high performance computing on graphics processing units: hgpu.org

The analysis of next-generation sequencing (NGS) data is a major topic in bioinformatics: short reads obtained from DNA, the molecule encoding the genome of living organisms, are processed to provide insight into biological or medical questions. This thesis provides novel solutions to major topics within the analysis of NGS data, focusing on parallelization, scalability and reproducibility. The read mapping problem is to find the origin of the short reads within a given reference genome. We contribute the q-group index, a novel data structure for read mapping with particularly small memory footprint. The q-group index comes with massively parallel build and query algorithms targeted towards modern graphics processing units (GPUs). On top, the read mapping software PEANUT is presented, which outperforms state of the art read mappers in speed while maintaining their accuracy. The variant calling problem is to infer (i.e., call) genetic variants of individuals compared to a reference genome using mapped reads. It is usually solved in a Bayesian way. Often, variant calling is followed by filtering variants of different biological samples against each other. With state of the art solutions, the filtering is decoupled from the calling, leading to difficulties in controlling the false discovery rate. In this work, we show how to integrate the filtering into the calling with an algebraic approach and provide an intuitive solution for controlling the false discovery rate along with solving other challenges of variant calling like scaling with a growing set of biological samples. For this, a hierarchical index data structure for storage of preprocessing results is presented and compression strategies are provided. The developed methods are implemented in the software ALPACA. Depending on the research question, the analysis of NGS data entails many other steps, typically involving diverse tools, data transformations and aggregation of results. These steps can be orchestrated by workflow management. We present the general purpose workflow system Snakemake, which provides an easy to read domain-specific language for defining and documenting workflows, thereby ensuring reproducibility of analyses. The language is complemented by an execution environment that allows to scale a workflow to available resources, including parallelization across CPU cores or cluster nodes, restricting memory usage or the number of available coprocessors like GPUs. The benefits of using Snakemake are exemplified by combining the presented approaches for read mapping and variant calling to a complete, scalable and reproducible NGS analysis.