Researchers at Profluent Bio, in partnership with international academic institutes, have developed OpenCRISPR-1, an innovative genetic editing system designed using AI models trained on a massive dataset of over one million CRISPR operons from 26 terabases of genomes and metagenomes. This approach leverages protein language models (LMs) that learn evolutionary and functional constraints of proteins to generate novel and diverse sequences.
Exceptional Diversification of CRISPR-Cas Proteins
By fine-tuning these models on various CRISPR-Cas families, the researchers generated approximately 4 million new protein sequences, increasing natural protein diversity by 4.8-fold, with marked enrichment in Cas9, Cas12a, and Cas13 families. OpenCRISPR-1, derived from this generation, diverges radically from SpCas9 with over 400 mutations while retaining effective and specific targeting of the human genome.
Demonstrated Performance and Specificity in Human Cells
OpenCRISPR-1 was functionally tested in HEK293T cells, showing editing efficiency comparable to or exceeding SpCas9 at multiple target sites. It stands out with a ~95% reduction in off-target effects, reducing unwanted edits. This increased specificity was confirmed via genome-wide off-target cleavage analyses (SITE-Seq), indicating that the off-target sites identified for OpenCRISPR-1 are a subset of those known for SpCas9, demonstrating a favorable safety profile.
Compatibility with Base Editing and Newly Designed Guide RNAs
A mutation conferring nickase activity enabled fusion of OpenCRISPR-1 with adenosine deaminases, thereby generating base editors capable of efficiently converting A to G without inducing double-strand breaks. Additionally, the researchers developed a generative model to design guide RNAs (sgRNAs) tailored to the new proteins, optimizing editing efficiency across multiple variants.
Reduced Immunogenicity and Robust Structural Modeling
Epitope analysis showed that OpenCRISPR-1 lacks major SpCas9 T-cell epitopes, suggesting reduced immunogenic potential. AlphaFold2 structural predictions revealed that mutations are predominantly surface-exposed, preserving essential catalytic and binding domains, and even incorporating insertions likely to stabilize protein–DNA/RNA interactions.
Toward Customizable and Safer Genome Editors
This work demonstrates that protein language models can circumvent natural evolutionary constraints to produce highly functional and safe genome editors, tailored to diverse biotechnological and therapeutic needs. The CRISPR–Cas Atlas resource developed in this effort further enriches this approach by providing an extensive database for designing future tools customizable by properties such as size, PAM preference, and optimal temperature.
Recommended Reading in Nature:
Design of Highly Functional Genome Editors by Modelling CRISPR–Cas Sequences, by Ruffolo et al., 2025 – DOI : 10.1038/s41586-025-09298-z
