NSAP: A Neighborhood Subgraph Aggregation Method for Drug-Disease Association Prediction

Exploring the association between drugs and diseases can help to accelerate the process of drug development to a certain extent. In order to investigate the association between drugs and diseases, this paper constructs a network composed of different types of nodes, and proposes a model NSAP based on neighborhood subgraph prediction. The model captures local and global information around the target node through metagraphs and contextual graphs, respectively, and can generate node representations with rich information. In addition, in metagraphs and context diagrams, the model takes advantage of graph structures to automatically generate weights for edges, which better reflects the degree of association of different neighbor nodes with the target node. At last, the attention mechanism is used to aggregate the nodal representations generated by different metapaths in the graph, so that the final representation of the nodes incorporates different semantic information. For the edge prediction, a correlation score between drug-disease node pairs is calculated by the decoder. The experimental results have confirmed that our model does have certain effect by comparing it with state of the art method. The data and code are available at: https://github.com/jqq125/NSAP.