A generalizable Cas9/sgRNA prediction model using machine transfer learning with small high-quality datasets.

Area of Science:

• Microbiology
• Molecular Biology
• Bioinformatics

Background:

• The CRISPR/Cas9 system from Streptococcus pyogenes (SpCas9) is a versatile tool for antimicrobial applications and bacterial genome engineering.
• Existing models for predicting bacterial single guide RNA (sgRNA) activity lack accuracy and generalizability, partly due to limitations in training datasets that conflate SpCas9/sgRNA activity with cellular toxicity.

Purpose of the Study:

• To develop a high-quality dataset for training SpCas9/sgRNA activity prediction models.
• To create a machine learning architecture, crisprHAL, for accurate and generalizable prediction of sgRNA activity in bacteria.
• To improve the design of sgRNAs for antimicrobial and genome engineering applications.

Main Methods:

• Utilized a two-plasmid positive selection system to generate high-fidelity data distinguishing SpCas9/sgRNA cleavage activity from toxicity.
• Developed the crisprHAL machine learning model, incorporating transfer learning capabilities.
• Validated the model on existing datasets and tested its generalizability across different bacterial species.

Main Results:

• Generated a novel, high-quality dataset for SpCas9/sgRNA activity assessment.
• The crisprHAL model demonstrated significant improvements in sgRNA activity prediction accuracy, especially when fine-tuned with limited high-quality data.
• crisprHAL successfully recapitulated known SpCas9/sgRNA-target DNA interactions and showed generalizability to various bacterial species.

Conclusions:

• The crisprHAL model offers a robust solution for predicting bacterial SpCas9/sgRNA activity, overcoming limitations of previous models.
• This tool enhances the reliability of sgRNA design for both sequence-specific antimicrobials and precise bacterial genome engineering.
• crisprHAL represents a significant step towards a universal prediction tool for bacterial CRISPR-based applications.