Standardisation and optimisation techniques in gut microbiome community analysis

With the emergence of high throughput next-generation sequencing the importance of the human gut microbiota as regulators, modulators and maintainers of human health and disease became more and more imminent. Advances in sequencing in the last two decades enabled the analysis of the composition and dynamics of the gut microbiome in unprecedented resolution and complexity. Investigations of this complex community by marker gene studies allowed assertions on presence, absence and ecological dynamics of gut bacteria. Several studies discovered strong relationships between the gut microbiota and human health. Some of these bacteria are shown to be essential for daily life processes like digestion, nutrition uptake, pathogen resistance and immune maturation. Likewise, disturbances of this close relationship, called dysbiosis, have been found to be associated with diseases like diabetes, obesity, colon cancer and inflammatory bowel disease. All this renders the gut microbiome as a highly relevant target of research in medical diagnostics and microbiome community analysis a valid hypothesis building tool. Nevertheless, the vast amount of different methodologies and lack of broadly accepted standards to create and handle gut microbiome abundance data complicates reproducible or replicable findings across studies. Especially in settings, where samples diverge significantly in their total biomass or microbial load, the analysis of the microbiome is hampered. Several efforts to allow accurate inter sample comparisons have been undertaken, including the use of relative abundances or random feature sub-sampling (rarefaction). While these methodologies are the most frequently used, they are not fully capable to correct for these sample-wide differences. To increase comparability between samples the use of exogenous spike-in bacteria is proposed to correct for sample specific differences in microbial load. The methodology is tested on a dilution experiment with known differences between samples and successfully applied on a clinical microbiome data set. These experiments suggest that current analysis methods lack a pivotal angle on the data, that is comparability between samples differing in microbial load. Meanwhile, the proposed spike-in based calibration to microbial load (SCML) allows for accurate estimation of ratios of absolute endogenous bacteria abundances. Furthermore, microbiome community analysis is heavily dependent on the resolution of the underlying read count data. While resolutions such as operational taxonomic units (OTUs) generally overestimate diversity and create highly redundant and sparse datasets, agglomerations to common taxonomy can obfuscate distinct read count patterns of possible sub-populations inside the given taxonomy. Even though the ladder agglomeration strategy might be valid for taxonomy with low phenotypical divergence, plenty taxonomic lineages in fact contain highly diverse sub-species. Thus, a more appropriate taxonomic unit would adapt its resolution for those densely populated branches, allowing for different count resolutions inside the same community. Here the concept of adaptive taxonomic units (ATUs) is introduced and applied on a perturbation experiment including mice receiving antibiotics. For this data set the different classical count resolutions (i.e. collapsed to order, family or genus etc.) produce highly contradictory results. Meanwhile, adaptive taxonomic units (ATUs) derived by hierarchical affinity merging (HAM) adapt the granularity of taxonomy to the underlying sequencing data. Branches of bacterial phylogeny that are highly covered in the data set receive a higher resolution than those that were infrequently observed. The algorithm hereby merges operational taxonomic units (OTUs) guided not only by sequence dissimilarity, but also by count distribution and OTU size. Due to the agglomeration the number of features is reduced significantly, lowering the complexity of the data, while preserving distributional patterns only observable at OTU level. Consequently, the sparsity of the count data is reduced significantly such that every ATU accumulates reasonable count number and can thus be reliably analysed. The algorithm is provided in the form of the R-Package dOTUClust