GL4SDA: Predicting snoRNA-disease associations using GNNs and LLM embeddings.

Small nucleolar RNAs (snoRNAs) play essential roles in various cellular processes, and their associations with diseases are increasingly recognized. Identifying these snoRNA-disease relationships is critical for advancing our understanding of their functional roles and potential therapeutic implications. This work presents a novel approach, called GL4SDA, to predict snoRNA-disease associations using Graph Neural Networks (GNN) and Large Language Models. Our methodology leverages the unique strengths of heterogeneous graph structures to model complex biological interactions. Differently from existing methods, we define a set of features able to capture deeper information content related to the inner attributes of both snoRNAs and diseases and design a GNN model based on highly performing layers, which can maximize results on this representation. We consider snoRNA secondary structures and disease embeddings derived from large language models to obtain snoRNAs and disease node features, respectively. By combining structural features of snoRNAs with rich semantic embeddings of diseases, we construct a feature-rich graph representation that improves the predictive performance of our model. We evaluate our approach using different architectures that exploit the capabilities of many graph convolutional layers and compare the results with three other state-of-the-art graph-based predictors. GL4SDA demonstrates improved scores in link prediction tasks and demonstrates its potential implication as a tool for exploring snoRNA-disease relationships. We also validate our findings through biological case studies about cancer diseases, highlighting the practical application of our method in real-world scenarios and obtaining the most important snoRNA features using explainable artificial intelligence methods.