Time series models of genome-wide expression patterns. (360G-Wellcome-080715_Z_06_Z)

Time series models of genome-wide expression patterns This research project will focus on using and developing time series models for modelling genome-wise expression patterns. In particular, state-space models will be investigated to analyse the dynamics of gene expression profiles. These models assume that the observed measurements depend on some hidden state variables that evolve according to Markovian dynamics. The hidden variables can capture effects that cannot be measured in a gene expression profiling experiment, e.g., genes that have not been included in the microarray, levels of regulatory proteins, the effects of messenger RNA and protein degradation, etc. System estimation and identification is usually achieved, in a maximum likelihood framework, using the EM algorithm or, in a fully Bayesian setting, using Markov chain Monte Carlo methods. An important aspect that will be considered relates to model choice and selection. For instance, if the true structure underlying the data changes, over time, then standard dynamic linear models may fail to accurately estimate the dynamics. To attack this problem, dynamic models with Markov switching will be considered. The statistical methodology developed above will be tested on two data sets that have been made available. The first data set is intended to explore the differences between tuberculin positive and tuberculin negative individuals consisting of 5 so-called "killers" and 4 "non-killers". Bacillus Calmette-Guerin (BCG) was added to blood samples and microarrays taken at 0, 2, 4, 8, 12, 24, 48, 72 and 96 hours. The analysis of these data sets is particularly problematic as a range of microarray layouts were used and the number of time points varies across individuals. The second data set consists of 9 patients, 3 with previous pulmonary tuberculosis, 3 with previous tuberculous meningitis and 3 mantoux positive controls. Microarrays were taken at 0, 2, 6, 12, and 24 hours. The data set is, by comparison, much cleaner with a single microarrray layout used throughout and no missing time points. The developed statistical methods will also be tested on publicly available data that have been previously analysed, and whose findings have been published, for comparative purposes.