SpliceSelectNet: a hierarchical Transformer-based deep learning model for splice site prediction.
Summary
Accurate RNA splicing is vital for gene expression, yet its mechanisms and the impact of mutations, which can cause diseases like cancer, are not fully understood. Current prediction tools struggle with long-range dependencies and often lack interpretability. SpliceSelectNet (SSNet) is a novel hierarchical Transformer-based deep learning model designed to predict splice sites from DNA sequences up to 100 kb. It integrates local and global attention mechanisms to efficiently capture both proximal and distal regulatory signals. SSNet achieves state-of-the-art performance in splice site prediction and aberrant splicing detection, offering a biologically interpretable framework for modeling long-range splicing regulation.
Analysis
This study introduces SpliceSelectNet (SSNet), a deep learning model based on a hierarchical Transformer architecture, representing a novel application for splice site prediction. It stands out for its ability to handle long-range dependencies (up to 100 kb) through integrated local and global attention mechanisms, which is an advancement over existing methods. The model was benchmarked on multiple datasets, where it achieved state-of-the-art performance in splice site prediction and aberrant splicing detection. Interpretability is a key strength, with attention scores reflecting functional sequence importance, which could facilitate the understanding of underlying biological mechanisms. If clinically deployed, SSNet could enhance genetic variant interpretation by more accurately identifying splicing-affecting mutations, potentially streamlining variant prioritization for molecular tumor boards.