A GNN Model for the Prediction of snoRNA-Disease Associations Based on snoRNA Secondary Structures and LLM Disease Embedding

Numerous studies have demonstrated the functional role of small nuclear RNAs (snoRNAs) in various biological processes associated with developing complex human disorders. Therefore, understanding the connections between different snoRNAs and diseases is essential for improving disease detection and therapy. In this work, we propose a graph neural network model to predict unknown snoRNA-disease associations. Our network consists of four layers, each constructed by a sequence of SAGEConv and GATConv layers. We take into account two class of features for both snoRNAs and disease never used by similar works. We generate the snoRNA node features according to substructures of varying sizes within their secondary structures; we obtain the disease node features from disease textual descriptions using two large language models, one trained on medical texts and the other trained on different text type, including medical text. Our results indicate that our model outperforms current state-of-the-ar model and in particular that the snoRNA features derived from smaller substructures are the most suitable for this problem, whereas the disease features that yielded the best performance were those extracted using a model specifically trained on medical texts.