Improved CRISPR/Cas9 off-target prediction with DNABERT and epigenetic features

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• CRISPR/Cas9 genome editing offers therapeutic potential but is limited by off-target effects.
• Accurate prediction of unintended edits is essential for clinical safety and efficacy.
• Existing deep learning models often lack comprehensive genomic knowledge due to task-specific training.

Purpose of the Study:

• To develop and evaluate a novel computational approach for predicting CRISPR/Cas9 off-target effects.
• To integrate a genome-pre-trained deep learning model (DNABERT) with epigenetic features.
• To assess the performance of the integrated model against state-of-the-art methods.

Main Methods:

• Utilized DNABERT, a deep learning model pre-trained on the human genome.
• Integrated epigenetic features including H3K4me3, H3K27ac, and ATAC-seq data.
• Benchmarked the DNABERT-Epi model against five leading methods on seven off-target datasets.
• Employed SHAP and Integrated Gradients for model interpretability.

Main Results:

• DNABERT-Epi achieved competitive or superior performance compared to existing methods.
• Ablation studies confirmed the critical contributions of genomic pre-training and epigenetic features to predictive accuracy.
• Interpretability techniques identified key epigenetic marks and sequence patterns influencing predictions.

Conclusions:

• This study demonstrates the significant potential of pre-trained DNA foundation models for CRISPR/Cas9 off-target prediction.
• Combining large-scale genomic knowledge with multi-modal data is a promising strategy for enhancing genome editing safety.
• The findings pave the way for developing more precise and safer genome editing therapeutics.