CrisprDA: A Data Augmentation Method Enhancing the Efficiency of sgRNA on-Target Activity Prediction

Area of Science:

• Genomics and Bioinformatics
• Molecular Biology
• Artificial Intelligence in Life Sciences

Background:

• CRISPR/Cas9 genome editing efficiency relies on single-guide RNA (sgRNA) activity.
• Predicting sgRNA activity using deep learning is hindered by data scarcity.
• Existing methods struggle with limited training data for accurate sgRNA activity prediction.

Purpose of the Study:

• To develop novel data augmentation and pseudo-label correction techniques to improve deep learning model performance for sgRNA activity prediction.
• To introduce CrisprDA, a new deep learning architecture for precise sgRNA activity prediction.
• To validate the effectiveness and generalizability of proposed methods on diverse datasets.

Main Methods:

• Automix: An autoencoder-based data augmentation method to address data scarcity.
• CNLC (Confidence-based Nearest Label Correction): A pseudo-label correction technique to enhance training data quality and quantity.
• CrisprDA: A parallel deep learning architecture integrating CNNs and attention mechanisms for sgRNA activity prediction.

Main Results:

• CrisprDA demonstrated superior predictive ability, outperforming five other methods across nine high-throughput and eight functional datasets.
• The application of Automix and CNLC improved the performance of other comparative methods, confirming their effectiveness.
• The study successfully enhanced the accuracy and reliability of sgRNA activity prediction models.

Conclusions:

• Automix and CNLC are effective strategies for overcoming data scarcity in deep learning for sgRNA activity prediction.
• CrisprDA offers a powerful new tool for precise sgRNA activity prediction, advancing CRISPR/Cas9 technology.
• The proposed methods provide a generalizable approach to improve genome editing efficiency through enhanced predictive modeling.