Fig. 6

Enformer attributes no meaningful impact to distal eQTLs and performs poorly on tasks where long-range information is crucial. A The measured and predicted changes in gene expression (expressed as an unsigned percentage) due to eQTL variants are plotted as a function of the distance between the (canonical) TSS and the eQTL. To account for linkage, we always take the maximal effect of all variants in the credible set. This predicted effect quickly decays with distance. B The GTEx eQTL normalized effect size is plotted against the distance to the TSS. We observe no systematic decay with distance. C Enformer struggles to predict the overall impact of the genomic environment on the expression of the hk1 promoter, as measured by TRIP-seq. Note that this is the promoter for which the model performs best. D Performance of Basenji2 and Enformer on the TRIP-Seq data. For Enformer, we computed predictions after restricting the input window, as previously. We find that most of the (limited) signal on this task once again derives from the proximal 20% of the input sequence