Fig. 3

Enformer has very similar predictive power even if we severely restrict its input window, partially because most strong regulators are proximal. A Fraction of variance in log-transformed expression, both between conditions and between genes, which Enformer can explain given varying amounts of sequence context. Values computed on Enformer held-out data. Most of the signal comes from the sequence immediately around the TSS, with the distal two-thirds of the input window contributing very little. B Distribution of the distance within 98 kb of TSS of bona fide regulatory elements (eQTL, purple, and CRISPRi validated enhancers, blue) and candidate elements (ENCODE CRE with enhancer-like signal, red) and CRISPRi tested but not validated enhancers (green). Most bona fide regulatory elements lie close to their target gene whereas candidate elements are uniformly distributed. We only consider elements within 98kb of a TSS, i.e., within the Enformer receptive field