Splice2Deep

Splice2Deep predicts donor and acceptor splice sites in genomic DNA using deep convolutional neural networks to improve exon–intron boundary annotation and the identification of alternative splice sites relevant to disease.

Key Features:

• Model architecture: An ensemble of deep convolutional neural networks (CNNs) is used for splice site prediction.
• Splice site types: Predicts donor and acceptor splice sites (SS) that demarcate exon and intron boundaries.
• Cross-organism generalization: Models can detect splice sites in organisms not included during training and enable selection of a taxonomically closest model.
• Evaluation species: Performance was evaluated across Homo sapiens, Oryza sativa japonica, Arabidopsis thaliana, Drosophila melanogaster, and Caenorhabditis elegans.
• Performance metrics: Reported average error rates of 41.97% for acceptor SS and 28.51% for donor SS.
• Implementation: Models are implemented in Python using the Keras API.

Scientific Applications:

• Gene annotation: Improves precise annotation of exon and intron boundaries in multi-exon genes.
• Alternative splicing analysis: Supports identification of alternative splice sites that may be linked to disease.
• Annotation of novel genomes: Facilitates annotation of splice sites in newly sequenced or poorly studied genomes by leveraging taxonomically related models.
• Conserved element detection: Aids detection of conserved genomic splice signals across diverse species.

Methodology:

An ensemble of deep convolutional neural networks was implemented in Python with the Keras API and trained and evaluated across Homo sapiens, Oryza sativa japonica, Arabidopsis thaliana, Drosophila melanogaster, and Caenorhabditis elegans, yielding reported average error rates of 41.97% (acceptor) and 28.51% (donor).