Deep Learning Based Models for CRISPR/Cas Off-Target Prediction

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• CRISPR/Cas genome editing offers precise genetic modification but faces challenges with off-target effects (OTS).
• Accurate prediction of OTS is crucial for safe clinical applications of CRISPR/Cas technology.
• In silico methods, especially deep learning, are emerging as powerful tools for OTS prediction due to their ability to learn complex sequence features.

Purpose of the Study:

• To review existing off-target site (OTS) prediction tools, focusing on deep learning methods.
• To characterize datasets used for training and testing deep learning models for OTS prediction.
• To evaluate and compare the performance of six prominent deep learning models for CRISPR/Cas OTS prediction.

Main Methods:

• Reviewed current OTS prediction tools, emphasizing deep learning approaches.
• Analyzed datasets utilized for deep learning model training and validation.
• Evaluated six deep learning models (CRISPR-Net, CRISPR-IP, R-CRISPR, CRISPR-M, CrisprDNT, Crispr-SGRU) using six public datasets and the CRISPRoffT database.
• Assessed model performance using metrics like Precision, Recall, F1 score, MCC, AUROC, and PRAUC.

Main Results:

• Deep learning models demonstrate potential for predicting CRISPR/Cas off-target sites.
• The inclusion of validated OTS datasets in training significantly enhanced model performance and prediction robustness, especially for imbalanced datasets.
• CRISPR-Net, R-CRISPR, and Crispr-SGRU exhibited strong overall performance across various evaluation scenarios.
• No single model consistently outperformed all others across all tested conditions.

Conclusions:

• High-quality, validated off-target site data is essential for improving the accuracy and reliability of deep learning-based predictions.
• Advanced deep learning architectures, when trained with appropriate data, can significantly enhance the prediction of CRISPR/Cas off-target sites.
• Integrating robust prediction tools is vital for ensuring the safety and efficacy of CRISPR/Cas genome editing in clinical settings.