RAG-GNN: retrieval-augmented graph neural networks for protein interaction network embeddings

Area of Science:

• Computational Biology
• Bioinformatics
• Network Science

Background:

• Protein interaction networks are crucial for understanding cellular processes.
• Graph Neural Networks (GNNs) are increasingly used to analyze these networks.
• Integrating external biological knowledge can potentially enhance GNN performance.

Purpose of the Study:

• To investigate the measurable benefits of retrieval augmentation for GNNs in protein interaction networks.
• To develop and evaluate the RAG-GNN framework, a novel end-to-end trainable model.
• To assess whether retrieval augmentation provides advantages beyond GNNs alone.

Main Methods:

• Developed RAG-GNN: a framework combining a GNN encoder with a dense retrieval module over a TF-IDF-indexed document corpus.
• Incorporated a gated fusion mechanism and contrastive alignment between node and document representations.
• Conducted a controlled methodological investigation using a cancer signaling protein interaction network.

Main Results:

• RAG-GNN improved functional clustering silhouette by +0.093 compared to a GNN-only ablation.
• The retrieval module achieved a mean precision at 10 of 0.242, a 152% relative improvement over a random baseline.
• Information decomposition revealed that topology and retrieval encode overwhelmingly shared information, with retrieval reorganizing the signal.

Conclusions:

• Retrieval augmentation offers measurable benefits for GNNs in analyzing protein interaction networks, particularly for functional clustering.
• The gains observed stem from general biological signals within the aggregated document corpus rather than solely node-specific semantic matching.
• RAG-GNN demonstrates task-specific complementarity, enhancing functional clustering while topology-focused methods excel at link prediction.