Transformers significantly improve splice site prediction
2024

Transformers Improve Splice Site Prediction

Sample size: 17848 publication 20 minutes Evidence: high

Author Information

Author(s): Jónsson Benedikt A., Halldórsson Gísli H., Árdal Steinþór, Rögnvaldsson Sölvi, Einarsson Eyþór, Sulem Patrick, Guðbjartsson Daníel F., Melsted Páll, Stefánsson Kári, Úlfarsson Magnús Ö.

Primary Institution: deCODE Genetics/Amgen Inc., Reykjavik, Iceland

Hypothesis

Large contexts are essential for tracking splice sites in pre-mRNA sequences.

Conclusion

The transformer-based method significantly outperforms existing tools in detecting splice sites and disease-related variants.

Supporting Evidence

  • The transformer model achieved a PR-AUC of 0.834, outperforming SpliceAI's 0.820.
  • Transformer-45k detected 744 more splice sites than SpliceAI-10k.
  • The model demonstrated superior performance in identifying disease-related splice variants in ClinVar.

Takeaway

This study shows that a new method using transformers can better find important parts of genes that help make proteins, which can help us understand diseases.

Methodology

The study used a transformer model trained on RNA sequencing data to predict splice sites from long nucleotide sequences.

Limitations

The model's performance may be affected by the quality of splice site annotations from RNA-Seq data.

Participant Demographics

The study involved an Icelandic cohort of 17,848 individuals, including 9,784 females and 8,064 males.

Statistical Information

Confidence Interval

[0.833, 0.835]

Digital Object Identifier (DOI)

10.1038/s42003-024-07298-9

Want to read the original?

Access the complete publication on the publisher's website

View Original Publication