Fig. 5

Adding pseudocounts leads to biased normalization. For each of the four microbiome count datasets (rows: Mouse, Lung, Diarrheal and Tara Oceans), we plot a CLR and b DESeq compositional scales obtained after adding a pseudo count value of 1, as a function of fraction of features that are zero in the samples (first column) and the sample depth (second column). The observed behavior was not sensitive to the value of pseudocount used. Refer Additional file 1: Figure S7 for the same plot for a pseudocount value of 10⁻⁷. c shows the total number of pseudocounts added, which is essentially the number of features observed in a dataset, and the total actual counts observed in the dataset divided by their sum i.e., the total implied sequencing depth after pseudocounts addition. A large fraction of sequencing depth in the new pseudocounted dataset is now arising from pseudocounts than the true experimental counts, when the data is excessively sparse. Indeed, if the pseudocount value is altered to a very low positive fraction value, the boxplots will reflect reversed locations, but this plot is only used to stress the level of alteration made to a dataset. Only in the Tara Oceans project, where the sample depth is 100K reads, do the boxplots shift. However, at a roughly median 90% features absent, that data when altered by pseudocounts, also leads to biased scaling factors as seen in a and b