ESAE-SDA: ensemble sparse autoencoder framework for epigenomics-informed snoRNA-disease associations prediction

Small nucleolar RNAs (snoRNAs), a class of non-coding RNAs broadly distributed in eukaryotes, are emerging as pivotal regulators in the field of epigenomics. In addition to guiding 2'-O-methylation and pseudouridylation modifications at specific rRNA sites to maintain ribosomal stability and support protein synthesis, snoRNAs have been increasingly implicated in epigenetic regulation, influencing gene expression, chromatin architecture, and RNA modification patterns. Accurate identification of potential snoRNA-disease associations (SDAs) is therefore essential for understanding epigenomic dysregulation in complex diseases and facilitating early intervention and drug repurposing. Although artificial intelligence (AI) methods have advanced SDA prediction, they are still hindered by issues such as sample imbalance and high false-negative rates. To address these challenges, we propose ESAE-SDA, a novel model integrating sparse autoencoders with an ensemble learning framework. ESAE-SDA first constructs a comprehensive snoRNA-disease representation using multi-source similarity metrics. It then applies k-means clustering to select high-confidence negative samples and employs a deep sparse autoencoder with sparsity constraints to learn compact, discriminative embeddings. Finally, multiple GNN-based learners are independently trained on dynamically resampled data, and ensemble inference is performed via weighted fusion, substantially enhancing robustness and generalization. Experiments on a public SDA dataset demonstrate that ESAE-SDA consistently outperforms state-of-the-art methods. Notably, a case study on ophthalmic diseases highlights the model's ability to uncover epigenetically relevant snoRNAs with potential regulatory and therapeutic significance, underscoring its value in epigenomics-driven disease research and target discovery.