Fig. 7
From: Analysis and correction of compositional bias in sparse sequencing count data
Wrench scales outperform competing approaches in reconstructing compositional changes and in differential abundance testing. Multiple iterations of two group simulations are simulated with various fractions of features perturbed across conditions (rows, f in figures), total number of reads. Their average accuracy metrics in reconstruction and differential abundance testing are plotted. The control proportions were set to those obtained from the mouse microbiome dataset. a Average log ratios of reconstructed to true absolute abundance changes. Each row corresponds to a particular setting of f, and each column a particular setting of average sequencing depth. Scran also suffered from being unable to provide scales for samples in each simulation set (sometimes as high as 60% of the samples at 4K and 10K average reads). b Average sensitivity, specificity and false discoveries at FDR.1 of detecting true differential absolute abundances. W0 is the regularized Wrench estimator without sparsity adjustments and W1,..W3 are various adjusted estimators compared here. For details on this and simulations, see Methods. Behavior was similar for other parameteric variations (variances of global and sample-wise fold change distributions, number of samples) of simulations