity between two GO terms.
GOAnnotator ranks the documents based on the extracted GO terms from the text and
their similarity to the GO terms present in the uncurated annotations (see Figure 2). Any
extracted GO term is an indication for the topic of the document, which is also taken from
the UniProt entry.
Figure 3. For each uncurated annotation, GOAnnotator shows the similar GO terms
extracted from a sentence of the selected document. If any of the sentences provides
correct evidence for the uncurated annotation, or if the evidence supports a GO term
similar to that present in the uncurated annotation, the curator can use the Add
option to store the annotation together with the document reference, the evidence
codes and additional comments.

292 Couto & Silva
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
GOAnnotator displays a table for each uncurated annotation with the GO terms that were
extracted from a document and found similar to the GO term present in the uncurated
annotation (see Figure 3). The sentences from which the GO terms were extracted are also
displayed. Words that have contributed to the extraction of the GO terms are highlighted.
GOAnnotator gives the curators the opportunity to manipulate the confidence and
similarity thresholds to modify the number of predictions.
Future Trends
The performance of text-mining tools that automatically annotate genes or proteins is still
not acceptable by curators. Gene or protein annotation is more subjective and requires
more expertise than simply finding relevant documents and recognizing biological
entities in texts. Moreover, an annotation tool can only perform well when it is using the
correct documents and the correct entities. Errors in the retrieval of documents or in the
recognition of entities will be the cause of errors in the annotation task.
Existing tools that retrieve relevant documents do not always provide what the curators
want. On the contrary, curators spend a large amount of their time finding the right
documents. This is probably the main reason why many curators are still not using text-
mining tools for gene or protein annotation. Another reason is that quite often the full
texts are not electronically available. Curators need additional information that is not
usually present in the abstracts, such as the type of experiments applied and the species
from which proteins originate. Finally, another reason is that most text-mining tools
depend on domain knowledge manually inserted by curators, which is also very time-
consuming.
Text-mining tools acquire domain knowledge from the curators in the form of rules or
cases. The identification of rules requires more effort to the curators than the evaluation
of a limited set of cases. However, a single rule can express knowledge not contained in
a large set of cases. Neither source of knowledge subsumes the other: the knowledge
represented by a rule is normally not well-represented by any set of cases, and it is
difficult to identify a set of rules representing all knowledge expressed by a set of cases.
Couto, Martins, and Silva (2004) proposed an approach to obtain the domain knowledge
that does not require human intervention. Instead of obtaining the domain knowledge
from curators, they propose acquiring it from publicly available databases that already
contain curated data. Text-mining systems could consider these databases as training
sets from which rules, patterns, or models can be automatically generated. Besides
avoiding direct human intervention, these automated training sets are usually much
larger than individually generated training sets. Another advantage is that the tools
training data does not become outdated as public databases can be tracked for updates
as they evolve.
Mining BioLiterature 293
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Summary
Bioinformatics aims at understanding living systems by inferring knowledge from
biological information, such as DNA and protein sequences. The role of Text Mining in
Bioinformatics is to automatically extract knowledge from BioLiterature. This field is new
and has evolved over the last decade, motivated by the opportunity brought by tapping
the large amount of information that has been published in BioLiterature with automatic
text processing software.
Researchers will tend to use databases to store and find facts, but the evidence
substantiating them will continue to be described as unstructured text. As a result, text-
mining tools will continue to have an important role in Bioinformatics. Recent advances
in Text Mining of BioLiterature are already promising, but many problems remain. In our
opinion, the future of text-mining tools for gene or protein annotation will mainly depend
on a better use of NLP techniques, and in an efficient integration of existing domain
knowledge available in biological databases and ontologies.
References
Andrade, M., & Valencia, A. (1998). Automatic extraction of keywords from scientific
text: Application to the knowledge domain of protein families. Bioinformatics,
14(7), 600-607.
Apweiler, R., Bairoch, A., Wu, C., Barker, W., Boeckmann, B., Ferro, S., et al. (2004).
UniProt: The universal protein knowledgebase. Nucleic Acids Research, 32
(Database issue), D115-D119.
Baker, C. (1989). English syntax. MIT Press.
Blaschke, C., Hirschman, L., & Valencia, A. (2002). Information extraction in molecular
biology. Briefings in BioInformatics, 3(2), 154-165.
Blaschke, C., Oliveros, J., & Valencia, A. (2004). Mining functional information associ-
ated to expression arrays. Functional and Integrative Genomics, 1(4), 256-268.
Chiang, J., & Yu, H. (2003). MeKE: Discovering the functions of gene products from
biomedical literature via sentence alignment. Bioinformatics, 19(11), 1417-1422.
Corney, D., Buxton, B., Langdon, W., & Jones, D. (2004). BioRAT: Extracting biological
information from full-length papers. Bioinformatics, 20(17), 3206-3213.
Couto, F., Martins, B., & Silva, M. (2004). Classifying biological articles using Web
resources. Proceedings of the 2004 ACM Symposium on Applied Computing (pp.
111-115).
Couto, F., Silva, M., & Coutinho, P. (2005). Finding genomic ontology terms in text using
evidence content. BMC Bioinformatics, 6(Suppl 1), S21.
Dickman, S. (2003). Tough mining: The challenges of searching the scientific literature.
PLoS Biology, 1(2), E48.
294 Couto & Silva
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Earley, J. (1970). An efficient context-free parsing algorithm. Communications of ACM,
13, 94-102.
GO-Consortium (2004). The gene ontology (GO) database and informatics resource.
Nucleic Acids Research, 32(Database issue), D258-D261.
Hand, D., Mannila, H., & Smyth, P. (2000). Principles of Data Mining. MIT Press.
Hearst, M. (1999). Untangling text data mining. Proceedings of the 37th ACL Meeting
of the Association for Computational Linguistics (pp. 3-10).
Hersh, W., Bhuptiraju, R., Ross, L., Johnson, P., Cohen, A., & Kraemer, D. (2004). TREC
2004 genomics track overview. Proceedings of the 13th Text REtrieval Conference
(TREC 2004) (pp. 14-24).
Hirschman, L., Park, J., Tsujii, J., Wong, L., & Wu, C. (2002). Accomplishments and
challenges in literature data mining for biology. Bioinformatics, 18(12), 1553-1561.
Hirschman, L., Yeh, A., Blaschke, C., & Valencia, A. (2005). Overview of BioCrEAtIvE:
critical assessment of information extraction for biology. BMC Bioinformatics,
6(Suppl 1), S1.
Kim, J., & Park, J. (2004). BioIE: Retargetable information extraction and ontological
annotation of biological interactions from literature. Journal of Bioinformatics and
Computational Biology, 2(3), 551-568.
Koike, A., Niwa, Y., & Takagi, T. (2005). Automatic extraction of gene/protein biological
functions from biomedical text. Bioinformatics, 21(7), 1227-1236.
Leake, D. (1996). Case-based reasoning: Experiences, lessons, and future directions.
AAAI Press/MIT Press.
Lin, D. (1998). An information-theoretic definition of similarity. Proceedings of the 15th
International Conference on Machine Learning (pp. 296-304).
Manning, C., & Schtze, H. (1999). Foundations of statistical natural language
processing. The MIT Press.
Mitchell, J., Aronson, A., Mork, J., Folk, L., Humphrey, S., & Ward, J. (2003). Gene
indexing: characterization and analysis of NLMs GeneRIFs. Paper presented at the
AMIA 2003 Annual Symposia, Washington, DC.
Mller, H., Kenny, E., & Sternberg, P. (2004). Textpresso: An ontology-based information
retrieval and extraction system for biological literature. PLOS Biology, 2(11), E309.
Palakal, M., Stephens, M., Mukhopadhyay, S., & Raje, R. (2003). Identification of
biological relationships from text documents using efficient computational meth-
ods. Journal of Bioinformatics and Computational Biology, 1(2), 307-342.
Prez, A., Perez-Iratxeta, C., Bork, P., Thode, G., & Andrade, M. (2004). Gene annotation
from scientific literature using mappings between keyword systems. Bioinformatics,
20(13), 2084-2091.
Rebholz-Schuhmann, D., Kirsch, H., & Couto, F. (2005). Facts from text - is text mining
ready to deliver? PLoS Biology, 3(2), e65.
Shatkay, H. & Feldman, R. (2003). Mining the biomedical literature in the genomic era:
An overview. Journal of Computational Biology, 10(6), 821-855.
Mining BioLiterature 295
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Smith, L., Rindflesch, T., & Wilbur, W. (2004). MedPost: A part-of-speech tagger for
biomedical text. Bioinformatics, 20(14), 2320-2321.
Spencer, A. (1991). Morphological theory. Oxford: Blackwell.
Wilks, Y., & Stevenson, M. (1997). The grammar of sense: Using part-of-speech tags as
a first step in semantic disambiguation. Journal of Natural Language Engineer-
ing, 4(3).
Wu, D. & Fung, P. (1994). Improving Chinese tokenization with linguistic filters on
statistical lexical acquisition. Proceedings of the 4th ACL Conference on Applied
Natural Language Processing (pp. 13-15).
Yeh, A., Hirschman, L., & Morgan, A. (2003). Evaluation of text data mining for database
curation: Lessons learned from the KDD challenge cup. Bioinformatics, 19(1),
i331-i339.
296 Choi & Kim
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Chapter XVI
Comparative Genome
Annotation Systems
Kwangmin Choi, Indiana University, USA
Sun Kim, Indiana University, USA
Abstract
Understanding the genetic content of a genome is a very important but challenging
task. One of the most effective methods to annotate a genome is to compare it to the
genomes that are already sequenced and annotated. This chapter is to survey systems
that can be used for annotating genomes by comparing multiple genomes and discusses
important issues in designing genome comparison systems such as extensibility,
scalability, reconfigurability, flexibility, usability, and data mining functionality. We
also discuss briefly further issues in developing genome comparison systems where
users can perform genome comparison flexibly on the sequence analysis level.
Introduction
Once a complete genome sequence becomes available, the next and more important goal
is to understand the content of the genome. The exponential accumulation of genomic
sequence data demands use of computational approaches to systematically analyze
huge amount of genomic data. The availability of such genome sequence data and
diverse computational techniques has made comparative genomics research activity
Comparative Genome Annotation Systems 297
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
to compare sequences of multiple genome sequences to become useful not only for
finding common features in different genomes, but also for understanding evolutionary
mechanisms among multiple genomes.
The process of assigning genomic functions to genes is called genome annotation,
which utilizes diverse domain knowledge sources from sequence data to the contextual
information of the whole genome. Currently there exist several methods for genome
annotation. Experimental approaches for genome annotation are probably most reliable,
but slow and labor-intensive. Genome annotation can be done computationally as well
by (1) assigning function(s) to a gene based on its sequence similarity to other genes
that are already annotated with well-defined gene functions, (2) assigning function(s)
to a gene based on its position in a conserved gene cluster through comparative analysis
of multiple genomes, and (3) inferring function via detecting functional coupling.
Genome annotation probably can be done most accurately by comparing a genome with
its phylogenetically-related genomes, which we termed as comparative genome anno-
tation. Comparing multiple genomes, however, is a very challenging task. First of all,
genome comparison requires handling complicated relationships of many entities. For
example, comparison of all protein coding genes in three genomes with 2,000 genes in
each genome can involve several million pairwise relationships among genes. The way
of genome selection raises another problem because the choice of genomes to be
compared is entirely subjective and there are numerous combinations of genomes to be
compared. Thus it is necessary to develop an information system for comparative
genome annotation which can deal with such challenges.
This chapter primarily aims to survey existing systems from the data mining perspective.
Well developed data mining tools will be very helpful in handling numerous functional
relationships among genes and genomes. For example, an accurate sequence clustering
tool can simplify the genome annotation task significantly, since the user can handle a
set of sequences as a single unit for the next analysis. Unfortunately, existing data mining
tools are not developed to handle a continuous character stream or genomic sequence,
thus our discussion from the data mining perspective is to propose what is needed, rather
than characterizing existing systems in the terms of traditional data mining perspective.
Here we propose the desirable features that comparative genome annotation systems
should have:
Extensibility: There are always new resources such as newly sequenced genomes
and computational tools. The system should be able to include them as they
become available.
Reconfigurability: There are numerous different ways to combine data and tools,
so no single system can meet all needs of users. When needed, it is desirable to
reconfigure the system for a specific task.
Genome selection flexibility: The choice of genomes to be compared is entirely
subjective. Thus users should be able to compare genomes of their choice with
different criteria for sequence comparison.
Usability: Genome comparison involves a huge amount of data, so the system
should be easy to use. In addition, it should be easy to port to other platforms.
298 Choi & Kim
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Data mining: Users need to perform a series of analyses to achieve a research goal
and the system should also provide high-level data mining tools to simplify genome
analysis task. To explain the data mining issue further, assume that a user wants
to analyze a sequence s against several genomes, G
1
, G
2
, G
3
. Obviously a user can
perform three different comparisons of s to each G
i
and then combine the results
into one, and then proceed for further analysis. A well developed system can easily
mitigate the burden of combining three different search results by providing some
system functions. Alternatively, it is possible to use a high performance sequence
clustering algorithm to generate the output in a single operation.
System integration in a distributed environment: The annotation of a large
genome is often performed collaboratively at different organizations. Thus, it is
challenging to maintain consistency of data and share information across the
multiple information systems.
This chapter will survey comparative genome annotation systems, the SEED, DAS,
CaWorkBench2.0, STRING, and PLATCOM. Then we will compare these systems with
the desirable features that we described above.
The SEED
The SEED project (Overbeek, Disz, & Stevens, 2004) aims to provide a suite of programs
which enable distributed users to annotate new genomes rapidly and cooperatively. By
Figure 1. Web interface of the SEED at http://theseed.uchicago.edu/FIG/index.cgi

Comparative Genome Annotation Systems 299
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
using this system, users can create, collect, and maintain sets of gene annotations
organized by groups of related biological functions (i.e., subsystems) over multiple
genomes. The subsystem is defined as a set of biological and biochemical functions
which together implement a specific process. By annotating one subsystem at a time, the
SEED supports the annotation of a single subsystem over multiple genomes simulta-
neously. So users may examine the relationship between a given gene and a group of other
genes by using contextual clues relevant to the determination of functions.
System Architecture and Data Sources
To maintain a scalable integration of genomes, the SEED consists of the essential and
basic elements, which are (1) a scalable infrastructure to support the construction of
large-scale integration genomic data, (2) the peer-to-peer (i.e., P2P) infrastructure which
supports the distributed annotation and the integrated database synchronization, and
(3) programs to enable extension of the integrated database environment to adopt new
applications and data.
For navigation and annotating genes, the SEED is providing eight core functions (1) to
localize and visualize gene clusters which are relevant to the analysis of a specific
subsystem, (2) to identify genes with sequence homology from other genomes, (3) to
visualize a gene neighborhood on the chromosome, (4) to comparatively compare one
gene neighborhood with others around corresponding genes in other genomes, (5) to
examine genes that implement closely related metabolic functions, (6) to modify existing
function assignments and annotations, (7) to detect inconsistent representation of
function, and (8) to generate packages of assignments and annotations corresponding
to a single subsystem.
The SEED also adopts a UNIX-like command-line environment where many tools are
direct descendants of the former WIT system. It provides convenient methods to easily
extract and manipulate genomic data. For example, a set of command, pegs 83333.1 |
fid2dna | translate, consists of three separate commands (pegs, fid2dna, and translate),
and redirections of their input (83333.1) and output are controlled by pipes as in UNIX
operating system. By using pipes, each program can be modular because a pipe can be
used between two programs to combine their functionality. This system design enables
the SEED to add new components and evolve easily by multiple participants and gives
reconfigurable feature to the system.
The SEED supports the information sharing and synchronization of data among indi-
vidual biologists who collaborate with others in a geographically distributed environ-
ment. Via Web interface, users can examine particularly interesting genes, define new
subsystems, modify annotations, and assign new functions. To synchronize all anno-
tations and assignments periodically, the SEED uses a basic peer-to-peer (i.e., P2P)
synchronization facility which supports the information sharing among the SEED
installations.
300 Choi & Kim
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Data Mining
Biologists are often interested in a specific subsystem or a set of genes that are known
to work together. A good example is the metabolic pathway subsystem. The SEED deals
with metabolic pathway subsystems where the gene functions are represented via EC
numbers
1
and users may enumerate the set of functional roles required by the subsystem.
Exploring a specific subsystem using the SEED system can be briefly described as
follows: To get metabolic overview, reference to KEGG database (http://
www.genome.ad.jp/kegg) is provided and users can rapidly extract a set of functional
roles (i.e.enzymes) represented with EC numbers. Then users are able to (1) build a table
representing which of the functional roles can be linked to genes in each of the completely
sequenced genomes, (2) modify assignments to make the table more accurate, (3) make
an assessment about which genomes have which versions of the subsystem to reveal genes
which are supposed to exist but cannot be located (i.e., missing genes), and (4) locate
candidates for the missing genes. Searching for missing genes by referring to such
contextual information is one of the most important subjects in comparative genomics.
DAS
DAS, or the Distributed Annotation System, is a genome annotation system which is
designed to deal with multiple third-party genome annotation and genomic data integra-
tion. The multiple third-party genome annotation introduces a combination of experimen-
tal and computational methods using diverse data types and analysis tools. Unfortu-
nately such a diverse annotating environment causes several critical problems. The most
serious problem is that multiple third-party annotation can make genomic information
fragmented, inconsistent, or redundant. As a result, to get some specific information,
users often should search multiple databases rather than one unified data source for
genomic annotation. In addition to this inconvenience, data formats are often inconsis-
tent, so users need to manually integrate heterogeneous data formats before using such
data.
One solution is to build a centralized system where all genome annotations are submitted
in a uniform format. The disadvantage of the centralized system is that it may impose users
too much restriction on the data input and retrieval. An alternative solution may be using
URL to link one entry to one or more Web-based genome annotation databases such as
GenBank and SwissProt. NCBI Linkout service is one example of this approach. Although
URL approach is already used in many existing databases, this is actually not an
integrated environment at all.
Instead, DAS provides a different solution called the client-server system which allows
genome annotation to be decentralized and integrated by client-side. This approach
enables a client to access genome annotation information from the multiple distant
reference and annotation servers, collect that information, and represent it in a single
and integrated view on a need basis. As a result, complicated coordination among the
multiple data sources is no longer necessary. DAS/1 servers are currently running at
WormBase (http://www.wormbase.org), FlyBase (http://flybase.bio.indiana.
Comparative Genome Annotation Systems 301
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
edu), Ensembl (http://www.ensembl.org/) and many others (See biodas Webpage). The
DAS project is still under development by many participants.
System Architecture and Data Sources
To implement a successful distributed annotation system, DAS adopts a standard data
format which deals with multiple levels of relative coordinates where annotations are
related to arbitrary hierarchical markers. This standard format must be easily generated,
parsed, and extended to include new types of annotations. Grouping annotations into
conceptual functional categories is introduced to make data more manageable and to
facilitate formulating biologically relevant queries on the annotation servers.
In the DAS system, annotations are based on common reference sequence which
consists of a set of entry points and their lengths. The entry points describe the top level
items on the reference sequence map and each entry point may have its substructure, i.e.,
a series of subsequences and their start and end points. The reference sequence can be
one of the entry points, or any of the subsequences within the entry point. Each
annotation can be clearly located by providing its start and stop positions relative to a
reference sequence.
DAS constructs a distributed annotation system by using a reference sequence server
and one or more annotation servers. The division between reference and annotation
servers is only conceptual and there exists no critical difference between them except that
reference sequence servers provide sequence map, raw DNA data, component annota-
tions describing how the sequence is assembled from the top to down, and super-
component annotations describing the assembly of the sequence from the bottom up,
whereas annotation servers are not required to provide such information.
When a reference server acts as a third-party annotation server, users may search against
one or more annotation servers to collate genomic information using a sequence browser.
Once the browser retrieves the information in a standard format, the annotations are
integrated and represented in tabular or graphical form. Annotation servers return the
lists of genome annotations across a given region of the genome. Each annotation is
linked to the genome map using positions (i.e., start-stop location) in relation to one of
the reference subsequences and the server stores a unique ID and a structured
description of each annotations attributes. Annotations include types (e.g., exon,
intron, CDS), methods (e.g., experiment, data mining tool), and categories (e.g., homol-
ogy, variation, transcribed).
Data Mining
Because DAS is a Web-based annotation system, clients-server interactions are carried
out via query as a formatted URL request to the annotation/reference servers. To
construct a DAS query, combine a genome assemblys base URL with the sequence entry
point and type specifiers available for the assembly. The entry point specifies chromo-
some position, and the type indicates the annotation table requested. That is, each URL
302 Choi & Kim
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
has a site-specific prefix, followed by a standardized path (i.e./das) and query string.
Clients should send their request to the server using the following format (Table 1).
Standard Request Format
Two prototype DAS client programs are provided. The first one is Geodesic which is a
stand alone Java application. It connects to multiple DAS servers upon the user
selection, collates annotations, and displays them in an integrated map. The user
identifies a segment of the genome by browsing through entry points or entering a region
name directly. Displayed data can be saved in FASTA, GFF, or DAS XML format.
Table 1. A standard URL query format and examples used in DAS system
Standard Request Format
http:// (Reference Server's URL) / (Standardized Path) / (Data Source) / (Command?Query)

e.g. A query that returns all the records in the WormBase at the location of 1000-2000 on the
chromosome I:
http://www.wormbase.org/db/das/elegans/features?segment=CHROMOSOME_I:1000,2000

e.g. A query that returns all the records in the refGene table for the chromosome position
chr1:1-100000 on the human hg16 assembly:
http://genome.ucsc.edu/cgi-bin/das/hg16/features?segment=1:1,100000;type=refGene
Figure 2. caWorkBench2.0 system interface (caWorkBench2.0 can be downloaded at
http://amdec-bioinfo.cu-genome.org/html/caWorkBench.htm)

Comparative Genome Annotation Systems 303
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Geodesic is appropriate for extensive, long-term use. The second one is the browser-less
server-side integration. A database is coupled with third-party servers behind the web
browser and these third-party data are then integrated into data displays of the database.
Thus no client-software would be needed. DasView is a Perl program working as a server-
side script for such purposes, which connects to multiple DAS servers, constructs an
integrated image, and displays the image as a set of image map on web browser. DasView
is suitable for casual use because it does not require the user to preinstall the software.
caWorkBench2.0
caWorkBench2.0 is an open source bioinformatics platform which evolved from a project
originally sponsored by the National Cancer Institute Center for Bioinformatics (NCICB).
caWorkBench2.0 is written in Java and provides sophisticated methods for data manage-
ment, analysis and visualization. This system especially emphasizes on the data and
algorithm integration, microarray data analysis, metabolic pathway analysis, sequence
analysis, reverse engineering, transcription factor binding site detection, and motif/
pattern discovery. All these features are fully implemented.
System Architecture and Data Sources
The key concept of caWorkBench2.0 system is using structured communication between
independent modules and their smooth integration into caWorkBench2.0 platform. Over
50 components are provided to combine diverse genome analyses using the communi-
cation model. All applications run using Suns Java WebStart.
caWorkBench2.0 deals with sequence data (e.g., DNA, RNA, protein), genetic data,
phenotype, gene expression data (e.g., Affymetrix, GenePix data format), and complex
multidimensional data types (e.g., metabolic pathway, gene regulatory pathway,
proteomics, phylogenetics data). External resources and services include cluster-accel-
erated version of BLAST, cluster implementation of pattern discovery algorithm, Golden
Path genome sequence retrieval, access to CGAP gene expression data (http://
cgap.nci.nih.gov/SAGE) and BioCarta metabolic pathway diagram (http://
cgap.nci.nih.gov/Pathways).
Data Mining
Functional modules in CaWorkBench2.0 can be roughly divided into nine categories: (1)
modules for microarray data analysis to perform data normalization, filtering, analysis,
and biclustering, (2) modules for microarray data visualization providing color-mosaic
display, dendrograms, expression profiles, tabular view, microarray view, expression on
pathways, and cytoscape expression, (3) modules for diverse sequence analyses which
perform BLAST, motifs search, promoter analysis, and synteny mapping, (4) modules for
sequence visualization providing alignment result visualization, motif visualization,
motif histogram, and dot plot, (5) modules for pathway analyses providing functions for
304 Choi & Kim
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
reverse engineering, Cytoscape (http://www.cytoscape.org/), BioCarta (http://
cgap.nci.nih.gov/Pathways/BioCarta_Pathways), caBIO (http://ncicb.nci.nih.gov/core/
caBIO), and pathway browser, (6) modules for genotypic data analysis to provide
association discovery tool, (7) modules for genotypic data visualization to provide
functions for color mosaic display, tabular view, microarray view, (8) modules for
ontology analysis providing gene ontology browser, and (9) other modules including
project panel, phenotype selection panel, gene selection panel, gene annotation, dataset
annotation, and dataset history.
For example, a sample gene expression analysis can be done as follows: data are imported
from Affymetrix and GenePix microarray results, then clustered and visualized based on
regulatory network reverse engineering. Sequence analysis includes BLAST/HMM/
Smith-Waterman search, pattern recognition (SPLASH), regulatory sequence mapping/
recognition, synteny calculation/visualization, sequence retrieval from DAS source
(UCSC GoldenPath, http://genome.ucsc.edu/). For gene annotation, GO (Gene Ontology,
http://www.geneontology.org/) annotation filtering is used.
STRING
STRING or Search Tool for the Retrieval of

Interacting Genes/Proteins is a database
which, under one common and simple framework, several categories of protein-protein
interaction information

are integrated and mapped onto a set of proteins or a large number
of genomes by using an integrated scoring scheme. STRING can be used for comparative
evolutionary studies.
Figure 3. STRING Web interface at http://string.embl.de

Comparative Genome Annotation Systems 305
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
System Architecture and Data Sources
STRING provides three types of proteinprotein association

relationships which include
(1) direct experimental data about protein-protein interaction, (2) functional classifica-
tion of metabolic, signaling, transcriptional pathways, and (3) de novo prediction of
protein-protein interaction.
The completely sequenced genome sequence and protein-protein association data in
STRING are

imported from several public databases, including ENSEMBLE (http://
www.ensembl.org/) and SwissProt (http://us.expasy.org/sprot/). STRING periodically
extracts genomic context associations from these databases such as conserved genomic
neighborhood,

gene fusion events, and co-occurrence of genes across genomes. Infor-
mation on functional grouping of proteins is collected from the KEGG database (http:/
/www.genome.jp/kegg/).
STRING also provides

a large collection of predicted associations that are produced de
novo. These de novo predictions are based on systematic genome comparisons

by
analyzing genomic context. Because STRING is fully pre-computed, access to informa-
tion is quite fast. Customization is achieved by enabling or disabling each evidence type
at run-time. The database contains a much larger number of associations than primary
interaction databases with diverse confidence scores. Thus it is recommended for users
to review all functional partners of a query protein if it is not yet completely annotated.
Data Mining
STRING employs two modes to transfer known

and predicted protein-protein associa-
tions over multiple genomes. Protein-mode uses

quantitative sequence similarity/
homology searches and no pre-assigned

orthology relationship is considered. The
transfer depends on a pre-computed

all-to-all similarity search against all proteins from
a sequence database. Smith-Waterman algorithm is used to compute all-to-all sequence
similarity. For each

association to be transferred, the algorithm searches for potential
orthologs of the interacting partners in other genomes. Given two proteins from different
genomes are considered orthologs if they form reciprocal best hits

and there are not
any close, second-best hits (paralogs) in

either genome. Instead, COG-mode uses
orthology-assignments imported from the COG database (http://www.ncbi.nlm.nih.gov/
COG/) and transfers protein-protein interactions in an all-or-none

fashion. In this mode,
all proteins in the same COG group are considered to be functionally

equivalent across
genomes.
Once the assignment of association scores and transfer between

genomes is completed,
the next step is to compute a final combined score between

any pair of proteins or pair
of COGs. If the combined score is higher than the individual subscores, it means the
association is supported by several types

of evidences with increased confidence.
After the predicted functional associations are retrieved, users can do in-depth inves-
tigation by changing view options to display different views to the genomic informa-
tion of retrieved result. These include gene neighborhood, gene fusion, occurrence, co-
expression, experiments, databases, text mining, and summary network.
306 Choi & Kim
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
PLATCOM
PLATCOM, or Platform for Computational Comparative Genomics on the Web, is a
Web-based computational environment where users can choose any combination of
genomes freely and compare them with a suite of data mining tools. Biologists can
perform diverse comparative genomic analyses over freely selected multiple genomes,
which include multiple genome alignments, genome-scale plot, synteny and gene
neighborhood reconstruction, gene fusion detection, protein sequence clustering,
metabolic pathway analysis, and so on. PLATCOM does not aims to be a sophisticated
annotation system, but a flexible, extensible, scalable, and reconfigurable system with
emphasis on high-performance data mining on the Web.
System Architecture and Data Sources
PLATCOM is designed to evolve through three development stages. Only its first stage
is complete building the underlying architecture and individual system modules. The
second stage is to integrate individual modules with genome analysis data types and
language. In the third stage, a number of specialized comparative analysis systems will
be developed on top of PLATCOM infrastructure. One example of such systems is
ComPath, a comparative pathway analysis system, which is currently being developed.
Although system modules in PLATCOM are designed to work in a cooperative manner
on the system level, single integrated interfaces for specific tasks need to be developed
to provide the integrated service on the Web. For example, genome-scale protein
sequence clustering is one of such interfaces where users can submit uncharacterized
Figure 4. PLATCOM Web interface at http://platcom.informatics.indiana.edu/platcom

Comparative Genome Annotation Systems 307
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
protein sequences to analyze them in comparison with protein sequences from multiple
genomes of users choice.
The ultimate goal is to provide a flexible and reconfigurable system where users can
combine different tools freely. This goal can be achieved by implementing the functional
composition of basic analysis modules. The system modules will be integrated by gluing
them together on the biological sequence level using high performance data mining tools,
e.g., BAG (Kim, 2003), and a language for genome analysis. As of July 2005, there are two
methods that integrate sequence analysis modules: the gene cluster search interface
where a set of genes in a single organism or a gene cluster can be searched in other
organisms and the gene set analysis interface where a set of sequences can be analyzed
with a number of sequence analysis tools.
The whole system is built on internal databases, which consist of GenBank (ftp://
ftp.ncbi.nlm.nih.gov/genomes), SwissProt (http://www.ebi.ac.uk/swissprot), COG (http:/
/www.ncbi.nlm.nih.gov/COG), KEGG (http://www.genome.ad.jp/kegg), and Pairwise
Comparison Database (PCDB). PCDB is designed to be used for the PLATCOM system
and incorporates newer genomes automatically, so that PLATCOM can evolve as new
genomes become available. FASTA and BLASTZ are used to compute all pairwise
comparisons of protein sequence files (.faa) and whole genome sequence files (.fna) of
more than 300 fully sequenced replicons. Multiple genome comparisons usually take too
much time to complete, but the pre-computed PCDB makes it possible to perform genome
analysis very fast even on the Web. In general, PLATCOM runs several hundred times
faster than a system without PCDB when several genomes are compared. In addition to
sequence data, PLATCOM will include more data types such as gene expression and
protein interaction data.
Data Mining
Implemented component tools are designed to be interconnected with each other,
internal databases, and high performance sequence analysis tools. These component
tools can be grouped into the following five categories.
Multiple genome comparison: (1) GenomePlot retrieves pairwise comparison data
from the pre-computed PCDB to generate 2-dimensional plot and its image map. (2)
mgALIGN (Choi, Choi, Cho, & Kim, 2005) and (3) GAME (Choi, Cho, & Kim, 2005)
are powerful multiple genome comparison tools aligning genomic DNA sequences.
(4) COMPAM is a genome browser implemented in Java, which is useful to detect
common genomic elements in multiple genomic sequences.
Gene neighborhood analysis: (1) OperonViz is a tool to generate graphical visu-
alization of gene neighborhood and useful to identify horizontal gene transfers,
functional coupling and functional hitchhiking. Two versions of OperonViz are
embedded in the system: OperonViz-COG, that uses the pre-classified COG data-
base to identify homologs, and OperonViz-BAG that uses PCDB and BAG cluster-
ing algorithm to compute de novo classification of protein families. OperonViz-BAG
can be used to analyze genomes that have not been annotated. (2) MCGS (Kim,
308 Choi & Kim
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Choi, & Yang, 2005) is a tool for mining correlated gene sets over selected multiple
genomes.
Gene fusion analysis: FuzFinder uses PCDB to identify putative gene fusion
events among a set of selected genomes with the following criteria: (1) Two genes
in a reference genome must match a single open reading frame (ORF) in a target
genome with a Z-score higher than a given value. (2) The two (disjoint) halves of
the target ORF must match back to the two genes in the reference genome with a
Z-score higher than a given value. (3) The two genes in the reference genome must
not be homologous to each other.
Metabolic pathway analysis: MetaPath is a metabolic pathway analysis tool. It
combines metabolic pathway information from KEGG and sequence information
from GenBank to reconstruct metabolic pathways over the selected multiple
genomes. This tool can be used to find missing genes in metabolic pathways by
comparing reference genome with a set of genome selection. ComPath is a
comparative pathway analysis system based on KEGG and allows users to perform
more sophisticated sequence analysis for missing enzyme analysis than MetaPath.
However, the list of genomes in ComPath is limited to those in KEGG.
Sequence clustering: Three protein sequence clustering tools are provided using
a scalable sequence clustering algorithm BAG (Kim, 2003). Users may upload a set
of protein sequences in FASTA format using FASTA-BAG and BLAST-BAG or
select genomes from the genome list using Genome-BAG for genome-scale clus-
tering analysis.
All component tools use the same interface design: a set of genomes selected by users
is submitted with parameter settings specific to each module via Web interface.
Summary
The systems that we surveyed have been developed and refined over a period of time
and have provided services or library modules which are valuable in comparing multiple
genomes. Existing systems for genome comparison can be classified into two categories:
(1) task-oriented and (i2) module-based systems. Task-oriented systems are those with
specific goals that these systems try to achieve.
2
For example, euGenes (Gilbert, 2002)
is a system that provides a comprehensive comparison of large eukaryotic genomes.
MBGD (Uchiyama, 2003) and STRING also belong to this category among those
mentioned in this chapter. Instead, module-based systems provide system modules that
can be combined in a flexible way. It can be viewed as wrapper approach that is commonly
used in Web-based systems. The SEED, DAS, CaWorkBench2.0 and PLATCOM can be
categorized into the module-based system. Of course, it is possible to develop systems
that are both task-oriented and module-based.
There is also an interesting issue in designing genome comparison systems. Comparing
genomes on the sequence level takes a huge amount of time, which becomes a major
Comparative Genome Annotation Systems 309
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
hurdle to the development of a genome comparison system. Several systems, such as
MBGD, the SEED, PLATCOM, STRING and euGenes, circumvent this hurdle by using
a pre-computed genome comparison database where all-pairwise comparison of genomes
on the protein and genomic DNA sequence level is pre-computed and stored. When the
users initiate genome comparison, the pre-computed results are retrieved whenever
possible.
Comparison of the Systems
We will compare each system in terms of desirable features we have discussed in the
introduction section.
Extensibility
The SEED: By the design of the system, the SEED is easy to incorporate new tools
and new genomes into the system. In particular, the UNIX-like command-line
environment allows new tools to be combined with other tools by writing appro-
priate wrappers for new tools.
DAS: It is designed in the object-oriented paradigm, so incorporation of new tools
and genomes should be easy.
CaWorkBench2.0: Its design philosophy is Plug&Play, so incorporation of new
tools and genomes should be easy.
STRING: It is a task-oriented system: For a given gene or sequence, it allows to
retrieve information about gene interaction, neighborhood, co-occurrence, co-
expression, etc. Thus incorporation of new tools or genomes may need substantial
programming.
PLATCOM: It is designed to compare genomes on the sequence analysis level, so
a limited number of tools are provided. If tools are consistent with the system
design, it is easy to incorporate them by providing wrappers. If not, it will require
substantial programming work. Incorporation of new genomes is easy by its
design.
Reconfigurability
The SEED: The system can be reconfigured relatively easily due to the command-
line environment.
DAS: Due to its object-oriented design, DAS should be reconfigurable. However,
it is probably the most comprehensive annotation system, so reconfiguration may
require substantial programming work.
CaWorkBench2.0: Its Plug&Play design enables the system to be reconfigured
easily.
310 Choi & Kim
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
STRING: The task-specific nature of the system may require substantial program-
ming work.
PLATCOM: The modular design of the system makes re-configurability easy for
sequence analysis.
Genome Selection Flexibility
The SEED: It allows users to select a genome from a list of over 1,000 genomes as
of June 2005.
DAS: It provides an environment for developing an annotation system for a
particular genome. Although the comparison with other genomes can be stored in
the system, it generally does not allow users to compare multiple genomes freely.
CaWorkBench2.0: It is a system providing a platform for annotating a particular
genome, and does not support multiple genome selection.
STRING: This system is not intended for genome annotation or comparison, so
there is no function to select multiple genomes.
PLATCOM: The basic design principle of the system is to allow users to select any
combinations of genomes (312 replicons as of June 2005) freely. It provides a very
flexible genome-level sequence analysis functionality.
Usability
The SEED: It provides both a Web-based interface as well as the UNIX-like
command environment. It is easy to use.
DAS: It is a Web-based distributed server-client system, so it is easy to use.
CaWorkBench2.0: This system focuses much more on the integration of each
module using a console style user interface. It is intuitive and easy to use.
STRING: Its goal is clear and focused, so it does provide an excellent Web interface
to get information about a given gene.
PLATCOM: It provides an easy-to-use genome comparison interface where users
can simply select any combination of genomes and then compare them.
Data Mining
Mining genomes requires data mining tools for genome sequence data. Unfortunately,
the availability of these tools is very limited since most standard data mining algorithms
that are developed in computer science and statistics community are not designed to
handle biological sequences.
Comparative Genome Annotation Systems 311
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
The SEED: The UNIX-like command-line environment allows users to combine
sequence analysis tools with much freedom. In addition, there is a concept, called
subsystems, that allows managing a set of sequences as a single unit. This makes
genome comparison much simpler and manageable.
DAS: The architecture that allows the user to designate a reference server
dynamically and the rest of servers as client servers greatly simplifies user
interaction in a distributed environment. The Web-based system design also
allows the user to navigate and annotate the genome in a convenient way. In
particular, users can access and retrieve multiple levels of sequences and subse-
quences very easily by using a structured URL query.
CaWorkBench2.0: The Plug&Play design of the system makes combination of
analysis tools easy, so the users can navigate genome information easily. This
system can also handle various types of information such as sequence, gene
expression, and genotype data.
STRING: The user can retrieve various information of a given gene in a very
convenient way on the Web. However, handling of a set of genes is quite limited.
PLATCOM: The design philosophy of this system is to simplify complicated
genome comparison using a set of sequence data mining tools such as clustering
or dynamic visualization tools. Since genomes are the basic units, users can explore
the genome relationship easily.
System Integration in a Distributed Environment
The SEED and DAS: These two systems support the information sharing and
synchronization of data among individual biologists who collaborate with distrib-
uted systems via Web interface. A peer-to-peer design (The SEED) and sever-client
system concept (DAS) are implemented respectively.
CaWorkBench2.0, STRING, and PLATCOM: These systems are not developed
for the third-party and distributed annotation. However, these systems also use
various public databases for data mining task, information from different sources
are internally integrated (and sometimes pre-computed) or provided using a URL
link-out to external Web-based genome annotation databases.
A Thought on Further Development
Comparing multiple genomes is quite a difficult task, and most systems have limited
capabilities, especially in performing multiple analyses in series, which will be referred
as multi-step analysis. The main challenge is that there are numerous ways to combine
tools and databases and each multi-step analysis should be provided as a separate
interface. A task-oriented system is sophisticated and information-rich, but it is not
flexible, limiting its service to those that are already provided. Thus the module-based
architecture is preferred when a flexible and reconfigurable system is needed. The
312 Choi & Kim
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
module-based systems try to address this problem by providing a set of library modules
so that users can combine them to provide a new multi-step analysis. However, this
approach has its own disadvantages. This system architecture requires substantial
programming practice by someone who is already familiar with the modules, which means
that the module-based approach is often limited only to bioinformatics experts.
The real challenge is how to provide an environment where users (biologists or medical
scientists) can perform multi-step analysis in a flexible way. Below we propose an
approach to developing such flexible and reconfigurable systems: defining abstract data
types for genome comparison, developing high performance data mining tools, and
designing and implementing a genome analysis language. We will briefly discuss these
issues:
Defining abstract data types for genome comparison: The definition of abstract
data type (ADT) is a set of data values and associated operations that are precisely
specified independent of any particular implementation. In a sense, some systems
reviewed in this chapter already adopted ADT such as subsystems (the SEED),
functional categories (DAS), and sets of sequences (PLATCOM). Use of ADT is
a first step to build a flexible comparative genome annotation system, because this
system architecture makes the system independent of any specific implementation
and so makes the system reconfigurable and flexible.
High performance genome data mining tools: Once the genome data types are
defined, it will be very helpful to develop high performance data mining tools since
well-developed data mining tools can simplify complicated genome comparison. It
should be also highly scalable due to the size of genomes.
Language for genome comparison: Given the genome data type and data mining
tools that are defined on top of the data type, it will be possible to design a genome
comparison language. At this level, we may consider to develop a genome algebra
and a genome analysis language. For example, if we define set operators, such as
union, intersection, and difference of genomes, we can easily compute a set of
unique genes in a certain organism compared to a set of particular genomes.
We believe that the systems we surveyed in this chapter have explored these concepts
implicitly or explicitly. Now it is the time to develop a genome comparison system in a more
formal way, incorporating ideas from many existing systems.
Acknowledgments
This work is partially supported by the National Science Foundation (USA) Career DBI-
0237901, INGEN (Indiana Genomics Initiative), and National Science Foundation (USA)
0116050. We thank Junguk Hur for his careful reading of the manuscript and suggestions.
Comparative Genome Annotation Systems 313
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
References
Calabrese, P., Chakravarty, S., & Vision, T.J. (2003). Fast identification and statistical
evaluation of segmental homologies in comparative maps. Bioinformatics, 19, 74-
80.
Choi, J.-H., Cho, H.-G., & Kim, S. (2005). A simple and efficient alignment method for
microbial whole genomes using maximal exact match filtering. Computational
Biology and Chemistry, 29(3), 244-253.
Choi, J.-H., Choi, K., Cho, H.-G., & Kim, S. (2005). Multiple genome alignment by
clustering pairwise matches. Proceedings of the 2nd RECOMB Comparative
Genomics Satellite Workshop, Lecture Notes in Bioinformatics, Bertinoro, Italy,
3388. Berlin: Springer-Verlag.
Choi, K., Ma, Y., Choi, J.-H., & Kim, S.(2005). PLATCOM: A Platform for Computational
Comparative Genomics. Bioinfomatics, 21(10), 2514-6.
Dowel, R. D., Jokerst, R. M., Eddy, S. R., & Stein, L. (2001). The distributed annotation
system. BMC Bioinformatics, 2(7).
Gilbert, D. G. (2002). euGene: A eukaryote genome information system. Nucleic Acids
Research, 30, 145-148.
Kim, S. (2003). Graph theoretic sequence clustering algorithms and their applications to
genome comparison. In C. H. Wu, P. Wang, & J. T. L. Wang, (Eds.), Computational
biology and genome informatics (Chapter 4). World Scientific.
Kim, S., Choi, J.-H., & Yang, J. (2005, August). Gene teams with relaxed proximity
constraint. IEEE Computational Systems Bioinformatics (CSB), San Francisco
(pp. 44-55).
Overbeek, R., Disz, T., & Stevens, R. (2004). The SEED: A peer-to-peer environment for
genome annotation. Communications of the ACM, 47, 47-50
Uchiyama, I. (2003). MBGD: Microbial genome database for comparative analysis.
Nucleic Acids Research, 31, 58-62.
von Mering, C., Huynen, M., Jaeggi, D., Schmidt, S., Bork, P. & Snel, B. (2003). STRING:
A database of predicted functional associations between proteins. Nucleic Acids
Research, 31, 258-261.
Endnotes
1
EC (Enzyme Commission) numbers assigned by IUPAC-IUBMB (http://
www.chem.qmw.ac.uk/iupac/jcbn/).
2
There are a number of task oriented systems and we could not include all of them
due to the space limitation.
314 About the Authors
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
About the Authors
Hui-Huang Hsu earned his PhD in electrical engineering from the University of Florida
(USA) in 1994. From 1995 to 2003, he was with Chung Kuo Institute of Technology (1995-
2000), Takming College (2000-2001), and Chinese Culture University (2001-2003), all
located in Taipei, Taiwan. He joined the Department of Computer Science and Information
Engineering, Tamkang University, Taipei, Taiwan in August 2003, where he is now an
associate professor. Dr. Hsu has participated in the organization of several international
conferences and workshops in the areas of information networking, multimedia, and e-
learning. His current research interest includes bioinformatics, multimedia processing,
data mining, and e-learning.
* * *
Tatsuya Akutsu received BEng and MEng in aeronautics and DEng in information
engineering from University of Tokyo (1984, 1986, and 1989, respectively). From 1989 to
1994, he was with the Mechanical Engineering Laboratory, Japan. From 1994 to 1996, he
was an associate professor in the Department of Computer Science at Gunma University.
From 1996 to 2001, he was an associate professor in Human Genome Center, Institute of
Medical Science, University of Tokyo. Since 2001, he has been a professor in the
Bioinformatics Center, Institute for Chemical Research, Kyoto University. His research
interests include bioinformatics and the design and analysis of algorithms.
About the Authors 315
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Marc Anderson is an assistant professor in the Department of Biological Sciences, North
Dakota State University (USA). He is interested in the changes in gene expression and
metabolism that occur in photosynthetic organisms during exposure to low temperature
stress. Specifically, the protective responses of cold acclimation in maize are compared
to the mechanisms of cold adaptation in cryophilic algae in order to identify novel
strategies for improvement of cold tolerance in agriculturally important plants. Dr.
Anderson received his PhD in plant physiology from Iowa State University, Ames, and
his MS in entomology from North Dakota State University, Fargo.
Christopher Besemann received his MS in computer science from North Dakota State
University, Fargo (2005), (USA). Currently, he works in data mining research topics
including association mining and relational data mining, with recent work in model
integration as a research assistant. He is accepted under a fellowship program for PhD
study at NDSU.
Hsuan T. Chang received his BS in electronic engineering from the National Taiwan
Institute of Technology (1991), and MS and PhD degrees in electrical engineering from
National Chung Cheng University, Taiwan (1993 and 1997, respectively). He was an
assistant professor in the Department of Electrical Engineering, National Yunlin Univer-
sity of Science and Technology, Douliu, Taiwan (2001-2002), where he currently is an
associate professor. His interests include image/video processing, optical information
processing/computing, and bioinformatics. He has published more than 100 journal and
conference papers. He has served as the reviewer of several international journals and
conferences, and on session chairs and program committees in domestic and interna-
tional conferences. He is a member of Society of Photo-Optical Instrumentation Engi-
neers (SPIE), Optical Society of America (OSA), Institute of Electrical and Electronic
Engineers (IEEE), International Whos Who (IWW), and The Chinese Image Processing
and Pattern Recognition Society.
Kwangmin Choi received BS in biology from Yonsei University, South Korea, an MA
in molecular genetics and microbiology at the University of Texas at Austin, and an MS
in bioinformatics, at Indiana University, Bloomington (USA), respectively. He worked
at Samsung Medical Center, Seoul, South Korea from 1997 to 1999. Since January 2004,
he is leading the PLATCOM (Platform for Computational Comparative Genomics on the
Web) project.
Francisco M. Couto received his degree in informatics and computer engineering at
Instituto Superior Tcnico da Universidade Tcnica de Lisboa, Portugal, with special-
ization in programming and systems information, and his masters degree in informatics
and computers engineering at the same institution, with a specialization in computer
science. He is currently a lecturer and a PhD student within the Informatics Department
at Faculdade de Cincias da Universidade de Lisboa, Portugal. His research interests
include bioinformatics, text mining, data mining and information systems. He is currently
working on the ReBIL (Relating Biological Information through Literature) project, which
316 About the Authors
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
aims at providing text-mining tools for biological literature that avoid the complex issues
of creating rules and patterns encompassing all possible cases and training sets that are
too specific to be extended to new domains.
Yun-Sheng Chung is a PhD candidate of the Department of Computer Science, National
Tsing Hua University, Hsin-Chu, Taiwan. He received his MS in computer science from
National Tsing Hua University (2004), and bachelors degree in finance from the National
Cheng Chi University, Taipei, Taiwan (2002). His research interests includes algorithms,
combinatorics, computational biology, and macro-informatics. He is currently devoted
to the theoretical foundations regarding data fusion such as properties of mappings
between rank space and score space, and general characterizations of the diversity
between or among scoring systems. Mr. Chung can be contacted by e-mail at
[email protected].
Phuongan Dam is a postdoctoral associate in the Biochemistry and Molecular Biology
Department at the University of Georgia, Athens. She received her PhD in the Cellular
and Molecular Biology Program at the University of Wisconsin at Madison (2002). Her
current research interests include computational inference and modeling of biological
pathways and networks, protein structure prediction and modeling, and large-scale
biological data mining.
Thomas Dandekar was born in 1960. He studied medicine and did his PhD in biochem-
istry. Today he is a professor with the Department of Bioinformatics at the University
of Wrzburg, Germany.
Anne Denton is an assistant professor in computer science at North Dakota State
University (USA). Her research interests are in data mining, bioinformatics, and knowl-
edge discovery in scientific data. Specific interests include data mining of diverse data,
in which objects are characterized by a variety of properties such as numerical and
categorical attributes, graphs, sequences, time-dependent attributes, and others. She
received her PhD in physics from the University of Mainz, Germany, and her MS in
computer science from North Dakota State University, Fargo.
Al Geist is a corporate research fellow at Oak Ridge National Laboratory (ORNL) (USA),
where he leads the 35-member Computer Science Research Group. He leads a national
scalable systems software effort, involving all the DOE and NSF supercomputer sites,
with the goal of defining standardized interfaces between system software components.
Al is a co-principal investigator of the national Genomes to Life (GTL) center. The goal
of this center is to develop new algorithms, and computational infrastructure for
understanding protein machines and regulatory pathways in cells. In his 20 years at
ORNL, he has published two books and more than 190 papers in areas ranging from
heterogeneous distributed computing, numerical linear algebra, parallel computing,
collaboration technologies, solar energy, materials science, biology, and solid state
physics.
About the Authors 317
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Richard Haney, a veteran software developer of Chapel Hill, NC, is a good programmer,
fine statistician, and a great dad.
Grant Heffelfinger, PhD, is a senior manager for Molecular and Computational Bio-
sciences in the Biological and Energy Sciences Center at Sandia National Laboratories
(USA). His graduate research in molecular physics led to a PhD from Cornell University
in 1988. He was appointed to a staff research position at Sandia National Laboratories
at that time where he co-invented Dual Control Volume Grand Canonical Molecular
Dynamics, a method for simulation of molecular phenomena in the presence of chemical
potential gradients such as diffusion through biomembranes. Subsequent responsibili-
ties have included principle authorship and technical leader for Accelerating Biology
with Advanced Algorithms and Massively Parallel Computing, a Cooperative Research
and Development Agreement (CRADA) between Sandia National Laboratories and
Celera Genomics signed January 19, 2001, and principle investigator of Carbon Seques-
tration in Synechococcus Sp.: From Molecular Machines to Hierarchical Modeling, one
of five initial projects funded by the US DOE Office of Sciences Genomes to Life
Program.
D. Frank Hsu is the Clavius Distinguished Professor of Science and a professor of
computer and information science and chair of the department at Fordham University in
New York. He received his PhD in combinatorial mathematics from the University of
Michigan (1979). He has served as visitor and on visiting faculty at Boston University,
JAIST (Kanazawa, Japan) (as Komatsu endowed chair professor), Keio University
(Tokyo, Japan) (as IBM endowed chair professor), MIT, Taiwan University, Tsing Hua
University (Hsin-Chu, Taiwan), and University of Paris-Sud (and CNRS). Dr. Hsus
research interests are in combinatorics, algorithms and optimization, network intercon-
nection and communications, informatics and intelligent systems, and information and
telecommunications technology and infrastructure. He is widely interested in both
microinformatics and macroinformatics with applications to a variety of domains such as
information retrieval, target recognition and tracking, biomedical informatics, and virtual
screening. Dr. Hsu has served on editorial boards of several journals including IEEE
Transactions on Computers, Networks, Inter. J. of Foundation of Computer Science,
Monograph on Combinatorial Optimization, and Journal of Interconnection Networks
(JOIN). He is currently editor-in-chief of JOIN. Dr. Hsu has served on program
committees of several conferences including I-SPAN, DIMACS workshop series, AINA
conferences, and IETC series. He is a senior member of IEEE, a foundation fellow of the
Institute of Combinatorics and Applications, and a fellow of the New York Academy of
Sciences. Dr. Hsu can be contacted by e-mail at [email protected].
Guan-Shieng Huang was born in Taipei in 1972. He obtained both his BS degree in
mathematics and PhD in computer science from the National Taiwan University (1994 and
1999, respectively). Then he joined the Institute of Information Science of Academia
Sinica, Department of Computer Science and Information Engineering of National Taiwan
University, and Laboratory of Research on Informatics of Universite de Paris Sud as a
postdoctoral fellowship (1999, 2000, and 2001, respectively). Afterward he became an
318 About the Authors
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
assistant professor in the Department of Computer Science and Information Engineering
of National Chi Nan University, Taiwan (2002). Dr. Huangs past research interests
include finite mathematics, satisfiability testing, and design and analysis of algorithms.
He currently focuses on computational problems related to bioinformatics.
Ron Hutchisons research involves investigating the changes in gene expression and
photosynthetic responses of cold adapted algae. The focus of this project is to identify
the differentially regulated genes or metabolic pathways that allow these organisms to
live and thrive very near 0
o
C. Dr. Hutchison is currently on the faculty at Stockton College
and began this project while at North Dakota State University. After receiving a PhD
degree in plant biology from the University of Illinois, he completed postdoctoral studies
at Ohio State University and the University of Minnesota. Dr. Hutchison received an
undergraduate degree in biology from Kenyon College.
Ching-Pin Kao was born on 1976 in Tainan, Taiwan, ROC. He received his BS from Yuan
Ze University in 1999 and his MS from the National Cheng Kung University (Taiwan) in
2001, both in computer science and engineering. He is currently a PhD student in
computer science and information engineering at National Cheng Kung University. His
research interests include data mining, clustering techniques, bioinformatics, and gene
expression analysis.
Arpad Kelemen received his PhD at the University of Memphis in 2002 and his MS at the
University of Szeged, Hungary in 1995, both in computer science. Currently he is an
assistant professor with the Department of Computer and Information Sciences, Niagara
University. He is also an adjunct faculty member at the department of Biostatistics, The
State University of New York at Buffalo. He appeared in Marquis Whos Who in Science
and Engineering and in AcademicKeys Whos Who in Health Sciences in 2005. His
research interests include bioinformatics, artificial intelligence, biometrics, machine
learning, neural networks, and cognitive science.
Takashi Kido is a senior researcher of HuBit Genomix Inc., Japan. He is currently working
on SNP analysis for discovering causative genes for common diseases and striving to
realize personal medicine and systematic drug discovery. He is interested in applying
Artificial Intelligence technology into bioinformatics domain. He received his PhD in
computer science from the Keio University, Japan. Before his current position, he had
been studying on as an intelligent agent in the NTT Laboratory in Japan and Malaysia.
Sun Kim received his BS, MS, and PhD degrees in computer science from Seoul National
University, Korea Advanced Institute of Science and Technology (KAIST), and the
University of Iowa respectively. After graduation, he was the director of bioinformatics
and postdoctoral fellow at the Biotechnology Center and a visiting assistant professor
of animal sciences at the University of Illinois at Urbana-Champaign from 1997 to 1998.
He then joined DuPont Central Research in 1998 and worked as a senior computer
About the Authors 319
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
scientist until August 2001. He is currently associate director of the Bioinformatics
Program, assistant professor of informatics at Indiana University, Bloomington (USA).
Sun Kim is a recipient of the Outstanding Junior Faculty Award at Indiana University
2004-2005, NSF CAREER Award DBI-0237901 from 2003 to 2008, and Achievement Award
at DuPont Central Research in 2000.
Bruce S. Kristal received a BS in life sciences from the Massachusetts Institute of
Technology (1986) and his PhD from Harvard University Graduate School of Arts and
Sciences, Division of Medical Sciences, Committee on Virology (1991). After being a
post-doctoral fellow, instructor, and research assistant professor at the University of
Texas Health Science Center at San Antonio (1991-1996), he joined the Dementia
Research Service of Burke Medical Research Institute in 1996 and subsequently the
Departments of Biochemistry (1997) and Neuroscience (1998) at Weill Medical College
of Cornell University as an Assistant Professor (Associate Professor in 2004). Dr.
Kristals research has been recognized by grant/meeting awards from the National
Institutes of Health, American Federation for Aging Research, Will Rogers Society,
Oxygen Society, American Institute for Cancer Research, Gordon Research Conferences,
and the Hereditary Disease Foundation. He has taught high data density informatics in
the short course in bioinformatics at the annual Society for Neuroscience meeting and
at the Research Update in Neuroscience for Neurosurgeons course. Dr. Kristal is the first
secretary and a member of the Board of Directors of the newly formed Metabolomics
Society. He is a contributing editor for Sciences Science of Aging Knowledge Environ-
ment. He serves as an ad hoc reviewer for NIH and other granting agencies.
Simon Lin is currently an associate director of bioinformatics at Northwestern Univer-
sity (USA). Besides teaching in proteomics and microarrays, his research focuses on
genomics data mining and medical applications. He lectured at Duke University and
managed the Duke Bioinformatics Shared Resource from 1999 to 2004. He co-chaired the
Critical Assessment of Microarray Data Analysis (CAMDA) Conference, and edited four
volumes of Methods in Microarray Data Analysis (Kluwer).
Li Liao is an assistant professor in the Department of Computer and Information
Sciences, University of Delaware (USA). His research interests and experience span a
wide range, including computer simulation of molecular systems, genome sequencing,
protein homology detection, and comparative genomics. He has published over 25
scientific papers and co-authored one book. He received his PhD from Peking University
and has since been on the faculty there before moving to the U.S. Prior to his current
position, he was a senior research scientist at DuPont Company. He is a member of the
ACM and the International Society for Computational Biology.
Shuang-Te Liao received his MS from the Department of Computer Science and
Information Engineering, Ming Chuan University, Taoyuan, Taiwan (2004). His research
interests include pattern recognition, machine learning, and bioinformatics.
320 About the Authors
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Yulan Liang received her masters and PhD degrees in applied statistics from the
University of Memphis. She joined the Department of Biostatistics, The State University
of New York at Buffalo (USA) as an assistant professor (2002). Her research interests
are in statistical genetics, bioinformatics, statistical pattern recognition and statistical
learning theory. She is also interested in multivariate statistics, Bayesian inference, data
mining, and neural networks. She was the president of the ASA Buffalo-Niagara chapter
in 2004. She is also a member of the American Statistical Association, the International
Biometric Society, the Bayesian Society, and the American Society of Human Genetics.
Patrick McConnells background is in neurosciences and computer science. He has
been working in bioinformatics for the last eight years. He spent his undergraduate time
working with the bioinformatics resource at Emory, where he spent time developing
bioinformatics training material and molecular modeling software. For the last four years,
he has been with the bioinformatics resource at the Duke Comprehensive Cancer Center,
where he developed applications for microarray and proteomics analysis. His main
interests lie with grid computing and data integration, and he is currently working on the
Cancer Biomedical Informatics Grid (caBIG
TM
) project.
Salvatore Mungal received a BS in biochemistry in 1984 and a MS in oral biology and
pathology in 1992 both from SUNY Stony Brook. He spent 18 years as a researcher in the
Life Sciences and eight years as a software engineer. He has authored and co-authored
research articles, abstracts and posters and has given many presentations. His interest
in computational biology has drawn him into the bioinformatics arena where he spends
his time developing software applications on a distributed grid for medical and research
domain experts in proteomics for the caBIG
TM
(Cancer Biomedical Informatics Grid)
project.
Tobias Mller was born in 1967. He studied mathematics and did his PhD in bioinformatics.
Today he is a lecturer at the Department of Bioinformatics at the University of Wrzburg,
Germany.
Chongle Pan is a PhD student in Genome Science and Technology Graduate Program
jointly offered by the University of Tennessee (UT) (USA) and Oak Ridge National
Laboratory (ORNL). He is mentored by Drs. Nagiza F. Samatova and Robert L. Hettich.
He has published in the areas of nucleic acid crystallography, intact proteins character-
ization with Fourier-transform ion cyclotron resonance mass spectrometry, and algo-
rithms development for tandem mass spectral data analysis. His current research
interests include prediction of the residue-residue contact with the co-evolution hypoth-
esis, development of analytical and computational tools for quantitative proteomics, and
characterization of Rhodopseudomonas palustris with its protein expression data.
Byung-Hoon Park, PhD, is a research scientist for the Computational Biology Institute,
Computer Science and Mathematics Division at Oak Ridge National Laboratory (ORNL).
About the Authors 321
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
He received his PhD in computer science from Washington State University in 2001. He
served in program committees of various data mining conferences, and was a guest editor
for a special issue on distributed and mobile data mining in the IEEE Transactions on
Systems, Man, Cybernetics, Part B. His research interests in bioinformatics include
computational modeling of protein-protein interactions and protein interface sites,
biological text mining using Markovian probabilistic approaches, and inference of gene
regulatory networks.
Edward F. Patz, Jr. is the James and Alice Chen professor in radiology, and a professor
of pharmacology and cancer biology at Duke University Medical Center (USA). He
currently is the director of a Molecular Diagnostics Laboratory with a focus on devel-
oping novel tumor imaging agents and protein biomarkers.
Sven Rahmann was born in 1974. He received his MS in mathematics at Heidelberg
University and his PhD in bioinformatics at Freie Universitt Berlin. Today he works at
the Faculty of Technology and leads the Algorithms and Statistics for Systems Biology
Group at the Center for Biotechnology at Bielefeld University in Germany.
Nagiza F. Samatova is a senior staff scientist for Computational Biology Institute,
Computer Science and Mathematics Division at Oak Ridge National Laboratory (ORNL)
(USA), where she leads a team of 15 researchers. She received her BS in applied
mathematics from Tashkent State University, Uzbekistan (1991) and her PhD in math-
ematics from the Russian Academy of Sciences, Moscow (1993) (supervised by Prof.
Pulatov). She also received an MS degree in computer science from the University of
Tennessee, Knoxville in 1998. She currently co-leads several multi-institutional research
projects aiming to develop bioinformatics tools for discovery and characterization of
biomolecular machines, inference and simulation of gene networks, and integration for
query and analysis of heterogeneous and dispersed biological data. She is the author of
over 50 publications, one book, and two patents.
Shyong-Jian Shyu received his BS degree in computer engineering from the National
Chiao Tung University in 1985, his MS degree in computer and decision sciences in 1987,
and his PhD degree in computer science in 1991 from the National Tsing Hua University,
Taiwan. From 1993 to 1994, he worked as a researcher at Academia Sinica Computer
Centre, Taiwan. Currently, he is a professor of the Department of Computer Science and
Information Engineering at Ming Chuan University, Taiwan. His research interests
include the design and analysis of algorithms, parallel computing, visual cryptography,
and bioinformatics.
Mrio J. Silva has a PhD in electrical engineering and computer science from the
University of California, Berkeley (1994). He held several industrial positions both in
Portugal and the U.S. He joined the Faculdade de Cincias da Universidade de Lisboa,
Portugal, in 1996, where he now leads a research group in data management at the Large-
322 About the Authors
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Scale Information Systems Laboratory (LASIGE). His research interests are in informa-
tion integration, information retrieval, and bioinformatics.
Vincent S. Tseng received his PhD in computer and information science from National
Chiao Tung University in 1997. From January 1998 and July 1999, he was an invited
postdoctoral research fellow in the Computer Science Division of University of Califor-
nia, Berkeley, Since August 1999, he has been on the faculty of the Department of
Computer Science and Information Engineering at National Cheng Kung University,
Taiwan. He has also been the director for Department of Medical Informatics at National
Cheng Kung University Hospital, Taiwan since February 2004. Dr. Tseng has a wide
variety of research specialties covering data mining, biomedical informatics, and multi-
media databases.
Matthias Wolf was born in 1972. He studied biology and did his PhD in phylogenetics.
Currently, he is an assistant professor at the Department of Bioinformatics at the
University of Wrzburg, Germany.
Ying Xu is an endowed chair professor of bioinformatics and computational biology in
the Biochemistry and Molecular Biology Department, and the director of the Institute of
Bioinformatics, University of Georgia (UGA), USA. Before joining UGA in 2003, he was
a senior staff scientist and group leader at Oak Ridge National Laboratory (ORNL), where
he still holds a joint position. He received his PhD in theoretical computer science from
the University of Colorado at Boulder (1991). His PhD thesis work was on development
of efficient algorithms for matroid intersection problems (supervised by Hal Gabow).
Between 1991 and 1993, he was a visiting assistant professor at the Colorado School of
Mines. He started his bioinformatics career in 1993 when he joined Ed Uberbachers
group at ORNL to work on the GRAIL project. His current research interests include (a)
computational inference and modeling of biological pathways and networks, (b) protein
structure prediction and modeling, (c) large-scale biological data mining, and (d)
microbial & cancer bioinformatics. He is interested in both bioinformatics tool develop-
ment and study of biological problems using in silico approaches.
Ajay Yekkirala is currently finishing an MS degree in biology from North Dakota State
University (USA). His research focus is in the field of functional genomics and
bioinformatics.
Peng-Yeng Yin received his BS, MS and PhD degrees in computer science from National
Chiao Tung University, Hsinchu, Taiwan. From 1993 to 1994 he was a visiting scholar at
the Department of Electrical Engineering, University of Maryland, College Park, and the
Department of Radiology, Georgetown University, Washington, DC. In 2000, he was a
visiting professor at the Department of Electrical Engineering, University of California,
Riverside. He is currently a professor and chairman of the Department of Information
Management, National Chi Nan University, Nantou, Taiwan. Dr. Yin received the
About the Authors 323
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Overseas Research Fellowship from Ministry of Education in 1993, and the Overseas
Research Fellowship from National Science Council in 2000. He has received the best
paper award from the Image Processing and Pattern Recognition Society. He is a member
of the Phi Tau Phi Scholastic Honor Society and listed in Whos Who in the World. His
current research interests include content-based image retrieval, machine learning,
computational intelligence, and computational biology.
Junying Zhang received her PhD degree in signal and information processing from
Xidian University, Xian, China (1998). From 2001 to 2002 and in 2004, she was a visiting
scholar in the U.S. and in Hong Kong, respectively. She is currently a professor in the
School of Computer Science and Engineering, and a part-time professor in the Institute
of Electronics Engineering, Xidian University. Her research interests focus on intelligent
information processing, including cancer-related bioinformatics and biochip data analy-
sis, machine learning and its application to cancer research, image processing and pattern
recognition, as well as automatic target recognition.
324 Index
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
Index
Symbols
2-dimensional (2-D) plots 88
A
a posteriori criterion 236
a priori 233
ab initio 11, 119
AbXtract 288
accuracy measures 240
affinity analysis 6
algorithms 3, 64, 221
alignment 119
alignment-less alignment 233
antecedent 238
artifacts 175
artificial intelligence (AI) 9
association 6
association rule 6, 235
association rule mining 271
autocorrelation 271
B
baseline correction 95
basic ICA model 196
Bayes theorem 8
Bayesian classifier 8
binning 97
BioCarta 304
bioinformatics 64, 284
bioinoformatics 15
BioLiterature 284
BioRAT 288
BLAST 16, 303
blind source separation 209
BLOCKS 16
bootstrap 105, 110
branch-and-bound 120
C
c-space feature selection 204
caBIO 304
cancer classification 175
case-based 286
CASP 133
caWorkBench2 303
Cayley network 36, 45
central limit theorem 196
chaos game representation 76
characterization 7
classification 5
cluster analysis of gene expression
(CAGED) 140
Index 325
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
clustering 5, 99, 145, 156
clustering algorithms 156, 231
clustering Methods 157
COG-mode 305
combinatorial algorithm 47
combinatorial fusion 34, 36
combinatorial fusion analysis (CFA) 32,
52
commercial databases 93
comparative genome annotation 297
comparative genomics 17, 296
computation reduction 162
computational algorithms 218
confidence 6
consensus scoring 33, 52, 53
constantly expressed genes (CEG) 175
contact map overlap (CMO) 132
core region 125
correction formulas 108
correlation search technique (CST) 161
CST Algorithm 164
curse of dimensionality 176
cycle representation 45
Cytoscape 304
D
DAS XML 302
DasView 303
data 87
data acquisition 89
data collection 4
data grid 36
data mining 2, 4, 285, 298, 300
data mining process 3
data mining tasks 5
data mining techniques 35, 269
data parsing 91
data preprocessing 4
data processing 89
data sources 299
data space 36
data warehousing 8
databases 3
DAVID-KEGG Analysis 148
decision trees (DT) 9
deviation analysis 5
differential ARM 273
differential association rule notation 274
differential association rules 269
diplotype analysis 227
disease association 214
distance estimation 109
distributed annotation system (DAS)
300
diversity rank function 49
diversity rank/score graph 49
diversity score function 48
divide-and-conquer 120
DNA microarray profiles 177
DNA microarray technology 176
DNA walks 68
domain-domain interactions 256
double dynamic programming 124
drug discovery 52
dynamic programming 121
E
efficiency 40
EMBL nucleotide sequence database 3
eMotif 16
European ribosomal rRNA database
112
evolutional approaches 47
evolutionary distances 107
evolutionary Markov process (EMP)
108
example of permutations 45
expectation maximization (EM) algo-
rithms 9
expression analysis systematic ex-
plorer (EASE) 147
extended phylogenetic profiles 26
extensibility 297
F
FASTA 16, 302
feature analysis 198
feature extraction 97
features 233
FiGO 291
figure of merit (FOM) 158
Fisher criterion 185
326 Index
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
FlyBase 300
fourier-transform ion cyclotron reso-
nance (FTICR) 87
frequent 237
frequent substructure ARM 272
frozen approximation 124
functional annotations 15
fuzzy systems 9
G
gap symbols 120
GemDOCK 53
GenBank 3
gene annotation 10
gene expression 156
gene expression profiles 175
gene fusion analysis 308
gene neighborhood analysis 307
gene ontology 288
gene space (G-space) 190
GeneRIF database 291
genetic algorithms (GA) 9
genome annotation 297
genome Reviews 3
genome selection flexibilit 297
genotype data 224
GFF 302
GOAnnotator 290
GOLD 53
goodness-of-hit (GH) 43
graph-based clustering 234, 242
graphical representation 64
H
h-curves 66
haplotype analysis 214
haplotyping 221
hierarchal profilings 15
hierarchical clustering 6
hierarchical merging 234
hierarchical merging method 241
hierarchical profile 15
histogram entropy 78
homology modeling 119
Huberts G (gamma) statistic 158
human proteome organization (HUPO)
86
hybrid clustering (C/C) 145
hybrid methods 138
I
in silico approaches 250
independent component analysis (ICA)
176, 194
independent subspace analysis (ISA)
195
information fusion 33
information retrieval 3, 33, 34, 51
inheritance pattern 25
Institute for System Biology (ISB) 89
instrumentation 89
integer program (IP) 127
integrated-condensed (I/C) 138
intensity image 160
interface sites prediction 253
introns 11
iterated function system (IFS) 76
iterative partitioning 234
J
Jaccard coefficient 158
joined-relation basis 274
jointly discriminatory genes (JDG) 175
K
k-means 234
k-means algorithm 6, 241
KEGG database 300
kernel functions 25
knowledge discovery in databases
(KDD) 2
L
laser capture microdissection (LCM)
194
LD block 218, 222
LDMiner 220
linear combination 46
linear programming 120
linear regression 9
Index 327
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
link analysis 6
link analysis or affinity analysis 6
linkage disequilibrium 217
liquid chromatography (LC) data 88
local evidence context (LEC) 291
Log Likelihood Test (LRT) 223
Long Branch Attraction (LBA) 114
longest arc-preserving common subse-
quence problem 130
lung cancer 93
M
m/z data 88
m/z values 86
machine learning 3, 9, 35
market basket analysis 6
mass spectrometry 85
master tree 19
maximum-likelihood 105
MEDLINE 284
MeKE 290
MEME 16
metabolic pathway analysis 308
metabolic pathways 11
MHC-Peptide Interaction Database 252
microarray 156
microarray and gene expression data
ontology (MGEDO) 86
microarray profiling 175
minimum support 237
morphology analysis 285
multi-locus haplotypes 223
multiplayer perceptron (MLP) 189
multiple approaches 34
multiple genome comparison 307
multiple sclerosis (MS) 139
multiple scoring systems 33, 36
mzXML 93
N
naphora (or co-reference) resolution
286
National Cancer Institute Center for
Bioinformatic 303
natural language processing (NLP) 7,
285
negative cases 52
neighbor-joining 105
network comparison 275
network topologies 261
neural networks (NN) 9
neutrality 40
non-negative matrix factorization (NMF)
176, 195
normalization 96
NP-hard 123
nucleotide sequence database 3
O
open source 93
optimal alignment 120
P
P-Tree approach 19
pairwise comparison database (PCDB)
307
parallel processing 3
paralogous verification method (PVM)
258
parsing 286
part-of-speech tagging 285
partial volume correction (PVC) 205
partially ordered set (Poset) 47
pattern differentiations 137
pattern formulations 137
pattern recognition 33
PAUP* package 106
PDZBase 252
peak finding 98
pearson correlation coefficient (PCC)
140
peer-to-peer (P2P) infrastructure 299
phylogenetic clustering 65
phylogenetic profiles 15
phylogenetic tree 11
phylogeny 104
PICA model 196
PINdb 252
PLATCOM 306
platform for computational comparative
genomics on 306
POINT 252
328 Index
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
polysyms 287
positive cases 52
predicting protein-protein interactions
254
prediction 5
prediction of protein functions 259
preprocessing 94
primary structure 10, 232
principal component analysis (PCA)
137, 178
probability Method 46
profile distances 114
profile neighbor-joining (PNJ) 107, 114
profile similarity 14
profile-profile 16
ProNIT 252
properties of interface sites 253
PROSPECT 125
protein 231
protein data bank (PDB) 3
protein function prediction 260
protein identification 87
protein interaction 270
protein interaction classification by
unlikely pro 256
protein interaction network 11, 259
protein profiling problems 87
protein structure 118
protein structure alignment 120
protein structure prediction 34, 54, 119
protein threading 119
protein threading with constraints 120
protein-clustering 231
proteomics 85
proteomics standards initiative (PSI) 86
PSI markup language (PSI-ML) 86
PSI ontology (PSI-Ont) 86
PSI-BLAST 16
Q
quasi-hyperplanes 181
quaternary structure 10
R
radical basis function network (RBFN)
55
Randics visualization methods 72
rank and score combination 41
rank combination 32, 42
rank function 32, 39
rank Statistics 46
rank/score function 32, 39
rank/score graph 33, 39
rank/score transfer 49, 50
ranking methods 185
RAPTOR 125
rate matrix 108
raw data 94
RCA (rank correlation analysis) 46
real microarray data 193
reconfigurability 297
regression 5, 9
relational ARM 271
reproducibility 138
RNA secondary structures 120
rule-based 286
S
Scalar measurements 158
score combination 32
score function 32, 39, 120
score matrix 49
secondary structure 10
secondary structure prediction 119
SEED 298
segmentation 5
sequence apriori algorithm 236
sequence clustering 308
sequence visualization methods 65
signal processing 98
significant analysis of microarray (SAM)
138
simple matching coefficient 158
single nucleotide polymorphisms
(SNPs) 215
single-locus analysis 224
singular value decomposition (SVD)
137
Smart-Cast 161, 167
smoothing 96
soft independent modeling of class
analogy (SIMCA) 35
Index 329
Copyright 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written
permission of Idea Group Inc. is prohibited.
spotted cDNA microarray 175
standard ARM 273
standard representation 45
standard request format 302
starting distribution 108
statistics 3, 8
STRING Web interface 304
structure-based virtual screening 52
summarization 7
supervised learning 5, 9
supervised machine learning methods
99
support 6
support vector machines (SVM) 9, 24,
189
SVM-Fisher 16
symmetric group 36
synonyms 287
system architecture 299
system integration in a distributed
environment 298
T
tertiary structure 10
text mining 7, 284
three-step method (TSM) 186
time-of-flight 87
Tokenization 285
topographic ICA (TICA) 195
tree distance 22
tree kernel approach 24, 27
U
unstable expressed genes (UEG) 175
unsupervised learning 5, 9
usability 297
V
virtual screening (VS) 35, 52
visualization 4, 41, 64
voting 46
W
Web content mining 7
Y
YCSKH-study 53, 54
Z
z-curves 70
I
n
t
r
o
d
u
c
i
n
g
I
n
t
r
o
d
u
c
i
n
g
The new IGI Teaching Case Collection is a full-text database contain-
ing hundreds of teaching cases related to the fields of information science,
technology, and management.
Key Features
Project background
information
Searches by keywords and
categories
Abstracts and citation
information
Full-text copies available for
each case
All cases are available in PDF
format
Cases are written by IT
educators, researchers, and
professionals worldwide
The Benefits of the IGI Teaching Case Collection
Frequent updates as new cases are available
Instant access to all full-text articles saves research time
No longer necessary to purchase individual cases
A Product Of
www.igi-online.com
For More Information Visit
www.igi-online.com
IGI Teaching Case Collection
Recommend to your librarian today!
View each case in full-text, PDF form.
Hundreds of cases provide a real-world
edge in information technology classes or
research!
InfoSci-Online
Experience the latest full-text research in the fields
of Information Science, Technology & Management
infosci-online.com
A PRODUCT OF
Publishers of Idea Group Publishing, Information Science Publishing, CyberTech Publishing, and IRM Press
The theoretical bent
of many of the titles
covered, and the ease
of adding chapters to
reading lists, makes it
particularly good for
institutions with strong
information science
curricula.
Issues in Science and
Technology Librarianship
To receive your free 30-day trial access subscription contact:
Andrew Bundy
Email: [email protected] Phone: 717/533-8845 x29
Web Address: www.infosci-online.com
InfoSci-Online is available to libraries to help keep students,
faculty and researchers up-to-date with the latest research in
the ever-growing field of information science, technology, and
management.
The InfoSci-Online collection includes:
Scholarly and scientific book chapters
Peer-reviewed journal articles
Comprehensive teaching cases
Conference proceeding papers
All entries have abstracts and citation information
The full text of every entry is downloadable in .pdf format
Some topics covered:
Business Management
Computer Science
Education Technologies
Electronic Commerce
Environmental IS
Healthcare Information Systems
Information Systems
Library Science
Multimedia Information Systems
Public Information Systems
Social Science and Technologies
InfoSci-Online
features:
Easy-to-use
6,000+ full-text
entries
Aggregated
Multi-user access
Single Journal Articles and Case Studies Are
Now Right at Your Fingertips!
www.idea-group.com
Business Process Reengineering Data and Database Management
Distance Learning E-commerce
Emerging and Innovative Technologies End User Computing
Healthcare Human Side of IT
Information Resource Management Internet-Based Technologies
IS/IT Planning IT Education
IT Management Knowledge Management
Organization Politics and Culture Software Engineering Tools
Systems Planning Decision Support Systems
Telecommunication and Networking Virtual Ofces
Client Server Technology Strategic Information Systems
Design, Implementation
You can now view the table of contents for each journal so it is easier to
locate and purchase one specic article from the journal of your choice.
Case studies are also available through XanEdu, to start building your
perfect coursepack, please visit www.xanedu.com.
For more information, contact [email protected] or 717-533-8845
ext. 10.
Purchase any single journal article or
teaching case for only $18.00!
Idea Group Publishing offers an extensive collection of research articles and
teaching cases that are available for electronic purchase by visiting
www.idea-group.com/articles. You will nd over 980 journal articles and
over 275 case studies from over 20 journals available for only $18.00. The
website also offers a new capability of searching journal articles and case
studies by category. To take advantage of this new feature, please use the
link above to search within these available categories: