Kevin McGregor, Nneka Okaeme, Reihane Khorasaniha, Simona Veniamin, Juan Jovel, Richard Miller, Ramsha Mahmood, Morag Graham, Christine Bonner, Charles N Bernstein, Douglas L Arnold, Amit Bar-Or, Ruth Ann Marrie, Julia O'Mahony, Eluen Ann Yeh, Yinshan Zhao, Brenda Banwell, Emmanuelle Waubant, Natalie Knox, Gary Van Domselaar, Feng Zhu, Ali I Mirza, Helen Tremlett, Heather Armstrong
Compositional data comprise vectors that describe the constituent parts of a whole. Data arising from various -omics platforms such as 16S and RNA sequencing are compositional in nature. In this kind of data, correlations between features on raw counts have no meaningful interpretation. Metrics of proportionality were formulated to address this problem. However, an inherent bias arises when these metrics are calculated empirically on count-based measures due to variability in read depths. We quantify the bias introduced by empirically calculating proportionality-based association metrics in count data. Additionally, we propose a means of estimating these metrics within a logit-normal multinomial model in pursuit of more accurate estimates. The model-based estimates are shown to outperform empirical estimates in simulated data and are applied to a mouse embryonic stem cell single-cell sequencing dataset, as well as a pediatric-onset multiple sclerosis metagenomic dataset.