Extended Data Fig. 5: Benchmark analysis of attribution variation for putative weak TF binding sites.
From: Interpreting cis-regulatory mechanisms from genomic deep neural networks using surrogate models
a, ISM attribution map given by BPNet for a representative genomic sequence containing multiple putative weak binding sites for the mouse TF Nanog (top). Blue represents the wild-type nucleotides while gray represents other nucleotides. PWM scores (given by PWM for Nanog motif) are displayed across the genomic sequence (bottom). Only positive PWM scores are shown. b, For each TF and DNN, plots show attribution variation values for 150 putative TF binding sites plotted against putative binding site strength. Bold lines indicate signals smoothed with a sliding window of 20 nt. Stars indicate P values computed using the one-sided Mann-Whitney U test: **, 0.001≤ p < 0.01; ***, p < 0.001. PWM scores for each of the 150 putative sites are shown above, along with a logo representation of the PWM used. Each site is represented by a gray bar shaded according to the number of mutations (0, 1, or 2) in the core of the putative site.
