Probabilistic Modeling in Bioinformatics and Medical

Informatics 1st Edition

Visit the link below to download the full version of this book:

https://medidownload.com/product/probabilistic-modeling-in-bioinformatics-and-me
dical-informatics-1st-edition/

Click Download Now

Preface

We are drowning in information,

but starved of knowledge.
– John Naisbitt, Megatrends

The turn of the millennium has been described as the dawn of a new scientific
revolution, which will have as great an impact on society as the industrial and
computer revolutions before. This revolution was heralded by a large-scale
DNA sequencing effort in July 1995, when the entire 1.8 million base pairs
of the genome of the bacterium Haemophilus influenzae was published – the
first of a free-living organism. Since then, the amount of DNA sequence data
in publicly accessible data bases has been growing exponentially, including a
working draft of the complete 3.3 billion base-pair DNA sequence of the entire
human genome, as pre-released by an international consortium of 16 institutes
on June 26, 2000.
Besides genomic sequences, new experimental technologies in molecu-
lar biology, like microarrays, have resulted in a rich abundance of further
data, related to the transcriptome, the spliceosome, the proteome, and the
metabolome. This explosion of the “omes” has led to a paradigm shift in
molecular biology. While pre-genomic biology followed a hypothesis-driven
reductionist approach, applying mainly qualitative methods to small, isolated
systems, modern post-genomic molecular biology takes a holistic, systems-
based approach, which is data-driven and increasingly relies on quantitative
methods. Consequently, in the last decade, the new scientific discipline of
bioinformatics has emerged in an attempt to interpret the increasing amount
of molecular biological data. The problems faced are essentially statistical,
due to the inherent complexity and stochasticity of biological systems, the
random processes intrinsic to evolution, and the unavoidable error-proneness
and variability of measurements in large-scale experimental procedures.
vi Preface

Since we lack a comprehensive theory of life’s organization at the molecular

level, our task is to learn the theory by induction, that is, to extract patterns
from large amounts of noisy data through a process of statistical inference
based on model fitting and learning from examples.
Medical informatics is the study, development, and implementation of al-
gorithms and systems to improve communication, understanding, and man-
agement of medical knowledge and data. It is a multi-disciplinary science
at the junction of medicine, mathematics, logic, and information technology,
which exists to improve the quality of health care.
In the 1970s, only a few computer-based systems were integrated with hos-
pital information. Today, computerized medical-record systems are the norm
within the developed countries. These systems enable fast retrieval of patient
data; however, for many years, there has been interest in providing additional
decision support through the introduction of knowledge-based systems and
statistical systems.
A problem with most of the early clinically-oriented knowledge-based sys-
tems was the adoption of ad hoc rules of inference, such as the use of certainty
factors by MYCIN. Another problem was the so-called knowledge-acquisition
bottleneck, which referred to the time-consuming process of eliciting knowl-
edge from domain experts. The renaissance in neural computation in the
1980s provided a purely data-based approach to probabilistic decision sup-
port, which circumvented the need for knowledge acquisition and augmented
the repertoire of traditional statistical techniques for creating probabilistic
models.
The 1990s saw the maturity of Bayesian networks. These networks pro-
vide a sound probabilistic framework for the development of medical decision-
support systems from knowledge, from data, or from a combination of the two;
consequently, they have become the focal point for many research groups con-
cerned with medical informatics.
As far as the methodology is concerned, the focus in this book is on proba-
bilistic graphical models and Bayesian networks. Many of the earlier methods
of data analysis, both in bioinformatics and in medical informatics, were quite
ad hoc. In recent years, however, substantial progress has been made in our
understanding of and experience with probabilistic modelling. Inference, de-
cision making, and hypothesis testing can all be achieved if we have access to
conditional probabilities. In real-world scenarios, however, it may not be clear
what the conditional relationships are between variables that are connected in
some way. Bayesian networks are a mixture of graph theory and probability
theory and offer an elegant formalism in which problems can be portrayed
and conditional relationships evaluated. Graph theory provides a framework
to represent complex structures of highly-interacting sets of variables. Proba-
bility theory provides a method to infer these structures from observations or
measurements in the presence of noise and uncertainty. This method allows
a system of interacting quantities to be visualized as being composed of sim-
 Preface vii

pler subsystems, which improves model transparency and facilitates system

interpretation and comprehension.
Many problems in computational molecular biology, bioinformatics, and
medical informatics can be treated as particular instances of the general prob-
lem of learning Bayesian networks from data, including such diverse problems
as DNA sequence alignment, phylogenetic analysis, reverse engineering of ge-
netic networks, respiration analysis, Brain-Computer Interfacing and human
sleep-stage classification as well as drug discovery.

Organization of This Book

The first part of this book provides a brief yet self-contained introduction to
the methodology of Bayesian networks. The following parts demonstrate how
these methods are applied in bioinformatics and medical informatics.
This book is by no means comprehensive. All three fields – the methodol-
ogy of probabilistic modeling, bioinformatics, and medical informatics – are
evolving very quickly. The text should therefore be seen as an introduction,
offering both elementary tutorials as well as more advanced applications and
case studies.
The first part introduces the methodology of statistical inference and prob-
abilistic modelling. Chapter 1 compares the two principle paradigms of statis-
tical inference: the frequentist versus the Bayesian approach. Chapter 2 pro-
vides a brief introduction to learning Bayesian networks from data. Chapter 3
interprets the methodology of feed-forward neural networks in a probabilistic
framework.
The second part describes how probabilistic modelling is applied to bioin-
formatics. Chapter 4 provides a self-contained introduction to molecular phy-
logenetic analysis, based on DNA sequence alignments, and it discusses the
advantages of a probabilistic approach over earlier algorithmic methods. Chap-
ter 5 describes how the probabilistic phylogenetic methods of Chapter 4 can
be applied to detect interspecific recombination between bacteria and viruses
from DNA sequence alignments. Chapter 6 generalizes and extends the stan-
dard phylogenetic methods for DNA so as to apply them to RNA sequence
alignments. Chapter 7 introduces the reader to microarrays and gene expres-
sion data and provides an overview of standard statistical pre-processing pro-
cedures for image processing and data normalization. Chapters 8 and 9 address
the challenging task of reverse-engineering genetic networks from microarray
gene expression data using dynamical Bayesian networks and state-space mod-
els.
The third part provides examples of how probabilistic models are applied
in medical informatics.
Chapter 10 illustrates the wide range of techniques that can be used to
develop probabilistic models for medical informatics, which include logistic
regression, neural networks, Bayesian networks, and class-probability trees.
viii Preface

The examples are supported with relevant theory, and the chapter emphasizes
the Bayesian approach to probabilistic modeling.
Chapter 11 discusses Bayesian models of groups of individuals who may
have taken several drug doses at various times throughout the course of a
clinical trial. The Bayesian approach helps the derivation of predictive distri-
butions that contribute to the optimization of treatments for different target
populations.
Variable selection is a common problem in regression, including neural-
network development. Chapter 12 demonstrates how Automatic Relevance
Determination, a Bayesian technique, successfully dealt with this problem for
the diagnosis of heart arrhythmia and the prognosis of lupus.
The development of a classifier is usually preceded by some form of data
preprocessing. In the Bayesian framework, the preprocessing stage and the
classifier-development stage are handled separately; however, Chapter 13 in-
troduces an approach that combines the two in a Bayesian setting. The ap-
proach is applied to the classification of electroencephalogram data.
There is growing interest in the application of the variational method to
model development, and Chapter 14 discusses the application of this emerging
technique to the development of hidden Markov models for biosignal analysis.
Chapter 15 describes the Treat decision-support system for the selection
of appropriate antibiotic therapy, a common problem in clinical microbiol-
ogy. Bayesian networks proved to be particularly effective at modelling this
problem task.
The medical-informatics part of the book ends with Chapter 16, a descrip-
tion of several software packages for model development. The chapter includes
example codes to illustrate how some of these packages can be used.
Finally, an appendix explains the conventions and notation used through-
out the book.

Intended Audience
The book has been written for researchers and students in statistics, machine
learning, and the biological sciences. While the chapters in Parts II and III
describe applications at the level of current cutting-edge research, the chapters
in Part I provide a more general introduction to the methodology for the
benefit of students and researchers from the biological sciences.
Chapters 1, 2, 4, 5, and 8 are based on a series of lectures given at the
Statistics Department of Dortmund University (Germany) between 2001 and
2003, at Indiana University School of Medicine (USA) in July 2002, and at
the “International School on Computational Biology”, in Le Havre (France)
in October 2002.
 Preface ix

Website
The website
http://robots.ox.ac.uk/∼parg/pmbmi.html
complements this book. The site contains links to relevant software, data,
discussion groups, and other useful sites. It also contains colored versions of
some of the figures within this book.

Acknowledgments

This book was put together with the generous support of many people.
Stephen Roberts would like to thank Peter Sykacek, Iead Rezek and
Richard Everson for their help towards this book. Particular thanks, with
much love, go to Clare Waterstone.
Richard Dybowski expresses his thanks to his parents, Victoria and Henry,
for their unfailing support of his endeavors, and to Wray Buntine, Paulo Lis-
boa, Ian Nabney, and Peter Weller for critical feedback on Chapters 3, 10,
and 16.
Dirk Husmeier is most grateful to David Allcroft, Lynn Broadfoot, Thorsten
Forster, Vivek Gowri-Shankar, Isabelle Grimmenstein, Marco Grzegorczyk,
Anja von Heydebreck, Florian Markowetz, Jochen Maydt, Magnus Rattray,
Jill Sales, Philip Smith, Wolfgang Urfer, and Joanna Wood for critical feed-
back on and proofreading of Chapters 1, 2, 4, 5, and 8. He would also like to
express his gratitude to his parents, Gerhild and Dieter; if it had not been for
their support in earlier years, this book would never have been written. His
special thanks, with love, go to Ulli for her support and tolerance of the extra
workload involved with the preparation of this book.

Edinburgh, London, Oxford Dirk Husmeier

UK Richard Dybowski
July 2003 Stephen Roberts
Contents

Part I Probabilistic Modeling

1 A Leisurely Look at Statistical Inference

Dirk Husmeier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 The Classical or Frequentist Approach . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 The Bayesian Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.4 Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2 Introduction to Learning Bayesian Networks from Data

Dirk Husmeier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.1 Introduction to Bayesian Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.1.1 The Structure of a Bayesian Network . . . . . . . . . . . . . . . . . 17
2.1.2 The Parameters of a Bayesian Network . . . . . . . . . . . . . . . . 25
2.2 Learning Bayesian Networks from Complete Data . . . . . . . . . . . . . . 25
2.2.1 The Basic Learning Paradigm . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2.2 Markov Chain Monte Carlo (MCMC) . . . . . . . . . . . . . . . . . 28
2.2.3 Equivalence Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.2.4 Causality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.3 Learning Bayesian Networks from Incomplete Data . . . . . . . . . . . . 41
2.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.3.2 Evidence Approximation and Bayesian Information
Criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.3.3 The EM Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.3.4 Hidden Markov Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.3.5 Application of the EM Algorithm to HMMs . . . . . . . . . . . . 49
2.3.6 Applying the EM Algorithm to More Complex Bayesian
Networks with Hidden States . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.3.7 Reversible Jump MCMC . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
2.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
xii Contents

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

3 A Casual View of Multi-Layer Perceptrons as Probability

Models
Richard Dybowski . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.1 A Brief History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.1.1 The McCulloch-Pitts Neuron . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.1.2 The Single-Layer Perceptron . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.1.3 Enter the Multi-Layer Perceptron . . . . . . . . . . . . . . . . . . . . . 62
3.1.4 A Statistical Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.2 Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.2.1 Maximum Likelihood Estimation . . . . . . . . . . . . . . . . . . . . . 65
3.3 From Regression to Probabilistic Classification . . . . . . . . . . . . . . . . 65
3.3.1 Multi-Layer Perceptrons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.4 Training a Multi-Layer Perceptron . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.4.1 The Error Back-Propagation Algorithm . . . . . . . . . . . . . . . 70
3.4.2 Alternative Training Strategies . . . . . . . . . . . . . . . . . . . . . . . 73
3.5 Some Practical Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.5.1 Over-Fitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.5.2 Local Minima . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
3.5.3 Number of Hidden Nodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
3.5.4 Preprocessing Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
3.5.5 Training Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.6 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

Part II Bioinformatics

4 Introduction to Statistical Phylogenetics

Dirk Husmeier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.1 Motivation and Background on Phylogenetic Trees . . . . . . . . . . . . . 84
4.2 Distance and Clustering Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.2.1 Evolutionary Distances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.2.2 A Naive Clustering Algorithm: UPGMA . . . . . . . . . . . . . . . 93
4.2.3 An Improved Clustering Algorithm: Neighbour Joining . . 96
4.2.4 Shortcomings of Distance and Clustering Methods . . . . . . 98
4.3 Parsimony . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4.3.2 Objection to Parsimony . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.4 Likelihood Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.4.1 A Mathematical Model of Nucleotide Substitution . . . . . . 104
4.4.2 Details of the Mathematical Model of Nucleotide
Substitution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
4.4.3 Likelihood of a Phylogenetic Tree . . . . . . . . . . . . . . . . . . . . . 111
4.4.4 A Comparison with Parsimony . . . . . . . . . . . . . . . . . . . . . . . 118
 Contents xiii

4.4.5 Maximum Likelihood . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120

4.4.6 Bootstrapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
4.4.7 Bayesian Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
4.4.8 Gaps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
4.4.9 Rate Heterogeneity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
4.4.10Protein and RNA Sequences . . . . . . . . . . . . . . . . . . . . . . . . . 138
4.4.11A Non-homogeneous and Non-stationary Markov Model
of Nucleotide Substitution . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142

5 Detecting Recombination in DNA Sequence Alignments

Dirk Husmeier, Frank Wright . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
5.2 Recombination in Bacteria and Viruses . . . . . . . . . . . . . . . . . . . . . . . 148
5.3 Phylogenetic Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
5.4 Maximum Chi-squared . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
5.5 PLATO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
5.6 TOPAL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
5.7 Probabilistic Divergence Method (PDM) . . . . . . . . . . . . . . . . . . . . . . 162
5.8 Empirical Comparison I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
5.9 RECPARS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
5.10 Combining Phylogenetic Trees with HMMs . . . . . . . . . . . . . . . . . . . 171
5.10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
5.10.2 Maximum Likelihood . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
5.10.3 Bayesian Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
5.10.4 Shortcomings of the HMM Approach . . . . . . . . . . . . . . . . . . 180
5.11 Empirical Comparison II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
5.11.1 Simulated Recombination . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
5.11.2 Gene Conversion in Maize . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
5.11.3 Recombination in Neisseria . . . . . . . . . . . . . . . . . . . . . . . . . . 184
5.12 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
5.13 Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188

6 RNA-Based Phylogenetic Methods

Magnus Rattray, Paul G. Higgs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
6.2 RNA Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
6.3 Substitution Processes in RNA Helices . . . . . . . . . . . . . . . . . . . . . . . 196
6.4 An Application: Mammalian Phylogeny . . . . . . . . . . . . . . . . . . . . . . . 201
6.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
xiv Contents

7 Statistical Methods in Microarray Gene Expression Data

Analysis
Claus-Dieter Mayer, Chris A. Glasbey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
7.1.1 Gene Expression in a Nutshell . . . . . . . . . . . . . . . . . . . . . . . . 211
7.1.2 Microarray Technologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
7.2 Image Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
7.2.1 Image Enhancement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
7.2.2 Gridding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
7.2.3 Estimators of Intensities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
7.3 Transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
7.4 Normalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
7.4.1 Explorative Analysis and Flagging of Data Points . . . . . . . 222
7.4.2 Linear Models and Experimental Design . . . . . . . . . . . . . . 225
7.4.3 Non-linear Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
7.4.4 Normalization of One-channel Data . . . . . . . . . . . . . . . . . . . 228
7.5 Differential Expression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
7.5.1 One-slide Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
7.5.2 Using Replicated Experiments . . . . . . . . . . . . . . . . . . . . . . . . 229
7.5.3 Multiple Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
7.6 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
8 Inferring Genetic Regulatory Networks from Microarray
Experiments with Bayesian Networks
Dirk Husmeier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
8.2 A Brief Revision of Bayesian Networks . . . . . . . . . . . . . . . . . . . . . . . 241
8.3 Learning Local Structures and Subnetworks . . . . . . . . . . . . . . . . . . . 244
8.4 Application to the Yeast Cell Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . 247
8.4.1 Biological Findings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
8.5 Shortcomings of Static Bayesian Networks . . . . . . . . . . . . . . . . . . . . 251
8.6 Dynamic Bayesian Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
8.7 Accuracy of Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
8.8 Evaluation on Synthetic Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
8.9 Evaluation on Realistic Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
8.10 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
9 Modeling Genetic Regulatory Networks using Gene
Expression Profiling and State-Space Models
Claudia Rangel, John Angus, Zoubin Ghahramani, David L. Wild . . . . . . 269
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
9.2 State-Space Models (Linear Dynamical Systems) . . . . . . . . . . . . . . . 272
9.2.1 State-Space Model with Inputs . . . . . . . . . . . . . . . . . . . . . . . 272
 Contents xv

9.2.2 EM Applied to SSM with Inputs . . . . . . . . . . . . . . . . . . . . . 274

9.2.3 Kalman Smoothing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
9.3 The SSM Model for Gene Expression . . . . . . . . . . . . . . . . . . . . . . . . . 277
9.3.1 Structural Properties of the Model . . . . . . . . . . . . . . . . . . . . 277
9.3.2 Identifiability and Stability Issues . . . . . . . . . . . . . . . . . . . . . 278
9.4 Model Selection by Bootstrapping . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
9.4.1 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
9.4.2 The Bootstrap Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
9.5 Experiments with Simulated Data . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
9.5.1 Model Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
9.5.2 Reconstructing the Original Network . . . . . . . . . . . . . . . . . . 283
9.5.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
9.6 Results from Experimental Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
9.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291

Part III Medical Informatics

10 An Anthology of Probabilistic Models for Medical

Informatics
Richard Dybowski, Stephen Roberts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
10.1 Probabilities in Medicine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
10.2 Desiderata for Probability Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
10.3 Bayesian Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
10.3.1 Parameter Averaging and Model Averaging . . . . . . . . . . . . 299
10.3.2 Computations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300
10.4 Logistic Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
10.5 Bayesian Logistic Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
10.5.1 Gibbs Sampling and GLIB . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
10.5.2 Hierarchical Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
10.6 Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
10.6.1 Multi-Layer Perceptrons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
10.6.2 Radial-Basis-Function Neural Networks . . . . . . . . . . . . . . . . 308
10.6.3 “Probabilistic Neural Networks” . . . . . . . . . . . . . . . . . . . . . . 309
10.6.4 Missing Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310
10.7 Bayesian Neural Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
10.7.1 Moderated Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
10.7.2 Hyperparameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312
10.7.3 Committees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313
10.7.4 Full Bayesian Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
10.8 The Naı̈ve Bayes Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316
10.9 Bayesian Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
10.9.1 Probabilistic Inference over BNs . . . . . . . . . . . . . . . . . . . . . . 318
10.9.2 Sigmoidal Belief Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
xvi Contents

10.9.3 Construction of BNs: Probabilities . . . . . . . . . . . . . . . . . . . . 321

10.9.4 Construction of BNs: Structures . . . . . . . . . . . . . . . . . . . . . . 322
10.9.5 Missing Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
10.10 Class-Probability Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
10.10.1 Missing Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
10.10.2 Bayesian Tree Induction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
10.11 Probabilistic Models for Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . 326
10.11.1 Data Conditioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327
10.11.2 Detection, Segmentation and Decisions . . . . . . . . . . . . . . . . 330
10.11.3 Cluster Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331
10.11.4 Hidden Markov Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
10.11.5 Novelty Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338

11 Bayesian Analysis of Population

Pharmacokinetic/Pharmacodynamic Models
David J. Lunn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
11.2 Deterministic Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
11.2.1 Pharmacokinetics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
11.2.2 Pharmacodynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359
11.3 Stochastic Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 360
11.3.1 Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 360
11.3.2 Priors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363
11.3.3 Parameterization Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 364
11.3.4 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
11.3.5 Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366
11.4 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367
11.4.1 PKBugs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367
11.4.2 WinBUGS Differential Interface . . . . . . . . . . . . . . . . . . . . . . 368
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369

12 Assessing the Effectiveness of Bayesian Feature Selection

Ian T. Nabney, David J. Evans, Yann Brulé, Caroline Gordon . . . . . . . . . 371
12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 371
12.2 Bayesian Feature Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372
12.2.1 Bayesian Techniques for Neural Networks . . . . . . . . . . . . . . 372
12.2.2 Automatic Relevance Determination . . . . . . . . . . . . . . . . . . 374
12.3 ARD in Arrhythmia Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
12.3.1 Clinical Context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
12.3.2 Benchmarking Classification Models . . . . . . . . . . . . . . . . . . 376
12.3.3 Variable Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
12.3.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380
12.4 ARD in Lupus Diagnosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381
12.4.1 Clinical Context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381
 Contents xvii

12.4.2 Linear Methods for Variable Selection . . . . . . . . . . . . . . . . . 383

12.4.3 Prognosis with Non-linear Models . . . . . . . . . . . . . . . . . . . . 383
12.4.4 Bayesian Variable Selection . . . . . . . . . . . . . . . . . . . . . . . . . . 385
12.4.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 386
12.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 387
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 388

13 Bayes Consistent Classification of EEG Data by

Approximate Marginalization
Peter Sykacek, Iead Rezek, and Stephen Roberts . . . . . . . . . . . . . . . . . . . . . . 391
13.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 391
13.2 Bayesian Lattice Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393
13.3 Spatial Fusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 396
13.4 Spatio-temporal Fusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 400
13.4.1 A Simple DAG Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . 401
13.4.2 A Likelihood Function for Sequence Models . . . . . . . . . . . . 402
13.4.3 An Augmented DAG for MCMC Sampling . . . . . . . . . . . . . 403
13.4.4 Specifying Priors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 404
13.4.5 MCMC Updates of Coefficients and Latent Variables . . . 405
13.4.6 Gibbs Updates for Hidden States and Class Labels . . . . . . 407
13.4.7 Approximate Updates of the Latent Feature Space . . . . . . 408
13.4.8 Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 409
13.5 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411
13.5.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412
13.5.2 Classification Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413
13.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 415
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 416
14 Ensemble Hidden Markov Models with Extended
Observation Densities for Biosignal Analysis
Iead Rezek, Stephen Roberts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419
14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419
14.2 Principles of Variational Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . 421
14.3 Variational Learning of Hidden Markov Models . . . . . . . . . . . . . . . . 423
14.3.1 Learning the HMM Hidden State Sequence . . . . . . . . . . . . 425
14.3.2 Learning HMM Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . 426
14.3.3 HMM Observation Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 427
14.3.4 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 431
14.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 435
14.4.1 Sleep EEG with Arousal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 435
14.4.2 Whole-Night Sleep EEG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 435
14.4.3 Periodic Respiration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 436
14.4.4 Heartbeat Intervals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 437
14.4.5 Segmentation of Cognitive Tasks . . . . . . . . . . . . . . . . . . . . . 439
14.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 440
xviii Contents

A Model Free Update Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 442

B Derivation of the Baum-Welch Recursions . . . . . . . . . . . . . . . . . . . . . 443
C Complete KL Divergences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
C.1 Negative Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 446
C.2 KL Divergences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 446
C.3 Gaussian Observation HMM . . . . . . . . . . . . . . . . . . . . . . . . . 447
C.4 Poisson Observation HMM . . . . . . . . . . . . . . . . . . . . . . . . . . . 448
C.5 Linear Observation Model HMM . . . . . . . . . . . . . . . . . . . . . 448
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 449

15 A Probabilistic Network for Fusion of Data and Knowledge

in Clinical Microbiology
Steen Andreassen, Leonard Leibovici, Mical Paul, Anders D. Nielsen,
Alina Zalounina, Leif E. Kristensen, Karsten Falborg, Brian
Kristensen, Uwe Frank, Henrik C. Schønheyder . . . . . . . . . . . . . . . . . . . . . . 451
15.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451
15.2 Institution of Antibiotic Therapy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
15.3 Calculation of Probabilities for Severity of Sepsis, Site of
Infection, and Pathogens . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 454
15.3.1 Patient Example (Part 1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 454
15.3.2 Fusion of Data and Knowledge for Calculation of
Probabilities for Sepsis and Pathogens . . . . . . . . . . . . . . . . . 456
15.4 Calculation of Coverage and Treatment Advice . . . . . . . . . . . . . . . . 461
15.4.1 Patient Example (Part 2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461
15.4.2 Fusion of Data and Knowledge for Calculation of
Coverage and Treatment Advice . . . . . . . . . . . . . . . . . . . . . . 466
15.5 Calibration Databases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467
15.6 Clinical Testing of Decision-support Systems . . . . . . . . . . . . . . . . . . 468
15.7 Test Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 468
15.8 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 469
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 470

16 Software for Probability Models in Medical Informatics

Richard Dybowski . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473
16.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473
16.2 Open-source Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 474
16.3 Logistic Regression Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 474
16.3.1 S-Plus and R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 475
16.3.2 BUGS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 476
16.4 Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 477
16.4.1 Netlab . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 477
16.4.2 The Stuttgart Neural Network Simulator . . . . . . . . . . . . . . 478
16.5 Bayesian Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 478
16.5.1 Hugin and Netica . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 481
16.5.2 The Bayes Net Toolbox . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 481