Ambryon Biosciences

High-fidelity genome and prime editing enabled by the AI-designed openCRISPR-1

Hwang et al. (2026). Genome MedResearch

High-fidelity Cas9 variants buy specificity by losing on-target activity or constraining guide design; OpenCRISPR-1, designed by generative AI, matched Cas9 across 28 loci while cutting off-targets up to 553-fold and tolerating every U6 guide format.

Key findings

  • Using the Cas9 route of nicking mutation plus engineered reverse transcriptase fusion, OpenCRISPR-PE2 matches PE2max with canonical pegRNAs and consistently exceeds it with tevo- and mpk-epegRNAs; OpenCRISPR-PE7 raises efficiency further in most conditions.

  • Across ten guide RNAs in multiplex Digenome-seq, OpenCRISPR-1 cleaves far fewer genome-wide candidate off-target sites than Cas9: 25.6 ± 13.3 versus 88.5 ± 33.2 sites, using recombinant protein on genomic DNA.

  • OpenCRISPR-1 cleavage is more confined to canonical PAMs (87.89% ± 11.50% at NGG) than Cas9's, which also cleaves noncanonical NGH (6.10%), NHG (11.75%) and NHH PAMs (0.79%) in Digenome-seq.

  • At HEK2 off-target site OT1 (two mismatches), OpenCRISPR-1 induces 0.01% ± 0.02% indels against 6.11% ± 0.18% for Cas9, a 553-fold improvement, while keeping higher HEK2 on-target activity than Cas9.

  • OpenCRISPR-1 also suppresses off-target editing at noncanonical PAMs: at VEGFA2 site OT2 its indels are undetectable versus 2.38% ± 0.16% for Cas9, and at double-mismatch EMX1 site OT3 it gives 1.90% ± 0.19% versus 51.58% ± 0.71%.

  • RNP delivery in engineered virus-like particles raises on-target editing over plasmid transfection at VEGFA1 and EMX1 and lowers each enzyme's off-target index; delivered this way, OpenCRISPR-1 reaches up to 257-fold greater specificity than Cas9.

Author abstract

RNA-guided nucleases such as CRISPR-Cas9 systems have revolutionized genome engineering by enabling programmable DNA modifications. Although structure-guided and evolution-derived high-fidelity Cas9 variants improve target specificity, they often compromise on-target activity or constrain guide RNA (gRNA) design. We performed head-to-head comparisons of OpenCRISPR-1 and Cas9 in human cells using amplicon sequencing, multiplex Digenome-seq, and off-target validation by targeted sequencing. Editing activity was assessed across 28 endogenous loci in HEK293T cells and further evaluated in human induced pluripotent stem cells (iPSCs) and MRC-5 fibroblasts. To test clinically relevant delivery, Cas9 and OpenCRISPR-1 ribonucleoproteins were delivered using engineered virus-like particles (eVLPs). We also generated OpenCRISPR-based prime editors, OpenCRISPR-PE2 and OpenCRISPR-PE7, and compared them with PE2max and PE7 using pegRNAs and engineered epegRNAs. Here, we show that OpenCRISPR-1, an AI-designed, Cas9-like nuclease, retains Cas9-level editing efficiency across multiple genomic loci while significantly reducing off-target mutations. Using multiplex Digenome-seq and targeted deep sequencing, OpenCRISPR-1 exhibits up to a 553-fold reduction in off-target mutations compared to Cas9 and achieves off-target indices that match or surpass those of high-fidelity Cas9 variants. OpenCRISPR-1 also sustains robust editing across diverse gRNA formats (GX19, gX19, and gX20), highlighting its enhanced versatility. Furthermore, converting OpenCRISPR-1 into a prime editor yields comparable editing efficiencies while lowering the relative specificity ratio by up to 97%. These findings establish generative AI-guided protein design as a powerful strategy to overcome the specificity-efficiency trade-off, expanding the genome editing toolkit for both research and therapeutic use, and ushering in a new era of rational protein design.

Identifiers & revision