Judging robustness and uncertainty in classifying sequencing reads in viral metagenomics

Classification of viral sequence reads by applying next generation sequencing (NGS) to the biological sample of an infected host can for example be done by mapping or machine learning approaches. The overall aim of this thesis is to improve methods
for quantifying the robustness and uncertainty of such classifications. In the context of mapping based classification, resampling approaches are presented that extend the original analysis result by better depicting the range of viral content in the sample. In addition, a supervised learning approach is presented to classify sequencing reads to biological orders and to correct class frequencies by incorporating prior information
about misclassification rates.
Most bioinformatics pipelines that rely on mapping result in a list of potential viruses present in the host sample. Such pipelines map sequencing reads or contigs, which are build from reads, to a set of known reference virus genomes and, if available, to the reference genome of the host. Several technical or biological processes can lead
to false positive or negative classifications. That means, viruses truly present in the sample can be omitted or, vice versa, viruses not in the sample can falsely be identified. Most existing computational methods for metagenomics data do not rate robustness of results. In this work however, two resampling algorithms to assess the robustness of mapping results in terms of false positive and negative findings are presented. To judge robustness, several indicators are derived from the resampling procedure such as the correlation between original and resampling read counts, or the statistical detection of outliers in the differences of read counts. Additionally, graphical illustrations of read count shifts via Sankey diagrams are provided. The resampling strategies include the generation of artificial FASTQ files whose read sequences and quality values are sampled based on statistical distributions of the
original mapping results. It is shown that, in contrast to an existing decoy strategy, this approach produces more realistic sequencing reads that share more statistical similarities with the original reads.
In contrast to the mapping approaches presented here, the learning approach makes inferences not on the level of the individual viruses but on the level of taxonomies. The viral taxa distribution is estimated by developing artificial neural networks as a special type of machine learning models and including different parameters derived from the read sequences as possible predictors. Afterwards, the statistical estimation is corrected by using estimated misclassification rates and applying probability calculus.