RAG-GNN: Integrating Retrieved Knowledge with Graph Neural Networks for Precision Medicine

TL;DR

RAG-GNN integrates topology-based graph neural networks with dynamically retrieved biomedical knowledge to produce joint embeddings for precision medicine. The framework reveals that network topology accounts for the majority of predictive information, while retrieved documents contribute a non-redundant 8.6% of information, enabling functional interpretation where topology alone falls short. A cancer signaling case study demonstrates DDR1 as a therapeutic target supported by retrieved evidence of synthetic lethality with KRAS mutations, validating practical impact in oncology. Overall, topology-focused methods excel at structural predictions, whereas retrieval-augmented approaches enhance functional interpretation, guiding task-specific method selection for biomedical applications.

Abstract

Network topology excels at structural predictions but fails to capture functional semantics encoded in biomedical literature. We present a retrieval-augmented generation (RAG) embedding framework that integrates graph neural network representations with dynamically retrieved literature-derived knowledge through contrastive learning. Benchmarking against ten embedding methods reveals task-specific complementarity: topology-focused methods achieve near-perfect link prediction (GCN: 0.983 AUROC), while RAG-GNN is the only method achieving positive silhouette scores for functional clustering (0.001 vs. negative scores for all baselines). Information-theoretic decomposition shows network topology contributes 77.3% of predictive information, while retrieved documents provide 8.6% unique information. Applied to cancer signaling networks (379 proteins, 3,498 interactions), the framework identifies DDR1 as a therapeutic target based on retrieved evidence of synthetic lethality with KRAS mutations. These results establish that topology-only and retrieval-augmented approaches serve complementary purposes: structural prediction tasks are solved by network topology alone, while functional interpretation uniquely benefits from retrieved knowledge.

Figures (6)

• Figure 1: RAG-GNN framework for precision medicine: Architecture overview. The framework integrates six interconnected components for knowledge-augmented biomedical prediction. (1) Biological network input: heterogeneous molecular interaction networks representing protein-protein interactions, signaling pathways, and regulatory relationships. (2) Neural conversion: network pruning transforms biological graphs into computationally tractable neural representations while preserving topological structure. (3) GNN encoder and retrieval module: graph attention (GAT) or graph convolutional (GCN) layers perform iterative message passing with edge updates and hierarchical pooling operations to generate node embeddings. (4) Knowledge corpus: the dense retriever queries diverse external knowledge sources including molecular docking predictions, immune cell interaction networks (featuring regulatory relationships among MDSC, Treg, Th2, and other immune populations), ADMET properties, archived literature (PubMed, DrugBank, pathway databases), and expert-curated domain knowledge. (5) Joint embedding: contrastive learning aligns network topology embeddings with retrieved knowledge representations in a unified semantic space. (6) Multimodal data integration: the framework incorporates diverse patient-level data including gene expression profiles, brain connectome information, multi-omic visualizations (volcano plots, circos diagrams), synthetic biology constructs, and clinical records. The prediction head enables downstream applications including therapeutic target identification (exemplified by DDR1) and drug response prediction. Biological validation through in vitro (cell-based assays) and in vivo (animal models) experiments with compound toxicity profiling provides iterative feedback for model refinement.
• Figure 2: RAG-GNN architecture for precision medicine. The complete system integrates network topology encoding, knowledge retrieval, and context fusion through six main components. The forward pass (solid arrows) begins with the input network $\mathcal{G}^{(p)} = (\mathbf{A}, \mathbf{X})$ representing patient-specific molecular interactions and node features. The GNN encoder applies $L$ layers of message passing to produce structural node embeddings $\mathbf{h}_i^{(L)}$ that capture network topology (\ref{['eq:gnn_update']}). These embeddings serve dual purposes: (1) querying the knowledge base through the dense retriever to identify top-$k$ relevant documents $\mathcal{D}_i^{(k)}$ from PubMed abstracts, pathway databases, and drug repositories (\ref{['eq:top_k_retrieval']}), and (2) providing structural context for fusion. The dense retriever employs quality-weighted semantic similarity to prioritize high-evidence documents. Retrieved documents are aggregated with attention weighting and fused with structural embeddings $\mathbf{h}_i^{(L)}$ through the context fusion module to produce final node representations $\mathbf{z}_i$ (\ref{['eq:fusion']}). These representations feed into task-specific prediction heads for therapeutic target scoring and drug response prediction. The training objective (bottom) jointly optimizes three components through gradient backpropagation (dashed arrows): task-specific loss $\mathcal{L}_{\text{task}}$ for prediction accuracy, retrieval quality loss $\mathcal{L}_{\text{retrieval}}$ ensuring relevant document selection, and contrastive embedding alignment loss $\mathcal{L}_{\text{contrastive}}$ coordinating node and document representations in shared semantic space (\ref{['eq:total_loss']}). The multi-objective formulation $\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{task}} + \lambda_1 \mathcal{L}_{\text{retrieval}} + \lambda_2 \mathcal{L}_{\text{contrastive}}$ enables end-to-end learning where retrieval and embedding components are optimized to support downstream prediction tasks. Curriculum learning stages the training process to ensure stable convergence and effective coordination between network encoding, document retrieval, and knowledge fusion mechanisms.
• Figure 3: Document retrieval performance for protein function queries. Precision-recall curves comparing retrieval methods for identifying functionally relevant documents across 379 protein queries. Ground truth relevance is determined by functional category matching (documents discussing proteins from the same pathway are considered relevant). TF-IDF achieves highest mAP (0.664) due to explicit protein name matching in the corpus. RAG-GNN (mAP=0.092) and topology-only (mAP=0.087) methods show similar embedding-based retrieval performance, both comparable to random baseline (mAP=0.096). The RAG-GNN approach demonstrates value not primarily through retrieval accuracy but through improved embedding quality for downstream tasks (see \ref{['Fig. 1']}C). Knowledge base contains 1,895 documents with functional annotations across 13 cancer pathway categories derived from curated pathway databases.
• Figure 4: RAG-GNN protein embeddings in cancer signaling networks using real STRING database interactions.(A) PCA projection of RAG-GNN embeddings: Two-dimensional visualization of 379 cancer-related proteins embedded in 128-dimensional space using GNN message passing combined with knowledge retrieval from functional annotations. Data source: STRING database (3,498 high-confidence interactions). Proteins are colored by functional pathway annotation across 13 major categories. Key oncogenes and tumor suppressors highlighted: TP53, EGFR, KRAS, MYC, BRCA1, PIK3CA, AKT1, and PTEN. Silhouette scores quantify functional clustering quality: RAG-GNN achieves $-0.046$ compared to $-0.133$ for GNN-only embeddings, representing a significant improvement of $+0.087$ in functional coherence. While both scores are negative (typical for complex biological networks with overlapping pathways), RAG-GNN substantially reduces intra-cluster dispersion. (B) Degree distribution: Node degree follows power-law distribution characteristic of scale-free biological networks, with hub proteins exceeding 60 connections. (C) Clustering quality comparison: Bar chart comparing silhouette scores between RAG-GNN and GNN-only methods, demonstrating the improvement from knowledge integration. (D) Top 20 hub proteins: Proteins ranked by betweenness centrality, identifying critical signaling bridges including CHUK, MAPK1/3, STAT3, and TP53. (E) Functional category distribution: Distribution of 379 proteins across categories, with transcription (44), DNA repair (40), apoptosis (35), and RTK signaling (35) as largest groups.
• Figure 5: DDR1 protein interaction subnetwork visualization with functional annotations and RAG-GNN embedding similarity.(A) Functional category representation: Network visualization of DDR1 (Discoidin Domain Receptor 1) and its interaction partners from STRING database, with nodes colored by functional pathway membership. The subnetwork contains 28 proteins connected by 128 edges, comprising 7 first-hop neighbors (direct interactors) and 20 second-hop neighbors (indirect interactors). DDR1 (purple, RTK signaling) functions as a receptor tyrosine kinase integrating extracellular matrix signals with intracellular signaling cascades. The network exhibits multi-pathway integration, with direct connections to PI3K-AKT-MTOR signaling (PIK3CA, PRKCA), RTK signaling components (SHC1, ABL1), and other kinase hubs. Node size reflects network centrality, with DDR1 shown as the largest node. The diverse functional categories demonstrate DDR1's role as a signaling hub coordinating proliferation, differentiation, and migration. (B) Embedding similarity landscape: Same network topology with nodes colored by cosine similarity to DDR1 in the 128-dimensional RAG-GNN embedding space. The embeddings capture functional relationships through fusion of GNN-derived topology features with retrieved functional knowledge. Similarity scores range from 0.0 (dark red, functionally distant) to 1.0 (dark green, functionally similar). Top 5 most similar proteins: PRKCA (0.391), SHC1 (RTK signaling, 0.261), ABL1 (RTK signaling, 0.235), ACVR2B (TGF-beta signaling, 0.157), and INSR (RTK signaling, 0.156). Notably, proteins from the same RTK signaling pathway (SHC1, ABL1, INSR) cluster with high similarity, while PRKCA shows cross-pathway relatedness reflecting shared downstream kinase signaling logic. The embedding similarity reveals biological relationships consistent with known signaling crosstalk, demonstrating the value of RAG-GNN representations for capturing functional relatedness.