SpliceRead: Improving Canonical and Non-Canonical Splice Site Prediction with Residual Blocks and Synthetic Data...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Accurate splice site prediction is crucial for understanding gene expression and associated disorders.
• Existing models often exhibit bias towards canonical splice sites, leading to poor detection of rare, non-canonical variants and high false-negative rates due to imbalanced datasets.

Purpose of the Study:

• To develop a novel deep learning model, SpliceRead, for improved classification of both canonical and non-canonical splice sites.
• To address the limitations of existing models in detecting rare splice site variants and improve overall splice site prediction accuracy.

Main Methods:

• Developed SpliceRead, a deep learning model incorporating residual convolutional blocks and synthetic data augmentation.
• Utilized a data augmentation strategy to generate diverse non-canonical splice site sequences.
• Employed residual connections to facilitate gradient flow and capture subtle genomic sequence features.

Main Results:

• SpliceRead demonstrated superior performance over state-of-the-art models across key metrics (F1-score, accuracy, precision, recall) on a multi-species dataset.
• Achieved a significantly lower misclassification rate for non-canonical splice sites compared to baseline methods.
• Validation through cross-validation, cross-species testing, and input-length generalization confirmed model robustness and adaptability.

Conclusions:

• SpliceRead provides a robust and generalizable framework for splice site prediction, particularly effective for low-frequency, non-canonical variants.
• The model enhances the accuracy of gene annotation in both model and non-model organisms.
• The open-sourced code facilitates further research and application in splice site analysis.