A deep learning model developed at the University of Tokyo can analyze DNA sequences spanning up to 100,000 base pairs to predict RNA splice sites with single-nucleotide resolution—addressing a key limitation of existing computational tools that struggle to capture regulatory signals located far from the splice sites they influence. The study was published in Nucleic Acids Research.
Accurate RNA splicing is essential for gene expression and human health, yet predicting how DNA sequence variations affect splicing remains difficult. While recent AI models have improved splice site prediction, most cannot account for regulatory signals located thousands of base pairs away—limiting their ability to interpret disease-causing mutations in conditions ranging from genetic diseases to cancer.
To address this, Kenta Nakai and Yuna Miyachi developed SpliceSelectNet (SSNet), a hierarchical Transformer-based deep learning framework. A central challenge in modeling long-range genomic interactions is that computational cost increases rapidly with sequence length. SSNet addresses this by dividing long DNA sequences into smaller blocks, analyzing local patterns within each block, and then integrating information across the full sequence through a hierarchical attention process. This design preserves dense attention while remaining computationally efficient.
“The key achievement of this work is that we successfully modeled ultra-long-range genomic interactions while preserving high computational efficiency and single-nucleotide resolution,” said Professor Nakai. “We also demonstrated that the regions highlighted by the model closely correspond to biologically meaningful regulatory elements, helping to bridge predictive accuracy and biological interpretability.”
SSNet was trained and benchmarked against leading splice prediction systems across multiple large genomic datasets, achieving strong performance for both splice site prediction and aberrant splicing detection. In simulations using the DMD gene and evaluations of pathogenic variants from ClinVar, the model maintained sensitivity to regulatory signals located many thousands of base pairs from the affected splice site—beyond the effective range of conventional approaches.
“Many existing AI models for DNA analysis were adapted from natural language processing, but DNA has fundamentally different properties,” said Miyachi. “By redesigning the architecture to account for long-range genomic interactions and strict sequence resolution, we aimed to create a system better suited to biological reality.”
The researchers suggest the hierarchical Transformer architecture could support future work in promoter-enhancer interactions, three-dimensional genome organization, and broader DNA language models. In clinical settings, the approach could help screen variants in non-coding regions of uncertain significance, and in pharmaceutical research it could assist in designing oligonucleotide therapeutics that target abnormal splicing.
Accurate RNA splicing is essential for gene expression and human health, yet predicting how DNA sequence variations affect splicing remains difficult. While recent AI models have improved splice site prediction, most cannot account for regulatory signals located thousands of base pairs away—limiting their ability to interpret disease-causing mutations in conditions ranging from genetic diseases to cancer.
To address this, Kenta Nakai and Yuna Miyachi developed SpliceSelectNet (SSNet), a hierarchical Transformer-based deep learning framework. A central challenge in modeling long-range genomic interactions is that computational cost increases rapidly with sequence length. SSNet addresses this by dividing long DNA sequences into smaller blocks, analyzing local patterns within each block, and then integrating information across the full sequence through a hierarchical attention process. This design preserves dense attention while remaining computationally efficient.
“The key achievement of this work is that we successfully modeled ultra-long-range genomic interactions while preserving high computational efficiency and single-nucleotide resolution,” said Professor Nakai. “We also demonstrated that the regions highlighted by the model closely correspond to biologically meaningful regulatory elements, helping to bridge predictive accuracy and biological interpretability.”
SSNet was trained and benchmarked against leading splice prediction systems across multiple large genomic datasets, achieving strong performance for both splice site prediction and aberrant splicing detection. In simulations using the DMD gene and evaluations of pathogenic variants from ClinVar, the model maintained sensitivity to regulatory signals located many thousands of base pairs from the affected splice site—beyond the effective range of conventional approaches.
“Many existing AI models for DNA analysis were adapted from natural language processing, but DNA has fundamentally different properties,” said Miyachi. “By redesigning the architecture to account for long-range genomic interactions and strict sequence resolution, we aimed to create a system better suited to biological reality.”
The researchers suggest the hierarchical Transformer architecture could support future work in promoter-enhancer interactions, three-dimensional genome organization, and broader DNA language models. In clinical settings, the approach could help screen variants in non-coding regions of uncertain significance, and in pharmaceutical research it could assist in designing oligonucleotide therapeutics that target abnormal splicing.