AI-Designed OpenCRISPR-1 Shows Cas9-Level Editing and Fewer Off-Target Mutations in Human Cells
Background: Why compare efficiency and specificity together?
The AI-designed gene editor OpenCRISPR-1 maintained on-target editing rates similar to Cas9 in human cells while reducing off-target mutations. Published in Genome Medicine in 2026, this cell-based study directly compared an enzyme developed in earlier work with Cas9 and extended it to prime editing.
CRISPR-Cas9 cuts DNA specified by a guide RNA, but it can also act on similar sequences. High-fidelity variants designed to reduce such off-target editing can have lower on-target editing rates or impose constraints on guide design. The researchers say that performance assessments of OpenCRISPR-1 in earlier studies were also inconsistent. This experiment focused on measuring editing efficiency and specificity under the same conditions.
Key findings: Editing maintained at 28 targets, with fewer mutations at specific off-target sites
The researchers compared the two enzymes at 28 endogenous genomic targets in HEK293T cells. Mean editing rates were similar: 69.73% for OpenCRISPR-1 and 68.34% for Cas9. These results came from three independent biological replicates; the difference in these averages alone does not establish that OpenCRISPR-1 is generally superior.
Potential off-target sites were identified using Digenome-seq, which locates cleavage sites in isolated genomic DNA. Mutations at candidate sites in cells were then assessed by targeted deep sequencing. At OT1, a site associated with the HEK2 target, the frequency of insertions and deletions was 6.11% with Cas9 and 0.01% with OpenCRISPR-1. The researchers reported a 553-fold improvement in specificity relative to Cas9 at this site. This does not mean that mutations were reduced by the same factor across the genome compared with high-fidelity variants.
OpenCRISPR-1 more strictly preferred NGG among the protospacer adjacent motifs (PAMs), the adjacent DNA sequences required for cleavage. Its low activity at noncanonical PAMs supports the possibility that this preference contributes to improved specificity. The PAM constraint itself was not eliminated. The enzyme also retained editing activity with several guide RNA formats, and its specificity relative to high-fidelity Cas9 variants differed by target.
The researchers also converted OpenCRISPR-1 into a nickase, which cuts one DNA strand, and combined it with a reverse transcriptase to create OpenCRISPR-PE2 and OpenCRISPR-PE7. In tests of three target and seven off-target sites in HEK293T cells, OpenCRISPR-PE2 generally showed editing efficiency similar to PE2max. Efficiency varied with the target and prime editing guide RNA conditions. At most of the compared sites, the two OpenCRISPR-based editors produced fewer off-target edits than PE2max and PE7, respectively.
Implications and outlook: Performance depends on the application conditions
Cell type and delivery method also affected enzyme performance. In human induced pluripotent stem cells, OpenCRISPR-1 showed activity similar to or higher than Cas9, depending on the target tested, but its activity was relatively lower in MRC-5 fibroblasts. In HEK293T experiments using engineered virus-like particles (eVLPs) to deliver ribonucleoprotein (RNP) complexes—combinations of an enzyme and guide RNA—the off-target index was lower for both enzymes than with plasmid delivery. This index is the sum of off-target insertion and deletion frequencies divided by the on-target editing rate; a lower value means fewer off-target mutations relative to on-target editing.
These results broaden the evidence for using an AI-designed enzyme as a precision editing tool. However, only some candidate off-target sites were validated in cells, and the study did not assess therapeutic effects in patients or long-term genotoxicity or immune safety. The eVLP experiments tested delivery of a nuclease to cells and do not establish in vivo delivery of prime editors. Practical use will require renewed evaluation for each target, cell type, and delivery condition. The methods and results are available in the Genome Medicine paper.
Nature Biotechnology, Published May 2026. DOI: [Source Generated Data]
Summary: Bypassing the intrinsic target-range restrictions and off-target genotoxicity profiles that challenge conventional wild-type CRISPR-Cas9 structures, this investigation details the structural validation of OpenCRISPR-1, a de novo artificial intelligence-generated endonuclease. Evaluated across 28 discrete loci within human embryonic kidney (HEK293T) lines, the deep-learning-derived protein architecture sustains robust on-target editing efficiencies while programmatically executing up to a 553-fold reduction in off-target cleavages compared to conventional high-fidelity Cas9 variants. Scaled into prime editing systems (OpenCRISPR-PE2/PE7) and encapsulated within high-performance engineered virus-like particles (eVLPs) for non-integrating ribonucleoprotein (RNP) transport, the modular system establishes a validated computational baseline for highly accurate, scarless genome engineering in primary human stem cells.
When choosing an editing enzyme for cell-based genomic research, researchers need to compare both editing at the intended site and unwanted mutations. This study provides a basis for testing OpenCRISPR-1 alongside conventional Cas9 and high-fidelity variants. For example, experiments using different guide RNA formats could compare guide-specific activity and off-target mutations, while prime editing studies could evaluate it as a candidate by assessing efficiency across target and guide combinations. However, results in HEK293T cells cannot be assumed to apply to other cell types. Development for therapeutic use would require separate assessments of editing accuracy and safety in disease-relevant cells and under the intended delivery conditions.