Finding Common Patterns In Heterogeneous Perturbation Data

Abstract

This work investigates and proposes statistical analysis methods for pattern detection in high-throughput data from perturbation experiments in biology and medicine. This is demonstrated in three examples. <br /> The first part of this thesis investigates the transcriptional responses of TGF-beta stimulation in different human and mouse cell types based on time-course microarray data from extensive experiments. The used statistical and bioinformatics methods enabled to identify commonly affected biological subsystems across different cell types. In particular the analysis suggests an important role of transcription factors, which appear to have a conserved influence across cell-types and species. Validation via an independent dataset confirms the findings and network analyses suggest explanations, how TGF-beta perturbation could lead to the observed effects.<br /> The second part investigates pro epileptic markers in microRNA expression profiling data from perturbation-induced pathogenic animal models. Experimental implica-tions resulting in incomplete and censored high-throughput qPCR data impairs the performance of analysis methods. A designated test procedure, which showed higher detection power at lower false positive rates base on simulated data, is proposed to resolve this issue. The method enabled the identification of novel pathogenic relevant miRNAs in epilepsy models. <br /> In the last part of this work a new method for drug-drug similarity assessment based on drug-proteins interaction network and drug pharmacological effects on disease related targets is proposed. The similarity measure, which does not require chemical structure information, is applied within a consensus clustering algorithm to detect useful patterns in a large compound dataset from different diseases. The method produced separated and stable clusters that could not be found using chemical structure-based approaches. Target proteins of compounds falling into one cluster suggested several new compound-target combinations, which could in several cases be confirmed by independent data.<br /> Altogether this thesis demonstrates that advanced analysis methods could help to extract common patterns from complex and seemingly heterogeneous data.