SpliceFinder: ab initio prediction of splice sites using convolutional neural network

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Accurate splice site identification is crucial for gene structure analysis.
• Common splice site dinucleotides (GT, AG) also appear in non-splice sites, leading to prediction challenges.
• Existing methods often miss non-canonical splice sites while reducing false positives.

Purpose of the Study:

• To develop an accurate ab initio splice site prediction tool.
• To address challenges of false positives and missing non-canonical splice sites.
• To create a versatile tool applicable across different species.

Main Methods:

• Developed SpliceFinder, a convolutional neural network (CNN) model.
• Utilized human genomic data for training.
• Employed an iterative dataset reconstruction to handle data imbalance.
• Implemented a sliding window approach for long sequence analysis.

Main Results:

• Achieved 90.25% classification accuracy, a 10% improvement over existing algorithms.
• Outperformed other methods in AUC, recall, precision, and F1 score.
• Reduced false positives by half while maintaining a recall above 0.8.
• Successfully identified non-canonical splice sites.
• Demonstrated cross-species transferability without retraining (Drosophila, mouse, rat, zebrafish).

Conclusions:

• SpliceFinder, a CNN-based tool, effectively predicts splice sites with reduced false positives.
• The tool accurately detects non-canonical splice sites.
• SpliceFinder shows significant potential for cross-species genomic analysis.