Interpretable CRISPR/Cas9 off-target activities with mismatches and indels prediction using BERT

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• CRISPR/Cas9 gene editing efficiency is limited by off-target effects.
• Existing deep learning models struggle to predict off-target activity with both mismatches and indels.
• Data imbalance and model interpretability pose challenges in off-target prediction.

Purpose of the Study:

• To develop a novel BERT-based model, CRISPR-BERT, for accurate off-target activity prediction including mismatches and indels.
• To address data imbalance using an adaptive batch-wise class balancing strategy.
• To enhance model interpretability through visualization of nucleotide position-dependent patterns.

Main Methods:

• Developed CRISPR-BERT, a BERT-based deep learning model for predicting off-target effects.
• Implemented an adaptive batch-wise class balancing strategy to handle imbalanced datasets.
• Utilized a visualization approach to analyze nucleotide-level patterns influencing off-target activity.

Main Results:

• CRISPR-BERT outperformed existing methods on multiple datasets for predicting off-target activities with mismatches and indels.
• Achieved superior performance based on AUROC and PRAUC metrics.
• Visualization analysis revealed generalizable patterns and demonstrated CRISPR-BERT's interpretability.

Conclusions:

• CRISPR-BERT offers an accurate and interpretable framework for predicting CRISPR/Cas9 off-target effects.
• The model aids in optimizing single-guide RNA (sgRNA) design for enhanced target specificity.
• This contributes to safer and more effective genome editing applications.