Fig. 5

RBPNet variant impact prediction. A Impact of the rs6981405 variant on the predicted RBPNet target signal. Reference (blue) and alternative (red) signal is shown in a 200-nt window around the variant position. The corresponding feature importance maps for reference and alternative sequence are shown below. A C-to-A transversion at the 5′ end of the QKI binding motif leads to a drastic change of the predicted signal compared to the reference signal, which can be quantified as the KL divergence between the reference and alternative allele predictions. B Comparison of impact scores of a systematic in silico mutagenesis of each position towards each of the 3 alternative bases within in a 200-nt sequence window around rs6981405. C RBPNet impact scores (standardized) on gnomAD variants as a function of sequence conservation, measured via the PhyloP score. Mutations are separated per biotype and their impact is evaluated for RBPs known for biotype-specific activity. Impact scores are higher in regions that are under strong evolutionary constraints. D Classification performance of allele-specific binding variants by RBPNet, measured in KLD and L2 norm (the “Methods” section), and DeepRiPe