Item

Abstract

Enrichment of methylated DNA followed by sequencing offers a reasonable compromise between experimental cost and genomic coverage, allowing genome- wide DNA methylation to be assessed for large numbers of samples, which is a common requirement for clinical studies. However, the computational analysis of these experiments is complex, and depends on specific normalization and statistical approaches. Furthermore, quantification of the enrichment signals in terms of absolute levels of methylation requires specific transformation. In this dissertation, I introduce specific computational methods for the individual steps of the analysis workflow. I assess the impact of sequencing library size, alterations in DNA copy number and CpG density on the local enrichment, and present a suitable normalization procedure. As the central part of the workflow, I developed a statistical model for the enrichment read counts, which is deployed in the Bayesian estimation of absolute levels of methylation. The model involves experimental parameters, such as sample specific enrichment characteristics. Accounting for different levels of prior knowledge, I suggest several calibration strategies for the model's parameters, which use either additional data or certain general assumptions. The transformation to absolute methylation levels greatly enhances interpretability and facilitates comparison with other methylation assays. By comparing the results with bisulfite sequencing validation data, I demonstrate the accuracy of the transformation, as well as the improvement over existing alternative methods. A common objective of methylome analysis is the detection of differentially methylated regions between groups of samples. I compare different statistical approaches for this task and discuss the inherent properties. I thereby identify likelihood ratio tests of nested generalized linear models to be well suited in terms of reliability and efficiency. The methods are implemented in two different R/bioconductor packages, MEDIPS and QSEA, which are easy to use and provide comprehensive functionality for the analysis of enrichment based experiments. All functions are documented and demonstrated by runnable examples, as well as detailed tutorials for specific practically relevant use cases. By presenting four representative studies published in peer-reviewed journals, I demonstrate the applicability and the versatility of the introduced methods. Taken together, this dissertation provides new computational methods for the analysis of enrichment based methylation experiments; these methods enhance the interpretability and reliability of the results from these experiments.