PAMmla: a new model for designing precision CRISPR tools

A new machine learning tool predicts safer, smarter CRISPR enzymes.

Researchers from Mass General Brigham (MA, USA) have unveiled a new approach that could significantly improve the accuracy and customizability of CRISPR-Cas9 enzymes, which could boost progress in the field of cell and gene therapy. By combining high-throughput protein engineering with machine learning, the team developed a model that predicts the properties of over 64 million Cas9 enzymes to enable researchers to design safer and more effective tools for gene editing.

CRISPR-Cas9 has revolutionized genome editing by allowing scientists to target and alter specific DNA sequences. However, one major limitation of current Cas9 systems is the risk of off-target effects, where the enzyme cuts unintended areas of the genome. These unintended edits can have serious consequences, particularly in therapeutic applications.

This research sought to overcome this issue by creating more targeted and efficient versions of the Cas9 enzyme through a combination of high-throughput protein engineering and machine learning.

The team started by performing saturation mutagenesis, which involves systematically mutating each amino acid in the protein, to create a large library of nearly 1,000 engineered variants of the widely used Streptococcus pyogenes Cas9 enzyme. They then characterized how each enzyme interacted with protospacer-adjacent motifs (PAMs); the short stretches of DNA that act like molecular signposts, telling Cas9 where to bind and cut. Using this dataset, they trained a neural network called PAMmla to predict how changes in the amino acid sequence of Cas9 would affect PAM recognition.

PAMmla was then used to analyze 64 million hypothetical Cas9 enzymes, identifying novel variants with highly specific and efficient activity profiles. The researchers validated several of these enzymes through lab-based proof-of-concept experiments, including in primary human cells and a mouse model of retinitis pigmentosa.

The new enzymes outperformed many existing Cas9 variants in terms of precision and efficiency. Not only did they demonstrate strong on-target activity, but they also showed significantly reduced off-target effects. Importantly, the study highlighted how PAMmla could be used to design enzymes for specific tasks, such as allele-selective targeting of the RHO P23H mutation, a genetic cause of blindness.

This research represents a major leap in the customization of genome editing tools. Rather than relying on broad, generalist enzymes, researchers can now predict and design bespoke Cas9 variants tuned for specific genetic sequences and therapeutic needs. The public release of the PAMmla web tool (available here) means the broader scientific community can begin applying this powerful method to their own gene-editing challenges.