Fig. 2

RBPNet prediction performance on ENCODE eCLIP datasets. A RBPNet predictions (total track) on the highest-count hold-out samples for TIA1, QKI, and U2AF2. B Pearson correlation coefficient (PCC) of RBPNet predictions (total track) with observed eCLIP crosslink counts on hold-out samples vs. PCC of observed counts between the two eCLIP replicates. C, D Mean auROC and AP of RBPNet (target track) and DeepRiPe predictions with respect to crosslink and non-crosslink positions called by PureCLIP across transcripts from hold-out chromosomes, respectively. As pre-trained DeepRiPe models are available only for 70 (out of 103) ENCODE HepG2 RBPs, performance comparison is shown only for those RBPs. E Distribution of RBPNet mixing coefficients of the top and bottom decile ENCODE narrow peaks, sorted by eCLIP signal fold-change over the SMInput. High-affinity ENCODE narrow peaks show on average higher mixing coefficients compared to low-affinity peaks. F, G Example RBPNet (target track) whole-transcript predictions on ENSG00000173207.12 and ENSG00000137955.15, together with observed eCLIP counts and called PureCLIP peaks, for QKI and HNRNPC, respectively