Fig. 5

The predicted impact of enhancers, particularly distal enhancers, is significantly smaller than experiments suggest. A Predicted promoter strength vs measured promoter strength. These strengths are determined by fitting linear models to the data/predictions. B Predicted enhancer strength vs measured enhancer strength. C The promoter can explain 54% of the variation in the measurements of Bergmann et al. [13], with enhancers explaining another 36%. However, 90% of the variation in Enformer predictions for the same sequences is driven by the promoter alone. Thus, at least in a plasmid context, Enformer strongly underestimates the importance of the enhancer for determining gene expression. D In Enformer, the predicted variation of expression induced by the enhancer also heavily depends on the promoter. Only promoters of intermediate predicted strength are sensitive to the choice of Enhancer. In the experimental data, strong and intermediate promoters show similar sensitivity. C The measured and predicted changes in gene expression (expressed as an unsigned percentage) due to enhancer knockout as a function of the distance between the gene and the enhancer. Values < 0.1% and > 100% are truncated. Shown are only validated enhancer-gene pairs from Fulco et al. [21] and Gasperini et al. [20]. Enformer attributes significantly less effect to most validated enhancers than the experiments suggest. This phenomenon is particularly strong for distal enhancers