Dual graph convolutional neural network for predicting chemical networks.

Area of Science:

• Bioinformatics
• Chemoinformatics
• Computational Chemistry

Background:

• Chemical compound prediction is crucial for drug discovery and metabolic engineering.
• Computational methods, including machine learning and graph-based approaches, are increasingly used.
• Existing methods struggle to integrate both chemical structure graphs and interaction graphs efficiently.

Purpose of the Study:

• To develop an end-to-end method for chemical network prediction that considers both compound structures and inter-compound interactions.
• To formulate chemical network prediction as a link prediction problem within a graph of graphs (GoG) framework.

Main Methods:

• Proposed a novel dual graph convolutional network (DGCN) architecture.
• Represented chemical networks using a graph of graphs (GoG) to capture hierarchical structures.
• Learned compound representations by processing both individual compound graphs and the inter-compound network simultaneously.

Main Results:

• The DGCN approach achieved high prediction performance on relatively dense chemical networks.
• Demonstrated the model's ability to learn from both structural and interaction data in an end-to-end fashion.
• Observed a decrease in performance on extremely sparse networks compared to denser ones.

Conclusions:

• The dual graph convolution approach is effective for chemical network prediction, particularly in dense networks.
• The GoG representation and DGCN architecture offer a novel way to integrate multi-relational graph data.
• Further research may be needed to optimize performance for highly sparse chemical networks.