ISBN 978-3-8439-0941-9

978-3-8439-0941-9, Reihe Informatik

Holger Franken
Machine Learning and Computational Mass Spectrometry for the Discovery of Metabolite Biomarkers Involved in Type 2 Diabetes

154 Seiten, Dissertation Eberhard-Karls-Universität Tübingen (2013), Softcover, A5

Zusammenfassung / Abstract

To postpone or even prevent the manifestation of type 2 diabetes by therapeutic intervention, effective and early diagnosis is of critical concern. But the available tests for diagnosis of early disease stages are insufficiently dependable and too inconvenient to be used in a routine manner.

The young discipline of metabolomics is engaged with the discovery of novel biomarkers to an increasing degree. However, the inherent complexity of metabolomic data sets requires adequate tools to handle and evaluate the huge amounts of acquired data.

The main focus of this thesis is the automated processing and interpretation of measurement data from mass spectrometry (MS) based metabolomics studies. It provides contributions to several stages of a typical metabolomics experiment for biomarker discovery.

The alterations resulting from different preanalytical sample treatments are elucidated using a typical approach in metabolomics data analysis. This results in suggestions regarding guidelines for pre-analyitcal sample handling and metabolite markers that indicate unreliable sample quality.

A novel algorithm for detection and label-free quantification of metabolite features from LC-MS data is presented. It is implemented as an open source software and based on standard data formats. Its applicability is therefore not limited to a certain type of MS instrumentation. The algorithm is intuitively adjustable by MS experimentalists and provides sensitive and reproducible detection of features.

The latent alterations in early stages of metabolic diseases, such as type 2 diabetes, may cause metabolic aberrations that do not exhibit a sufficient amount of variance to be captured by conventional data analysis methods. The field of machine learning comprises a variety of techniques for the solution of such problems. Supervised wrapper- and ensemble-based feature selection techniques can unravel subliminal metabolic differences and identify compound subsets that can robustly and accurately classify samples from highly similar proband groups.

To improve the understanding of metabolic alterations that contribute to the development of type 2 diabetes a method for simultaneous selection of discriminative feature subsets and inference of feature dependencies is presented. Metabolite interconnection networks are derived from these dependencies. These interconnections can be mapped to physiological pathways, that are confirmed by type 2 diabetes-related literature.