Predicting ncRNA-Protein interactions with a graph attention model exploiting personalized subgraphs

Predicting interactions between ncRNAs and proteins is crucial for advancing our understanding of gene regulation, disease mechanisms, targeted drug design, and biomarker discovery, thereby driving innovation in research and therapeutic development. Numerous computational methods, particularly those employing machine learning and deep learning, have been proposed to address this challenge. Recent studies show that graph neural networks (GNNs) enhance ncRNA-protein interaction prediction accuracy by capturing intricate relationships and structural details in molecular data. However, current GNN approaches frequently rely on fixed-hop subgraphs for structural analysis, limiting their capacity to capture diverse interaction patterns fully. This fixed-hop approach may omit crucial nodes and edges outside the predefined neighborhood, potentially reducing prediction accuracy. To overcome this constraint, we introduce a novel method for ncRNA-protein interaction prediction by extracting the most informative subgraphs around each interaction using the personalized subgraph selection framework. These subgraphs are then utilized in a graph attention network (GAT) to learn node representations. K-mer frequencies are used to capture sequence-level features, while node2vec embeddings capture structural information, providing the GNN with a robust set of features. Experimental results on relevant datasets indicate a significant improvement in predicting ncRNA-protein interactions, with the algorithm maintaining an acceptable level of computational complexity even on large datasets. By integrating both sequence and structural insights through personalized subgraphs, this approach delivers a more accurate and scalable solution for predicting ncRNA-protein interactions.