Abstract
      The present thesis addresses the statistical analysis of single and multiple series Time Course Microarray (TCM) data. This type of data comes from studies in which gene expression evolution is analysed throughout time for one or several experimental conditions of interest. The work describes the development, application and evaluation of novel specific methodologies which take into consideration particular aspects and problematic that this type of data causes, both from a gene selection and from a functional point of views. The developed algorithms are compared to other state-of-the-art methodologies, evaluating the different approaches in terms of performance and biological meaning of the results.
      The thesis has been structured in two main blocks. First, the relevant literature is revised and summarized in an introductory part. A general overview of microarray technology and a discussed review of statistical methods applied to microarray data are presented in Chapters 1 and 2. By using data from a multifactor microarray experiment we show how the application of general methods to time series microarray data suffers from a number of limitations. This indicates the need for the development of specific methods for the analysis of TCM. Chapter 3 ends this first block with a dedicated review on the up-to-date statistical methods for the analysis of time series data. Most of the methodologies presented in this chapter have been published during the time span of this thesis.
      In the second block, the novel methodologies for TCM developed within the present research are introduced and discussed. Chapter 4 introduces the first novel approach to TCM analysis: maSigPro (microarray Significant Profile) methodology. The maSigPro strategy uses linear regression analysis to model gene expression and follows a two-step strategy for the selection of differentially expressed genes (d.e.g): a first step identifies responsive genes while the second discloses the patterns of significant differential time evolution, in a gene-by-gene fashion. In Chapter 5 a multivariate technique ASCA (ANOVA Simultaneous Component Analysis) is adapted to TCM, resulting in the ASCA-genes method. This new methodology combines multivariate exploration of time course data with a selection procedure for the identification of relevant changing genes. In Chapter 6 the ability of ASCA to dissect expression signals and exploit the coordinative behaviour of gene expression is combined with the strong ability of maSigPro to model time series data and identify significant d.e.g. Our results show that, especially when high structural noise is present in the data, the use of ASCA as pre-processing strategy greatly improves maSigPro results. We also demonstrated that this data filtering strategy can be applied as well to other methods of TCM analysis improving false and negative discovery rates. These approaches, as others in microarray time series data analysis, provide results as lists of differentially expressed genes. However, in the study of gene expression, a much more interpretable and useful result appears when gene regulation is indicated as cellular functions or processes. In most cases, this translation is done subsequent to the generation of a list of differentially expressed genes. This implies in many cases limitations in discovery power for the need of an arbitrary calling of d.e.g. and the ignorance of the coordination between biological functions. The last methodological chapter of this thesis (Chapter 7) deals with the development of statistical approaches for an integrated or direct assessment of the alterations in gene functions embedded in time series expression data. To this end, maSigPro, ASCA and PCA have been adapted to incorporate functional data resulting in the novel methodologies maSigFun, PCA-maSigFun and ASCA-functional.
      This dissertation is ended with Chapter 8 which includes conclusions and some proposals deserving future research.