DNA shape features improve prediction of CRISPR/Cas9 activity

Area of Science:

• Genomics
• Bioinformatics
• Molecular Biology

Background:

• CRISPR/Cas9 genome editing is revolutionizing biological and medical research.
• Off-target effects and incomplete understanding of Cas9 mechanisms limit its application.
• Machine learning models for predicting Cas9 activity require effective feature engineering.

Purpose of the Study:

• To evaluate the impact of incorporating epigenetic and DNA shape features into sequence-based datasets for predicting Cas9 activity.
• To assess performance improvements in machine learning predictors for CRISPR/Cas9 off-target effects.

Main Methods:

• Utilized neural networks trained on datasets with varying feature groups: sequence only, sequence and epigenetic features, and sequence, epigenetic, and DNA shape features.
• Systematically assessed performance increases using Akaike and Bayesian information criteria.
• Employed permutation and LIME methods to evaluate individual feature importance.

Main Results:

• Inclusion of DNA shape features significantly enhanced predictive performance.
• Sequence features like mismatches and nucleotide composition remained important.
• DNA-RNA hybrid distortion parameters (opening, stretch) also influenced model outcomes.

Conclusions:

• Epigenetic and DNA shape features are crucial for improving the accuracy of Cas9 off-target prediction models.
• A comprehensive feature set, including sequence, epigenetic, and DNA shape data, leads to more reliable CRISPR/Cas9 activity predictions.