Figure 3. Benchmark analysis across TFs, DNNs, and flank sizes.

a, Attribution variation analysis for various TFs and genomic DNNs. Results are visualized as in Figure 2b. Each test used $n=50$ sequences from either the human genome (for ResidualBind-32), the mouse genome (for BPNet), or the fly genome (for DeepSTARR), together with consensus TF binding site listed in Supplementary Table 1, and flanked by 50 nt of DNA. b, Mean attribution scores computed as in panel a, but using variable lengths of flanking DNA. The boxes above each plot indicate the largest (i.e., least significant) p-value, computed as in Figure 2b. n.s., p ≥ 0.05; *, 0.01 ≤ p < 0.05; **, 0.001 ≤ p < 0.01; ***, p < 0.001. Dashed line, flank size used in panel a. c, Binding motifs computed for Nanog via attribution analysis of BPNet. Left, Attribution maps averaged across 50 mouse loci containing (and centered on) the consensus Nanog binding site AGCCATCAA. Right, Attribution map observed for a single such locus. Gray bars indicate 10.5 nt periodicity on either side of the consensus Nanog binding site.