Extended Data Fig. 7: Occlusion analysis of AP-1 binding site effects.
From: Interpreting cis-regulatory mechanisms from genomic deep neural networks using surrogate models
Occlusion study based on the wild-type sequence investigated in Figures 6d and 6e. Occlusions were performed for every combination of one, two, or three occluded motifs, with the DNN prediction taken independently for each instance (n = 100 for each boxplot). In each occluded sequence, the corresponding AP-1 core (7-mer) sites were scrambled using a uniform probability of nucleotides at each position. The baseline score (CTRL) was calculated from the median of predictions corresponding to n = 100 instances of a dinucleotide shuffle over the full (2048 nt) sequence. The DNN prediction rapidly approaches the genomic baseline as additional binding sites are occluded. Boxplot lines represent median, upper quartile, and lower quartile; whiskers represent 1.5 × the inter-quartile range.
