CROTON: an automated and variant-aware deep learning framework for predicting CRISPR/Cas9 editing outcomes

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• CRISPR/Cas9 gene editing is crucial in biology and medicine.
• Predicting CRISPR/Cas9 editing outcomes relies on local DNA sequences.
• Current prediction methods are limited by manual feature and model engineering.

Purpose of the Study:

• To develop an automated deep-learning framework for predicting CRISPR/Cas9 editing outcomes.
• To overcome limitations of existing prediction methods dependent on expert knowledge.
• To assess the impact of single nucleotide variants on gene editing in clinically relevant genes.

Main Methods:

• Utilized deep multi-task convolutional neural networks (CNNs) and neural architecture search (NAS).
• Developed an end-to-end deep-learning framework named CROTON (CRISPR Outcomes Through cONvolutional neural networks).
• Tuned CROTON architecture with NAS on a large synthetic dataset and validated on primary T cell data.

Main Results:

• CROTON outperformed existing expert-designed models and non-NAS CNNs in predicting insertion/deletion probabilities and frameshift frequency.
• The model identified local sequence determinants influencing diverse editing outcomes.
• CROTON analysis revealed significant SNV-induced differences in editing outcomes for ACE2, CCR5, CTLA4, and PDCD1.

Conclusions:

• CROTON provides an automated and accurate approach for predicting CRISPR/Cas9 editing outcomes.
• The findings highlight the importance of considering local sequence features for precise gene editing.
• Single nucleotide variants can substantially alter genome editing efficiency, necessitating their consideration in gRNA design for clinical applications.