SkipGNN: predicting molecular interactions with skip-graph networks.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning

Background:

• Molecular interaction networks are crucial for biological discovery and are increasingly analyzed using machine learning.
• Graph neural networks (GNNs) have advanced interaction prediction but primarily focus on direct node similarities.
• Indirect relationships, or 'skip similarity,' in biological networks offer valuable predictive information.

Purpose of the Study:

• To introduce SkipGNN, a novel graph neural network (GNN) designed to predict molecular interactions by leveraging both direct and second-order (skip) similarities.
• To enhance the predictive power of GNNs by incorporating information from non-adjacent nodes in molecular interaction networks.

Main Methods:

• Developed SkipGNN, a GNN that aggregates information from immediate neighbors and two-hop neighbors ('skip similarity').
• Constructed a 'skip graph' by modifying the original network to incorporate second-order interactions.
• Implemented an iterative fusion scheme to optimize the GNN using both the original and skip graphs.

Main Results:

• SkipGNN demonstrated superior and robust performance across four diverse interaction networks: drug-drug, drug-target, protein-protein, and gene-disease.
• The model successfully learned biologically meaningful node embeddings, outperforming existing GNNs.
• SkipGNN showed particular effectiveness in handling noisy and incomplete biological interaction data.

Conclusions:

• SkipGNN offers a significant advancement in predicting molecular interactions by effectively utilizing skip similarity.
• The approach enhances the interpretability of GNNs by learning meaningful biological embeddings.
• SkipGNN provides a robust solution for analyzing complex and imperfect biological networks.