Feature selection and modelling methods for microarray data from acute coronary syndrome

Acute coronary syndrome (ACS) represents a leading cause of mortality and morbidity worldwide. Providing better diagnostic solutions and developing therapeutic strategies customized to the individual patient represent societal and economical urgencies. Progressive improvement in diagnosis and treatment procedures require a thorough understanding of the underlying genetic mechanisms of the disease. Recent advances in microarray technologies together with the decreasing costs of the specialized equipment enabled affordable harvesting of time-course gene expression data. The high-dimensional data generated demands for computational tools able to extract the underlying biological knowledge.

This thesis is concerned with developing new methods for analysing time-course gene expression data, focused on identifying differentially expressed genes, deconvolving heterogeneous gene expression measurements and inferring
dynamic gene regulatory interactions. The main contributions include: a novel multi-stage feature selection method, a new deconvolution approach for estimating cell-type specific signatures and quantifying the contribution of each cell type to the variance of the gene expression patters, a novel approach to identify the
cellular sources of differential gene expression, a new approach to model gene expression dynamics using sums of exponentials and a novel method to estimate stable linear dynamical systems from noisy and unequally spaced time series data.

The performance of the proposed methods was demonstrated on a time-course dataset consisting of microarray gene expression levels collected from the blood samples of patients with ACS and associated blood count measurements. The results of the feature selection study are of significant biological relevance. For the
first time is was reported high diagnostic performance of the ACS subtypes up to three months after hospital admission. The deconvolution study exposed features
of within and between groups variation in expression measurements and identified potential cell type markers and cellular sources of differential gene expression.
It was shown that the dynamics of post-admission gene expression data can be accurately modelled using sums of exponentials, suggesting that gene expression
levels undergo a transient response to the ACS events before returning to equilibrium. The linear dynamical models capturing the gene regulatory interactions
exhibit high predictive performance and can serve as platforms for system-level analysis, numerical simulations and intervention studies.