Machine learning customises CRISPR-Cas PAM recognition

The team built a training dataset of over 45,000 CRISPR-Cas PAMs by mining 26.2 terabases of microbial genomic data, and then trained transformer-based models to predict PAM specificity directly from Cas protein sequences across the types I, II, and V systems. The models achieved prediction accuracies ranging from 0.868 to 0.955 and successfully identified PAMs for 92% of tested operons – nearly threefold more than existing spacer-based methods. Using in silico mutagenesis, Protein2PAM pinpointed PAM-interacting residues in Cas9 without requiring structural information, with 80% of high-effect mutations occurring at residues that form hydrogen bonds with the PAM DNA.

The researchers then used Markov chain Monte Carlo optimisation to computationally evolve Nme1Cas9 towards specific target PAMs. Experimental validation using high-throughput PAM determination assays confirmed that engineered variants recognised substantially broader PAM repertoires – for example, NmeN4G.1 cleaved 42 different N₄G PAMs compared to just seven for wild-type Nme1Cas9, whilst maintaining comparable on-target editing efficiency to SpCas9 in human cells. The machine learning framework enables single-step customisation of PAM specificity without the need for laboratory-derived training data, potentially expanding the targetable sequences for personalised genome-editing applications.

The study was led by Stephen Nayfach and Ali Madani at Profluent Bio in Emeryville, California. It was published in Nature Biotechnology on 2 February 2026.