Fig. 1

RBPNet overview. A Schematic outline of data preparation, RBPNet training, and downstream applications. B RBPNet model architecture. The one-hot encoded RNA input sequence is first passed through a 1D convolutional layer, followed by several residual blocks, each consisting of a dilated convolution, batch normalization, ReLU, and dropout, respectively. Probability vectors of the target and control tracks are predicted from the output of the last residual block via a transposed convolutional layer while the total track is given by an additive mixture of target and control tracks. Given the predictions, a loss is computed by taking the sum of the negative log-likelihoods of the observed total and control counts. C Example prediction of an RBPNet model trained on eCLIP data of QKI showing observed counts (top) and predicted count distributions for the total (blue), control (red) and target (green) tracks. Integrated gradients feature attribution maps with respect to each predict track are shown below