Fig. 1

The illustrative flowchart of DeepRKE learning framework. First, we use RNAShapes to predict the RNA secondary structure from primary sequences. Second, word embedding algorithm is used to learn the distributed representations of 3-mers from primary sequences and secondary structures. Third, the learned distributed representations are fed into two CNNs (one is for RNA sequence and the other for secondary structures) to transform sequence and structure features, which are in turn input into a CNN module and a bidirectional LSTM layer followed by two fully connected layer. Finally, we use a sigmoid classifier to predict the probability of being RBP binding sites