Fig. 7
From: MBE: model-based enrichment estimation and prediction for differential sequencing data
Real experimental prediction results. Comparison of Spearman correlation between wLER or MBE predictions and n experimental property measurements from the SARS-CoV-2-derived peptide [17], GB1 double site saturation [12], Chorismate mutase homolog [21], Bgl3 random mutagenesis [23], and AAV5 insertion [20] libraries. Each method is trained on real pre- and post-selection sequencing data, then used to predict the fitness of the n unobserved test sequences. The 100-unit NN model architecture is used for all datasets except the Bgl3 dataset, for which the linear architecture with IS features is used. The average performance improvement of MBE over wLER over all five experimental datasets, jointly, is statistically significant (\(p < 0.03\)). For comparisons of model predictions and property measurements for each experimentally measured sequence for each dataset, see Additional file 1: Fig. S12