Extended Data Fig. 5: Human enhancer design by in silico evolution.

a, Distribution of GC content in GC-adjusted random sequences (green) and human genomic regions (red). b, Average number of motif hits at each mutational step compared to genomic enhancers. c, Delta number of motifs in each mutational step. The TF-Modisco patterns and the most similar position weight matrices from the cisTarget motif database are shown at the top of each plot. ZEB2 upside-down pattern is contributing negatively to the model’s prediction and is destroyed by the model on each step. d, Bar plot showing the mean luciferase signal (log₂ fold-change over Renilla) in a MES melanoma line (MM047) of the synthetic MEL enhancers (generated by in silico sequence evolution), showing no activity compared to positive control genomic MES enhancers. The bar shows the mean (n = 2 biological replicates). e, MM001 (left) and MM099 (right) ATAC-seq profiles of all integrated lentiviral EFS reporters. f, MM001 ATAC-seq profile of 3 integrated EFS reporters: initial (top), evolved (middle) and post-evolution with repressive sites (bottom). g, DeepMEL2 prediction score (left), luciferase activity levels in MM001 (middle) and correlation between prediction score and activity (right) for EFS-1 (top) and EFS-8 (bottom) sequences after incremental mutation steps. In e, f, red dashed lines indicate boundaries of the enhancer. In g, the error bars show the standard error of the mean (n = 3 biological replicates).