Fig. 6

Simulated sequence selectivity prediction results. Comparison of wLER and MBE (using 100-unit NNs) for identifying selective test sequences over three simulated datasets: a 21-mer insertion (\(4.6 \times 10^7\) short reads), b avGFP mutagenesis (\(4.6 \times 10^5\) long reads), and c AAV recombination (\(4.6 \times 10^5\) long reads). Colored points show the true positive and negative fitness of the top ten test sequences identified from each of the three test folds from threefold cross-validation according to each model’s predicted selectivity (i.e., difference in predicted positive and negative fitness values). To gauge overall performance, the average point from each method is also plotted in black-and-white, as is a theoretical optimally selective sequence (star) with the maximum positive fitness and minimum negative fitness among all sequences in the relevant dataset. Distance from optimal to average is conveyed by a circular contour line through the average point for each method; the size of the gap between the two circles is indicative of how much closer MBE is to the optimum than wLER. On all three datasets, MBE is significantly more accurate than wLER at identifying the \(1\%\) of test sequences with the highest true selectivity (\(p < 10^{-3}\))