Benchmarking deep learning methods for predicting CRISPR/Cas9 sgRNA on- and off-target activities.

Area of Science:

• Genomics and Molecular Biology
• Bioinformatics and Computational Biology

Background:

• CRISPR/Cas9 gene editing relies on effective single guide RNA (sgRNA) design for specificity and efficiency.
• Deep learning methods have shown promise in predicting sgRNA on-target activity and off-target mutations.

Approach:

• Systematic survey and performance evaluation of 10 deep learning on-target predictors across nine public datasets.
• Unbiased experimental comparison of eight off-target prediction methods on 12 public datasets with varying sample imbalance.

Key Points:

• Deep learning models generally perform better on large- and medium-scale datasets for on-target prediction.
• Most off-target prediction methods perform well on balanced datasets but struggle with imbalanced data.
• Performance of predictive models is sensitive to dataset size and sample balance.

Conclusions:

• This review offers comprehensive insights into the current state of CRISPR/Cas9 sgRNA activity prediction.
• Identifies areas for improvement in deep learning model development for both on- and off-target prediction.
• Highlights the need for robust methods that can handle imbalanced datasets in off-target analysis.