Preview

PDF, English Download (33MB) | Terms of use

Abstract

Recent developments of experimental techniques and instrumentation allow life scientists to acquire enormous volumes of data at unprecedented resolution. While this new data brings much deeper insight into cellular processes, it renders manual analysis infeasible and calls for the development of new, automated analysis procedures. This thesis describes how methods of pattern recognition can be used to automate three popular data analysis protocols: Chapter 1 proposes a method to automatically locate bimodal isotope distribution patterns in Hydrogen Deuterium Exchange Mass Spectrometry experiments. The method is based on L1-regularized linear regression and allows for easy quantitative analysis of co-populations with different exchange behavior. The sensitivity of the method is tested on a set of manually identified peptides, while its applicability to exploratory data analysis is validated by targeted follow-up peptide identification. Chapter 2 develops a technique to automate peptide quantification for mass spectrometry experiments, based on 16O/18O labeling of peptides. Two different spectrum segmentation algorithms are proposed: one based on image processing and applicable to low resolution data and one exploiting the sparsity of high resolution data. The quantification accuracy is validated on calibration datasets, produced by mixing a set of proteins in pre-defined ratios. Chapter 3 provides a method for automated detection and segmentation of synapses in electron microscopy images of neural tissue. For images acquired by scanning electron microscopy with nearly isotropic resolution, the algorithm is based on geometric features computed in 3D pixel neighborhoods. For transmission electron microscopy images with poor z-resolution, the algorithm uses additional regularization by performing several rounds of pixel classification with features computed on the probability maps of the previous classification round. The validation is performed by comparing the set of synapses detected by the algorithm against a gold standard detection by human experts. For data with nearly isotropic resolution, the algorithm performance is comparable to that of the human experts.