Fig. 4

Simulated library results. Spearman correlation between ground truth fitness and cLE, wLER, and MBE estimates on full-length sequences of interest for the tasks of a prediction and b estimation. The cLE approach can only be used for estimation (not prediction), and additionally, only for experiments where the sequencing reads were long enough to cover the sequences of interest (“Cover" = 100). Thus, cLE is absent from some experiments. All differences are statistically significant (\(p < 10^{-10}\)). For wLER and MBE, the results shown are for the cross-validation-selected architecture for each approach, as described in the main text, and are presented in the order of Table 1. For comprehensive results across all model architectures see Additional file 1: Fig. S1. For comparisons using Pearson correlation and MSE see Additional file 1: Figs. S2–S3, and see Additional file 1: Fig. S4 for comparisons to DEseq2