DRANetSplicer: A Splice Site Prediction Model Based on Deep Residual Attention Networks.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Accurate splice site identification is crucial for gene annotation and understanding gene structure/function.
• Existing methods for splice site prediction have limitations in accuracy and cross-organism applicability.

Purpose of the Study:

• To develop and evaluate DRANetSplicer, a novel deep learning model for precise splice site identification.
• To assess DRANetSplicer's performance against benchmark methods across multiple organisms.
• To investigate DRANetSplicer's cross-organism predictive capabilities.

Main Methods:

• Developed DRANetSplicer, a deep learning model integrating residual learning and attention mechanisms.
• Constructed training datasets using recent genomic data from *Oryza sativa japonica*, *Arabidopsis thaliana*, and *Homo sapiens*.
• Performed comparative analyses with established splice site prediction tools (SpliceFinder, Splice2Deep, Deep Splicer, EnsembleSplice, DNABERT).

Main Results:

• DRANetSplicer achieved high average accuracies (96.57% donor, 95.82% acceptor) across three organisms.
• Demonstrated superior predictive performance, reducing average error rates by 4.2%-11.6% compared to benchmarks.
• Showcased strong cross-organism prediction capabilities, outperforming benchmarks even in non-cross-organism predictions.

Conclusions:

• DRANetSplicer offers enhanced accuracy and robustness for splice site identification.
• The model's cross-organism predictive power supports its use in gene annotation for diverse species, including understudied ones.
• Model interpretability confirmed its ability to learn key biological features, validating its practical application.