SPGA: graph representation learning and attention fusion for enhanced disease-associated snoRNA prediction

BACKGROUND: Small nucleolar RNAs (snoRNAs) are increasingly recognized for their involvement in human diseases. Accurate and robust prediction of disease-associated snoRNAs is crucial for accelerating drug discovery and disease treatment. However, the limited availability of biomedical data in this domain poses a significant challenge to the generalization ability of machine learning models. While existing computational methods have made progress, their performance is often constrained by data scarcity. RESULTS: To overcome these limitations, we introduce SPGA, a novel graph representation learning framework designed to enhance snoRNA-disease association prediction. SPGA leverages intrinsic structural features of snoRNAs and diseases to construct interaction-aware graph representations. It then employs residual graph convolutional networks augmented with a hierarchical attention mechanism to learn robust node embeddings for association predictions. Comprehensive experiments demonstrate that SPGA effectively alleviates the data scarcity problem for model generalization and significantly improves state-of-the-art methods in terms of prediction accuracy, achieving an AUC of 0.9812 and an AUPR of 0.9749 under five-fold cross-validation setting. Case studies further validate its efficacy in identifying novel snoRNA-disease associations. CONCLUSIONS: Our study provides a potent computational tool for prioritizing candidate disease-related snoRNAs, thereby facilitating the discovery of potential diagnostic biomarkers and therapeutic targets.