Fig. 2: PrismNet predicts RBP binding in cellular conditions more accurately than methods which use only RNA sequence.
From: Predicting dynamic cellular protein–RNA interactions by deep learning using in vivo RNA structures
a Model architecture of PrismNet. The input features include RNA sequence encoded in the 4-dimensional one-hot encoding, and the use of icSHAPE structural scores as the fifth-dimension. The neural network consists of multiple convolutional layers, squeeze-and-excitation (SE) networks, and residual blocks to capture the joint sequence-and-structural determinants of RBP binding. The green arrows indicate the data flow during network training, and the red arrows indicate the data flow during inference of HARs. b Predicted vs observed binding sites of IGF2BP1 on the EIF3F transcript. Green/black, observed binding sites in K562/HepG2 cells by eCLIP, used as the training/ground truth reference data; Blue and red indicate, respectively, true positive and false positive predictions in HepG2 cells, based on the models trained using K562 data. c “Circos and violin” plot of the respective and overall AUC scores of PrismNet vs other methods, including RCK, GraphProt, and DeepBind, for all 256 of the PrismNet models representing 168 human RBPs. ***P < 0.001 (one-sided paired t-test). d Violin plot of the overall AUPRC scores of PrismNet vs other methods for all 256 of the PrismNet models for 168 human RBPs. ***P < 0.001 (one-sided paired t-test). e Violin plot of the overall AUC scores of PrismNet models using different types of input data in all 256 PrismNet models of 168 human RBPs. ***P < 0.001 (one-sided paired t-test). f Scatter plot of AUC scores of PrismNet models using in vivo structures vs computationally predicted structures. Each dot represents an RBP. g Scatter plot of AUC improvements of PrismNet vs AUC scores of PrismNet models using only sequence information for 256 RBP models. Each dot represents an RBP model. RBPs are colored with their RNA-binding domains. RRM, RNA Recognition Motif; KH, K-Homology domain; Zinc, zinc finger domain; dRBM, double-stranded RNA binding motif. h Density map of binding scores of IGF2BP1 predicted using PrismNet vs the observed binding scores from eCLIP experiments in K562 cells. i Violin plot of PrismNet-predicted binding probabilities at the binding sites in K562 cells only, HepG2 cells only, or both. ***P < 0.001 (unpaired t-test).
