Feature representation for explainable CRISPR off-target prediction and base editing efficiency

Area of Science:

• Molecular Biology
• Bioinformatics
• Gene Editing Technologies

Background:

• CRISPR/Cas9 gene editing efficacy relies on guide RNA-target DNA interactions.
• Accurate prediction models require biological insight beyond high accuracy.
• Off-target prediction is crucial for safe and effective CRISPR applications.

Purpose of the Study:

• To analyze the impact of feature representation on accuracy and interpretability in CRISPR/Cas9 off-target prediction.
• To compare different sequence encoding methods for gene knockout and base editing.
• To evaluate model performance using classification and regression tasks.

Main Methods:

• Utilized CHANGE-seq, GUIDE-seq, and Hanna screening datasets for gene knockout (KO) and base editing (BE) applications.
• Employed XGBoost models for both classification and regression prediction tasks.
• Evaluated paired sequence representations and K-mer encodings.

Main Results:

• For KO, One-Hot (OH, OH5C) excelled on GUIDE-seq, while Bulges representation was best for CHANGE-seq.
• One-hot encoding consistently outperformed K-mer for BE prediction accuracy in both tasks.
• SHapley Additive exPlanations (SHAP) revealed the Protospacer Adjacent Motif (PAM)-proximal region is critical for both KO and BE.

Conclusions:

• Predictive performance is comparable across gene knockout and base editing, showing framework robustness.
• Feature representation significantly influences prediction accuracy and interpretability.
• The PAM-proximal region is a key determinant for CRISPR/Cas9 editing outcomes.