Prediction of the Association between Transfer RNA and Diseases: A Deep Learning Approach Combining Multi-View Graph Convolution and Attention Mechanisms

Background: Identifying the associations between transfer RNA (tRNA) and diseases is critical for disease diagnosis and treatment. Computational methods offer an efficient approach for exploring these associations. Methods: We proposed an MGC2ATDA model, which integrated a multiview graph convolutional network, a scaled attention fusion module, and a cross-attention mechanism to identify tRNA-disease associations. First, tRNA sequence data were compiled into the MNDR4.0 data set, and similarity matrices of tRNAs and diseases were integrated separately with the tRNA-disease association network to construct a multiview network. Next, graph convolutional networks were employed to extract node embeddings from the network, which were further integrated via a scaled attention fusion module, generating high-quality node representations. The cross-attention mechanism then refined these representations and achieved tRNA-disease association prediction. Results: On the tRNA-disease association data set, the MGC2ATDA model achieved AUC and AUPR scores of 0.8786 and 0.3657, respectively, outperforming five comparison methods. When applied to the piRNA-disease association data set, the MGC2ATDA model attains AUC and AUPR scores of 0.9353 and 0.6105, demonstrating strong generalization ability. Ablation experiments validated the scaled attention fusion module's effectiveness. Conclusions: As an efficient and accurate computational method, MGC2ATDA provides a critical tool for identifying potential tRNA-disease associations in biomedical research.