LUNER Prediction method of drug target interaction based on heterogeneous network

LUNER is innovative to combine the graph convolutional neural network based on the spatial domain with the attention mechanism. When representing node-level embedding in a heterogeneous network, our model can reflect the impact of different types of interactions on node embedding and get better and more interpretable embedding. The flow of our model is as follows: we first formed a heterogeneous network of 12 different types of relationship networks, including drug-drug interaction, drug structure similarity, drug-disease association, drug-side effect association, drug-protein interaction, disease-drug association, and side-effect-drug association. Then, the graph convolutional neural network is used to automatically learning the neighborhood information of the complex heterogeneous relationship network and integrate the attention mechanism to reflect the importance of the sum of different types of neighborhood information to obtain the representation characteristics of each node. After that, LUNAR applied the method of network topology reconstruction to extract the feature representation in the relational heterogeneous network. In this way, we can get the repositioning network whose edge weight indicates the strength of the relationship. Finally, predict new drug-target interactions based on the predicted scores between Chinese medicine and target nodes in the relocation network.