Design and Validation of Proteome Measurements

Proteomics is a branch in biology that aims to comprehensively characterize a proteome. Mass spectrometry based proteomics has proven to be the most powerful approach to achieve this goal. This thesis introduces statistical concepts to optimally design and validate shotgun proteomics experiments and thereby enables to efficiently achieve reliable and extensive proteome coverage. The first part reports methods to estimate false discovery rates for peptide and protein identifications. These approaches enabled to reliably and comprehensively identify unusually modified protein variants. It turned out that these variants contribute to a significant fraction of the spectral evidence. This work presents a generalized target-decoy approach to estimate false discovery rates for protein identifications. This work shows evidence that the reliability of protein identifications in large studies has so far been largely overestimated and provides guidelines to compile identifications at well defined confidence. This part concludes with formulating a generic framework to compare protein inference engines based on protein identification false discovery rates. A systematic comparison of thousands of protein inference variants revealed that simple approaches yield optimal inference performance. The second part develops a nonparametric Bayesian approach to optimally design shotgun proteomics studies. Therefore the proteome coverage prediction task is introduced. An extended infinite Markov model is presented to perform proteome coverage prediction for simple shotgun proteomics experiments is presented. To capture the intricate similarities among peptide distributions arising in integrated shotgun proteomics studies, this work developed the general concept of the fractal Dirichlet process that augments the hierarchical Dirichlet process by introducing self-referential base measures. The fractal process is successfully applied to predict proteome coverage for integrated shotgun proteomics datasets. Rational stop criteria for these studies are discussed and evaluated by means of the proteome coverage prediction approaches. Finally the proteome coverage