Finding the Centre: Compositional Asymmetry in High-Throughput Sequencing Datasets

High-throughput sequencing datasets comprise millions of reads of genomic data and can be modelled as count compositions. These data are used for transcription profiles, microbial diversity, or relative cellular abundance in culture. The data are sparse and high dimensional. Moreover, they are often unbalanced, i.e. there is often systematic variation between groups due to presence or absence of features, and this variation is important to the biological interpretation of the data. The imbalance causes samples in the comparison groups to exhibit varying centres contributing to false positive and false negative identifications. Here, we extend the centred log-ratio transformation method used for the comparison of differential relative abundance between two groups in a Bayesian compositional context. We demonstrate the pathology in modelled and real unbalanced experimental designs to show how this causes both false negative and false positive inference. We examined four approaches to identify denominator features, and tested them with different proportions of modelled asymmetry; two were relatively robust, and recommended. We recommend the ‘LVHA’ transformation for asymmetric transcriptome datasets, and the ‘IQLR’ method for all other datasets when using the ALDEx2 tool available on Bioconductor.