Improving splice site usage prediction with SPLAIRE.

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Alternative splicing impacts over 95% of human genes and is crucial in disease.
• The spliceosome identifies splice sites using sequence motifs at exon-intron junctions.
• Deep learning models have advanced splice site prediction and pathogenic variant identification.

Purpose of the Study:

• To evaluate the performance of current splice site prediction models.
• To develop a novel splicing model optimized for specific cell types.
• To address limitations in predicting low-usage and tissue-specific splice sites.

Main Methods:

• Leveraged one of the largest paired RNA and genotyping datasets.
• Trained a dilated convolutional neural network on human airway epithelial cell data from 100 donors.
• Evaluated model performance on splice site identification and usage quantification.

Main Results:

• Current deep learning models show substantial gaps in predicting low-usage and tissue-specific splice sites.
• The novel model significantly outperforms state-of-the-art methods in splice site identification.
• The model demonstrates superior splice site usage quantification, even on tissues not used in training.

Conclusions:

• A comprehensive evaluation reveals both strengths and weaknesses in current splicing models.
• The developed model offers improved accuracy for splice site prediction and usage quantification.
• Identified key areas for future development in splicing prediction models.