Protein Language Model Reprograms CRISPR PAM Recognition, Making Cutting 50 Times Faster

CRISPR-Cas enzymes cut DNA only when a specific short sequence, the PAM (protospacer-adjacent motif), sits next to the target. This constraint limits the sites in the genome that can be touched, and changing PAM recognition required repeated experiments in which countless variants were made and tested. Researchers at Profluent Bio and Massachusetts General Hospital in the United States reported in Nature Biotechnology a model that replaces this process with computation.

Predicting the PAM from the protein sequence alone

The researchers collected more than 45,000 PAMs and trained a deep learning model based on a protein language model, 'Protein2PAM.' The model predicts PAM preference from the amino acid sequence of a Cas protein alone, without structural information, and it was applied to type I, type II and type V CRISPR systems. Through in silico mutation experiments, it also pointed to residues in Cas9 that are decisive for PAM recognition.

Next, using designs the model proposed, they 'computationally evolved' Nme1Cas9 and obtained variants that recognize a wider range of PAMs and cut up to 50 times faster in vitro than the wild type. This means targets that were out of reach because of the PAM constraint can be brought into the range of editing.

Limitations

This is basic research at the level of test tubes and cells, and because a wider target range can also raise the risk of off-target editing, accuracy needs to be assessed separately.