Abstract
Head and neck cancers represent a critical global health issue, contributing to substantial morbidity and mortality. Recent research has explored the role of microRNAs (miRNAs) in these cancers by constructing miRNA-associated disease networks using bipartite graphs. Graph attention networks (GATs) have emerged as a powerful tool for predicting disease associations within such biological networks, offering enhanced accuracy in identifying potential miRNA-disease relationships. This study employs GATs to uncover and predict potential miRNA contributors to head and neck cancers. Data on miRNA-disease associations were sourced from the HMDD v4.0 database, a platform based on SQLite and Django. The head and neck neoplasms dataset included miRNA, disease, causality, category, and PubMed ID (PMID). GATs were applied to analyze the network, leveraging their ability to capture the significance and interdependencies of nodes and edges. The model used a learnable weight matrix to compute attention coefficients, normalize them, and aggregate information from neighboring nodes for edge prediction. The GAT model, integrating graph neural networks with attention mechanisms, achieved an accuracy of 83% in predicting miRNA-disease associations for head and neck neoplasms. This study highlights the potential of graph-based deep learning models, particularly GATs, in accurately predicting miRNA-disease associations. A functional enrichment analysis revealed significant involvement of miRNAs in oral cancer pathways, notably highlighting the critical roles of the TGF-beta and PI3K-Akt signaling pathways in tumor progression and cell survival. These findings offer a pathway to better understanding the molecular mechanisms underlying head and neck cancers. Future improvements in dataset size, model evaluation, and interpretability could further enhance prediction accuracy, potentially advancing diagnostic and therapeutic strategies for these cancers.