A comprehensive graph neural network method for predicting triplet motifs in disease-drug-gene interactions

Area of Science:

• Biomedical Informatics
• Graph Machine Learning
• Pharmacology

Background:

• Biomedical knowledge graphs integrate drug-disease, gene-disease, and drug-gene relationships, crucial for understanding complex biological processes.
• Existing methods for analyzing (disease, drug, gene) triplets primarily focus on triangle motifs, neglecting other significant structural patterns.
• A comprehensive approach is needed to predict diverse motifs within triplets for uncovering novel pharmacological mechanisms and improving disease-gene-drug interaction insights.

Purpose of the Study:

• To develop a novel method for predicting various graph motifs within (disease, drug, gene) triplets.
• To enhance the discrimination of different triplet motifs by addressing issues like redundant context and motif imbalance.
• To provide a comprehensive analysis of triplet motifs and reveal new pharmacological insights.

Main Methods:

• Propose TriMoGCL, a graph contrastive learning-based method for triplet motif prediction.
• Utilize a graph convolutional encoder for node feature extraction and employ node/edge pooling for context information.
• Implement node and class-prototype contrastive learning to denoise features and improve motif discrimination.

Main Results:

• TriMoGCL demonstrates effectiveness and reliability in identifying seven typical motifs within triplets across two knowledge graphs.
• The method successfully denoises triplet features and enhances discrimination between different motif types.
• New pharmacological mechanisms were revealed through a comprehensive analysis of triplet motifs.

Conclusions:

• TriMoGCL offers a robust solution for predicting diverse motifs in (disease, drug, gene) triplets.
• The approach advances the understanding of complex interactions within biomedical knowledge graphs.
• The findings contribute to uncovering novel pharmacological mechanisms and improving drug discovery processes.