Fig. 6

A flowchart of DeepLPI. It begins with a multimodal deep learning neural network (MDLNN) that uses embedding layers, convolutional layers, LSTM layers and other layers of Keras to extract features from the sequence and structure data of lncRNAs and protein isoforms, and calculate initial interaction scores. Weighted correlation network analysis (WGCNA) is used to construct co-expression networks from expression data of lncRNAs and protein isoforms. Based on the pairwise potentials and unary potentials inferred from the co-expression relationship and the initial interaction scores, respectively, a conditional random field (CRF) optimization is used to predict the interactions between lncRNAs and protein isoforms. The whole model is trained using an iterative semi-supervised learning algorithm based on multiple instance learning (MIL)