Fig. 4

Enformer accurately predicts genetic perturbations of promoters. A Correlation between model predictions and measurements across synthetic promoters of the Weingarten-Gabbay et al. [25] parallel reporter assay. Enformer outperforms preceding models, notably in targeted perturbation experiments (transcription factor motif ablation and core promoter motif perturbations). B Enformer can detect and often correctly interpret the impact on expression of subtle locational shifts of the TATA box in the RPLP0 promoter background. C In silico mutagenesis analysis suggests that the large drop of expression around position −50 from panel B likely is due to the disruption of a CAT-box at this location, rather than the positional preference of the TATA box. D Pearson correlation between predicted and observed variant effects for the different loci tested by Kircher et al. [14] Enformer outperforms other models for most loci. E Measured vs. predicted variant effects for two loci (F9, LDLR) individually and then both loci combined for CAGE (top) and DNase (bottom). The CAGE predictions appear to be miscalibrated across those two loci