REVIEW PAPER ON APLICATIONS OF GRAPH NEURAL NETWORKS AND BIOMEDICAL LINK PREDICTION

Abstract: Graph neural networks (GNN) provide a natural framework for learning from relational biomedical data, where drugs, proteins, diseases and adverse events are connected through heterogeneous interactions. This review examines twelve papers spanning foundational Graph Neural Network (GNN) architectures, drug-drug interaction (DDI) prediction models, biomedical knowledge graphs and recent projects in attention, multimodal learning, contrastive learning and cold-start prediction. Architectures like Graph Convoluted Networks (GCN) [1] establish an efficient neighborhood propagation method; Graph Attention Network (GAT) [2] introduces learned neighbor attention; architectures like Relational Graph Convoluted Network (R-GCN) [3] and Relational Graph Attention Networks (r-GAT) [12] incorporate relation-specific transformations. Decagon [4] extends relational graph learning to polypharmacy side effects; and Multimodal Deep Neural Network (MDNN) [5] combines graph and heterogeneous drug features. More recent approaches further incorporate multiple features to current approaches like the Deep Attention Neural Network for Drug- Drug Interations (DANN-DDI) [6], GNN which uses a novel approach of Simplified Molecular Input Line Entry System (SMILES) strings (SmileGNN) [7], or a cross-attention GNN [8], a flow based GNN for emerging drug interations (EmerGNN) [9], a Knowledge Graph based DDI model with contrastive learning (KG-CLDDI) [10], and models which use large-scale biomedical knowledge graphs such as PrimeKG [11]. Rather than comparing numerical scores directly across incompatible datasets, the review focuses on how each architectural choice addresses a specific limitation of earlier approaches. The comparison highlights four recurring requirements for adverse drug event prediction: preserving relation semantics, integrating complementary biological views, handling sparse or unseen entities, and maintaining interpretability and scalability. These observations motivate a three-branch GAT architecture for the proposed ADE project.

Keywords: Drug-Drug Interaction Prediction, Graph Neural Networks, Knowledge Graphs, Link Prediction, Multimodal Learning.