This invention discloses a multi-view structural prior negative sampling method for predicting
drug-target interactions, belonging to the field of
bioinformatics. The method includes: acquiring a
drug-target
bipartite graph, corresponding
drug features, and
target protein features (as well as the corresponding drug SMILES sequences and the
complete protein structure PDB file); applying a multi-view negative sampling strategy to the drug-target
bipartite graph, corresponding drug features, and
target protein features to generate drug-view
negative sample sets,
protein-view
negative sample sets, and random-view
negative sample sets, and then fusing them to obtain a negative sample set; using a heterogeneous graph convolutional network to perform
feature learning on the drug-target
bipartite graph and calculating the interaction
prediction probability corresponding to each negative sample. This invention, through a multi-view structural prior-guided negative sampling mechanism, effectively utilizes the similarity knowledge between the
chemical structure of the drug and the
spatial structure of the
target protein, reducing the risk of false negative samples and improving the quality of negative samples.