Extracting information from high-throughput gene expression data with pathway analysis and deconvolution

Abstract

Modern technologies allow for the collection of large biological datasets that can be utilised for diverse health-related applications. However, to extract useful information from such data, computational methods are needed. The field that develops and explores methods to analyse biological data is called bioinformatics. In this thesis I evaluate different bioinformatic methods and introduce novel ones related to processing gene expression data. Gene expression data reflects how active different genes are in a set of measured biological samples. These samples can be for example blood from human individuals, tissue samples from tumours and the corresponding healthy tissue, or brain samples from mice with different neural diseases. This thesis covers two topics, pathway analysis and deconvolution, related to downstream analysis of gene expression data. Notably, this summary does not repeat in detail the same points made in the original publications, but aims to provide a comprehensive overview of the current knowledge of the two wider topics. The original publications focus on comparing and evaluating the available methods as well as presenting new ones that cover some previously untouched features.

While the terms ’pathway analysis’ and ’deconvolution’ have been used with alternative definitions in other fields, in the context of this thesis, pathway analysis refers to estimating the activity of pathways, i.e. interaction networks body uses to react to different signals, based on given gene expression data and structural information of the relevant pathways. I focus on different types of analysis methods and their varying goals, requirements, and underlying statistical approaches. In addition, the strengths and weaknesses of the concept of pathway analysis are briefly discussed. The first two original publications I and II empirically compare different types of pathway methods and introduce a novel one. In the paper I, the tested methods are evaluated from different perspectives, and in the paper II, a novel method is introduced and its performance demonstrated against alternative tools.

Many biological samples contain a variety of cell types and here, deconvolution means computationally extracting cell type composition or cell type specific expression from bulk samples. The deconvolution sections of this thesis also focus on a general overview of the topic and the available computational methodology. As deconvolution is challenging, I discuss the factors affecting its accuracy as well as alternative wet lab approaches to obtain cell type specific information. The first original publication about deconvolution (publication III) introduces a novel method and evaluates it against the other available tools. The second (publication IV) focuses on identifying cell type specific differences between sample groups, which is a particularly difficult task.