DAHNGC: A Graph Convolution Model for Drug-Disease Association Prediction by Using Heterogeneous Network

Area of Science:

• Computational biology
• Bioinformatics
• Drug discovery

Background:

• Predicting drug-disease associations is crucial for drug development and repositioning.
• Existing graph convolution methods primarily use homogeneous network features, neglecting valuable heterogeneous network information.

Purpose of the Study:

• To propose a novel drug-disease association prediction model, DAHNGC.
• To enhance prediction accuracy by incorporating attribute information from both homogeneous and heterogeneous networks.

Main Methods:

• Developed DAHNGC, a graph convolutional neural network model.
• Implemented DropEdge technique to address oversmoothing in homogeneous networks.
• Designed an automatic feature extraction method for heterogeneous networks.
• Utilized bilinear decoding for predicting potential drug-disease pairs.

Main Results:

• The DAHNGC model demonstrated strong predictive performance for drug-disease associations.
• The integration of heterogeneous network features significantly improved prediction insights.

Conclusions:

• DAHNGC offers a more effective approach to predicting drug-disease associations.
• The model's ability to leverage diverse network information advances drug discovery and repositioning strategies.