Fig. 6
From: Improved modeling of RNA-binding protein motifs in an interpretable neural model of RNA splicing
Examples of regulatory landscapes of FM and AM models for annotated and decoy exons. Predicted 3′SS and 5′SS are shown as orange and purple lollipops, respectively; height represents score, normalized so that − 1 is the top-k threshold for the model for each class; only splice sites scored above − 5 in the normalized score by at least one model are shown. Black triangles on the x-axis correspond to annotated splice sites, while circled predictions correspond to the predicted splice sites for each model. Motif predictions are shown as blue lollipops on the FM and AM plots, in relative affinity units. Filled circles represent motifs that are present in both models; empty circles are motifs that appear in only one. Solid arrows connect any motif to the corresponding splice site if an in silico knockout of the motif increases the splice site prediction probability by 1.5-fold or more (green) or decreases it by a factor of 2/3 or less (red arrows). At least the top 5 effects are shown for each exon, using dashed arrows to indicate those that do not meet the cutoff for solid line