Machine Learning Unlocks Custom CRISPR Enzymes for Precision Editing

In research published last week in Nature, Rachel Silverstein, Ben Kleinstiver (Massachusetts General Hospital (MGH) and colleagues at Mass General Brigham and Harvard Medical School addressed a major limitation of traditional CRISPR-Cas9 technology: the need for specific protospacer adjacent motifs (PAMs) in close proximity to editing targets. Rather than developing so-called relaxed PAM endonucleases that recognise more PAMs but sacrifice specificity, they took a different approach.

»Previous approaches to engineer CRISPR-Cas9 enzymes are generally lower throughput and typically yield few enzymes, « explains Ben Kleinstiver. »We developed a more scalable methodology to engineer and characterise the biochemical properties of hundreds or thousands of novel Cas9 proteins.«

The team mutated six key amino acids in Cas9's PAM-interacting domain, creating a library of 64 million potential variant enzymes. After characterising almost 1,000 engineered enzymes, they used the resulting data to train a neural network that relates amino acid sequence to PAM specificity. The result was PAMmla, a machine learning algorithm that predicts PAM recognition properties for any of the 64 million enzymes.

Importantly, the researchers showed that the bespoke Cas9 enzymes don't just work in theory. They demonstrated that engineered PAMmla-predicted enzymes could selectively target the P23H mutation in the Rhodopsin gene, which leads to retinitis pigmentosa (RP), effectively editing the disease-causing allele in vitro and in vivo in a humanised heterozygous model of RP, while leaving the healthy allele intact.

The team also demonstrated that their custom enzymes could be incorporated into base-editing systems (based on ABE8e and TadCBEd architectures), where they outperformed conventional PAM-relaxed variants with superior editing efficiency at target sites harbouring their preferred PAMs.

Unlike generalist, PAM-relaxed Cas9 variants that can modify DNA at undesired sites in the genome, the custom Cas9 enzymes showed reduced off-target effects and improved editing efficiency in human cells and mice, which are critical prerequisites for potential therapeutic applications.

»A major outcome of this work is that we now have an enormous toolbox of safe and precise Cas9 proteins,« says first author Rachel Silverstein, who recently obtained a PhD from Harvard University under the supervision of Ben Kleinstiver at MGH.

On the broader impact of combining machine learning with protein engineering, Kleinstiver adds: »We hope this motivates a transition away from PAM-relaxed enzymes and towards these more active and specific tools, which ultimately are safer and more effective options for clinical translation.«

The team has made their model publicly accessible through a web interface (https://pammla.streamlit.app/), allowing researchers to design enzymes tailored to specific genomic targets.

PAMmla represents a significant advance towards more precise genome editing, especially for treating genetic disorders that call for highly specific targeting.

Read the full article entitled ‘Custom CRISPR—Cas9 PAM variants via scalable engineering and machine learning' here.