EDeepSSP: Explainable deep neural networks for exact splice sites prediction

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Splice site prediction is vital for gene regulation, function, and genome annotation.
• Existing computational methods often lack interpretability and require manual feature engineering.
• Deep learning offers a potential solution for automated feature extraction and improved prediction accuracy.

Purpose of the Study:

• To introduce EDeepSSP, a deep learning model for splice site prediction.
• To address the challenges of interpretability and manual feature extraction in current methods.
• To automatically identify and explain the significance of motifs in splice site prediction.

Main Methods:

• Utilized convolutional neural networks (CNNs) for automatic feature extraction.
• Developed the EDeepSSP model for predicting splice sites.
• Conducted experiments on human, cress, and fly splice site datasets (acceptors and donors).

Main Results:

• EDeepSSP demonstrated superior performance compared to state-of-the-art methods.
• Achieved high AUC_ROC (99.32%) and AUC_PR (99.26%) on human donor datasets.
• Successfully extracted and validated significant motifs against the JASPAR database.

Conclusions:

• EDeepSSP provides an interpretable deep learning approach for splice site prediction.
• The model enhances gene annotation and understanding of gene regulation.
• Automatic feature extraction via CNNs significantly improves prediction accuracy and efficiency.