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

We present RAG-GNN, an end-to-end trainable framework that augments a graph neural network (GNN) encoder for protein interaction networks with a jointly optimized dense retrieval module over a TF-IDF-indexed document corpus, a gated fusion mechanism, and contrastive alignment between node and document representations. The study is positioned as a controlled methodological investigation of whether retrieval augmentation provides measurable benefit beyond a matched GNN-only ablation, rather than as a precision-medicine or therapeutic-target discovery tool. On a cancer signaling case study (379 proteins, 3,498 interactions, 14 pathway categories), RAG-GNN improves functional clustering silhouette from -0.237 ± 0.065 (GNN-only) to -0.144 ± 0.066 (+0.093 ± 0.022 across 10 seeds; ARI +0.021 ± 0.015), while the learned retrieval projection attains mean precision at 10 = 0.242, a 152% relative improvement over a random baseline (0.096). Counterfactual experiments confirm that random and absent retrieval contexts degrade performance, but a shuffled-document control (real documents reassigned to incorrect proteins) performs comparably to proper retrieval, indicating that the gain reflects general biological signal in the aggregate TF-IDF corpus rather than node-specific semantic matching. A heuristic information decomposition with bootstrap confidence intervals shows that topology and retrieval encode overwhelmingly shared information (95.6%), with retrieval contributing primarily by reorganizing this shared signal. Benchmarking against eight embedding methods reveals task-specific complementarity: Topology-focused methods are stronger for link prediction, while retrieval augmentation improves functional clustering within the controlled ablation. A DDR1 subnetwork analysis is reported as confirmatory recovery of biology already present in the input corpus, not as prospective discovery.