Improving spliced alignment by modeling splice sites with deep learning.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Spliced alignment of RNA or protein sequences to eukaryotic genomes is crucial for gene annotation and function studies.
• Current methods often use simplified splice site models, limiting accuracy with dissimilar sequences.

Purpose of the Study:

• To develop a sophisticated model for splice site recognition to enhance spliced alignment accuracy.
• To improve the analysis of gene functions and annotations, particularly for challenging datasets.

Main Methods:

• Implemented minisplice, a tool employing a 1D-CNN to learn splice signals from genomic data.
• Trained a 7026-parameter model on vertebrate and insect genomes, capturing conserved and species-specific splice signals.
• Integrated empirical splicing probabilities into minimap2 and miniprot aligners.

Main Results:

• The minisplice model identified conserved splice signals across phyla and GC-rich introns in mammals and birds.
• Empirical splicing probabilities were estimated for all GT/AG dinucleotides.
• Modified aligners demonstrated significantly improved junction accuracy on human long-read RNA-seq and cross-species protein data.

Conclusions:

• The 1D-CNN approach effectively models splice signals, outperforming traditional methods.
• Leveraging empirical splicing probabilities enhances alignment accuracy for noisy and evolutionarily distant sequences.
• This method offers a substantial improvement for gene annotation and functional genomics research.