Multivariate Analysis of Tumour Gene Expression Profiles Applying Regularisation and Bayesian Variable Selection Techniques
Author(s)
Zucknick, Manuela
Type
Thesis
Abstract
High-throughput microarray technology is here to stay, e.g. in oncology for tumour classification
and gene expression profiling to predict cancer pathology and clinical outcome. The global
objective of this thesis is to investigate multivariate methods that are suitable for this task.
After introducing the problem and the biological background, an overview of multivariate
regularisation methods is given in Chapter 3 and the binary classification problem is outlined
(Chapter 4). The focus of applications presented in Chapters 5 to 7 is on sparse binary classifiers
that are both parsimonious and interpretable. Particular emphasis is on sparse penalised
likelihood and Bayesian variable selection models, all in the context of logistic regression. The
thesis concludes with a final discussion chapter.
The variable selection problem is particularly challenging here, since the number of variables
is much larger than the sample size, which results in an ill-conditioned problem with
many equally good solutions. Thus, one open problem is the stability of gene expression profiles.
In a resampling study, various characteristics including stability are compared between a
variety of classifiers applied to five gene expression data sets and validated on two independent
data sets.
Bayesian variable selection provides an alternative to resampling for estimating the uncertainty
in the selection of genes. MCMC methods are used for model space exploration, but
because of the high dimensionality standard algorithms are computationally expensive and/or
result in poor Markov chain mixing. A novel MCMC algorithm is presented that uses the
dependence structure between input variables for finding blocks of variables to be updated together.
This drastically improves mixing while keeping the computational burden acceptable.
Several algorithms are compared in a simulation study. In an ovarian cancer application in
Chapter 7, the best-performing MCMC algorithms are combined with parallel tempering and
compared with an alternative method.
and gene expression profiling to predict cancer pathology and clinical outcome. The global
objective of this thesis is to investigate multivariate methods that are suitable for this task.
After introducing the problem and the biological background, an overview of multivariate
regularisation methods is given in Chapter 3 and the binary classification problem is outlined
(Chapter 4). The focus of applications presented in Chapters 5 to 7 is on sparse binary classifiers
that are both parsimonious and interpretable. Particular emphasis is on sparse penalised
likelihood and Bayesian variable selection models, all in the context of logistic regression. The
thesis concludes with a final discussion chapter.
The variable selection problem is particularly challenging here, since the number of variables
is much larger than the sample size, which results in an ill-conditioned problem with
many equally good solutions. Thus, one open problem is the stability of gene expression profiles.
In a resampling study, various characteristics including stability are compared between a
variety of classifiers applied to five gene expression data sets and validated on two independent
data sets.
Bayesian variable selection provides an alternative to resampling for estimating the uncertainty
in the selection of genes. MCMC methods are used for model space exploration, but
because of the high dimensionality standard algorithms are computationally expensive and/or
result in poor Markov chain mixing. A novel MCMC algorithm is presented that uses the
dependence structure between input variables for finding blocks of variables to be updated together.
This drastically improves mixing while keeping the computational burden acceptable.
Several algorithms are compared in a simulation study. In an ovarian cancer application in
Chapter 7, the best-performing MCMC algorithms are combined with parallel tempering and
compared with an alternative method.
Date Issued
2008-12
Date Awarded
2009-03
Copyright Statement
Attribution NoDerivatives 4.0 International Licence (CC BY-ND)
Advisor
Gabra, Hani
Richardson, Sylvia
Sponsor
Wellcome Trust
Creator
Zucknick, Manuela
Publisher Department
Epidemiology and Public Health
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)