Transformers significantly improve splice site prediction.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Mutations affecting RNA splicing are crucial in human diversity and disease.
• Accurate detection of splicing events is vital for genetic research and diagnostics.

Purpose of the Study:

• To introduce a novel machine learning method for detecting RNA splicing from long nucleotide sequences.
• To evaluate the performance of this new method against existing state-of-the-art tools.

Main Methods:

• Utilized transformers, a type of machine learning model, for splicing detection.
• Generated embeddings with residual neural networks and applied hard attention for efficient long-sequence training.
• Tested the method on GENCODE and ENSEMBL annotations, and RNA sequencing data from Icelandic and GTEx cohorts.

Main Results:

• The new method demonstrated superior performance in detecting splice sites compared to the leading tool, SpliceAI.
• Achieved higher precision-recall AUC (0.834 vs. 0.820) for splice junction detection in large cohorts.
• Showed greater effectiveness in identifying disease-related splice variants in ClinVar (PR-AUC = 0.997 vs. 0.996).

Conclusions:

• The developed transformer-based method offers enhanced accuracy for RNA splicing detection.
• This advancement has significant potential for improving genetic research and clinical diagnostics for splicing-related disorders.