piCRISPR: Physically informed deep learning models for CRISPR/Cas9 off-target cleavage prediction

Area of Science:

• Molecular Biology
• Bioinformatics
• Gene Editing Technologies

Background:

• CRISPR/Cas systems are powerful gene editing tools with therapeutic potential.
• Off-target edits pose a significant safety concern for in vivo gene editing applications.
• Existing prediction algorithms often underutilize physically informed features.

Purpose of the Study:

• To develop an advanced off-target prediction model incorporating physically informed features.
• To enhance the accuracy of predicting CRISPR/Cas nuclease cleavage activity.
• To improve the safety and efficacy of gene editing for therapeutic applications.

Main Methods:

• Implementation of state-of-the-art deep learning algorithms.
• Utilization of physically informed features capturing the biological environment of cleavage sites.
• Training and evaluation on the comprehensive crisprSQL off-target cleavage dataset.

Main Results:

• The piCRISPR model demonstrates high performance in off-target prediction.
• Physically informed features, including sequence context and chromatin accessibility, are critical for accurate predictions.
• Novel features significantly improve prediction accuracy for sequence-identical locus pairs.

Conclusions:

• piCRISPR offers a more accurate method for predicting CRISPR/Cas off-target effects.
• The model's environmentally sensitive features are vital for clinical guide design.
• This approach advances the safety of in vivo gene editing therapies.