A systematic evaluation of data processing and problem formulation of CRISPR off-target site prediction

Area of Science:

• Genomics
• Molecular Biology
• Bioinformatics

Background:

• CRISPR/Cas9 gene editing is powerful but can cause unintended off-target mutations.
• Existing computational methods for predicting off-target sites (OTSs) are limited by small, low-quality datasets.
• New high-throughput experimental techniques like CHANGE-seq and GUIDE-seq provide larger, higher-quality datasets.

Purpose of the Study:

• To systematically evaluate data processing and model formulation for CRISPR off-target prediction using large-scale datasets.
• To develop improved computational models for predicting off-target cleavage sites.
• To assess the impact of data pre-processing and feature engineering on prediction accuracy.

Main Methods:

• Utilized large-scale in vitro (CHANGE-seq) and in cellula (GUIDE-seq) datasets for CRISPR off-target site detection.
• Systematically evaluated data transformation techniques as a critical pre-processing step.
• Incorporated potential inactive OTSs and the number of mismatches as features in predictive models.
• Developed and compared new predictive models against state-of-the-art methods.

Main Results:

• Data transformation during pre-processing significantly improves model training for OTS prediction.
• Including potential inactive OTSs in training datasets enhances prediction performance.
• The number of mismatches between guide RNA and OTS is an important predictive feature.
• Developed predictive models based on in vitro and in cellula data show competitive performance.

Conclusions:

• High-throughput datasets are crucial for developing accurate CRISPR off-target predictors.
• Effective data processing, feature selection (e.g., mismatch count), and inclusion of diverse data (active/inactive sites) are key to improving prediction accuracy.
• The developed models provide a strong foundation for future advancements in CRISPR safety and specificity assessment.