Genome editing has advanced at a rapid pace with promising results for treating genetic conditions—but there is always room for improvement. A new paper by investigators from Mass General Brigham published in Nature showcases the power of scalable protein engineering combined with machine learning to boost progress in the field of gene and cell therapy. In their study, authors developed a machine learning algorithm—known as PAMmla—that can predict the properties of about 64 million genome editing enzymes. The work could help reduce off-target effects and improve editing safety, enhance editing efficiency, and enable researchers to predict customized enzymes for new therapeutic targets. Their results are published in Nature .
"Our study is a first step in dramatically expanding our repertoire of effective and safe CRISPR-Cas9 enzymes. In our manuscript we demonstrate the utility of these PAMmla-predicted enzymes to precisely edit disease-causing sequences in primary human cells and in mice," said corresponding author Ben Kleinstiver, PhD, Kayden-Lambert MGH Research Scholar associate investigator at Massachusetts General Hospital (MGH), a founding member of the Mass General Brigham healthcare system. "Building on these findings, we are excited to have these tools utilized by the community and also apply this framework to other properties and enzymes in the genome editing repertoire."
CRISPR-Cas9 enzymes can be used to edit genes at locations throughout the genomes, but there are limitations to this technology. Traditional CRISPR-Cas9 enzymes can have off-target effects, cleaving or otherwise modifying DNA at unintended sites in the genome. The newly published study aims to improve this by using machine learning to better predict and tailor enzymes to hit their targets with greater specificity. The approach also offers a scalable solution—other attempts at engineering enzymes have had a lower throughput and typically yielded orders of magnitude fewer enzymes.
One of the key elements of utilizing CRISPR-Cas9 technologies is that the enzymes must locate and bind to a short DNA sequence called a protospacer adjacent motif (PAM). Researchers used machine learning to predict the PAMs of millions of Cas9 enzymes, identifying a set of novel engineered Cas9 enzymes that would have the best on-target activity and specificity. The researchers conducted proof-of-concept experiments in human cells and a mouse model of retinitis pigmentosa, finding that the bespoke enzymes had greater specificity.
"A major outcome of this work is the creation of this PAMmla model that can now be used by researchers to predict customized enzymes that are uniquely tuned for their specific use cases," said lead author Rachel A. Silverstein, PhD candidate, NSERC postgraduate scholar and 2024 Albert J. Ryan Fellow in the Kleinstiver lab at MGH. "The result of this model is that we now have an enormous toolbox of safe and precise Cas9 proteins that can be utilized for a variety of research and therapeutic applications."
The researchers have made a web tool to allow others to use the PAMmla model, which is available at https://pammla.streamlit.app/
Authorship: In addition to Kleinstiver and Silverstein, Mass General Brigham authors include Nahye Kim, Ann-Sophie Kroell, Russell T. Walton, Justin Delano, Rossano M. Butcher, Blaire K. Smith, Kathleen A. Christie, Leillani L. Ha, Luca Pinello, and Qin Liu. Additional authors include Martin Pacesa, Ronald J. Meis, Aaron B. Clark, Aviv D. Spinner, Cicera R. Lazzarotto. Yichao Li, Azusa Matsubara, Elizabeth O. Urbina, Gary A. Dahl, Bruno E. Correia, Debora S. Marks, Shengdar Q. Tsai, and Suk See De Ravin.
Paper cited: Silverstein R et al. "Custom CRISPR—Cas9 PAM variants via scalable engineering and machine learning" Nature DOI: 10.1038/s41586-025-08880-9
