The advances of high-throughput technologies in genomics and proteomics have revolutionized the biological research field. The increased resolution capabilities have strengthened the focus on quantification analysis and the massively parallel nature of the instruments ultimately enables quantification at a genome- and proteome-wide scale. The field has experienced an explosion of applications, which is accompanied by many computational challenges and a strong demand for novel quantification analysis tools. Quantitative workflows are complex, involving many steps from sample preparation to data acquisition, processing, and the inference of quantitative estimates. Multiple sources within this process cause different bias in quantification. The goal to retrieve best estimates of the true underlying quantities from a sample remains a difficult task. It leaves a constant need for new method development to reduce systematic errors and biases in the data. In this thesis, new statistical and computational strategies are presented to improve the quantification accuracy of data types from high-throughput omics applications. The aim is to correct biases and minimize the overall variance; therefore understanding the potential error sources and data characteristics is essential. One focus in this work is to identify biases and solutions which are common to different omics workflows and data types. A second aim is to assess the statistical confidence of resulting quantitative measures. A general lack persists in high-throughput quantification on how to measure and report reliability of quantitative estimates, especially in quantitative proteomics research. A lot of knowledge on statistical methodology for large-scale data analysis has been acquired in the microarray era. Generally, independent of underlying technologies, final quantitative values often exhibit similar properties from a statistical point of view. Hence, a strong potential lies in revealing parallels between different omics fields and in the transfer of established statistical concepts. In addition, however, it is equally important to precisely integrate and account for specific data characteristics and technique-induced biases. Overall, quantitative analyses are highly heterogeneous and one-fit-all methods are not appropriate. The contribution of this thesis comprises three major projects which address different biological objectives and different data types based on three quantitative highthroughput techniques. New approaches concerning data pre-processing, quantification inference and resolution, and quantitative comparisons, are introduced. In the first project, affinity purification is coupled with mass spectrometry (APMS) aiming to identify protein-protein-interactions. Here, quantitative counts of proteins obtained from pull-down experiments are compared with counts from negative controls in order to separate true interactions from false-positive hits. Current methods for AP-MS analysis mainly rely on scoring systems to rank potential interaction proteins. However, uncertainty on where to set the cutoff score remains for candidate selection and also no estimation on the expected number of false positives is given. Statistical pre-and postprocessing is an underrepresented topic in AP-MS analysis. A thorough statistical framework is introduced, which can embed any scoring method and enables to replace scores by statistical p-values using a permutation principle. In addition, a two-stage poisson model adapted from RNA-Seq to AP-MS data is proposed as an alternative method for assessing interactions. For pre-processing, different normalization methods and statistical filtering, adjusted to AP-MS data, are investigated. Several experiments demonstrate how the number of true interactions can be significantly increased while controlling a false detection rate. The second project concerns the accurate inference of protein quantities. In mass spectrometry, measurements are assessed at the peptide spectrum level. Although all peptide spectra assigned to the same protein are assumed to share similar intensity values, in fact, a substantial heterogeneity exists due to random and systematic biases. Clever summarization strategies are needed. Current methods rely exclusively on peptide quantitative information. However, this work hypothesizes that a wealth of other peptide features are available that reflect spectra reliability. Several features are correlated with the observed variance heterogeneity and their relation to quantification accuracy in the spectra is investigated. As a result, a new peptide-to-protein summarization method is presented, referred to as iPQF (isobaric Protein Quantification based on Features), which integrates peptide features with quantitative values for protein quantification. As a novelty, peptide spectra are weighted according to their feature reliability. Extensive evaluation of iPQF in comparison to nine other summarization methods proves the added value of feature information to enhance protein ratio accuracy. NGS-based quantification equivalently relies on shotgun measurements and requires summarization strategies. The third project focuses on accurate inference of quantities on strain level in NGS-based metagenomics data. Specific challenges arise on strain level due to the presence of highly similar reference sequences, which underlie a strong quantification bias due to shared read mappings. There is increasing demand for analyzing microbial communities at higher resolution, but only few tools provide quantitative profiling beyond species level. In this work, DiTASiC (Differential Taxa Abundance including Similarity Correction) is presented as a novel tool for abundance estimation and also for differential analysis applicable down to exact genome level in metagenomics samples. A new generalized linear model framework is introduced for the resolution of shared read counts, additionally including an error term to assess abundance estimation uncertainties. In a new statistical approach, the abundance variances are integrated to infer abundance distributions for differential testing sensitive to strain level. Performance evaluations on latest benchmark studies show highly accurate abundance estimations down to sub-strain level and improved detection of differentially abundant taxa. Altogether, these three contributions improve the current repertoire of computational methods in high-throughput quantification of omics data. This work intends to raise awareness for the complexity of quantification analysis. On one side, it highlights the comprehensive usage and transfer of established statistical concepts across different omics techniques. Equally, it aims to emphasize the importance to specifically address underlying data characteristics and the need to offer individualized strategies in order to achieve high quantification accuracy.
