Fig. 4

RBPNet feature attribution maps and binding motif discovery. A Example integrated gradient attribution maps with respect to the target track for RBFOX2, HNRNPK, TIA1, and QKI with corresponding motifs taken from the RBPmap database. B Distributions of similarity scores between 5-mers extracted from RBPNet attribution maps and PWMs of motifs reported in RBPmap for control, target, and total. C Average RBPmap PWM similarity across RBPs for 5-mers extracted from target and total track attributions. While similarities of the target track on par or higher compared to the total track for the vast majority of RBPs, the improvement is more pronounced for some RBPs. D Recall of the top 20 in vitro 5-mers recovered by the top 20 5-mers extracted from attribution maps of RBPNet control, target, and total tracks as well as DeepRiPe and the PEKA motif finder. RBPNet target track and PEKA show comparable performance, outperforming DeepRiPe and the RBPNet total track. E Consensus motifs constructed from extracted 5-mers of the RBPNet target track. Consensus motifs with highest (primary) and second highest (secondary) k-mer support are shown. The corresponding k-mer support is shown as a fraction of the total number of extracted 5-mers next to the consensus motif logos