Preview

PDF, English Download (8MB) | Terms of use

Abstract

Epigenetics, particularly the study of DNA methylation, is a cornerstone field for our understanding of human development and disease. DNA methylation has been included in the "hallmarks of cancer" due to its important function as a biomarker and its contribution to carcinogenesis and cancer cell plasticity. Long-read sequencing technologies, such as the Oxford Nanopore Technologies platform, have evolved the study of structural variations, while at the same time allowing direct measurement of DNA methylation on the same reads. With this, new avenues of analysis have opened up, such as long-range allele-specific methylation analysis, methylation analysis on structural variations, or relating nearby epigenetic modalities on the same read to another.

Basecalling and methylation calling of Nanopore reads is a computationally expensive task which requires complex machine learning architectures. Read-level methylation calls require different approaches to data management and analysis than ones developed for methylation frequencies measured from short-read technologies or array data. The 2-dimensional nature of read and genome associated DNA methylation calls, including methylation caller uncertainties, are much more storage costly than 1-dimensional methylation frequencies. Methods for storage, retrieval, and analysis of such data therefore require careful consideration. Downstream analysis tasks, such as methylation segmentation or differential methylation calling, have the potential of benefiting from read information and allow uncertainty propagation. These avenues had not been considered in existing tools.

In my work, I explored the potential of long-read DNA methylation analysis and tackled some of the challenges of data management and downstream analysis using state of the art software architecture and machine learning methods.

I defined a storage standard for reference anchored and read assigned DNA methylation calls, including methylation calling uncertainties and read annotations such as haplotype or sample information. This storage container is defined as a schema for the hierarchical data format version 5, includes an index for rapid access to genomic coordinates, and is optimized for parallel computing with even load balancing. It further includes a python API for creation, modification, and data access, including convenience functions for the extraction of important quality statistics via a command line interface. Furthermore, I developed software solutions for the segmentation and differential methylation testing of DNA methylation calls from Nanopore sequencing. This implementation takes advantage of the performance benefits provided by my high performance storage container. It includes a Bayesian methylome segmentation algorithm which allows for the consensus instance segmentation of multiple sample and/or haplotype assigned DNA methylation profiles, while considering methylation calling uncertainties. Based on this segmentation, the software can then perform differential methylation testing and provides a large number of options for statistical testing and multiple testing correction.

I benchmarked all tools on both simulated and publicly available real data, and show the performance benefits compared to previously existing and concurrently developed solutions. Next, I applied the methods to a cancer study on a chromothriptic cancer sample from a patient with Sonic Hedgehog Medulloblastoma. I here report regulatory genomic regions differentially methylated before and after treatment, allele-specific methylation in the tumor, as well as methylation on chromothriptic structures.

Finally, I developed specialized methylation callers for the combined DNA methylation profiling of CpG, GpC, and context-free adenine methylation. These callers can be used to measure chromatin accessibility in a NOMe-seq like setup, showing the potential of long-read sequencing for the profiling of transcription factor co-binding.

In conclusion, this thesis presents and subsequently benchmarks new algorithmic and infrastructural solutions for the analysis of DNA methylation data from long-read sequencing.