Molecular Structure-Driven Multi-Relation DGI Prediction With High-Low-Order Attention Denoise.

Area of Science:

• Computational biology
• Bioinformatics
• Drug discovery

Background:

• Drug-Gene Interaction (DGI) is vital for drug discovery and personalized medicine.
• Traditional experimental methods for DGI prediction are costly and time-consuming.
• Computational approaches are needed to efficiently predict complex drug-gene relationships, but existing methods face data scarcity and generalization issues.

Purpose of the Study:

• To propose a novel computational method for multi-relation DGI prediction.
• To address data scarcity and poor generalization challenges in existing DGI prediction models.
• To improve the efficiency and accuracy of predicting drug-gene relationships.

Main Methods:

• A multi-relation DGI prediction framework integrating molecular structure information via atom and bond channels.
• Enhancement of network structure analysis using both low-order graph convolutional networks and high-order hypergraph-based message propagation.
• Application of consistency information loss and inter-channel attention mechanisms to refine feature representations.

Main Results:

• The proposed model demonstrates superior performance on three drug-gene datasets.
• Significant F1 score improvements of 4.06% on DrugBank and 5.67% on DGIdb were observed, particularly on sparse datasets.
• The model effectively captures molecular structural information and refines network features.

Conclusions:

• The developed computational approach offers a powerful tool for DGI prediction, outperforming existing methods.
• The framework's ability to handle sparse data makes it valuable for drug discovery and personalized medicine.
• Publicly available implementations will facilitate further research and application in the field.